URISA’s 46

th
Annual Conference &
Exposition
October 7–10, 2008 — New Orleans, LA
URISA Leadership Academy
December 8–12, 2008 — Seattle, WA
13
th
Annual GIS/CAMA Technologies
Conference
February 8–11, 2009 — Charleston, SC
URISA’s Second GIS in Public Health
Conference
June 5–8, 2009 — Providence, RI
URISA/NENA Addressing Conference
TBD
URISA’s 47
th
Annual Conference &
Exposition
September 29–October 2, 2009 — Anaheim, CA
GIS in Transit Conference
November 11–13, 2009 — St Petersburg, FL
w
w
w
.
u
r
i
s
a
.
o
r
g
Upcoming

conferences
Volume 20 • No. 1 • 2008
Journal of the Urban and Regional Information Systems Association
Contents
REFEREED
5 Automatic Generation of High-Quality Three-Dimensional Urban Buildings
from Aerial Images
Ahmed F. Elaksher and James S. Bethel
15 Robust Principal Component Analysis and Geographically Weighted Regression:
Urbanization in the Twin Cities Metropolitan Area of Minnesota
Debarchana Ghosh and Steven M. Manson
27 Where Are They? A Spatial Inquiry of Sex Offenders in Brazos County
Praveen Maghelal, Miriam Olivares, Douglas Wunneburger, and Gustavo Roman
35 Tools And Methods For A Transportation Household Survey
Martin Trépanier, Robert Chapleau, and Catherine Morency
45 Mapping Land-Use/Land-Cover Change in the Olomouc Region, Czech
Republic
Tomáš Václavík
PLUS
53 Mapping the Future Success of Public Education
2 URISA Journal • Vol. 20, No. 1 • 2008
Journal
EDITORIAL OFFICE: Urban and Regional Information Systems Association, 1460 Renaissance Drive, Suite 305, Park Ridge, Illinois 60068-1348;
Voice (847) 824-6300; Fax (847) 824-6363; E-mail info@urisa.org.
SUBMISSIONS: Tis publication accepts from authors an exclusive right of frst publication to their article plus an accompanying grant of non-
exclusive full rights. Te publisher requires that full credit for frst publication in the URISA Journal is provided in any subsequent electronic or
print publications. For more information, the “Manuscript Submission Guidelines for Refereed Articles” is available on our website, www.urisa.
org, or by calling (847) 824-6300.
SUBSCRIPTION AND ADVERTISING: All correspondence about advertising, subscriptions, and URISA memberships should be directed to:
Urban and Regional Information Systems Association, 1460 Renaissance Dr., Suite 305, Park Ridge, Illinois, 60068-1348; Voice (847) 824-6300;
Fax (847) 824-6363; E-mail info@urisa.org.
URISA Journal is published two times a year by the Urban and Regional Information Systems Association.
© 2008 by the Urban and Regional Information Systems Association. Authorization to photocopy items for internal or personal use, or the internal
or personal use of specifc clients, is granted by permission of the Urban and Regional Information Systems Association.
Educational programs planned and presented by URISA provide attendees with relevant and rewarding continuing education experience. However,
neither the content (whether written or oral) of any course, seminar, or other presentation, nor the use of a specifc product in conjunction there-
with, nor the exhibition of any materials by any party coincident with the educational event, should be construed as indicating endorsement or
approval of the views presented, the products used, or the materials exhibited by URISA, or by its committees, Special Interest Groups, Chapters,
or other commissions.
SUBSCRIPTION RATE: One year: $295 business, libraries, government agencies, and public institutions. Individuals interested in subscriptions
should contact URISA for membership information.
US ISSN 1045-8077
Publisher: Urban and Regional Information Systems Association
Editor-in-Chief: Jochen Albrecht
Journal Coordinator: Scott A. Grams
Electronic Journal: http://www.urisa.org/journal.htm

URISA Journal • Vol. 20, No. 1 • 2008 3
URISA Journal Editor
Editor-in-Chief
Jochen Albrecht, Department of
Geography, Hunter College City University
of New York
Tematic Editors
Editor-Urban and Regional Information
Science
Vacant
Editor-Applications Research
Lyna Wiggins, Department of Planning,
Rutgers University
Editor-Social, Organizational, Legal,
and Economic Sciences
Ian Masser, Department of Urban Planning
and Management, ITC (Netherlands)
Editor-Geographic Information Science
Mark Harrower, Department of Geography,
University of Wisconsin Madison
Editor-Information and Media Sciences
Michael Shifer, Department of Planning,
Massachusetts Institute of Technology
Editor-Spatial Data Acquisition and
Integration
Gary Hunter, Department of Geomatics,
University of Melbourne (Australia)
Editor-Geography, Cartography, and
Cognitive Science
Vacant
Editor-Education
Karen Kemp, Director, International Masters
Program in GIS, University of Redlands
Section Editors
Software Review Editor
Jay Lee, Department of Geography, Kent State
University
Book Review Editor
David Tulloch, Department of Landscape
Architecture, Rutgers University
Article Review Board
Peggy Agouris, Department of Spatial Information
Science and Engineering, University of Maine
Grenville Barnes, Geomatics Program, University
of Florida
Michael Batty, Centre for Advanced Spatial Analysis,
University College London (United Kingdom)
Kate Beard, Department of Spatial
Information Science and Engineering,
University of Maine
Yvan Bédard, Centre for Research in Geomatics,
Laval University (Canada)
Barbara P. Buttenfeld, Department of
Geography, University of Colorado
Keith C. Clarke, Department of Geography,
University of California-Santa Barbara
David Coleman, Department of Geodesy and
Geomatics Engineering, University of New
Brunswick (Canada)
David J. Cowen, Department of Geography,
University of South Carolina
Massimo Craglia, Department of Town &
Regional Planning, University of Shefeld
(United Kingdom)
William J. Craig, Center for Urban and
Regional Afairs, University of Minnesota
Robert G. Cromley, Department of Geography,
University of Connecticut
Kenneth J. Dueker, Urban Studies and
Planning, Portland State University
Geofrey Dutton, Spatial Efects
Max J. Egenhofer, Department of Spatial Information
Science and Engineering, University of Maine
Manfred Ehlers, Research Center for
Geoinformatics and Remote Sensing, University of
Osnabrueck (Germany)
Manfred M. Fischer, Economics, Geography &
Geoinformatics, Vienna University of Economics
and Business Administration (Austria)
Myke Gluck, Department of Math and
Computer Science, Virginia Military Institute
Michael Goodchild, Department of Geography,
University of California-Santa Barbara
Michael Gould, Department of Information
Systems Universitat Jaume I (Spain)
Daniel A. Grifth, Department of Geography,
Syracuse University
Francis J. Harvey, Department of Geography,
University of Minnesota
Kingsley E. Haynes, Public Policy and
Geography, George Mason University
Eric J. Heikkila, School of Policy, Planning, and
Development, University of Southern California
Stephen C. Hirtle, Department of Information
Science and Telecommunications, University of
Pittsburgh
Gary Jefress, Department of Geographical
Information Science, Texas A&M University-
Corpus Christi
Richard E. Klosterman, Department of
Geography and Planning, University of Akron
Robert Laurini, Claude Bernard University of
Lyon (France)
Tomas M. Lillesand, Environmental
Remote Sensing Center, University of Wisconsin-
Madison
Paul Longley, Centre for Advanced Spatial Analysis,
University College, London (United Kingdom)
Xavier R. Lopez, Oracle Corporation
David Maguire, Environmental Systems Research
Institute
Harvey J. Miller, Department of Geography,
University of Utah
Zorica Nedovic-Budic, Department of Urban
and Regional Planning,University of Illinois-
Champaign/Urbana
Atsuyuki Okabe, Department of Urban
Engineering, University of Tokyo (Japan)
Harlan Onsrud, Spatial Information Science
and Engineering, University of Maine
Jefrey K. Pinto, School of Business, Penn State Erie
Gerard Rushton, Department of Geography,
University of Iowa
Jie Shan, School of Civil Engineering,
Purdue University
Bruce D. Spear, Federal Highway Administration
Jonathan Sperling, Policy Development &
Research, U.S. Department of Housing and
Urban Development
David J. Unwin, School of Geography, Birkbeck
College, London (United Kingdom)
Stephen J. Ventura, Department of
Environmental Studies and Soil Science,
University of Wisconsin-Madison
Nancy von Meyer, Fairview Industries
Barry Wellar, Department of Geography,
University of Ottawa (Canada)
Michael F. Worboys, Department of Computer
Science, Keele University (United Kingdom)
F. Benjamin Zhan, Department of Geography,
Texas State University-San Marcos
EDITORS AND REVIEW BOARD
URISA Journal • Elaksher, Bethel 5
IntroductIon
Tree-dimension building information is required for a variety
of applications, such as urban planning, mobile communication,
visual simulation, visualization, and cartography. Automatic
generation of this information is one of the most challenging
problems in photogrammetry, image understanding, computer
vision, and GIS communities. Current automated algorithms
have shown some progress in this area. However, some defcien-
cies still remain in these algorithms. Tis is particularly apparent
in comparison to manual extraction techniques, which, although
slow, are essentially perfect in accuracy and completeness.
Recent research covers extracting building information from
high-resolution satellite imageries, high-quality digital elevation
models (DEMs), and aerial images. For example, QuickBird and
IKONOS high-resolution satellite imageries are used to acquire
planemetric building information with one-meter horizontal ac-
curacy (Theng 2006, Lee et al. 2003). However, aerial images are
the primary source used to acquire accurate and reliable geospatial
information. Lin and Nevatia (1998) proposed an algorithm to
extract building wireframes from a single image. However, a
single image does not provide any depth information. A pair of
stereo images could also be used to extract building information
(Avrahami et al. 2004, Chein and Hsu 2000). Using one pair
of images is insuffcient to extract the entire building because of
hidden features that are not projected into the image pair.
Kim et al. (2001) presented a model-based approach to
generate buildings from multiple images. Three-dimensional
rooftop hypotheses are generated using three-dimensional roof
boundaries and corners extracted from multiple images. The gen-
erated hypotheses then are employed to extract buildings using an
expandable Bayesian network. Wang and Tseng (2004) proposed
a semiautomatic approach to extract buildings from multiple
views. They proposed an abstract foating model to represent real
objects. Each model has several pose and shape parameters. The
parameters are estimated by ftting the model to the images using
least-squares adjustment. The algorithm is limited to parametric
models only. In Shmid and Zisserman (2000), lines are extracted
in the images and matched over multiple views in a pair-wise
mode. Each line then is assigned two planes, one plane on each
side. The planes are rotated, and the best-ftting plane is found.
Planes then are intersected to fnd the intersection lines.
Henricsson et al. (1996) presented another system to extract
suburban roofs from aerial images by combining two-dimensional
edge information together with photometric and chromatic
attributes. Edges are extracted in the images and aggregated to
form coherent contour line segments. Photometric and chromatic
contour attributes for adjacent regions around each contour are
assigned to it. For each contour, attributes are computed-based
on the luminance, color, proximity, and orientation, and saved
for the next step. Contour segments then are matched using their
attributes. Segments in three dimension are grouped and merged
according to an initial set of plane hypotheses.
Fischer et al. (1997) extracted three-dimensional buildings
from aerial images using a generic modeling approach that depends
on combining building parts. The process starts by extracting
low-level image features: points, lines, regions, and their mutual
relations. These features are used to generate three-dimensinal
building part hypotheses in a bottom-up process. A step-wise
model-driven aggregation process combines the three-dimensional
building feature aggregates to three-dimensional parameterized
building parts and then to a more complex building descriptor.
The resulting complex three-dimensional building hypothesis
then is back-projected to the images to allow component-based
hypothesis verifcation.
A semiautomated approach is used in Förstner (1999) to
solve the building-extraction problem. First, the user has to defne
the building model and fnd the building elements in one image
Automatic Generation of High-Quality three-dimensional
urban Buildings from Aerial Images
Ahmed F. Elaksher and James S. Bethel
Abstract: High-quality three-dimensional building databases are essential inputs for urban area geographic information systems.
Because manual generation of these databases is extremely costly and time-consuming, the development of automated algorithms
is greatly needed. This article presents a new algorithm to automatically extract accurate and reliable three-dimensional build-
ing information. High overlapping aerial images are used as input to the algorithm. Radiometric and geometric properties of
buildings are utilized to distinguish building roof regions in the images. This is accomplished with image segmentation and
neural network techniques. A rule-based system is employed to extract the vertices of the roof polygons in all images. Photogram-
metric mathematical models are used to generate the roof topology and compute the three-dimensional coordinates of the roof
vertices. The algorithm is tested on 30 buildings in a complex urban scene. Results showed that 95 percent of the building roofs
are extracted correctly. The root-mean-square error for the extracted building vertices is 0.35 meter using 1:4000 scale aerial
photographs scanned at 30 microns.
6 URISA Journal • Vol. 20, No. 1 • 2008
by a number of mouse clicks. Then the algorithm fnds the cor-
responding features in other images and matches them to build
the three-dimensional wireframe of the building. This approach
supports the extraction of more complex buildings; however, it
requires the user to spend a great amount of time interacting
with the system. Rottensteiner (2000) presented a semiautomated
building-extraction technique in which the user can select an
appropriate building primitive from a database and then adjust
the parameters of the primitive to the actual data by interactively
measuring points in the digital images, and determine the fnal
building parameters by automated matching tools.
High-quality DEMs such as those available from light
detection and ranging (LIDAR) have been used to generate
three-dimensional building models. Tse et al. (2006) proposed an
algorithm based on segmenting the raw data into high and low
regions, and then modeling the walls and roofs by extruding the
triangulated terrain surface (TIN) using CAD-type Euler opera-
tors. Tarsha-Kurdi et al. (2006) proposed another algorithm that
discriminates terrain and off-terrain clouds in a LIDAR point
cloud. Then it categorizes the off-terrain points to building and
vegetation subclasses. Building points then are detected via seg-
menting the original LIDAR three-dimensional point cloud. In
Brunn and Weidner (1997), the digital surface models (DSMs)
are extracted using mathematical morphology. The differences
between the DEMs and the DSMs are computed and building
points are detected by thresholding these differences. In the next
step, building wireframes are generated using parametric and
prismatic models depending on the complexity of the detected
building. Morgan and Habib (2002) proposed another algorithm
for building detection that has the following steps: segmenta-
tion of laser points, classifcation of laser segments, generation
of building hypothesis, verifcation of building hypothesis, and
extraction of building parameters.
Several researchers worked on integrating LIDAR and
aerial images for building extraction. The approach presented
in Hongjian and Shiqiang (2006) is based on aerial images and
sparse laser scanning sample points. Linear features are extracted
in the aerial images frst. Bidirection projection histogram and
line matching then are used to extract the contours of buildings.
The height of the building is determined from sparse laser sample
points that are within the contours of the buildings extracted
from the images.
The presented systems display many defciencies. Satellite
imageries still do not provide high-quality elevation data. Several
systems require human interaction. Using a parametric model
to represent buildings limits some systems to specifc building
models. Another problem is the excessive reliance on primitive
features such as corner points or line segments. Naive matching
of such primitive features yields numerous false matches, and
misses many correct ones. Systems using more than a pair of im-
ages perform the feature matching in a pair-wise mode. Systems
utilizing only DEM in building extraction start by segmenting the
DEMs. This process is problematic because of outliers and spikes.
In addition, such DEMs are expensive to collect and insuffcient
to provide surface texture.
In this article, a new algorithm to extract building wire-
frames using more than two images is presented. The algorithm
has the capability to extract a wide range of buildings with dif-
ferent shapes, orientations, and heights. The human operator
needs only to select an image patch for the building in the frst
image, specify the location of the input data fles, and set up the
thresholds. The algorithm will then fnd corresponding patches
in other images, using the input data, and start the extraction
process. The algorithm is implemented using computer vision
techniques, artifcial intelligence algorithms, and rigorous pho-
togrammetric mathematical models. Computer vision techniques
are employed to extract image regions using a modifed version
of the split-and-merge image segmentation technique. Artifcial
intelligent algorithms are used to discriminate roof regions and to
convert the image regions to two-dimensional polygons. Photo-
grammetric mathematical models are employed to simultaneously
and rigorously match image polygons and vertices across all views.
One of the powerful tools that photogrammetry provides is to
simultaneously match features across multiple images. Inputs
are four images for the building. The algorithm has been tested
on a large sample of buildings selected quasi-randomly from the
Purdue University campus. Four images are used for each build-
ing and the automatically extracted wireframes of the extracted
buildings are presented. Results show signifcant improvement in
the detection rate and accuracy and suggest the completeness and
accuracy of the proposed algorithm. The remainder of this paper
is organized as follows. First, the process of extracting building
polygons in aerial images is presented. Then the generation of
three-dimensional building models is proposed. Results are given
in the next section, followed by discussions and conclusions.
ExtrActInG BuIldInG
PolyGons In AErIAl ImAGEs
Image Region Extraction
Several researchers worked on segmenting aerial and satellite im-
ages in urban environments. Muller and Zaum (2005) proposed
an algorithm to detect and classify buildings from a single aerial
image using a region-growing algorithm. Lari and Ebadi (2007)
proposed another segmentation algorithm to detect building re-
gions in satellite images. Te results, although applied to a single
image, showed the signifcance of implementing segmentation
strategies to detect buildings in aerial and satellite images. In this
research, a modifed split-and-merge image segmentation process,
Horowitz and Pavlidis (1974) and Samet (1982), is applied to
segment the aerial images. Tis technique obtains good results if a
high contrast exists between the foreground and the background
objects and if the segmented objects are internally homogenous.
For urban aerial images this is not always the case. Objects such
as roads, cars, trees, and buildings are common to a typical urban
aerial image. Although this wide variety of objects is expected to
be seen in aerial images, building roofs have an important attri-
URISA Journal • Elaksher, Bethel 7
bute that distinguishes them from other spatial features. Building
roofs often are homogenous objects that can be segmented from
other image features. Homogeneity is a scale-dependent attribute,
but for small-scale images (1:4000), roof regions often appear
homogeneous. However, utility pipelines and ducts can disturb
the building roof homogeneity. Texture is another problem that
can decrease the ability to segment building roofs. To account for
these problems, several modifcations are proposed to the original
split-and-merge algorithm. Te algorithm presented difers from
the conventional split-and-merge algorithm in its capability to
join neighboring regions based on their intensity and size difer-
ences and in its potential for detecting and flling region gaps.
The segmentation process is implemented as follows. The
image frst is divided into smaller regions until a homogeneity
condition is satisfed. This is implemented by constructing a
quadtree of the image, progressing down the tree, splitting as
necessary when inhomogeneities exist. Then a merging algorithm
is implemented. The merging is carried out in two steps. Adjacent
regions are frst merged based on the differences between the mini-
mum and maximum intensities. Large regions then are merged
with their small neighbors if the differences in their intensities are
smaller than a given threshold. Intensity thresholds for splitting
and merging range between 10 and 15, while the size threshold
is kept fxed relative to the image size.
One of the problems noticed after segmenting the image is
the presence of holes inside the segmented regions. This can oc-
cur because of texture and/or utility features. Holes are detected
and removed as follows. First for each region its pixels are located
and copied to a template image with a background intensity of
zero. A region-growing algorithm then is used to connect all
the pixels that do not belong to either the background or the
region. These pixels are described as holes and are attached to
the original region. Another problem also observed in the results
is the splitting of some roof patches into two or more regions.
To overcome this problem, an average intensity is computed for
each region and any two neighboring regions are merged if the
difference between their average intensities is smaller than a given
threshold, regardless of their sizes. Small regions, very dark regions,
and very bright regions are eliminated. Thresholds for the region
merging and elimination are kept fxed relative, at this stage, to
the intensity range of the images and the image size. The results
of the segmentation process for one sample building are shown
in Figure 1. Figure 2 shows the results of segmenting the images
of fve other buildings.
Region Classifcation
Buildings usually are elevated blobs in the DEMs. On the other
hand, building regions possess linear borders. Other features
such as trees also are elevated; however, they do not have linear
borders. Roads and sidewalks have linear borders, but are not
elevated. Terefore, in this research the high elevation and border
linearity attributes are used to discriminate building roof regions
from other regions using a neural network. Each image region is
assigned two attributes for the discrimination process. Te frst
attribute measures the linearity of the region boundaries, while
the second attribute measures the percentage of the points in the
region that are above a certain height.
Border linearity is measured using a modifed version of
Hough transformation (Hough 1962). First, border points are
extracted and sorted so they traverse the border clockwise. For
each border point, the previous fve points and the next fve points
are found to form a local line at each point. The adjusted line
parameters (α
a
, P
a
) and the quadratic form of the residuals for
the local line at each point are computed using the least-squares
estimation technique. The algorithm is implemented in three runs
at each point; the frst run is when the point is in the middle of the
line; in the second run, the local line is shifted so that the point
is at the end. In the third run, the local line is shifted so that the
specifed point is the frst point. If the minimum quadratic form

value is small, the parameter space cell at the location of the local
line parameters, i.e., α
a
, P
a
, is increased by one. The parameter
(a) (b) (c)
Figure 1. Split-and-merge segmentation results for the image of one
building: (a) original image, (b) image after splitting, (c) fnal regions
(a) (b)
Figure 2. Split-and-merge segmentation results for the images of fve buildings: (a) original images, (b) fnal regions
8 URISA Journal • Vol. 20, No. 1 • 2008
(a) (b)
Figure 3. The parameter space for a roof region (a) and a nonroof region (b)
(a) (b)
Figure 4. DEM (a) and DSM (b) for an area with several buildings, same vertical scale
space then is searched and analyzed to determine a measure for
border linearity. Border linearity is measured as the percentage of
the number of points in the larger four cells to the total number of
border points. Figure 3 shows a parameter space for a roof region
and another parameter space for a nonroof region.
A digital elevation model (DEM) is used to quantify the
height of each region. First, the digital surface model (DSM),
i.e., representing bare ground, is extracted. Minimum flters are
used to perform this task (Masaharu and Ohtsubo 2002, Wack
and Wimmer 2002). The fltering process detects and conse-
quently removes points above the ground surface to recognize
high points in the data set. The minimum flter size should be
large enough to include data points that are not noise. However,
iterative approaches could be used to avoid the effect of noise.
In this research, the size of the flter is 9x9 pixels. The fltering is
repeated iteratively until the DSM is extracted. The differences
between the DEM and the DSM then are computed and used to
represent height information. The use of the height information
in preference to the elevations makes the algorithm applicable for
both fat and slope terrains. Figure 4 shows the DEM and DSM
of an area with several buildings.
Each point in the image then is assigned a height value by
projecting the differences between the DEM and the DSM back
to the image using the image registration information, the pixel
location in the image, and the DEM elevation. For each image
point a ray is generated, starting from the exposure station of the
camera and directed toward the point. The intersection between
the ray and the DEM defnes the location of the corresponding
DEM post. The height information at this location then is used
as the height of the corresponding image point. The region height
URISA Journal • Elaksher, Bethel 9
measure is defned as the percentage of the number of points in
the region that are above a certain height to the total number of
points in the same region.
The neural network implemented in this research is a feed-
forward back-propagation network (see Figure 5). The network
consists of three layers: an input layer, one hidden layer, and an
output layer. The number of neurons in the frst layer is two.
The number of neurons in the second layer is selected to be ten.
The number of neurons in the third, i.e., last, layer is one. The
output of this neuron is either one in case the region is a roof
region or zero in case the region is not a roof region. The activa-
tion function for all neurons in the frst and second layers is the
sigmoid functions (Principe et al. 1999). For the output neuron,
the step function is chosen as the activation function. To study the
performance of the neural network, a variety of training data sets
are used with different sizes: 20, 50, 100, 200, and 400 samples,
including 2, 5, 10, 20, and 40 roof samples, respectively, while
the other samples are nonroof samples. The average detection
rate and false alarm rate for each training data set is recorded and
shown in Figure 6. Results show that increasing the size of the
training data set does not affect the detection rate signifcantly.
However, increasing the size of the training data set does have a
signifcant effect on the false alarm rate.
Image Polygon Extraction
Te two-dimensional modifed Hough space is utilized in extract-
ing the borderlines for the roof regions. Given all points contrib-
uting to a certain cell, a nonlinear least-squares estimation model
is used to adjust the line parameters of that cell. Lines then are
grouped recursively until no more lines with similar parameters
are left and short lines are rejected. Te next step is to convert the
extracted lines to polygons via a rule-based system. Te rules are
designed as complex as possible to cover a wide range of polygons.
Te mechanism works in three steps. Te frst step is to fnd all
the possible intersections between the borderlines. However, if
two lines are almost parallel, i.e., intersecting angle out of the
range (30
o
—150
o
), the intersection point is not considered. Te
next step is to generate a number of polygons from all recorded
intersections. Each combination of three and four intersection
points is considered a polygon hypothesis. Hypotheses are ignored
if the diference in area between the region and the hypothesized
polygon is large. Te best polygon that represents the region is
chosen from the remaining hypotheses using a template-matching
process. Te template is chosen to be the region itself, while it is
matched across all polygon hypotheses. Te hypothesis with the
largest correlation and minimum number of vertices is chosen to
be the best-ftting polygon. Te extracted polygons for six sample
buildings are shown in Figure 7. Tis algorithm succeeded in
overcoming several limitations of the segmentation process, such
as partially occluded regions, overshooting and undershooting
borders, and incomplete regions. Tis is observed by comparing
Figures 1, 2, and 7.
tHrEE-dImEnsIonAl PolyGon
ExtrActIon
Polygon Correspondence
After extracting the building roof polygons from the images, poly-
gon correspondence should be established. A new technique to
fnd corresponding polygons based on their geometrical properties
(a) (b)
Figure 6. The detection rate (a) and the false alarm rate (b)
Figure 5. The implemented neural network
10 URISA Journal • Vol. 20, No. 1 • 2008
equations (Mikhail et al. 2001) can be written using the least-
squares estimation technique. After determining the object space
coordinates of the intersecting point, the quadratic form of the
residuals is computed. If the four polygons across all views are
corresponding, then the quadratic form should be small, otherwise
it would be large. Terefore the quadratic form for each polygon
correspondence set serves as its cost. A four-dimensional array
is constructed to store the corresponding cost values. Te array
dimension is the same as the number of images. Each axis in the
array represents the polygons of one image. Te residual cost value
is stored at the corresponding location in the four-dimensional
array. Te axis with the largest number of polygons is used as the
reference axis and for each polygon on this axis; its subspaces array
is used to fnd its corresponding polygons in the other images.
Tis subspace array is searched and its minimum value defnes the
corresponded sets. Te minimum value indexes for the subspace
array are used as the corresponding polygons in other images. If
a polygon corresponds to more than one set, the total residual
cost is used to solve such cases.
Computing the Three-Dimensional Polygon
Coordinates
After fnding the corresponding polygons, the three-dimensional
coordinates of each roof polygon center are computed; however,
the correspondence problem between the polygon vertices is not
solved. To solve this problem, another least-squares adjustment
model is implemented. Te input observations are the image
coordinates of the polygon vertices and the unknowns are the
object space coordinates for the three-dimensional polygon ver-
tices. To solve the correspondence relations between the vertices
within a group of corresponding polygons, the following process
is implemented. For each polygon correspondence set, all vertex
combinations are considered, and for each combination, the three-
dimensional coordinates for the polygon vertices are computed.
Each set includes a number of subsets (three subsets for triangles
and four for quadrilaterals); each subset includes a group of hy-
pothesized matching vertices. Te quadratic form is computed
for each subset, then they are added to create the total residual
Figure 8. The roof vertices of one building, before reconstructing the
roof topology
Figure 9. The roof vertices of one building, after reconstructing the
roof topology
Figure 7. Extracted image polygons for six sample buildings
is developed and presented in this section. All possible polygon
corresponding combination sets are considered, and for each
combination a correspondence cost is computed. Te computed
cost is used to determine the best corresponding polygon set.
First, the coordinates of each two-dimensional polygon center
are computed from the coordinates of its vertices. All possible
polygon correspondence combinations are exhaustively enu-
merated and considered, and for each combination the polygon
centers are matched across all available views. Because more than
one pair of images is available, a function of the residuals of the
image coordinates can be calculated and used as the matching
cost. For each combination of four polygons, four collinearity
URISA Journal • Elaksher, Bethel 11
of the set. Te quadratic forms of the sets then are compared and
the set with the minimum quadratic form is selected.
Roof Topology Reconstruction
After fnding the corresponding vertices, the three-dimensional
coordinates for each vertex are computed; however, the building
topology is not yet constructed as shown in Figure 8. A geometri-
cally constrained least-squares model is implemented to refne the
locations of the polygon vertices and to construct the building
topology. Te input observations are the image coordinates of the
polygon vertices; the unknowns are the object space coordinates
for the three-dimensional vertices. Te aim of this step is to convert
groups of neighboring vertices into one vertex, adjust the elevations
of horizontal points, and reconstruct the correct relation between
adjacent polygons. Te following constraints are used:
1. Nearby vertices should be grouped into one vertex: The resultant
vertices from the previous step need to be grouped, as shown
in Figure 8, for the extracted three-dimensional polygons may
not be contiguous. This results in having more than one point
in the position of a single roof vertex. If the distance between
any two or more vertices is smaller than a fxed threshold (0.50
meter), their coordinates are constrained to be equal.
2. The polygon vertices should be planer: It is assumed that
complex building roofs are built up from planar surfaces. The
aim of this constraint is to force the vertices of any polygon
to lie in the same plane. This constraint is used only if the
number of vertices in the polygon is larger than three.
3. Points that are almost in a horizontal plane are constrained
to have the same elevation: If a group of points has a small
difference in elevation (0.10 meter), they are constrained to
have the same (Z) coordinate.
4. Symmetric polygons should be constrained to have symmetric
parameters: If the computed parameters for any two planes
indicate that the two planes are approximately symmetric,
the two planes are constrained to be symmetric.
Thresholds for applying the constraints are fxed for all build-
ings. The results of this step are shown in Figure 9.
Building Topology Reconstruction
After determining the topology of the roofs, the last step is to
reconstruct the perimeter facets. Tis is achieved by the following
algorithm. First, border vertices are determined, using a point in
polygon algorithm, and sorted clockwise. A facet is generated for
any two successive border vertices. Te two vertices are assumed to
be the upper points of the facet. Te horizontal coordinates of the
lower two points are taken exactly as the horizontal coordinates
of the upper two points, while the elevations are automatically
measured from the nearest DSM post.
Results
A sample of 30 buildings extracted using the presented algorithm
is shown in Figure 10. Te results show the completeness and
Figure 10. The wireframes of the extracted buildings (repositioned)
accuracy of the three-dimensional buildings that can be extracted
using the presented system. To evaluate the accuracy of the
extracted buildings, the three-dimensional coordinates of the
extracted building vertices were extracted manually and compared
with the automatically extracted coordinates. Te average root-
mean-square errors for the horizontal and vertical coordinates of
all building vertices are 0.30 meter and 0.40 meter. Out of about
150 roof regions, 141 roof polygons were detected.
dIscussIons And
conclusIons
Tis paper presents a new algorithm to generate high-quality
three-dimensional building information. Results show the great
improvement that the algorithm adds. Te user only has to select
an image patch for the building in the frst image. Te algorithm
utilizes radiometric and geometric properties of urban build-
12 URISA Journal • Vol. 20, No. 1 • 2008
ings. Image segmentations, neural networks, rule-based systems,
photogrammetric mathematical models, and rigorous geometric
constraints are used. Te paper shows that at the employed im-
age scale (1:4000), segmentation provides high-quality image
regions that can be used in the building extraction process. Te
DEM is only used to provide evidence about the height of the
regions. Te DEM is generated from the same data set and the
only requirement on its quality is that at least one DEM post
is inside each roof region. Regions are classifed using border
linearity and region height. Classifcation results show that these
attributes are adequate to discriminate roof regions. However,
other attributes, such as color, could be used. Although the
rule-based system extracted either triangle or quadrilateral roof
patches, the system could be used to extract any roof shape. Te
rigorous photogrammetric mathematical models succeeded in
fnding conjugate polygons, simultaneously matching across
several views, and implementing the constraint equations for to-
pology reconstruction. Although several thresholds are used, their
values were fxed for all buildings, except the intensity diference
threshold that ranged from 10 to 15. Future work will focus on
using color information in the segmentation. Te model provided
an average RMS of about 0.35 meter. With comparison to the
image scale, this represents an average accuracy of three pixels in
the image coordinates for the vertices. Te algorithm succeeds
in extracting a wide range of urban building. Te tested data set
includes simple buildings with one rectangular roof, gabled roof
buildings, multistore buildings with large relief, and a variety of
complex buildings.
About the Authors
Ahmed F. Elaksher is an assistant professor at the engineering
faculty at Cairo University. He earned his Ph.D. in August
of 2002 from Purdue University. His areas of interest include
object recognition, feature extraction, three-dimensional
modeling from remote-sensing images, management of
spatial information databases, and decision making using
GIS, Internet, and wireless technologies.
Corresponding Address:
Department of Public Works
Faculty of Engineering
Cairo University
Giza, Egypt
ahmedelaksher@yahoo.com
James Bethel is an associate professor at the School of Civil
Engineering, Purdue University. His areas of interest include
photogrammetry, remote sensing, data adjustment, and
digital-image processing.
Corresponding Address:
Purdue University
School of Civil Engineering
550 Stadium Mall Drive
West Lafayette, IN 47907-2051
bethel@ecn.purdue.edu
References
Avrahami, Y., Y. Raizman, and Y. Doytsher. 2004. Semiautomatic
3D mapping of buildings from medium scale (1:40,000) aerial
photographs. The International Archives of Photogrammetry,
Remote Sensing, and Spatial Information Sciences, Istanbul,
Turkey, July 2004, XXXV (B3), 774-9.
Brunn, A., and U. Weidner. 1997. Extracting buildings from
digital surface models, The International Archives of
Photogrammetry, Remote Sensing, and Spatial Information
Sciences. Stuttgart, Germany, September 1997, XXXII
(3-4W2), 27-34.
Chein, L., and W. Hsu. 2000. Extraction of man-made buildings
in multispectral stereoscopic images. The International
Archives of Photogrammetry, Remote Sensing, and Spatial
Information Sciences. Amsterdam, The Netherlands, July
2000, XXXIII (B3/1), 169-76.
Fischer, A., T. Kolbe, F. Lang, A. Cremers, W. Förstner, L.
Plümer, and V. Steinhage. 1998. Extracting buildings from
aerial images using hierarchical aggregation in 2D and 3D.
Computer Vision and Image Understanding 72, no. 2:
185-203.
Förstner, W. 1999. 3D-city models: Automatic and semiautomatic
acquisition methods. Photogrammetric Week, Stuttgart
(September 1999).
Henricsson, O., F. Bignone, W. Willuhn, F. Ade, O. Kubler, E.
Baltsavias, S. Mason, and A. Grun. 1996. Project amobe:
Strategies, current status, and future work. The International
Archives of Photogrammetry, Remote Sensing, and Spatial
Information Sciences, Vienna, Austria, July 1996, XXXI
(B3), 321-30.
Hongjian, Y., and Z. Shiqiang. 2006. 3D building reconstruction
from aerial CCD image and sparse laser sample data. Optics
and Lasers in Engineering 44, no. 6: 555-66.
Horowitz, S. L., and T. Pavlidis. 1974. Picture segmentation
by a direct split and merge procedure. In Proceedings of
2nd International Conference on Pattern Recognition,
Copenhagen, Denmark, August 1974, 424-33.
Hough, P. V. C. 1962. Method and means for recognizing
complex patterns. U.S. Patent 3,069,654.
Kim, A. 2001. Multi-view 3D object description with uncertain
reasoning and machine learning. Ph.D. Thesis, University
of Southern California, August 2001.
Lari, Z., and H. Ebadi. 2007. Automatic extraction of building
features from high resolution satellite images using artifcial
neural networks. In Proceedings of ISPRS Conference on
Information Extraction from SAR and Optical Data, with
Emphasis on Developing Countries, Istanbul, Turkey, May
2007.
URISA Journal • Elaksher, Bethel 13
Lee, D. S., J. Shan, and J. S. Bethel. 2003. Class-guided building
extraction from IKONOS imagery. Photogrammetric
Engineering and Remote Sensing 69, no. 2: 143-50.
Lin, C., and R. Nevatia. 1998. Building detection and description
from a single intensity image. Computer Vision and Image
Understanding, 72, no. 2: 101-21.
Masaharu, H., and K. Ohtsubo. 2002. A filtering method
of airborne laser scanner data for complex terrain. The
International Archives of Photogrammetry, Remote Sensing,
and Spatial Information Sciences, Graz, Austria, XXXIV
(3B), 165-9.
Mikhail, E., J. Bethel, and J. McGlone. 2001. Introduction
to modern photogrammetry. New York: John Wiley and
Sons.
Morgan, M., and A. Habib. 2002. Interpolation of LIDAR and
automatic building extraction. In Proceedings of the 2002
ASPRS, Washington, D.C., April 2002 (on CD-ROM).
Muller, S., and D. W. Zaum. 2005. Robust building detection in
aerial images. The International Archives of Photogrammetry,
Remote Sensing, and Spatial Information Sciences, XXXVI
(3/W24), Vienna, Austria, August 2005 (on CD-ROM).
Principe, J., N. Euliano, and W. Lefebvre. 1999. Neural and
adaptive systems: Fundamentals through simulation. New
York: John Wiley and Sons.
Rottensteiner, F. 2000. Semiautomatic building recognition
integrated in strict bundle block adjustment. The International
Archives of Photogrammetry, Remote Sensing, and Spatial
Information Sciences, Amsterdam, The Netherlands, July
2000, XXXIII (B2), 463-8.
Samet, H. 1982. Neighbor finding techniques for images
represented by quadtrees. Computer Graphics and Image
Processing 18, no. 1: 37-57.
Shmid, C., and A. Zisserman. 2000. The geometry and matching
of lines and curves over multiple views. IJCV 40, no. 3:
199-233.
Tarsha-Kurdi, F., T. Landes, P. Grussenmeyer, and E. Smigiel.
2006. New approach for automatic detection of buildings
in airborne laser scanner data using frst echo only. The
International Archives of Photogrammetry, Remote
Sensing, and Spatial Information Sciences, Bonn, Germany,
September 2006, XXXV (I/3) (on CD-ROM).
Theng, L. B. 2006. Semi-automatic building extraction utilizing
Quickbird imagery. Journal of Engineering Letters 13, no.
3 (November 2006), http://www/engineeringletters.com/
issues_v13/is sue_3/.
Tse, R. O. C., C. M. Gold, and D. Kidner. 2006. A new approach
to urban modeling based on LIDAR. In Proceedings of
Winter School of Computer Sciences, the 14th International
Conference in Central Europe on Computer Graphics,
Visualization and Computer Vision, Pilsen, Czech Republic,
January-February 2006, 279-86.
Wang, S. D., and Y. H. Tseng. 2004. Semi-automated CSG
model-based building extraction from photogrammetric
images. The International Archives of Photogrammetry,
Remote Sensing, and Spatial Information Sciences, Istanbul,
Turkey, July 2004, XXXV (B3) (on CD-ROM).
Wack, R., and A. Wimmer. 2002. Digital terrain models from
airborne laser scanner data—a grid based approach. The
International Archives of Photogrammetry, Remote Sensing,
and Spatial Information Sciences, Graz, Austria, XXXIV
(3B), 293-6.
TM
URISA Journal • Ghosh, Manson 15
Keywords: Land use, urbanization, robust principal component
analysis, geographically weighted regression
IntroductIon
We have long altered the land by clearing forests, farming, and
building settlements. Tis land change has serious social and
environmental impacts, many of which are increasingly evident
in urban areas that now host the majority of the world’s popula-
tion. In the United States, urbanization is driven primarily by
suburbanization or decentralized, low-density residential land
use, and creation of far-fung suburbs or exurbanization. While
suburbanization ofers important benefts such as afordable
housing, it also has negative impacts on systems ranging from
transportation to natural habitat to infrastructure efciencies to
inner-city economies (Burchell et al. 1998, Daniels 1999, EPA
2001).
The magnitude and nature of urbanization impacts are tied
not only to the amount of land converted to urban use but also
to its spatial confguration and pattern (IGBP-IHDP 1995).
Dispersed urbanization, for example, creates infrastructure inef-
fciency by spreading out roads or sewer networks. Despite the
importance of spatial patterning in determining impacts of urban-
ization, a good deal of urban research focuses on aggregate mea-
sures such as commute time or population density (Galster et al.
2001). Though this synoptic view is a critical avenue for research,
it may not capture the temporal and fne-scaled spatial patterns
and processes of urbanization (Hasse and Lathrop 2003).
A variety of approaches meet the need to examine and
model land use at fne spatial scales, and to these we add a new
one. Methodologies range from simple mathematical formulas
and gravity models to sophisticated spatiotemporal simulations
(Kaimowitz and Angelsen 1998, Lambin 1994, Parker et al.
2003). In this paper, we present a hybrid approach—robust
robust Principal component Analysis and Geographically
Weighted regression: urbanization in the twin cities
metropolitan Area of minnesota
Debarchana Ghosh and Steven M. Manson
Abstract: In this paper, we present a hybrid approach, robust principal component geographically weighted regression (RP-
CGWR), in examining urbanization as a function of both extant urban land use and the effect of social and environmental
factors in the Twin Cities Metropolitan Area (TCMA) of Minnesota. We used remotely sensed data to treat urbanization via
the proxy of impervious surface. We then integrated two different methods, robust principal component analysis (RPCA) and
geographically weighted regression (GWR) to create an innovative approach to model urbanization. The RPCGWR results show
signifcant spatial heterogeneity in the relationships between proportion of impervious surface and the explanatory factors in
the TCMA. We link this heterogeneity to the “sprawling” nature of urban land use that has moved outward from the core Twin
Cities through to their suburbs and exurbs.
principal component geographically weighted regression
(RPCGWR)—to examine both the location of urban land use
and the relative infuence of socioeconomic, demographic, policy,
and environmental factors. We integrate two different methods,
robust principal component analysis (RPCA) and geographically
weighted regression (GWR) to create a novel alternative to stan-
dard statistical approaches. First, to reduce the dimensions and
number of primary regressors, we applied principal component
analysis (PCA) to the explanatory variables. To account for the
infuence of outliers in standard PCA, we conducted a robust
principal component analysis (RPCA) by employing a projection
pursuit approach. Second, to capture spatial heterogeneity in the
urban landscape, we conducted GWR on the robust principal
components (RPCs). We compared the results of the RPCGWR
with a standard global principal component regression (RPCGR)
and used a series of visual and statistical comparisons to better
understand how RPCGWR lends insight into the complex dy-
namics of urban land use.
study ArEA And BAckGround
Urbanization has profound implications for the environmental
and socioeconomic sustainability of communities such as the Twin
Cities Metropolitan Area (TCMA) of Minnesota (see Figure 1).
Tis 7,700 km
2
seven-county area is the economic hub of a mul-
tistate region. Home to 2.8 million people, it is forecasted to top
3.5 million by 2020. It is also a major center of sprawl, the rapid
expansion of low-density suburbs into formerly rural areas and the
creation of urban, suburban, and exurb agglomerations bufered
from others by undeveloped land. Te metropolitan region also
has seen a marked increase in sprawl and associated aspects such
as trafc congestion (CEE 1999, Schrank and Lomax 2004).
The TCMA is an ideal setting for examining land use. The
region exemplifes the spatial and temporal dynamics of urban-
16 URISA Journal • Vol. 20, No. 1 • 2008
ization in the United States. It serves as the hub for a large geo-
graphic area and stands in relative isolation from other large urban
agglomerations, making it easier to extract land-use dynamics at
the metropolitan scale. It also is important to understand the
role of the region’s distinctive policy setting in shaping land-use
patterns because jurisdictions nationwide are wrestling with the
balance between local laissez-faire dynamics and tightly controlled
regional development (Pendall et al. 2002).
mEtHodoloGy
Urban researchers and policy makers have long used statistical
regression analysis to understand factors important to urbaniza-
tion. Tis approach creates a mathematical model of the relation-
ship between some measure, termed the response variable, and a
series of explanatory variables. In the case of the TCMA, we can
assess the relationship between urbanization in a given location
(percentage of impervious surface) as a function of potential
environmental, socioeconomic, and demographic factors, and
policy variables for that location (measured by the explanatory
variables). We expanded on this general approach by using robust
PCA (RPCA), which acts on the large number of explanatory
variables to create several key robust principal components (RPCs)
that serve as composite variables (Li 1985). We then conducted
geographically weighted regression (GWR) with the RPCs to
capture spatial heterogeneity in the relationship between urban
land use and the selected principal components.
Data
We identifed a broad slate of explanatory factors important to
understanding urbanization in the TCMA through a theoretical
analysis of the land-use literature and consultation with experts in
the community, in particular the staf of the Metropolitan Council
(see Table 1). Te council is a comprehensive regional planning
framework that coordinates the land-use activities of the region’s
272 local units of government, including 188 townships.
Theories of relative space focus on the broader spatial orga-
nization of social and environmental factors that affect decision
making, particularly through returns to land that vary with their
distance to phenomena that, in turn, affect input costs or output
prices (e.g., bid-rent Alonso and Von Thünen circles) (Alonso
1964, Bockstael 1996). Such phenomena include proximity to
employment centers, infrastructure, and locations with aesthetic
or recreational value. For the TCMA, of particular importance is
access to key infrastructure including sewerage, primary highways,
and surface roads. Also critical is cost-distance to the core Twin
Cities (Minneapolis and St. Paul), distance to the nearest park,
water bodies, shopping centers, and urban agglomeration. We
included proximity to airplane noise as a nuisance factor that
discourages urban development. Euclidian distance to a feature
is denoted (D); for example, SEWERD is distance to the nearest
sewerage. We also calculated cost-distance surface to the nearest
feature as a function of highways (D1) and surface roads (D2).
In this case, we calculated two different cost-distance surfaces that
vary according to whether highways or surface streets are used to
arrive at the feature of interest. SHOPD1, for example, denotes
the cost-distance (measured in time as a function of distance trav-
eled on the network) to the nearest shopping center via highways,
while SHOPD2 denotes cost-distance via surface streets.
On the other hand, related theories of absolute space consider
local neighborhood characteristics such as population density,
tax rates, existing land cover, lot size, and school district quality
(e.g., the Ricardian view of economic activity) (Bockstael 1996,
Irwin 2002). Absolute demographic and socioeconomic factors
that bear on urbanization in the TCMA include median income,
nonwhite population, population density, school test scores, and
prevalence of subsidized school lunches as a measure of neighbor-
hood characteristics (Bayoh et al. 2006, Soliani and Rossi 1992,
Vincent 2006). Key political and policy institutions include
agricultural and natural areas protection programs, county gov-
ernments, which control some development costs, and the Met-
ropolitan Council’s Metropolitan Urban Services Area (MUSA),
which enforces planned growth policies. Environmental factors
include soil quality, bedrock depth, elevation, and slope, which,
in turn, affect ease of construction, the potential for aesthetic
views, and competition for land. Table 1 includes the variable
name, description, and data sources for all the explanatory factors
included in the study.
roBust PrIncIPAl
comPonEnt AnAlysIs (rPcA)
We used an RPCA to reduce the number of explanatory variables
examined in the GWR. Tere are two advantages in regressing the
response variable (impervious surface) against RPCs rather than
directly on the explanatory variables. First, explanatory variables
often are highly correlated with one another (multicollinearity),
which may cause inaccurate estimations of regression coefcients
or poorly behaved covariance matrices when estimating a stan-
dard regression model. One solution to this problem, dropping
variables, can be at odds with the need to keep theoretically valid
and distinct explanatory variables in the model, such as distance
to parks and distance to water, which are conceptually separate
but often highly correlated. Because the RPCs are uncorrelated,
multicollinearity can be avoided by using the RPCs in place of
Figure 1. Twin Cities Metropolitan study site and percentage of
impervious surface
URISA Journal • Ghosh, Manson 17
the original explanatory variables. Second, extracting a subset of
RPCs for prediction reduces the dimensionality of the regressors.
In the case of the TCMA, ongoing collaboration with local ofcials
and researchers has identifed a large number of potential explana-
tory variables (see Table 1). Tus, we used an RPCA to reduce
the dimensionality of the problem—the number of explanatory
variables—and accommodate for multicollinearity.
Classical PCA is vulnerable to outlying observations. As
even a single large outlier can heavily infuence the parameter
estimates of PCA, we used a method termed projection pursuit,
or the notion that while most projections (combinations of de-
rivatives of data) are of low-order complexity or largely normal
Gaussian distribution, a few combinations or derivatives will
offer far-from-Gaussian distributions or high-order complexity
(Croux 1996, Li 1985).
For given observations x
1
,…..,x
n
Є IR
p
, collected in the rows
of the data matrix X, a coeffcient vector b Є IR
p
is defned for
one-dimensional projection of the data. We assume that the frst
k – 1 projection directions y
1
,……,y
k – 1
(eigenvectors) ( k > 1)
have already been measured. For fnding the kth eigenvalue, a
projection matrix is defned as
(1)
for projection on the orthogonal complement of the space
spanned by the frst k – 1 eigenvectors (for k = 1 we can take P

k
= I
p
). The kth eigenvector then is defned by maximizing the
function b
®
S (X P
k
b) under the conditions b
T
b.
To extract RPCs, we have to select a subset k < p of com-
ponents through a three-part process. First, through sequential
selection, we include k RPCs that explain approximately 90
percent variation of the data set. Second, we select k RPCs with
eigenvalues greater than one. Third, we analyze the “scree plot,”
which graphs the eigenvalues (expressed as explained variance) by
each RPCs as a line diagram (detailed later in this paper).
GEoGrAPHIcAlly WEIGHtEd
rEGrEssIon of roBust
PrIncIPAl comPonEnts
With the selected k RPCs, we conducted a geographically
weighted regression (GWR) analysis of land use. GWR ofers a
number of advantages over standard regression. A typical least-
squares regression model of the form:
(2)
is a “global” regression, which assumes that the relationship
between the explanatory variables and the response variable is
constant everywhere in the study area. In many situations, this is
not necessarily true, especially with spatial varying variables. GWR
extends the traditional global regression framework (equation 2)
by allowing local rather than global parameters to be estimated.
In this case, the model in equation 2 is rewritten as:
Table 1. Response and explanatory variables in the study
Variables Description Source*
TCMAIMP
Impervious surfaces in the TCMA
(dependent variable)
MC
AGRIPROT
Enrollment in the agricultural pro-
tection program
MC
BEDROCK Bedrock height MC
COUNTY County (7 counties, dummy variables) NHGIS
ELEV Elevation DNR
HWYD1
Highway, cost-distance to nearest
via main freeways
MC
HWYD2
Highway, cost-distance via surface
streets
MC
INCOME
Median income by census block
group
NHGIS
MAC1995D
Airport 65db noise contour, dis-
tance to (1995)
MAC
MAC2006D
Airport 65db noise contour, dis-
tance to (2006)
MAC
MPLSD1
Minneapolis, cost-distance to via
main freeways
MC
MPLSD2
Minneapolis, cost-distance to via
surface streets
MC
MUSA Metropolitan Urban Services Area MC
NONWHITE
Nonwhite population by census
block group
NHGIS
PARKD1 Park, cost-distance to nearest via
main freeways
DNR,
DOT
PARKD2
Park, cost-distance to nearest via
surface streets
DNR,
DOT
POP
Population density by census block
group (persons/km2)
NHGIS
PROTECTD Protected areas, distance to nearest MC
SCHENG English test scores, by school district MC
SCHLNCH
Subsidized lunch programs, by
school district (%)
MC
SCHMATH Math test scores, by school district MC
SEWERD Sewerage, distance to nearest MC
SHOPD1
Shopping center, cost-distance to
nearest via main freeways
MC
SHOPD2
Shopping center cost-distance to
nearest via surface streets
MC
SLOPE Slope MC
SOIL Soil types (3 types, dummy variables) DNR,
USGS
STPAULD1
St. Paul, cost-distance to via main
freeways
MC
STPAULD2
St. Paul, cost-distance to via surface
streets
MC
TCD1
City, cost-distance to via main
freeways
MC
TCD2
City, cost-distance to nearest via
surface streets
MC
WATERD Water body > 3 acres (distance to) DNR
*Metropolitan Airports Commission (MAC), Metropolitan Council
(MC), Minnesota Department of Natural Resources (DNR), Min-
nesota Department of Transport (DOT), National Historical GIS
(NHGIS), Remote Sensing and Geospatial Analysis Laboratory (RS-
GAL), and U.S. Geological Survey (USGS)
18 URISA Journal • Vol. 20, No. 1 • 2008
(3)
where (u
i,
v
i
) denotes the coordinates of the ith point in space
and β
k
(u
i
,v
i
) is a realization of the continuous function β
k
(u,v)
at point i. Equation 3 creates a continuous surface of estimated
parameter values, and measurements of this surface are taken at
certain points to denote the spatial variability of the surface. This
spatial variability is estimated through the geographical weight-
ing scheme, W(u
i,
v
i
), defned such that data points nearer to (u
i,
v
i
) will be assigned higher weights in the model than data points
farther away. That is,
(3)
where the bold type denotes a matrix that represents an
estimate of , and W(u
i,
v
i
) is an n by n matrix whose off-diagonal
elements are zero and whose diagonal elements denote the geo-
graphical weighting of each of the observed data for regression
point i (Fotheringham 2002). The resulting parameter estimates
then can be mapped to analyze local variations in the estimated
parameter relationships. Various diagnostic measures further
increase the analytical capability of GWR, such as the Akaike
Information Criterion (AIC), local standard errors, local measures
of infuence, and local goodness of ft. As examined later on, the
parameter estimates also are tested for evidence of signifcant
spatial variation relative to the global model. Figure 2 summarizes
the steps involved in the methodology in a schematic diagram.
rEsults And dIscussIon
As mentioned previously, we engaged in a three-step process.
We frst extracted RPCs, and analyzed component loadings
and clustering of initial explanatory variables in the component
space. Second, we used the RPCs in a standard global regression
in what we term a robust principal component global regression
(RPCGR), where we model the response variable, proportion of
impervious surface, against the selected RPCs. Tird, we exam-
ined diferences between the results of RPCGR and a robust prin-
cipal component geographically weighted regression (RPCGWR).
We used the GWR 3.0 software package (Fotheringham 2002)
and R statistical software (Robust PCA and Projection Pursuit,
pcaPP package) for statistical analysis and Arc GIS 9.1 (ESRI)
for calibrating the RPCGWR model components and visualiza-
tion of the results.

Figure 2. Schematic diagram showing methodological framework
Table 2. Total variance explained by the robust principal components
Principal
Component
Eigenvalue Variation
(%)
Cumulative
Variation
(%)
1 7.32 74.80 74.80
2 1.10 13.07 87.87
3 0.45 4.72 92.59
4 0.41 2.89 95.48
5 0.22 1.96 97.44
6 0.15 1.01 98.45
7 0.09 0.63 99.08
Figure 3. Scree plot for RPCA on TCMA variables
URISA Journal • Ghosh, Manson 19
percent, the second 13 percent, and the third explains 5 percent
of the variance (see Table 2). Tere is, therefore, a steep drop in
the percentage of explained variance after the frst RPC.
This drop also is evident in a scree diagram, which plots
the eigenvalues (variances) of the RPCs on the y-axis against the
RPC number on the x-axis (see Figure 3). The term scree refers
to the fact that the explained-variance curve resembles the side of
a mountain with a scree, or rock debris, at the base. When read
left to right across the abscissa, this plot shows a clear separation
between RPCs with high-explained variance versus low-explained
variance. The point of separation is termed the elbow for obvi-
ous reasons that nonetheless invite a justifed charge of mixed
metaphors.
In concordance with Table 2 and Figure 3, the frst RPC
explains the large majority of variation, the second less so, and
the third a small amount. The elbow occurs at the third RPC,
indicating the separation of the most important RPCs from less
important RPCs, namely the fourth onward. Thus, we retained
the frst three RPCs as explanatory variables for further analy-
sis.
The key opportunity, and challenge, of PCA is determin-
ing what the components actually mean in a real-world setting.
Component loadings indicate the relative contribution of the
variables to each component (seeTable 3).
In addition to examining the degree of correspondence
between components and individual variables (Table 3), we also
can look for clustering in component space. Figure 4 is a three-
dimensional graph that shows the position of explanatory variables
with high component loadings in the component space. The
graph identifes how these variables relate to both the principal
Table 3. Explanatory variable loadings onto individual selected robust
principal components
Variable PC1 PC2 PC3
AGRIPROT 0.000 0.000 0.000
BEDROCK 0.000 0.000 -0.002
COUNTY 0.000 0.000 0.000
ELEV 0.000 0.000 0.000
HWYD1 0.222 0.098 -0.379
HWYD2 0.217 0.121 -0.384
INCOME -0.070 0.976 0.058
MAC1995D 0.211 0.054 0.508
MAC2006D 0.209 0.052 0.52
MPLSD1 0.323 0.069 -0.255
MPLSD2 0.488 0.038 -0.178
MUSA 0.000 0.000 0.000
NONWHITE 0.001 0.002 0.002
PARKD1 0.056 -0.003 -0.107
PARKD2 0.055 -0.007 -0.096
POP -0.009 -0.011 -0.003
PROTECTD 0.000 0.000 0.000
SCHENG 0.000 0.000 0.000
SCHLNCH 0.000 0.000 0.000
SCHMATH 0.000 0.000 0.000
SEWERD 0.127 -0.002 -0.083
SHOPD1 0.266 0.032 -0.253
SHOPD2 0.312 -0.037 -0.157
SLOPE 0.000 0.000 0.000
SOIL 0.000 0.000 0.000
STPAULD1 0.34 0.029 -0.137
STPAULD2 0.538 -0.064 0.178
TCD1 0.333 0.048 -0.21
TCD2 0.418 -0.019 -0.049
WATERD 0.012 -0.016 -0.022
Robust Principal Component Analysis Results
RPCA using the projection pursuit approach extracted three un-
derlying dimensions from the 30 explanatory variables expected
to infuence urban development in the TCMA (see Table 1).
Table 2 shows both the eigenvalue and the raw and cumulative
percentage of variance explained by the extracted RPCs that ac-
count for 99 percent of the total variation. Te frst three RPCs
account for 93 percent of the total variation. Te frst explains 75
Figure 4. Variables in three-dimensional component space
20 URISA Journal • Vol. 20, No. 1 • 2008
components and other input variables.
The frst component, RPC1, has a number of variables with
high loadings. The variables, in descending order, are cost-dis-
tance surfaces by the second-order roads to St. Paul (STPAULD2),
Minneapolis (MPLSD2), and to the nearest of the two cities
(TCD2). These are followed by cost-distance surface by frst-order
roads to St. Paul (STPAULD1), Minneapolis (MPLSD1), and to
the nearest of the two cities by highways (TCD1). The seventh
and eighth variables are the two cost-distances to the nearest
shopping center by highway and surface streets (SHOPD1 and
SHOPD2). These variables are positively related to RPC1, or, in
other words, observations with higher values of RPC1 also will
indicate higher values of all the variables mentioned previously
and vice versa. Because these variables measure cost-distances to
key urban centers, markets, and infrastructure, we termed RPC1
as the cost-distance factor.
Figure 5 illustrates the spatial variation of the cost-distance
factor (RPC1) in the TCMA region along with the explanatory
variables that load highly onto this component. Figure 5b shows
the infuence of cost-distances to the Twin Cities and major shop-
ping centers. The cost-distance factor is least near the center or the
two central business districts (CBDs) of the TCMA and increases
gradually outward from the center to the suburbs and then to the
outer suburbs. Not surprisingly, the initial cost-distance variables
also show a similar spatial pattern (see Figure 5a).
The second component, RPC2, has median income by
block group (INCOME) with a very high component loading
of 0.976, almost double that of any other variable/component
loading combination. INCOME is positively related to RPC2,
indicating that higher values of RPC2 are associated with higher
values of INCOME, leading us to term RPC2 as the “income
factor.” Figure 6 demonstrates the strong correspondence in
spatial variation between the initial variable, INCOME (Figure
6a), and the income-factor, PC2 (Figure 6b). Other indicators
that often (but not always) map onto socioeconomic status and
neighborhood characteristics, such as school lunch programs or
ethnicity, are almost completely subsumed by INCOME.
The third component, RPC3, captures a more complicated
situation than those associated with RPC1 and RPC2. Unlike
the other components, RPC3 has both high positive and nega-
tive loadings (see Figure 7). The two strongest positive loadings
are past and present distance to the airport 65db noise contour
(MAC1995D, MAC2006D), followed by a lower loading
on cost-distance to St. Paul via surface streets (STPAULD2).
The remaining variables have small negative loadings on PC3.
These include cost-distances to the nearest highway (HWYD1
and HWYD2), cost-distances to Minneapolis (MPLSD1
Figure 5. Cost-distance factor (RPC1)
Figure 6. Income factor (RPC2)
Figure 7. Infrastructure factor (RPC3)
URISA Journal • Ghosh, Manson 21
and MPLSD2), cost-distances to the nearest shopping center
(SHOPD1 and SHOPD2), cost-distances to the nearest of the
two cities (TCD1), and cost-distance to the nearest park via
surface streets (PARKD2).
RPC3 is more diffcult to interpret than the frst two com-
ponents because of the small amount of variance explained by
RPC3 and the mixed nature of explanatory variables contributing
to it. Although it is not immediately obvious from the loadings
on this component, we defne PC3 as the “infrastructure factor,”
because as will be explored in the following section, this factor
captures path-dependence on urban form exerted by exiting
infrastructure.
Robust Principal Component Global Regression
(RPCGR)
We used the three RPCs as explanatory variables in a robust
principal component global regression (RPCGR) to form a basis
for comparison to robust principal component geographically
weighted regression (RPCGWR). Table 4 shows the results of
a standard multiple linear regression model (again, termed a
global model to distinguish it from GWR). We estimated the
model (and the RPCGWR that follows) with 7,000 observations
sampled across the TCMA. Te RPCGR model is signifcant for
all components (R
2
= 0.217, p > 0.01 for PC1-PC3).
The proportion of impervious surface in TCMA is nega-
tively related to the cost-distance factor (RPC1) and income
factor (RPC2) but positively related to the infrastructure factor
(RPC3). Regions with a higher proportion of impervious surface
have lower values of the cost-distance factor and the income fac-
tor. In contrast, the association between RPC3 and impervious
surface is positive; as PC3 increases, the percentage of impervious
surface also increases.
The global regression model explains only 22 percent of the
variance in the percentage of impervious surface as a function
of the RPCs, which indicates that the model does not account
for all the factors infuencing urban development in the TCMA.
Beyond missing variables, however, the low explained variance
also can be attributed at least in part to the fact that the estimated
parameters represent global averages of relationships between im-
pervious surface and the RPCs that may exhibit spatial variation
(Fotheringham 2002). In other words, some of the unexplained
variance may be associated with the assumption of spatial station-
arity underlying the global regression model. Theories of relative
space, mentioned previously, contend that the intensity of urban
development declines with increasing distance from central cities,
in this case, Minneapolis and St. Paul. While the global regres-
sion identifed this relationship between urban development and
distance from a central city, it may fail to identify local variations
in the power of this relationship. Thus, if the relationship be-
tween an explanatory variable and urban development is spatially
nonstationary, then the global multiple regression approach can
misspecify the actual relationship (Fotheringham 2002). One
approach for analyzing these local variations is RPCGWR.
Robust Principal Component Geographically
Weighted Regression
As noted previously, we term the combination of RPCA and
GWR as robust principal component geographically weighted
regression (RPCGWR). To assess the efectiveness of RPCGWR,
we compared its performance to that of global regression analysis
with the same variables, proportion of impervious surface versus
the three RPCs that in turn condensed 30 explanatory variables
shown in Table 1.
With GWR, the parameter estimation at any of the sample
observations depends not only on the input data but also on the
kernel (model form) chosen and the kernel’s bandwidth (spatial
extent of the sample). We used a Gaussian model because the
response variable is continuous. The bandwidth was optimized
as a part of the GWR calibration using the AIC method, which
balances the complexity of the estimated model (defned by how
specifc the bandwidth becomes) with the extent to which the
model fts the data (defned by explained variance). GWR uses
a Monte Carlo simulation to test the following hypotheses: (1)
whether the data may be described by a GWR model rather than
a stationary one and (2) whether individual regression coeffcients
are stable over geographic space (Fotheringham 2002).
A comparison of regression parameters illustrates that RP-
CGWR outperforms RPCGR for this study. The AIC is reduced
from 64,554.63 for the global regression model to 63,871.10 for
the RPCGWR model, where a lower AIC indicates a more ef-
fcient model. RPCGWR has an R
2
of 0.49, which is reasonably
high, especially compared to an R
2
of 0.22 for RPCGR. RP-
Table 4. Global robust principal component regression (RPCGR)
analysis results
Parameter Coefcients
Value
Std. Error T-Value
Intercept 16.033 0.566 28.346
PC1 -4.852 0.146 -33.119
PC2 -2.356 0.247 -9.530
PC3 1.601 0.416 3.847
Figure 8. Spatial variation in infuence of cost-distance factor (RPC1)
22 URISA Journal • Vol. 20, No. 1 • 2008
CGWR also has a much lower residual sum of squares (284,179
versus 3,282,543). Finally, we can compare RPCGWR against
RPCGWR via an ANOVA, where an F of 6.69 at the p < 0.01
level indicates that RPCGWR signifcantly improves on the global
regression model.
In spatial terms, the Monte Carlo test of the local estimates
for each of the three RPCs (cost-distance, income, and infrastruc-
ture factors) indicates signifcant spatial variation for a bandwidth
of approximately three miles. The spatial pattern of the estimated
cost-distance factor (RPC1) is shown in Figure 8.
Areas marked by higher absolute values indicate regions
where the explanatory variables with higher component loading
under RPC1 have a greater infuence, while areas of lower values
indicate where the explanatory variables are less infuential. As
noted previously, the global relationship between RPC1 and the
proportion of impervious surface is negative, which suggests that
urban development is more likely to occur closer (lower cost-
distance) to the major cites of St. Paul and Minneapolis, either
considered singly or in terms of whichever is closest to a given
location (STPAULD1/2, MPLSD1/2, and TCD1/2, respec-
tively), the nearest shopping center (SHOPD1/2), or highway
(HWYD1/2).
RPCGWR shows that the contribution of RPC1 parameter
in the regression equation varies over the study region, including
a change in sign from negative to positive, which indicates that
this relationship is more complex than is suggested by the global
regression results. We can assess the statistical signifcance of the
spatial variation by examining t-values at the observations. Values
falling beyond 1.96 (i.e., the 95 percent confdence interval) are
considered signifcant (Fotheringham 2002). According to this
rubric, there are several areas in the TCMA where a signifcant
positive relationship exits between impervious surfaces and the
cost-distance factor. Figure 8b identifes the areas of the TCMA
in which the range of values in Figure 8a constitute signifcant
spatial variation. Areas of particular signifcance are located in
Anoka and Washington counties in the northwest and Carver
and Scott counties in the southwest (also see Figure 1). This said,
within these counties, there are areas where the inverse is true,
particularly northern Anoka County, central Carver County, and
northwestern Washington County. These subareas show develop-
ment in the suburbs away from the CBD and Twin Cities and
confrm the “sprawling” nature of urbanization in the TCMA.
The global relationship between the proportion of impervi-
ous surface and the income factor (RPC2) is signifcantly positive.
Figure 9a shows that this positive relationship holds over most of
the study area because the majority of the local parameters also are
positive. However, as we move away from the central region of the
TCMA to the peripheral regions, especially in the northwestern
corner of the Washington County (Figure 9 and Figure 1) and
the inner-city region of the Twin Cities of Minneapolis and St.
Paul, the relationship becomes negative. This negative relation of
income factor and percentage of impervious surface highlights a
typical inner-city scenario witnessed in almost all metropolitan
regions of the country. Growing population, cheap housing poli-
cies, and the dense nature of transportation network are some of
the factors that changed the land to concrete urban impervious
surface but have simultaneously discouraged the reinvestment
and redevelopment of older inner-city communities. This creates
a situation of “negative growth,” affecting the processes of land
use and environment around the metropolitan perimeter.
Figure 9b indicates that the negative relation between
percentage of impervious surface and the income factor in the
inner-city region of the Twin Cities and parts of northwestern
Washington County are statistically signifcant.
The spatial pattern of RPC3, the infrastructure factor, is
shown in Figure 10a. As noted previously in terms of the load-
Figure 9. Spatial variation in infuence of income factor (RPC2)
URISA Journal • Ghosh, Manson 23
Figure 10. Spatial variation in infuence historical high-density factor (RPC3)
ings on RPC3, there are strong positive association with both
the past and present distance to the airport’s 65db noise contour
(MAC1995D, MAC2006D) and weak positive associations
with cost-distance to St. Paul via surface streets (STPAULD2).
Conversely, there are weak negative associations between pro-
portion of impervious surface and cost-distances to the nearest
highway (HWYD1 and HWYD2), Minneapolis (MPLSD1 and
MPLSD2), nearest shopping center (SHOPD1 and SHOPD2),
nearest of the two cities (TCD1), cost-distance to St. Paul via
freeways (STPAULD2), and cost-distance to the nearest park via
surface streets (PARKD2). An examination of local parameters
from RPCGWR, however, shows a distinct spatial pattern in the
changing nature of this relationship. PC3 is positively related to
impervious surface in the eastern part of Hennepin County, the
entirety of Ramsey and Anoka counties, and the northern part
of Dakota County (Figure 10a). These regions are characterized
by high values of airport noise (MAC1995D, MAC2006D)
and low values for cost-distances to infrastructure and markets
(HWYD1/2, MPLSD1/2, SHOPD1/2, TCD1).
We originally included airplane noise given anecdotal
evidence in the TCMA, supported by studies elsewhere (Nelson
et al. 2004), that land development is less likely to increase in
areas of high noise, limited indirectly through consumer prefer-
ences and directly through policy limits on land use. While the
negative impact of airplane noise on urban development may be
true in many other study areas, the region around the airport on
the border between Hennepin and Dakota Counties contained
some of the longest settled neighborhoods of the TCMA. These
areas and the adjacent suburbs were developed decades before
the airport was built in 1962. These areas also contributed to
further development of key infrastructure facilities, including rail,
tram, and road networks and are home to the highest density of
detached single-family dwellings, which, in turn, feature smaller
backyards than elsewhere in the TCMA and an attendant dense
street grid and alley network.
The spatial patterning of RPC3, while driven in large part
by areas of past development that map onto the spatial pattern
of airport noise, also refects negative loadings for several of the
cost-distance measures. In particular, the majority of the areas
with an apparently negative relationship with cost-distance fall in
largely rural areas that lie far from current infrastructure (highways
or markets) or in recent suburban developments that boast lower
amounts of imperviousness with their large lot sizes and irregular
street grids. As seen in Figure 10a, these areas lie in the northern
and western parts of the TCMA.
conclusIon
Tis study has two types of conclusions, specifc and applied.
Te specifc conclusions are derived directly from the RPCGWR
model and its interpretations in the context of the TCMA. Tese
conclusions contribute to the literature of spatial analysis and
modeling of urbanization in geography. On the other hand, the
applied conclusions are the inputs to the policy community of the
Metropolitan Council, the Minnesota Pollution Control Agency,
and other governmental agencies in the TCMA. Te following
paragraphs will expand on the specifc conclusions frst, followed
by the applied conclusions.
We introduce a hybrid approach, RPCGWR, to examine
the relationships between urban development, as approximated
by impervious surface, and myriad explanatory social and envi-
ronmental factors in the TCMA. This approach allows us to sift
through a large number of potential factors and identify several
key composite factors in the form of RPCs. RPCA, in addition to
accounting for outliers in the original data set, is useful for reduc-
ing the complexity of explanatory variables, as seen in reducing
the initial 30 explanatory variables to three robust components
24 URISA Journal • Vol. 20, No. 1 • 2008
that account for 93 percent of the total variation.
It is important to note, however, that RPCA does not have
to divorce the components from the original factors, but, instead,
can serve as a way of aggregating and understanding relationships
among these factors. The frst RPC was defned as the cost-
distance factor because it aggregates cost-distance measures for city
centers, the nearest highways, and shopping centers. The second
RPC collapses myriad socioeconomic and demographic factors
into a single measure, income. The third component tells a more
complicated story because, in addition to contributing a small
amount of explained variance, it captures the effect of historical
infrastructure development in heavily urban areas and relative
paucity of infrastructure in other, more rural areas.
We then go one step further by using GWR to better under-
stand the spatial variation in the RPCs as such and their relation-
ship with both urban development and the original explanatory
factors. We demonstrate that, for the TCMA, RPCGWR out-
performed RPCGR as a tool for examining spatially varying
relationships. RPCGWR sheds light on how the importance of
socioeconomic and environmental factors can vary over space,
in essence adding context to theories of land use by highlighting
areas where these factors play out differently. This approach also
identifes areas for further investigation. The spatial pattern of
the cost-distance component identifed areas away from the Twin
Cities, for example, where signifcant positive relationships exist
between cost-distances and the proportion of impervious surface
contrary to theoretical expectations and trends elsewhere in the
region. Similarly, even though the global relationship between the
proportion of impervious surface and the income factor (RPC2) is
signifcantly negative, RPCGWR located patches outside the Twin
Cities where there is a signifcant positive relationship between
impervious surfaces and income. This spatial heterogeneity is
linked to the differences in the nature of urban land use between
the core Twin Cities and their suburbs and exurbs.
From the application point of view, this work is the frst step
in developing trajectories of future land change for the Twin Cit-
ies. This aspect of the research is tied to the policy community
through ongoing consultation with experts from the Metropolitan
Council. The mission of this body is to develop, in coordina-
tion with local communities, a comprehensive regional planning
framework that focuses on transportation and aviation, wastewater
treatment and water resources, regional parks, and regional devel-
opment and infrastructure. Its chief goal is to plan and coordinate
effcient and sustainable growth of the metropolitan area. We are
exploring two key linkages to the Metropolitan Council. First,
we have been working with the Metropolitan Council and the
Minnesota Pollution Control Agency, for these agencies fnd the
maps to be useful inputs to hydrology and runoff models and
planning future development. Satellite remote sensing provides
a cost-effective alternative for obtaining such information when
the costs of traditional mapping approaches are increasing and
budgets are declining. Second, working with staff at the Metro-
politan Council and with the archives maintained by the council,
we could identify likely scenarios of growth, tied to population
and socioeconomic forecasts, paying particular attention to the
role of policy instruments such as zoning, transportation, and
infrastructure provision.
Acknowledgments
Tis work is supported in part by the National Aeronautics and
Space Administration New Investigator Program in Earth-Sun
System Science (NNX06AE85G), the University of Minnesota’s
College of Liberal Arts Graduate Research Partnership Program,
and the McKnight Land-Grant Professorship Program. Te
authors gratefully acknowledge the assistance of the editor and
anonymous reviewers. Responsibility for the opinions expressed
herein is solely that of the authors.
About the Authors
Debarchana Ghosh is a Ph.D. candidate in the Department of
Geography, University of Minnesota. Her research interests
include quantitative analysis, spatial statistics and modeling,
GIS, health, and environmental geography.
Corresponding Address:
University of Minnesota
Department of Geography
414 Social Sciences
267 - 19th Avenue South
Minneapolis, MN 55455
Phone: (612) 625-6080
Fax: (612) 624-1044
ghos0033@umn.edu
Dr. Steven M. Manson is an associate professor in the
Department of Geography, University of Minnesota. He
combines environmental research, social science, and
geographic information science to understand changing
urban and rural landscapes in the United States and Mexico.
This work is part of his longer-term research on global
environmental change, decision making, and understanding
complex human-environment systems.
URISA Journal • Ghosh, Manson 25
References
Alonso, W. 1964. Location and land use. Cambridge, MA:
Harvard University Press.
Bayoh, I., E. G. Irwin, and T. Haab. 2006. Determinants of
residential location choice: How important are local public
goods in attracting homeowners to central city locations?
Journal of Regional Science 46: 97-120.
Bockstael, N. E. 1996. Modeling economics and space: the
importance of a spatial perspective. American Journal of
Agricultural Economics 78: 1168-80.
Burchell, R. W., N. A. Shad, D. Listokin, H. Phillips, A. Downs,
S. Seskin, J. S. Davis, T. Moore, D. Helton, and M. Gall.
1998. Costs of sprawl: Revisited. Washington, D.C.:
National Academy Press.
CEE. 1999. Two roads diverge: Analyzing growth scenarios
for the Twin Cities. Minneapolis: Center for Energy and
Environment.
Croux, C., and A. Ruiz-Gazen. 1996. A fast algorithm for
robust principal components based on projection pursuit.
In Computational statistics. Heidelberg: Physica-Verlag:
211-16.
Daniels, T. 1999. When city and country collide: Managing
growth in the metropolitan fringe. Washington, D.C.:
Island Press.
EPA. 2001. Our built and natural environments. Washington,
D.C.
Fotheringham, A. S., C. Brundson, and M. Charlton. 2002.
Geographically weighted regression—the analysis of spatially
varying relationship. West Sussex, England: John Wiley &
Sons Ltd.
Galster, G., R. Hanson, M. R. Ratcliffe, H. Wolman, S. Coleman,
and J. Freihage. 2001. Wrestling sprawl to the ground:
Defning and measuring an elusive concept. Housing policy
debate 12: 681.
Hasse, J. E., and R. G. Lathrop. 2003. Land resource impact
indicators of urban sprawl. Applied Geography 23:
159-75.
IGBP-IHDP. 1995. Land-use and land-cover change: Science/
research plan. Stockholm: International Geosphere-
Biosphere Programme and International Human Dimensions
Programme.
Irwin, E. 2002. The effects of open space on residential property
values. Land Economics 78: 465-81.
Kaimowitz, D., and A. Angelsen. 1998. Economic models
of tropical deforestation: A review. Jakarta: Centre for
International Forestry Research.
Lambin, E. F. 1994. Modelling deforestation processes: A review.
Luxembourg: European Commission.
Li, G., and Z. Chen. 1985. Projection-pursuit approach to robust
dispersion matrices and principal components: Primary
theory and Monte Carlo. Journal of the American Statitstical
Association 80: 759-66.
Nelson, A. C., R. J. Burby, E. Feser, C. J. Dawkins, E. E. Malizia,
and R. Quercia. 2004. Urban containment and central city
revitalization. Journal of the American Planning Association
70: 411-25.
Parker, D. C., S. M. Manson, M. Janssen, M. J. Hoffmann, and
P. J. Deadman. 2003. Multi-agent systems for the simulation
of land use and land cover change: a review. Annals of the
Association of American Geographers 93: 316-40.
Pendall, R., J. Martin, and W. Fulton. 2002. Holding the line:
Urban containment in the United States. Washington, D.C.:
Brookings Institute.
Schrank, D., and T. Lomax. 2004. The 2004 urban mobility report.
College Station, TX: Texas Transportation Institute.
Soliani, L., and O. Rossi. 1992. Demographic factors and
land-use planning in the small islands of southern Europe.
Environmental Management 16: 603-11.
Vincent, J. M. 2006. Public schools as public infrastructure—roles
for planning researchers. Journal of Planning Education and
Research 25: 433-37.
URISA Journal • Maghelal, Olivares, Wunneburger, Roman 27
IntroductIon
Several studies in the past decade have analyzed sexual abuse
on males and females under the age of 18 in the United States
(Tjaden and Toennes 1998, Greenfeld 1997). Finkelhor
(1994) informed that one in fve females and one in seven males
are sexually abused by the age of 18. Te fear, both personal
and altruistic, of becoming a victim of sexual abuse consciously
or semiconsciously exists in the community. Tis fear has been
rekindled even more with the recent unfortunate events of sex-
related crimes throughout the nation.
In an attempt to redeem neighborhoods of these mishaps,
law-enforcement agencies have regulated various sex offender
restriction statutes that can help manage the risk posed by sex
offenders. While numerous statutes have been in place for about
four decades now, the Jacob Wetterling Crimes Against Children
and Sexually Violent Offenders Registration Program
1
(42 U.S.C.
14071 et seq.) of 1994 reshaped the way law enforcers managed
Registered Sex Offenders (RSOs) in the United States. This law
required convicted sex offenders to register and notify their law
administrators of their movement. Information about offenders
such as each offender’s name, age, gender, height, weight, race,
and details of the offense are provided to the state authorities such
as the State Department of Public Safety.
After the death of Megan Kanka at the hands of a convicted
sex offender living across the street in New Jersey, President
Clinton signed an amendment to this law, requiring all states to
make the information about pedophiles and rapists available to
the general public (Beck and Travis 2004, Engeler 2005). When
this law was signed in May of 1996, the local citizenry was and
continues to be informed of the whereabouts of sex offenders in
1. Details of the Jacob Wetterling Crimes Against Children and
Sexually Violent Offenders Registration Program can be obtained
at the Cornell University Law School U.S. Code collection
at http://www.law.cornell.edu/uscode/html/uscode42/usc_
sec_42_00014071----000-.html.
Where Are they? A spatial Inquiry of sex offenders in
Brazos county
Praveen Maghelal, Miriam Olivares, Douglas Wunneburger, and Gustavo Roman
Abstract: The United States has several laws that restrict the movement of registered sex offenders. The majority of these laws
are spatial in nature. However, only a few studies have investigated the use of spatial technology (geographic information system)
to analyze the implications of these laws. This study uses GIS to map and identify the offenders who violate the laws policed by
federal agencies. Also, the perceived risk because of the sex offenders’ existence in the community is mapped to assist law-enforcement
agencies identify the potential suspects when sex crimes are reported. Spatial analysis revealed that more than 50 percent of the
offenders resided within the restriction distance of the places where children congregate. Immediate and lateral risk zones were
created around each offender to measure the risk each offender brings to the community in which he or she resides. Finally, this
study proposes the use of spatial technology to communicate sex-related crimes to increase the awareness of communities at risk
of sex-crime victimization.
their community. This notifcation system exists in all the states,
and makes it mandatory for the offenders to inform the respec-
tive state authorities about their movements anywhere in the
United States. This information then is made public to notify
the communities of the offenders’ details. The Jacob Wetterling
act sets minimum standards by federal administration for states.
Individual states, on the other hand, can impose more stringent
requirements on the offenders. In Texas, the Code of Criminal
Procedure, SB1054, Article 42.12, Section 13B (Texas Legislature
Online, 78th Legislature) mandates the Child Safety Zone (CSZ)
for the state of Texas to be “within 1,000 feet of premises such
as school, day-care facility, playground, public or private youth
center, public swimming pool, or video arcade facility, places
where children generally gather.” Currently, the state of Texas
stipulates anywhere from 200 feet to 1,000 feet for this zone,
which follows the drug-free zone restrictions used in the state.
This study investigates the locations of sex offenders’ residences
with respect to the CSZ using a standard 1,000-foot buffer (as
mandated by the Texas legislature and currently under discussion
in the legislature
2
) around the child facilities on proximity to an
RSO, area of risk owing to the RSO presence, and how such
information can be communicated to the general public.
The movement of RSOs within and between different states
with varying restriction laws makes it diffcult for the offenders
and the supervising authorities to exactly determine the distance
between the residences of the offenders and the CSZ. However,
current trends in modern technology such as using a geographic
information system (GIS) have made it feasible to closely super-
vise the mobility restrictions of the registered sex offenders. GIS
2. The existing Texas legislature requires a distance of 1,000 feet.
However, the Board of Pardon and the judiciary system can decide
this distance case by case. However, Martha Wong, Texas state
representative, has moved for an amendment that mandates that all
offenders be subjected to a 1,000-foot distance throughout the state
of Texas. Details regarding the House Bill 1828 can be found on the
Texas legislature Website at http://www.capitol.state.tx.us/.
28 URISA Journal • Vol. 20, No. 1 • 2008
provides a powerful tool to map these locations, effciently update
the data, and frequently check for violators residing in the CSZ.
This study uses the current restriction laws stipulated by the Texas
legislature to inquire: (1) How many sex offenders reside within
the Child Safety Zone (1,000-foot buffer)? (2) how to identify
the known offenders in closest proximity to where a victim is
reported missing? and (3) how can this digital mapping system
help notify and bring awareness to the local community?
BAckGround
Felson and Clark (1998), in their analysis, reported that all crimes
are a result of available opportunities. Te laws that have been
enacted to reduce sex crimes attempt to reduce these opportuni-
ties as much as possible. Cohen and Felson (1979) stated that
the increase in opportunities of crime is because of the presence
of three elements: (1) a suitable target, (2) absence of a guard-
ian, and (3) a motivated ofender. Tis is defned as the Routine
Activity Teory (RAT). Terefore, occurrence of sex crimes can
be reduced or checked when a guardian is aware of the pres-
ence of a motivated ofender who may attack a suitable victim.
Megan’s Law, in a sense, performs a similar task for society. Te
law-enforcement agencies inform the respective communities of
the presence of a sex ofender through the notifcation system to
increase the awareness of individuals living in close proximity to
the ofender. As a result, these acts help to avoid or reduce the
chances of a motivated ofender meeting a suitable target in space
and time. Extensive literature that analyzes proximity and sex
crime are reported in Walker, Golden, and VanHouten (2001).
Similarly, to monitor the adjudicated offenders, agencies
such as the Parole Board and the Board of Pardon have devel-
oped restrictions that control the mobility of the offenders in
the community. These restrictions have received wide attention
recently.
Robertson (2000, 109) suggested that:
… an understanding of geographic trends of registered sex
offenders, especially as they relate to schools and daycare
facilities, may help police narrow their suspect lists in open
cases, to those individuals contained within their registration
database, living within a close proximity to the victim
procurement site, who pose a high risk of recidivism.
Although RAT indicates that there is a relationship between
proximity and repeated sex crimes, Levenson and Cotter (2005)
reported that there is lack of empirical research in this area.
Canter and Larkin (1993) proposed the commuter and marauder
hypothesis based on the components of proximity and crime.
They proposed that commuters travel to commit crimes in other
areas beyond their homes and the marauder criminals use the areas
around their homes to commit crimes. Crime and proximity in
relation to Child Safety Zone (commuters) has been investigated
spatially by Walker et al. (2001). Their study analyzed the prox-
imity of sex offenders and the potential victims using GIS. They
used the Arkansas Code of 1,000-foot buffer and identifed the
offenders who lived within this distance. They also investigated
the number of offenders who resided within 1,000 feet from
such premises as schools, parks, and day-care centers. In several
cases, sex offenders were found to be living in close proximity to
the premises where children congregate. Also, just about half of
the offenders (47 percent) were reported to be living within the
1,000-foot restriction area.
Although the impact of Megan’s Law, though spatial in na-
ture, has not been investigated spatially in relation to proximity
and crime (marauder), its effect has been reviewed for its mode
of notifcation by Thomas (2003). His study reported that the
methods of disseminating the community notifcation (Megan’s
Law) was through leafets or fyers, community notifcation meet-
ings, and other means such as the media, “need-to-know” basis,
and marked-car licenses in various communities across the United
States. While the exact distance traveled by an offender to commit
a sex crime remains to be investigated, Megan’s Law is based on
the premise that an individual in immediate and close proximity
to a sex offender is at risk of victimization and thus needs to be
notifed about sex offender presence in the community. In Texas,
this act requires notifcation
3
(Texas Criminal Procedures Code
Annotated Section 62.201(a)) by mail to households within the
zone of infuence, defned in two levels depending on the location
of the urban or rural setting of an RSO’s dwelling: immediate
risk—within three city blocks (or approximately 0.33 miles) in
urban or subdivided settings, and lateral risk—within one mile
in rural settings (areas outside subdivisions).
Conversely, Terry Thomas (2001) reported that the Sex
Offender Act of 1997 in the United Kingdom had unforeseen
consequences. The implication of this law required setting up
a mechanism of registration to help the police decide the risk
caused by a dangerous sex offender. This idea of registration of
sex offenders was based on three arguments: (1) to help the police
identify the suspects after a crime, (2) to prevent crime, and (3)
to act as a deterrent (Home Offce, 107). The registration laws
in United States serve the same purpose. However, residential
restrictions for the sex offenders vary from state to state. In Illinois,
the least restrictive distance is 500 feet, while California restricts
the dwelling of sex offenders within a quarter mile of schools
(Levenson and Cotter 2005). Texas law allows these restrictions
to vary for each offender, depending on the type of offense com-
mitted. Presently, the parole board assigns this distance based
on the individual’s offense. This distance can be anywhere up
to 1,000 feet. This makes it diffcult for the law-enforcement
offcials to keep a check on the movement of these offenders. To
alleviate these diffculties, the Texas legislature has been forwarded
a petition to make the restrictive distance of 1,000 feet uniform
throughout the state of Texas.
Nonetheless, the undisputed fact remains that the primary
aim of the notifcation system is to increase the vigilantism of
communities. One of the more recent methods of increasing
vigilantism against crime is through Web-based technologies.
3. Details regarding the notifcation system in Texas are available at
http://tlo2.tlc.state.tx.us/statutes/cr.toc.htm, titled “Article 62.201.
Additional public notice for individuals subject to civil commitment.”
URISA Journal • Maghelal, Olivares, Wunneburger, Roman 29
Spatial technology, in addition to being used as a mapping tool
(e.g., Walker et al. 2001, Foote and Crum 1995), can be effectively
used as a communication tool using Web GIS. While Web GIS
has been used as a communication tool for some time now (e.g.,
Ramasubramanian 1995), its application to communicate sex
crimes has been increasingly advocated (e.g., Albrecht and Pingel
2005, Shyy, Stimson, Western, Murray, and Mazerolle 2005). It
is important to communicate information regarding sex crimes
because, as suggested by Sampson, Raudenbush, and Earls (1997),
it increases the “collective effcacy” of the community. They state
that an increase in collective effcacy of a community results in
lower expected levels of crime.
Therefore, this study provides a methodological approach to
identify and notify individuals about the sex offenders in their
community using GIS. Deriving from the RAT, sex crime, as a
result of proximity, can occur in two ways. In the absence of a
guardian by (1) proximity to probable or suitable victims and
(2) proximity to motivated offenders. While a detailed literature
review by Walker et al. (2001) assesses the issues of proximity to
suitable victims using GIS, proximity to motivated offenders has
been neglected as a cause of sex crime. The following investiga-
tion will help (1) identify the percentage of offenders who violate
the residential restrictions if and when 1,000-foot restriction is
approved; (2) identify the areas of probable crime in close vicin-
ity to the offenders’ homes; and (3) measure the effectiveness of
developing a Web-based GIS notifcation system that maps the
offenders and the restriction zones.
mEtHodoloGy
Study Area
Te area under study, Brazos County, Texas, has a population of
about 152,000 (U.S. Census 2000), comprised of the cities of
Bryan, College Station, and Wellborn. About 88 percent of the
population of Brazos County resides in the twin cities of Bryan
and College Station. Te Texas Department of Public Safety
(TXDPS) lists 164 registered sex ofenders in the zip codes of
Brazos County. Also included on this list are the name of each
ofender (including alias names), date of birth, gender, race, cur-
rent residential address, information pertinent to the ofense, and
latest photograph with other information. Without a geographical
system in place to track registered sex ofenders, the cities of Col-
lege Station and Bryan have not been able to check the violators
who reside within the Child Safety Zones for a long time. Tus,
it was necessary to provide the law-enforcement authorities with
tools to help them locate such violators residing in the neighbor-
hoods within child safety zones.
Spatial Inquiry into the Location of Sex Offenders
Te spatial information for Brazos County was provided by the
city of Bryan Information Technology (IT) Department. Te two
main themes created for this analysis were (1) the Child Safety
Zone and (2) the location of residence of each ofender.
This study required the geocoding of day-care facilities in
Brazos County obtained from the Department of Family and Pro-
tective Services, and schools and parks in Brazos County. Schools,
parks, and day-care centers in Brazos County were geocoded
using parcel-level data. These are the locations where children
generally gather and were used as basic themes to develop the
CSZ. The address information of the RSOs was obtained from
the Texas Department of Public Safety’s Sex Offender Database.
The spatial data of Brazos County parcels was used to geocode
(single-feld (fle)) the “USaddress” feld with the address database
fles of the day-care centers and the registered sex offenders in
Brazos County.
Matching interactively, the unmatched addresses of the of-
fenders were searched and selected. About 12 of the 164 addresses
of the sex offenders were either located out of the Brazos County
or could not be located in the Brazos parcels fle and thus were
not used for further analysis. The layers with information on
parks and schools in Brazos County were buffered for a distance
of 1,000 feet. These layers were appended and merged together
to form the new dissolved layer of all the buffers that formed the
Child Safety Zone. Playgrounds, public or private youth centers,
and public swimming pools are a part of the schools in Bryan and
College Station. Once the sex offender locations were geocoded,
spatial query was made to locate the offenders residing in the
CSZ (see Figure 1).

Locating Known Offenders
Risk-assessment tools to predict risk have been investigated for
some time now (Hanson and Tornton 2000, Tornton et al.
2003). Te Static 99 risk-assessment tool generates four categories
of risk: low, low-medium, medium-high, and high, and has been
validated by Beech, Friendship, Erikson, and Hanson (2002)
and Tornton (2002). By using the Texas Case Classifcation
and Risk-Assessment tool, community supervision ofcers have
established three diferent levels of risks associated with each
Figure 1. Methodology used to develop the tool to locate sex
offenders
30 URISA Journal • Vol. 20, No. 1 • 2008
registered sex ofender. Te risk levels based on the nature of the
crime are high, medium, and low (Texas Department of Criminal
Justice Website). Tese risk levels are assigned by the Department
of Corrections, the Department of Social and Health Service, and
the Sentence Review Board. For high-risk ofenders, the TXDPS
is required to send postcards to residents in the one-mile radius
of a nonsubdivided area and a three-block radius of a subdivided
neighborhood within seven days of release and ten days of move
of a sex ofender to their neighborhood. Tis is because high-risk
ofenders are considered the most probable convicts to reofend.
Even though only the moving in of a high-risk ofender requires
notifcation to the community, every ofender induces a certain
level of perception of risk in the community where he or she lives.
Terefore, the Critical Risk Zones (CRZs) are classifed based on
the risk level of an ofender as high, moderate, and low and by
proximity as immediate and lateral risk.
Currently, law-enforcement agencies buffer the location
where a victim is reported missing and identify the offenders
within the buffer to check the possible reoffenders of the reported
crime (e.g., Hubbs 2003). This method, however productive, does
not allow the law-enforcement offcials to identify the closest
offender with the highest level of risk. Including the dimension
of proximity and the level of risk in this search can help the of-
fcials search for the suspect beginning with the closest high-risk
offender to the farthest low-risk offender, possibly minimizing
the time to fnd the suspects most probable of committing the
reported crime.
This study utilized the standards required by Megan’s Law
as the baseline to geographically analyze the victim procurement
site. The Critical Risk Zone of each sex offender was based on
the distances specifed in Megan’s Law. Two zones: immediate risk
zone, a three-block distance from the residence of the offender,
and lateral risk zone, a one-mile distance from the residence of the
registered sex offender, were created as the “area of infuence” for
each offender. These were termed the “Critical Risk Zones.”
Community Notifcation
Te spatial mapping technique to locate sex ofenders can be
a useful community notifcation tool. Terefore, a Web-based
GIS interface was developed and launched at the city of Bryan
police department Website. Tis Website can aid the individuals
of the community to access the spatial georeferenced information
regarding registered sex ofenders and child safety zones. Public
access to this service was monitored for number of hits to assess
if individuals of the Brazos County used this service, resulting in
improved collective efcacy of the community.
rEsult
Te descriptive analysis of the registered sex ofenders residing in
Brazos County, Texas, showed that 73 percent of the ofenders
were white. More than 10 percent of ofenders have committed a
crime at least two or more times; about 5 percent of the ofenders
were females; and more than 10 percent of the ofenders were
high-risk ofenders. About 44.2 percent of the ofenders were 15
to 30 years old when the crime was committed, and 35 percent
of them were 30 to 45 years old. More than 50 percent of the
victims were 15 to 25 years old, and more than 80 percent (131)
of the victims were females.
Spatial Inquiry into Location of Sex Offenders
Te ofenders in the CSZ (77 of 164) were categorized based on
their risk levels (see Figure 2). Tirty-eight were identifed as low-
risk ofenders, 27 as moderate-risk ofenders, and 12 as high-risk
ofenders. An investigation, similar to that conducted by Walker
et al. (2001), revealed that six (50 percent) of these high-risk
ofenders were within a 1,000-foot proximity of at least one day
care, four were near schools, and 11 were in close proximity to
parks in the cities of Bryan and College Station. Four ofenders
were identifed within 1,000 feet of at least one day care and one
park in Bryan. One high-risk ofender, who had been charged
with indecency with females age 12 and 14 in 1982 and females
age 13 in 1990, resides within 1,000 feet of one day care, one
park, and one school in Bryan.
The spatial query showed that an alarmingly high percent-
age (55.41 percent) of offenders resided within the CSZ. The
proximity, as shown in Figure 2, of the offenders to the schools,
Figure 2. Mapping of residential location of sex offenders in relation
to schools, parks, and day-care centers in Brazos County
URISA Journal • Maghelal, Olivares, Wunneburger, Roman 31
parks, and day-care centers in Bryan/College Station was not
in adherence with the state restriction of 1,000 feet in Texas.
Although this percentage may vary with continuous moving in
or out of the offenders in Brazos County, the fndings at this
snapshot of time reveal that a high percentage of these offenders
reside within the CSZ.
Locating Known Offenders
Te zones were classifed based on the risk level and proximity,
according to the following six divisions: (i) Low Immediate Risk
Zones, (ii) Low Lateral Risk Zones, (iii) Moderate Immediate
Risk Zones, (iv) Moderate Lateral Risk Zones, (v) High Immedi-
ate Risk Zones, and (vi) High Lateral Risk Zones (see Figure 3).
Tese zones may or may not overlap for two or more ofenders
based on the distance between each other. Using GIS, the loca-
tion where the child was reported missing can be georeferenced.
Upon identifying that location, a list of registered sex ofenders
who lie within the immediate risk zone and lateral risk zone can
be generated for the purpose of investigation.
This risk-level analysis provides a platform from where the
authorities can identify the registered offender residing in the
closest proximity of a reported victim, or provide some indication
as to where to direct the investigations after a victim is reported
missing.
Community Notifcation
Te press release of the new Web-based GIS service was an-
nounced on May 5
,
2005. Access to this Website was monitored
and automatically recorded to measure the number and sources
of hits (see Figure 4). Te hits on the city of Bryan RSO Website
shared about 50 percent of all the hits on the city of Bryan Website
immediately after the press release of the new Web-based tool.
Also, May of 2005 reported three times the total hits (67,777)
compared to all other months except the holiday month of
December 2004. Although the large increase in access to this
Website can be attributed to the press release, the monitoring and
assessment of total number of hits on the Website in future can
show if the individuals of the community accessed the Website
to constantly update themselves. Such assessment can indicate
increased communication of sex crime–related information
through the Web-based GIS service. Except for December 2004
(holiday season), there was an increase in access to the Bryan
County Website in May of 2005 when the Web-based service
was launched.
dIscussIon
Sex-crime analysis, like any other crime analysis, is associated with
the notion of place with a geographical location. Occurrence of
crime has a spatial dimension that has been explored since the
1970s (Chainey and Ratclife 2005). Using GIS to analyze the
sex-crime occurrence now is advocated more than ever (Grubesic,
Mack, and Murray 2007). Tus, it is important to use spatial
technology efciently to analyze the occurrence of crime and as
a communication tool to disseminate sex crime–related informa-
tion. Tis study used GIS to analyze the location of sex ofenders
within the CSZ. Enforcement of a 1,000-foot bufer for CSZ
would result in a high percentage of ofenders being in violation
of the law and require relocation by the authorities to avoid hav-
ing them in close proximity to children. However, irrespective
of where the ofenders reside, they bring some amount of risk to
the community. GIS can help map that risk using existing law
as a framework to help locate potential ofenders in proximity
to the location where a sex crime was committed. Advancement
in spatial technology allows dissemination of sex crime–related
information to individuals of the community to increase their
awareness about existing ofenders in the neighborhood.
The high percentage (more than 55 percent) of violators liv-
ing in the CSZ can be a concern for the local community. It has to
be noted that this high percentage was due to the fact that these
Figure 3. Critical risk zones (both immediate and lateral) for each offender based on each one’s risk
32 URISA Journal • Vol. 20, No. 1 • 2008
Summary by Month
Month
Daily Average Monthly Totals
Hits Files Pages Visits Sites KBytes Visits Pages Files Hits
May 2005 67777 47263 15354 459 4939 14997006 8739 291740 898001 1287773
Apr 2005 24242 15410 5354 213 3219 11878090 6410 160623 462325 727261
Mar 2005 23776 14236 6270 226 3495 9754452 7033 194376 441340 737066
Feb 2005 22161 13352 5563 262 3456 10149913 7362 155774 373879 620522
Jan 2005 21700 13195 5621 356 4625 9163086 11064 174271 409075 672723
Dec 2004 58305 23736 27928 957 16659 18585619 29684 865796 735838 1807461
Nov 2004 15938 10612 3633 146 1814 6621826 4399 109006 318375 478166
Oct 2004 16401 11292 3709 144 2001 6963752 4478 114988 350074 508444
Sep 2004 15797 11035 3564 127 2062 7617430 3816 106937 331051 473912
Aug 2004 15989 10820 3642 127 1926 42513587 3942 112922 335437 495687
Jul 2004 15598 10979 3533 135 1866 7155400 4209 109529 340363 483545
Jun 2004 14743 10325 3455 120 1845 6884369 3623 103675 309772 442290
Totals 152284530 94759 2499637 5305530 8734850
Figure 4. Chart and table of total number of hits to access the sex offender Web service
offenders currently reside based on the restrictions that do not
follow the 1,000-foot distance. Cross-verifcation of the ethnicity
of offenders with their current photos on the Website revealed that
a high percentage of “White” offenders were reported because the
general classifcation was categorized as Black or White. Hispanic-
looking individuals were classifed as White as well. Nevertheless,
at the discretion of the authorities, the violators can be notifed
to relocate and also suggested to locate outside the CSZ, yet still
be accessible to their jobs.
The CRZ can help law-enforcement offcers identify the
registered offenders in the closest proximity to a victim reported
missing or at the location of crime. This spatial service can be
made available to the law-enforcement agencies to investigate a
reported crime. The categories of risk used here are not empiri-
cally evolved, but have been used commonly by several states in
the United States.
The Web-based GIS service was available for local com-
munity members to increase awareness among the community
about registered sex offenders’ residence locations in reference to
their homes, workplaces, as well as places they visit on a regular
basis. The Website provides a sense of geographical reference to
law enforcers and to the local community. Increased access to
the Website indicates increased public awareness and interest in
using Web-based mapping services as a communication tool.
This Web-based service helps individuals relate the location of
sex offenders with their residences and the paths their children
take to commute to school.
conclusIon
Tis study investigated the use of spatial technology to map and
communicate sex crime–related information. GIS was used to
map the locations of places where children congregate, such as
parks, schools, and day-care centers, and the residences of regis-
tered sex ofenders in Brazos County, Texas. Based on the locations
where children congregate, more than half of the ofenders were
found to be in violation of the restriction distance mandated by
the Texas legislature. Also, ofenders were mapped for the risk
they are perceived to bring to the community. Tis area under risk
for each ofender was mapped based on the notifcation system
and level of risk, allowing law enforcers to identify the ofenders
with greatest risk and closest proximity to a reported sex crime.
While spatial technology can be used to map a sex crime, it also
can help communicate this information visually to the individuals
of the community. Tis is evident from the analysis in this study
when the Web-based sex-ofender service was monitored for the
number of hits after its induction. Te methodology proposed in
this article also can be an efective tool and more importantly cost-
efective and time-efective in managing risk. With the availability
of GIS, geocoding of ofenders’ residences, and the availability of
themes such as parks and schools, sex crimes related to children
can be efciently managed. Terefore, the present study hopes
to encourage the use of GIS technology in investigating crime-
management strategies, especially related to sex crimes.
One primary limitation of this study is the use of a common
1,000-foot buffer to create the CSZ. The restriction distance var-
ies between each state. In Brazos County, the restriction distance
URISA Journal • Maghelal, Olivares, Wunneburger, Roman 33
varies with the individual offender. The signifcance of the distance
is more perceived than empirically tested. Therefore, a future
scope of this study would include empirical examination of the
infuence of restriction distance on the risk to the community us-
ing GIS. Nonetheless, this study demonstrated the use of GIS to
effciently map and communicate the information related to sex
crimes to the residents of the community and assist law enforcers
to regulate sex crimes in their jurisdictions.
About The Authors
Dr. Praveen Maghelal is an assistant professor in the
Department of Urban and Regional Planning at Florida
Atlantic University and has educational background in
civil engineering, architecture, and planning. His research
interest includes spatial planning, physical activity and built
environment, and transportation planning.
Corresponding Address:
111 East Las Olas Boulevard
Department of Urban and Regional Planning
Florida Atlantic University
Fort Lauderdale, Florida 33301
Phone: (954) 726-5030
Fax: (954) 762-5673
pmaghela@fau.edu
Miriam Olivares is a Ph.D. candidate in Urban and Regional
Science at Texas A&M University, from where she earned a
master’s degree in land development. She holds a bachelor’s
degree in architecture with emphasis on planning from
Monterrey Tech, Mexico. Her research interest is in
sustainable development. Currently, she is working on her
dissertation regarding sustainable communities and sex-crime
management.
Dr. Douglas Wunneburger is a senior lecturer in the Department
of Landscape Architecture and Urban Planning at Texas
A&M University. His primary research interests include the
integration of spatial and information technology for studies
in landscape ecology-based planning and management.
Gustavo Roman is the Director of Information Technology for
the City of Bryan, Texas. He holds master’s and bachelor’s
degrees from Texas A&M University and has more than
12 years of municipal government experience, including
seven-plus years in the implementation and management
of GIS systems.
References
Albrecht, J., and J. Pingel. 2005. GIS as a communication process:
Experiences from the Milwaukee COMPASS project. In
F. Wang, ed., Geographic information systems and crime
analysis. Hershey, PA: Idea Group Publishing, 1-23.
Beck, V. S., and L. F. Travis III. 2004. Sex offender notifcation
and fear of victimization. Journal of Criminal Justice 32:
455-63.
Beech, A., C. Friendship, M. Erikson, and R. Hanson. 2002.
The relationship between static and dynamic risk factors and
reconviction in a sample of UK child abusers. Sexual Abuse:
A Journal of Research and Treatment 14: 155-67.
Canter, D. V., and P. Larkin. 1993. The environmental range
of serial rapists. Journal of Environmental Psychology 13:
63-69.
Chainey, S., and J. Ratcliffe. 2005. GIS and crime mapping.
Chichester: John Wiley and Sons.
Cohen, L. E., and M. Felson. 1979. Social change and crime rate
trends: A routine activity approach. American Sociological
Review 44: 588-608.
Engeler, A. 2005. Is your child a target? The sex offender next
door. Good Housekeeping 240, no. 5: 192-97.
Finkelhor, D. 1994. Current information on the scope and nature
of child sexual abuse. Child Abuse and Neglect 4: 31-53.
Felson, M., and R. V. Clarke. 1998. Opportunity makes the
thief: Practical theory for crime prevention. Police Research
Series, Paper 98. Home Offce, Policing and Reducing Crime
Unit. Great Britain: Research, Development and Statistics
Directorate.
Foote, K., and S. Crum. 1995. Cartographic communication
(November 17, 2007), http://www.colorado.edu/geography/
gcraft/notes/cartocom/cartocom_f.html.
Greenfeld, L. 1997. Sex offenses and offenders: An analysis of
data on rape and sexual assault. Washington D.C.: U.S.
Department of Justice, Bureau of Justice Statistics.
Grubesic, T. H., E. Mack, and A. T. Murray. 2007. Geographic
exclusion: Spatial analysis for evaluating the implications
of Megan’s Law. Social Science Computer Review 25, no.
2: 143-62.
Hanson, R., and D. Thornton. 2000. Improving risk assessments
for sex offenders: A comparison of three actuarial scales. Law
and Human Behavior 24: 119-36.
Hubbs, R. 2003. Mapping crime and community problems in
Knoxville, Tennessee. In M. R. Leipnik and D. R. Albert,
eds., GIS in law enforcement: Implementation issues and
case studies. London: Taylor & Francis.
Levenson, J. S., and L. P. Cotter. 2005. The impact of sex offender
residence restrictions: 1000 feet from danger or one step
from absurd? International Journal of Offender Therapy and
Comparative Criminology 49, no. 2: 168-78.
Ramasubramanian, L. 1995. Building communities: GIS and
participatory decision making. Journal of Urban Technology
3, no. 3: 67-79.
Robertson, L. S. 2000. Assessment of the proximity of registered
sex offenders to schools and day cares: St. Louis City,
Missouri, 1998. MO: Department of Administration of
Justice.
Sampson, R., S. Raudenbush, and F. Earls. 1997. Neighborhoods
and violent crime: A multilevel study of collective effcacy.
Science 277: 918-24.
34 URISA Journal • Vol. 20, No. 1 • 2008
Shyy, T-K., R. J. Stimson, J. Western, A. T. Murray, and L.
Mazerolle. 2005. WebGIS for mapping community crime
rates: Approaches and challenges. In F. Wang, ed., Geographic
information systems and crime analysis. Hershey, PA: Idea
Group Publishing, 14, 236-53.
Texas Department of Family and Protective Services, Child Care
Licensing Division (November 20, 2004), http://www.
tdprs.state.tx.us/Child_Care/About_Child_Care_Licensing/
default.asp.
Texas Department of Criminal Justice Website (January 10,
2005), http://www.tdcj.state.tx.us/cjad/cjad-who-sa-tool.
htm.
Texas Department of Public Safety (November 20, 2004), http://
www.txdps.state.tx.us/.
Texas Legislature Online, 78th Legislature (November 20,
2004), http://www.tdh.state.tx.us/hcqs/plc/csotlaws.
htm#legislation.
Tjaden, P., and N. Thoennes. 1998. Prevalence, incidence, and
consequences of violence against women: Findings from the
national violence against women survey. Washington, D.C.:
U.S. Department of Justice, National Institute of Justice.
Thomas, T. 2003. Sex offender community notification:
experiences from America. The Howard Journal 42, no. 3:
217-28.
Thomas, T. 2001. Sex offenders, the home offce, and the Sunday
papers. Journal of Social Welfare and Family Law 23, no.
1: 103-8.
Thornton, D., R. Mann, S. Webster, L. Blud, R. Travers, C.,
Friendship, and M. Erickson. 2003. Distinguishing and
combining risks for sexual and violent recidivism. Annals of
the New York Academy of Sciences 989: 225-35.
Thornton, D. 2002. Constructing and testing a framework for
dynamic risk assessment. Sexual Abuse: A Journal of Research
and Treatment 14: 139-53.
U.S. Census Bureau, http://www.census.gov.
Walker, J. T., J. W. Golden, and A. C. VanHouten. 2001. The
geographical link between sex offenders and potential
victims: A routine activities approach. Justice Research and
Policy 3, no. 22: 15-33.
URISA Journal • Trépanier, Chapleau, Morency 35
IntroductIon
Large household surveys always have presented a methodological
challenge for transportation planners and authorities. Conduct-
ing a survey of more than 70,000 households is not a simple task
because of the sample size and the complexity of the survey itself.
Every planner knows that transportation data are strongly related
to the spatial elements of a territory and to the transportation
network (roads and public transit), and that the survey tool must
take these specifcities into account. Today, even though intelligent
transportation systems (ITSs) have provided new ways to collect
data, large transportation surveys still are needed. Data collected
from these operations now are well integrated in the felds of
transportation planning, fnance, and management.
This paper presents the information technologies that were
used for the 2003 Greater Montreal Area Household Survey
(Quebec, Canada). It also emphasizes the technological back-
ground and architectures that were required to yield the best
results possible from the survey. Following a recounting of the
history of the household survey in the Montreal area, the totally
disaggregate approach and transportation object-oriented model-
ing, two key elements that helped support and develop the 2003
tools, are presented in the background section. The third part of
the paper, “Survey Information System Framework,” describes
the methodology that was used to prepare and synchronize the
various software programs and databases. The “Implementation”
section is aimed at demonstrating the functions of the software
that was used for the survey. The conclusion reports some fndings
on the 2003 experience in Montreal.
BAckGround
In the past, travel surveys were conducted mainly by mail or
face-to-face interviews. Tey basically provided data for the de-
velopment of aggregated travel forecasting models. Richardson
et al. (1995) propose a thorough description of classical Survey
tools And mEtHods for A trAnsPortAtIon
HousEHold surVEy
Martin Trépanier, Robert Chapleau, and Catherine Morency
Abstract: Nowadays, large transportation household surveys cannot be conducted without the help of powerful management
and support tools, and the information technologies are useful for preparing, conducting, and postanalyzing such surveys. In the
Greater Montreal Area (GMA), the 2003 household survey followed the general methodology that has been developed over the
past 20 years to integrate the fnest software, databases, and methods. The tools making up the household survey information
system (HSIS) are based on the totally disaggregate approach and its object-oriented extension. This paper presents the background
and the fundamentals of the Montreal 2003 survey information system, and describes the way in which it has been assembled,
illustrating the functional and technical architectures that were used. It also emphasizes the transposition of the method to other
transportation survey activities and planning tools. The fnal discussion stresses the “winning” elements involved in conducting
a modern transportation household survey successfully.
Methods for Transport planning. With the advent of new tech-
nologies, combining spatial information systems and computa-
tion capacities, travel surveys have become an integral part of the
continuing transportation planning process and assist many types
of transportation studies.
In the Transportation Research Board Millennium Paper of
the Committee on Travel Surveys Methods, Griffths et al. (2000)
identify future directions for travel survey methods:
The improvement of the quality standards of travel surveys •
through full and honest documentation of the survey process.
The need to document all stages of the survey process also
appears as the most overriding conclusion of a conference
held in 1997 on raising the standards of travel surveys
(Richardson 2002).
The use of mixed-mode survey designs to meet the data •
needs of the surveyor in ways that create the least burden
and the greatest fexibility for the respondents. The concept
of common cognitive space between an interviewer and a
respondent was outlined by Brög (2000). The purpose of
survey tools is to maximize this common space to facilitate
the exchange of information between the two agents and to
lessen the respondent burden.
A move toward a more continuous survey to provide more •
timely data in an economical manner, which also would
develop and preserve technical and managerial skills in the
conduct of complex surveys.
The judicious use of new technologies to augment existing •
survey techniques.
In this regard, computer-assisted telephone interviewing
(CATI) is one of the main felds of development regarding travel
surveys. It allows interviewers to administer a survey question-
naire via telephone and capture responses electronically. CATI
“employs interactive computing systems to assist interviewers
and their supervisors in performing the basic data-collection
36 URISA Journal • Vol. 20, No. 1 • 2008
tasks of telephone interview surveys” (Nicholls II 1988). It can
be viewed as a tool to facilitate or expedite telephone surveys,
to enhance and control survey data quality, and to allow new
types of surveys. Jones and Polak (1992) point to the ability to
combine the data collection and management functions as one
of the key advantages shown by CATI. For recent discussions,
the reader can refer to a report on Survey Automation Tools by
the National Research Council (2003) or to Couper et al. (1998)
who discuss Computer Assisted Survey Information Collection
(CASIC) methods.
As will be discussed in the following sections, the House-
hold Survey Information System of the 2003 Montreal Survey
integrates tools and functions to address these issues. Automatic
documentation of the survey process, synchronous/asynchronous
monitoring of interviews, transposition of the tools to other sur-
vey methods (postcoding of onboard surveys or self-completion
of survey questionnaires through private access via Internet),
and integration with implemented planning and operational
tools are some of the features of the presented household survey
information system.
Household Surveys in the Greater Montreal Area
Te history of origin-destination household surveys in Montreal
begins in the 1960s, when the frst large-scale survey in the re-
gion was conducted. Since 1970, eight large surveys have been
conducted, at fve- to six-year intervals (see Table 1). Te standard
survey method relies on the following principles:
The interviews are conducted by telephone, by agents •
specially recruited and trained for this purpose. The
telephone remains an effcient way to survey people, even
though there are problems of reach ability and nonresponse
with privacy protection systems and homes not equipped
with fxed phones (Westrick and Mount 2007, Link and
Kresnow 2006).
All the trips made on the previous day are collected for •
every person residing at the contacted household; details
regarding trip ends, times of departure, mode sequence,
and trip purposes are gathered. Surveys are trip-based and
relate to a single week-day. Even if emerging issues regarding
the substitution of out-of-home activities by in-home
activities are discussed in the literature, the metropolitan
steering committee on travel surveys sees no need to move
toward an activity-based survey for the main purpose of the
origin-destination surveys is to precisely measure the use of
transportation networks. Thus, it appears more important to
preserve comparability between successive surveys to measure
the evolution of trip patterns over the area. Moreover, totally
disaggregate functions allow the construction of activity
patterns from individual travel behaviors.
Generally, a unique respondent provides all the information •
regarding the trips made by all the members of the household.
Comparability issues also are dictating the continuity of this
methodological choice. As noted by Liss (2005) regarding
the National Household Travel Survey (NHTS), “the Proxy
reporting yields a lower trip rate than that of respondents
who are interviewed personally.” Badoe and Steuart (2002)
also discuss the potential bias caused by interviewing by proxy
respondents. Incidentally, the effects of proxy respondents on
trip rates are evaluated cyclically for the Montreal surveys.
The sample is approximately 5 percent of the residing •
population. This sample is drawn from a set of residential
phone numbers. Sampling strata are defned to monitor the
construction of the fnal sample through the overall interview
process (four months of survey).
The standard questionnaire is organized into three sections: •
households, people, and trips.
More details regarding the origin-destination surveys held
in the Montreal area can be found on the Web site of the Met-
ropolitan Information Centre on Urban Transportation (http://
www.cimtu.qc.ca/index.asp). The next survey in the Montreal
area will be conducted during the fall of 2008, with a proposed
budget of 1.8 millions CAD$ (Bergeron 2008).
Through the years, this process has evolved both technically
and methodologically.
Technical Evolution
Technically, the surveys have benefted from the evolution of
computer technology. Te frst survey data were posttreated with
computer programs running during weekends on large comput-
ers rented to sizable organizations such as the Montreal School
Board. In 1982, data was validated using computer procedures
that now form the basis for the well-known MADITUC (Modèle
D’analyse Désagrégée des Itinéraires de Transport Urbain Collec-
tif ) system and the totally disaggregate approach (both of which
are defned in the following sections). In 1987, survey data was
coded, geocoded, and validated with the help of microcomputers.
In 1993, a survey frm was contracted to conduct the survey, and
data was captured by means of its own in-house software, based
on the VAX system. Because of data postvalidation and survey
quality concerns, it was decided that computer-assisted interview-
ing software (CATI) would be developed for future surveys. In
1998 and 2003, a software suite was used, which combined the
best practices and procedures from past surveys. It also took ad-
vantage of the evolution of computer technology (both hardware
and software) that had occurred during this period, especially the
multi-tasking capabilities of Microsoft Windows.
Methodological Evolution
Most of the advances of the 1970–2000 period were method-
ological. Many aspects of survey methodology have evolved
since 1970:
Spatial zoning. • Prior to 1987, the territory was divided into
several zones (the transit equivalent of traffc analysis zones, or
TAZ), refecting the general usage in transportation planning
of synthesized and aggregate models for which little precision
is needed. This was also because of a lack of spatial search
engine capabilities in the survey tools. With the advent of the
URISA Journal • Trépanier, Chapleau, Morency 37
totally disaggregate approach, zones have been abandoned at
the coding level for a much higher spatial level of resolution.
In 1987, the Canadian postal code (corresponding to block
faces in urban areas), then considered the best means of
location defnition, was used. Now, every trip end is coded
at the X-Y coordinate level (in meters), which is the best
means available currently. In 1998 and 2003, every location
was stored and treated as well. For example, a trip generator
now can be identifed under many names.
Transit network defnition. • In the Greater Montreal Area,
household surveys tend to be oriented towards transit
planning usage and have been remarkably successful in
this feld. Moreover, transit network data for analysis have
become more and more precise over the years. Early on, the
network was specially coded for the survey. Now, a more
“real” representation of the transit network is derived from
operational data fles obtained from transit authorities.
Sampling and weighting (expansion). • Signifcant changes
have been made to sampling methods over the years. In the
beginning, expansion was based only on people. Today, it
is categorized by both people and households, and different
weights are given to people and households, depending on
their attributes (age and size, respectively).
Survey execution. • Initially, a survey would be conducted by
employees of the transit authorities because of the absence of
CATI software and the complexity of the task. Since 1993,
a specialized survey frm has been mandated to do this. The
frm provides expertise in conducting surveys, and in staff
and telephone infrastructure, but uses the CATI software
selected by the survey board committee.
The question as to whether or not to use specialized CATI
survey software has long been decided in the GMA. CATI pro-
vides the fexibility and the power that are needed in conducting
such a complex survey. It is complex because questions on house-
holds, people, and trips are interleaved with looping; every trip
end needs to be geocoded; online transit trip declarations must
be validated; and the validity of trip chains within the household
must be checked.
Totally Disaggregate Approach
Te totally disaggregate approach (TDA) was developed in the
1980s in the Greater Montreal Area for the validation, process-
ing, and modeling of large computer-assisted household origin-
destination surveys conducted by telephone interview. It was used
in particular to process transit usage declarations, but then was
extended to include other survey information. Typically, in 1998,
a telephone survey would involve more than 65,000 households
(5 percent overall sampling). To use such a quantity of data, even
a 1,500-zone system and its aggregate approach could not satisfy
planners (Chapleau 1986), so a new method had to be developed
Table 1. Comparative statistics for the past seven household surveys in Montreal
Year 1974 1978 1982 1987 1993 1998 2003
Total area 2,331 km2 2,331 km2 3,341 km2 3,350 km2 4,500 km2 5,300 km2 6,445 km2
Population 2,824 000 2,954,000 2,895,000 2,900,000 3,263,000 3,493,000 3,505,810
Sampling rate 4.8% 5.3% 7.0% 5.0% 4.7% 4.5% 4.8%
Surveyed
households
43,000 50,000 75,000 54,000 61,000 65,000 70,000
Surveyed trips 265,000 305,000 492,000 338,000 350,000 380,000 388,000
Zoning sys-
tem/ Geocodes
1,192 zones 1,264 zones 1,496 zones 70,000 PC 30,000 TG
70,000 PC
9,000 SN
40,000 IN
44,600 TG
100,000 AR
89,000 PC
34,000 SN
191,000 IN
77,800 TG
160,000 AR
119,000 PC
40,200 SN
201,000 IN
AR: Address ranges, IN: Intersections, PC: Postal codes, SN: Street names, TG: Trip generators
Figure 1. Three-dimensional transit load profle of A.M. peak period,
Montreal 1998 household survey
38 URISA Journal • Vol. 20, No. 1 • 2008
to store and process data on households, people, and trips. Set-
ting aside its many features and special functionalities, the TDA
is briefy defned here by its two essential elements:
Individual trip data processing throughout the transportation •
analysis process, maintaining all trip characteristics (time,
purpose, modes, itinerary) with their associated person and
household;
Use of X-Y coordinates, monuments, and place declarations •
as the basic spatially referenced system for origin, destination,
residential, and intermodal junction locations for each trip
and other spatialized objects in the system.
In terms of data completeness, the TDA does not use an
origin-destination matrix, which would aggregate and dissolve
information, but rather maintains origin-destination survey trip
fles containing information on trips, people, and households
intact. The use of the most fully defned information improves
the level of resolution of the system, while at the same time
preserving any possible aggregation. As reported by O’Donnell
and Smith David (2000), possibilities are widened because the
number of dimensions distinguished by the information system
is increased. The use of special analysis modules, combined with
the presence of an underlying GIS, provides useful tools to the
planner such as three-dimensional load profles (see Figure 1).
These load profles help to calibrate the modules and validate the
results of the survey using ground counts and other observed data
over the transit network.
Transportation Object-Oriented Modeling
Transportation object-oriented modeling (TOOM) is based on
the use of transportation objects, which are special components
intended for the modeling, observation, planning, and analysis
of a transportation system. For this purpose, these objects have a
variable state in time and space, and are characterized by special
properties and methods. A road link object, for example, has
common road properties (length, name, number), but also can
have time-varying properties (such as pavement condition).
Four metaclasses of transportation objects are involved in
dynamic and spatialized relations:
Immobile (static) objects • have fxed locations in time and
space. Their roles are to describe the territory and serve as
transportation movement beacons. Some examples are the
trip generator, postal code, census tract, and zone objects.
Dynamic objects • are the transportation actors. These objects
“decide” and contribute to their movements. They represent a
group of persons (household, person), a moving object (bus,
car), or even moved objects (goods).
Kinetic objects • are the movement describers. Some examples
are the trip, transit link (simple kinetics), or the path and
transit route (compound kinetics) objects.
System objects • are groups of embedded objects, with their
set of relations. They can be operational (transit network,
road network), informational (survey, census), or globally
comprehensive (urban system).
Figure 2. Object model for 2003 household survey in Montreal
URISA Journal • Trépanier, Chapleau, Morency 39
A transportation method is an “intelligent” sequence of proce-
dures used to manipulate and transform one or more transportation
objects. It blends models with information, creating “infomodels”
to be reapplied to similar objects. It is important to mention that
transportation object-oriented modeling is not primarily aimed at
software design or database structure, and is not a database issue.
First and foremost, TOOM is a “way of thinking” about the role
and specifc use of every piece of information in the system. With
adequate object diagrams, objects can be rapidly identifed, along
with their properties and methods that are engaged in the analysis
(Chapleau et al. 1998). The software implementation can easily
integrate these underlying concepts, but not all software languages
are adapted to this methodology. TOOM was recently applied to
smart card data analysis (Morency et al. 2007).
In the object-model of the Montreal survey, there is an ob-
vious link between household, person, and trip objects, which
constitute the core model of the interview (see Figure 2). But
derived objects, such as car, parking spaces, activity, and status,
also can be defned and analyzed with the help of the other objects,
even though they were not clearly declared in the survey. To bet-
ter understand the links between household surveys, TDA, and
TOOM, please refer to Trépanier and Chapleau (2001).
surVEy InformAtIon systEm
frAmEWork
Because the Montreal household survey is a short-term endeavor
(September to January), it must be well prepared at the begin-
ning; most of this preparation involves the assembly (“montage”)
of information systems, which requires mounting data structures
and collecting, normalizing, and storing data using a convenient
software technology.
Assembly of the Geographical Information
System for Transportation
Undoubtedly, there exists a need for a geographical information
system for transportation (GIS-T) to support CATI during the
interview. Te GIS-T is mainly used to geocode trip end and
junction locations, but also is called on to validate walking dis-
tances for transit access or to geocode places of work and home
locations. GIS-T is used because of its awareness of transportation
specifcities, which are diferent from those of classical GIS usage
(Trépanier et al. 2002).
A comprehensive road network database is developed frst
to ensure:
Adequate identifcation of all streets within the region. This •
includes the various aliases (alternative street names) used
by inhabitants and also considers the language differences
between French and English.
Integrity of the list of civic numbers, which is based on street •
arc geometry and refers to the street-name database.
Automatic building of the intersection list from the geometry •
of the street network. (There is also a need to generate all
possible identifcation combinations.)
Normalization of the postal code list. When possible, postal •
codes are linked to the road network to ensure a better
identifcation (always on the same multiple-alias basis).
The trip generator database is critical because most respon-
dents give trip generators as trip ends and mistakes can easily be
made when choosing such locations in the database. Research
projects by Trépanier et al. (2003) have identifed important
issues about trip generators, such as the fact that a “good” trip
generator database must not be too narrowly defned because
of possible mismatches between two places (for example, two
franchise locations), but must contain all “major” generators.
To discover what these major generators are, data from previous
surveys were analyzed. In constructing the trip generator database,
it is important not to simply amass lists of companies provided
by commercial data vendors, because these data have not been
validated (they may contain double entries, spelling errors, deleted
entries, or incomplete entries), usually are not well geocoded, and
are not classifed. For the 2003 survey, every trip generator was
well characterized (class, exact location, named with aliases) and
uniquely identifed.
Each location in all the tables (civic addresses, intersections,
postal codes, and trip generators) is characterized and positioned at
the fnest level of resolution, although variable defnitions can be
used in surrounding regions. This ensures good geocoding during
the survey. However, the CATI software also accepts locations that
are not so well defned, as is often the case in household surveys
because respondents do not know, or do not want to give, precise
information. For example, a street name alone can be given if the
street is not too long, or a municipality name alone is acceptable
for places outside a territory, and so on.
For the household survey, the GIS-T also integrates the best
possible defnition of the transit network. An analytical transit
network (ATN), built up from the information provided by each
operator, also is required. When respondents describe transit trips,
they give the sequence of routes taken. With the help of the ATN,
the CATI software immediately validates the information, while
at the same time rejecting bad sequences and asking for precision.
Mistakes often are made when operators’ routes have the same
number. Also, some trips may include too much walking distance
to the stops, in which cases the CATI software fags the problems
and asks for second validations.
Assembly of the Household Survey Information
System
Te household survey information system (HSIS) gathers all
the information necessary to conduct the survey. Te main data
tables are:
Households. • This table contains all the information gathered
about the households surveyed (respondent name, size, car
ownership).
Persons. • This table stores data on people, such as age, gender,
status (worker, retired, student, etc.), and possession of a
driver’s license.
40 URISA Journal • Vol. 20, No. 1 • 2008
Trips. • This table describes each trip collected during the
survey (purpose, time of departure, origin and destination
locations). It also contains the sequence of modes (car,
transit routes, bike, foot, paratransit, school bus, taxi, etc.),
bridges crossed (if any), parking information, and freeways
traveled (if any).
Locations. • This table groups together all the described
locations in a single structure that identifes them by civic
number, street, intersection, generator’s name, coordinates,
region, and categories.
Calls. • Each call made by an interviewer is stored in this table
and is characterized by a status (completed, refused, busy
signal, voice mail, language problem, etc.).
Sample. • This table contains household home location data
(civic number, street, municipality, postal code, telephone
number). To ensure spatial portioning of the sample, every
household is geocoded at the fnest level with the help of
the civic address. Experience has shown that the postal code
is not precise enough and is prone to error and a precise
location is needed in the case of boundary streets between
two sampling districts.
Stratum. • This table describes the group of sampled
households within a given area of the territory, ensuring a
uniform sampling rate during the survey.
Batches. • This table describes the group of sampled households
that must be chosen for a single survey day, ensuring
uniformity of interviews over time.
Interviewers. • Each interviewer is a user in the system, and his
or her user name serves to analyze his or her performance.
Queries. • Prestored queries are used by survey monitoring
staff and by transportation planners before, during, and
after the survey.
GIS-T tables. • These all are part of the HSIS (streets, addresses,
intersections, postal codes, generators, routes, road geometry,
transit geometry).
Technical Architecture
Te survey software suite works on Microsoft Windows and is
installed on standard personal computers. In 2003, the survey
foor was composed of 50 interviewer stations, fve supervisor
stations (supervisors also could use interviewer stations when
needed), and a server station. All were equipped with Microsoft
Windows 2000 (workstation and server). Te database architec-
ture refects the needs of survey operation. Large GIS-T tables
are placed directly on workstations to make CATI more powerful
by accessing its own tables directly. In fact, these tables are not
updated often during the survey, and so a centralized database
management system is not needed.
Survey management tables containing the sample data and
state of each household are stored on the server to ensure integrity.
To facilitate data exchange, requests between workstations and
the server are made with XML fles, as these provide fexibility
and, more importantly, variable data structures and length in cases
where all the declaration information associated with a household
has to be transferred.
The software was developed with Microsoft Visual Foxpro,
using tools provided by the Windows technologies: Microsoft’s In-
ternet Information Server (IIS), the Microsoft Extended Markup
Language (MSXML) component model, and the Microsoft Offce
(MSO) component model. Applications also involved the use
of Adobe Portable Document Files (PDF) and Microsoft Excel
spreadsheets.
ImPlEmEntAtIon
Tis section presents the features of the tools that were used in the
Montreal 2003 survey. Te intention here is not to focus on the
software itself (which is an in-house product and not commercially
available) but rather on the various procedures employed, the
conduct of the survey itself, and some associated statistics.
Three software components were used in the 1998 and
2003 surveys. The technology evolved, but their principal roles
remained the same:
CATI software used by interviewers (MADQUOI, Module •
Questionneur Utilisé pour l’Obtention d’Information),
Real-time survey management software (MADASARE, •
Module d’Application de Suivi et d’Analyse Rigoureux de
l’Échantillon), and
Survey surveillance and statistics software (MADVIJIE, •
Module de Validation Incontournable Journalière des
Informations d’Enquête).
Computer-Assisted Telephone Interview
CATI is the core component of the survey software. CATI is
one of the oldest computer-assisted interviewing methods used
in travel surveys. Its primary role is to employ interactive com-
puting systems to assist interviewers in performing the basic
data-collection tasks of telephone interview surveys (Nicholls II
1988, Wermuth et al. 2003). It guides the interviewer through the
interview process—gathering, validating, and storing information.
Following are its main functions:
To display questions to the interviewer in suitable order, •
according to the survey protocol. CATI also constructs the
necessary loops of questions when, for example, many people
are interviewed for many trips. The answer to each question
is stored with the entry, according to the authorized domain
of answers (especially for locations). Then it quickly analyzes
the answers and prompts the next question.
To geocode every location related to the interview (home •
location, place of work, origins, destinations, junction points)
with a special interface supported by the GIS-T database.
CATI needs, and has, an “intelligent” way of dealing with
locations. A location list is made up for each household and
so a previously searched location can be easily reused. The
spatial logic of locations for a trip chain must be ensured: The
origin of a trip is the destination of the preceding one.
To proceed to immediate answer validation in the case of •
spatially driven questions such as transit routes and bridges
taken. In the case of a transit route, the walking distance to
URISA Journal • Trépanier, Chapleau, Morency 41
access the network is checked. Then the sequence of routes
is checked with the help of the transit network geometry.
Finally, to proceed to overall interview validation. As needed, •
the software checks the integrity of all answers using a
special procedure. This is usually performed following each
respondent’s answer and after the whole interview has been
completed. Warnings and error messages are displayed to the
interviewer. Because of productivity concerns, the interview
can be accepted even with such messages; fnal decisions are
made by supervisors.
In addition to the integrity of the interview itself, CATI, used
by all the interviewers, ensures the uniformity of the survey, for
it provides a single database for all locations, declarations, route
sequences, and so on, and facilitates postsurvey analysis. Figure 3
presents the trip declaration screen of the Montreal 2003 CATI
software. Parts A and B show the household activity summary.
Locations and other circumstances of the declared trip are gath-
ered in the C section of the screen. The sequence of modes of
the trip is displayed in D for the current person. A detailed trip
summary also is available in E. Finally, the F section is used to
display question and answers for the currently selected feld in
other sections of the screen.
rEAl-tImE surVEy
mAnAGEmEnt
To ensure real-time survey management, extended markup lan-
guage (XML) fles were used for information exchange between
workstations and the server. XML fles were helpful because their
data structure can change when needed. Te server application
controlled the distribution of the sample to the interviewer. Tis
describes the process:
From a workstation, CATI requests individual household •
information as needed by an interviewer for an interview.
The request is written in an XML fle.
The server processes the XML fle in sequential order to avoid •
data collision and double sample distribution. Many criteria
are evaluated in the choice of a household: language (some
interviewers can speak a foreign language), appointments
made with respondents, batches that must be treated on a
priority basis, and so on. Then it sends the information to
CATI in the same XML fle (header and workstation data
are kept).
CATI processes households one at a time. It completes •
the XML fle with the information that is gathered by the
interviewer. This ensures data integrity, because the XML
fle also is used for validation. Whether the interview is
postponed, stopped, or completed, or an appointment is
made, the XML fle is updated and sent back to the server.
The XML fle also contains performance indicators for the
interview, such as duration, number of errors, etc.
The server receives the XML fles from the workstations. •
It processes the data to transform and store them in the
centralized survey database.
Meanwhile, the workstation stores a copy of each interview.
CATI also logs every entry made by the interviewer for later use.
A model was estimated using these numerical logs to detect the
measurable variables that have a signifcant impact on interview
duration (Morency 2008, Chapleau 2003). In addition, the server
stores all raw XML fles.
Survey Statistics Management
When conducting large household surveys, planners must be able
to follow survey activities on a day-to-day basis. Traditionally, this
is accomplished through daily reports that are distributed among
them. Tis generates large amounts of paper and is not always
suited to specifc needs.
In 2003, the online survey statistics application provided
information on:
Global productivity per day or per week, or of the whole •
survey (number of completed calls, trip rate per household,
person, overall call status);
Productivity of each interviewer per day (completed calls, •
average interview duration, call status evolution);
Sample productivity (noncompleted calls per stratum, batch •
progress, stratum household statistics, batch household
statistics);
Technical maintenance for the survey (list and types of errors, •
list of households to be completed, locations to be geocoded,
list of new trip generators, comments, error rates);
Daily maintenance (calls to a single household, interviewers •
on duty, list of current appointments);
Reporting software management (list of reports, user accesses, •
log fle).
To generate reports, the application stores queries based on
Structured Query Language (SQL), parameterized when needed
to input text, date, or number. Reports are available in four for-
mats, depending on their structure: paper-like reports in a PDF
Figure 3. CATI software interface for trip declaration
42 URISA Journal • Vol. 20, No. 1 • 2008
fle, on-screen table-format reports, on-demand charts, or Excel
spreadsheets.
conclusIon
Nowadays, it is our view that a large telephone household sur-
vey in transportation cannot be conducted without the help of
powerful software components that will support all the functions
surrounding this labor-intensive and costly activity. Tere is a
particular need for an adequate geographical information sys-
tem for both territorial and operational data. Tis GIS-T must
be comprehensive and its entities should be well identifed for
searching for address, intersection, and postal code. Transit stop
locations and trip generators must be carefully selected to avoid
bias in location choice. In the Montreal region, this architecture
has been dictated by the use of household survey data: transpor-
tation planning, transit network development, user behavior,
transit fnancing, road network usage, trip generator analysis, and
so on. Te same approach will be used for the survey that will be
conducted in the Greater Montreal Area in the Fall of 2008.
The contributions of such tools are multiple:
Better uniformity in interactions between interviewers and •
respondents because the information is equally available
among interviewers.
Immediate validation of most collected data, from single- •
feld values to complete transit itinerary.
Autocorrection of sampling rates during survey. With the •
help of the statistical tool, the sampling strategies can be
modifed anytime to reduce underreporting and spatial
biases.
Improved interviewers monitoring and training. Rapid •
identifcation of failures and mistakes.
Integrated call management: voice mails, appointments, •
privacy concerns.
Improved overall quality of the household survey process, •
from preparation to postvalidation. Better data helps
to minimize the need for technical resources, especially
transportation planners.
In the future, many elements will affect the way in which
household surveys are conducted. The use of smart card payment
systems on a large scale will provide fresh data to update and
complete those obtained via telephone survey. The ever-growing
diffculties encountered in reaching people by telephone at home,
the increasing number of single-person households (these are hard
to reach), and the growing use of cell phones that sometimes
replace home telephone service will challenge the traditional
ways of conducting surveys and will prompt a rethinking of the
methods used.
Acknowledgments
Te authors especially acknowledge the major contribution of
Bruno Allard, research assistant, who implemented the software
and the methods. Tis work has been supported by the partners
who conducted the 2003 survey: the Agence Métropolitaine de
Montréal (Montreal Transportation Agency), the Société de Trans-
port de Montréal (Montreal Transit Commission), the Société de
Transport de la Ville de Laval (Laval Transit Commission), the
Réseau de Transport de Longueuil (Longueuil Transportation
Network), and the Ministère des Transports du Québec (Quebec
Ministry of Transportation).
About the Authors
Martin Trépanier, P. Eng., Ph.D., is a professor of industrial
engineering at the Ecole Polytechnique de Montreal
(Logistics and Transportation). His research is mainly related
to logistics, information systems, object-oriented modeling,
GIS development, and Internet applications. He worked
along with the MADITUC group on this research project.
Corresponding Address :
École Polytechnique de Montréal
P.O. Box 6079
Station Centre-Ville
Montreal, Quebec, H3C 3A7, Canada
Phone: (514) 340-4711, #4911
Fax: (514) 340-4173
mtrepanier@polymtl.ca
Robert Chapleau, P. Eng., Ph.D., is a professor of civil
engineering (Transportation Planning) and founder-director
of the MADITUC group, Civil Engineering Department,
Ecole Polytechnique de Montreal. He developed the totally
disaggregate approach in transportation and is participating
in several research projects in the Montreal area.
Catherine Morency, P. Eng., Ph.D., is a professor of civil
engineering (Transportation Planning) at the Ecole
Polytechnique de Montreal. Her research work covers
household survey data analysis. She also is interested in urban
dynamics related to transportation. She is participating in
MADITUC research projects.
References
Badoe, D. A., and G. N. Steuart. 2002. Impact of interviewing
by proxy in travel survey conducted by telephone. Journal
of Advanced Transportation 6, no. 1.
Bergeron, D. 2008. Enquêtes origine—destination métropolitaine:
bilan, expérimentation et choix pour l’avenir. 43e Congrès
de l’Association Québécoise du Transport et des Routes,
Québec.
URISA Journal • Trépanier, Chapleau, Morency 43
Brög, W. 1997. Raising the standard! Transport survey quality
and innovation. Keynote Paper for the 8th Meeting of
the International Association of Travel Behavior Research,
Austin, Texas.
CERTU. 1998. L’enquête ménages déplacements méthode
standard. Collection du CERTU, Transport et mobilité.
Chapleau, R. 2003. Measuring the internal quality of a CATI
travel household survey. In Peter Stopher and Peter Jones,
Transport survey quality and innovation. Kruger, South
Africa: Pergamon: 69-87.
Chapleau, R. 2000. Conducting telephone origin-destination
household survey with an integrated informational approach:
Transport surveys: Raising the standard —Proceedings of
an international conference on transport survey quality and
innovation, Grainau, Germany, Transportation Research
Circular EC-008.
Chapleau, R. 1995. Symphonie d’usages des grandes enquêtes
origine-destination, en totalement désagrégé majeur. Opus
Montréal 87 et 93, Huitièmes Entretiens du Centre Jacques-
Cartier, Lyon, France.
Chapleau, R. 1986. Transit network analysis and evaluation
with a totally disaggregate approach. World Conference on
Transportation Research, Vancouver.
Chapleau, R., B. Allard, M. Trépanier, and C. Morency. 2001.
Les logiciels d’enquête transport comme instruments
incontournables de planifcation analytique. Recherche,
Transport, Sécurité, Paris 70 (Janvier-Mars 2001): 59-77.
Chapleau, R., M. Trépanier, P. Lavigueur, and B. Allard.
1997. Origin-destination survey data dissemination
in a metropolitan context: A multimedia experience.
Transportation Research Record, Washington 1551:
26-36.
Couper, M. P., R. P. Baker, J. Bethlehem, C. Z. F. Clark, J. Martin,
W. L. Nicholls II, and J. M. O’Reilly. 1998. Computer
assisted survey information collection. Wiley.
Griffths, R., A. J. Richardson, and M. E. H. Lee-Gosselin.
2000. Travel surveys, transportation in the new millennium.
Transportation Research Board, http://gulliver.trb.org/
publications/millennium/00135.pdf.
Jones, P., and J. Polak. 1992. Collecting complex household travel
data by computer. In selected readings in E. S. Ampt, A.
J. Richardson, and A. H. Meyburg, eds., Transport survey
methodology. Eucalyptus Press: 283-305.
Link, M. W., and M. J. Kresnow. 2006. The future of random-
digit-dial surveys for injury prevention and violence research.
American Journal of Preventive Medicine 31, no. 5: 444-
50.
Liss, S. 2005. Planning for the next national household travel
survey, Part 1. In Data for understanding our nation’s
travel, Transportation Research Circular, Number E-C071,
Transportation Research Board.
Morency, C., M. Trépanier, and B. Agard. 2007. Measuring
transit use variability with smart-card data. Transport Policy
14, no. 3: 193-203.
Morency, C. 2008. Enhancing the travel survey process and
data using the CATI system. Transportation Planning and
Technology 31, no. 2 (April 2008): 229-48.
National Research Council. 2003. Survey automation: Report
and workshop proceedings. Oversight Committee for the
Workshop on Survey Automation. In D. L. Cork, M. L.
Cohen, R. Groves, and W. Kalsbeek, eds., Committee
on national statistics. Washington, D.C.: The National
Academies Press.
Nicholls II, W. L. 1988. Computer-assisted telephone
interviewing: A general introduction. In R. M. Groves,
P. P. Biemer, L. E. Lyberg, J. T. Massey, W. L. Nicholls II,
and J. Waksberg, eds., Telephone survey methodology. NY:
Wiley: 377-85.
O’Donnell, E., and J. Smith David. 2000. How information
systems infuence user decision: A research framework and
a literature review. International Journal of Accounting
Information Systems 1(2000): 178-203.
Richardson, A. J. 2002. Current issues in travel and activity
surveys. In H. Mahmassani, ed., In perpetual motion: travel
behaviour research opportunities and application challenges.
Oxford UK: Pergamon: 341-57.
Richardson, A. J., E. S. Ampt, and A. H. Meyburg. 1995. Survey
methods for transport planning. Eucalyptus Press.
Trépanier, M., R. Chapleau, B. Allard, and C. Morency. 2003.
Trip generator relocation impact analysis methodology based
on household surveys. Institute of Transportation Engineers
Journal (on the Web) Washington 73, no. 10.
Trépanier, M., R. Chapleau, and B. Allard. 2002. Geographic
information system for transportation operations: models
and specifcity. Compte-rendus de la 30e conférence annuelle
de la Société canadienne de génie civil, Juin 2002, Montréal,
559-68.
Trépanier, M., and R. Chapleau. 2001. Linking transit operational
data to road network with a transportation object-oriented
GIS. Journal of the Urban and Regional Information Systems
Association 13, no.2: 23-27.
Wermuth, M., C. Summer, and M. Kreitz. 2003. Impact of new
technologies in travel surveys. In P. Stopher and P. Jones.
2003. Transport survey quality and innovation. Kruger,
South Africa: Pergamon: 455-81.
Westrick, S. C., and J. K. Mount. 2007. Evaluating telephone
follow-up of a mail survey of community pharmacies.
Research in Social and Administrative Pharmacy 3, no. 2:
160-82.
URISA Journal • Václavík 45
IntroductIon
Te Czech Republic currently is undergoing transformation from
the centralized regime of a communist dictatorship towards a
modern democratic state. Fanta et al. (2005) recognizes three main
events in the last half century that had profound consequences for
the country and its land use. First, the communist coup d’état and
the following collectivization of land in the 1950s that introduced
large-scale collective farming, especially intense in the Olomouc
region, which aimed at the maximum production of agricultural
commodities. Second, the abolition of the totalitarian political
system in 1989, which was followed by the restitution of private
land ownership in the 1990s, the reintroduction of democracy
and a market economy, and the development of market-driven
forms of land use. Tird, the preparation of the Czech Republic
for ingression into the European Union in 2004, including its
complete association with the EU environmental and agricultural
policies, and its search for appropriate methods and forms of
land use.
This research pays closer attention to specifc trends in land-
use changes within the past 25 years: changes in agricultural areas,
forest areas, and residential development. These particular trends
can be described as followed.
Agricultural areas. Political transition in the Czech Republic
lead to marginalization of intensive agricultural areas, i.e., a pro-
cess driven by a combination of socioeconomic and environmental
factors caused by farming that ceased to be viable at many places,
resulting in frequent abandonment of the agricultural land (Fanta
et al. 2004). Extensive areas of previously cultivated land in the
country now are laying fallow or were converted to secondary
grasslands—meadows and pastures.
Forested areas. Since the time of their minimum extent at
the end of the 18th century, forested areas have been increasing,
mapping land-use/land-cover change in the
olomouc region, czech republic
Tomáš Václavík
Abstract: The Olomouc region in the Czech Republic has undergone signifcant changes in the past several decades, including
the change in political system of the country in 1989. Although the political and cultural transformation generally is recognized
as an important driver of land use (Ptáček 2000), few studies were conducted that would empirically assess and quantify land-
use/land-cover changes in the Czech Republic, especially in the context of the postsocialistic transformation (Fanta et al. 2004,
Zemek et al. 2005). This study presents an approach for identifying major land-use/land-cover changes in the Olomouc region,
applying remote-sensing techniques to compare data from multispectral satellite sensors acquired 12 years before and 12 years
after the revolution in 1989. The study closely covers specifc trends in land-cover changes: changes in agricultural areas, forested
areas, and residential development. The results support initial assumptions that the land cover will refect the changes in the hu-
man perception of landscape and natural resources, such as a smaller need for intensive agriculture, a shift to an environmentally
friendly management of forested areas, or increased development and suburbanization.
reaching the present 33 percent of the total vegetation cover in the
country (ÚHÚL 2006). Most of the forest is far from its natural
composition, for it was converted to monocultures of Norway
spruce (Picea abies), serving predominantly a productive function.
However, since the boom of environmental consciousness in the
1990s, a distinctive tendency has grown towards alternative ap-
proaches in forest management considering the natural species
composition and potential vegetation (Neuhäuslová 1998).
Residential development. As in other parts of Europe, the issue
of suburbanization was well identifed in the Czech Republic in
the 1990s (Ptáček 1998; Jackson 2002). However, it is repre-
sented by a relatively small extent of residential development in
vicinities of larger cities, and does not bear the typical traits and
negative effects of the American-type large-scale suburban sprawl
as described by Václavík (2004).
The main objective of this study is to analyze relevant remote-
sensing data from 1976 and 2001 and to identify the locations,
types, and trends of the main land-use and land-cover changes in
the past 25 years. Although the issue of land change is examined
based on the background of political transformation of the coun-
try, this article does not explicitly address the effect of political
transitions on land-cover change. However, it was assumed that
the land cover will refect some changes in the human perception
of landscape and natural resources, such as the decreased need
for intensive agriculture, the shift to an environmentally friendly
management of forested areas, or the increased development and
suburbanization. The hypothesis is that the later satellite image
of the Olomouc region study site will exhibit a smaller total area
of intensive agriculture and more meadows and pastures, fewer
coniferous forests, and more mixed or deciduous tree cover, as
well as an overall higher residential development.
46 URISA Journal • Vol. 20, No. 1 • 2008
mEtHods
Study Site
Te study area chosen for this project is the Olomouc region in
the eastern Moravian part of the Czech Republic (see Figure 1).
Te study area of 5,012 km
2
covers most of the Olomouc County
administration unit, one of the 14 administration units in the CR,
but the northeastern part overlaps to Moravskoslezsky County.
Te central part is formed by the wide alluvial plane of the up-
per stream of the Morava River, surrounded by the undulated
hills of the Zabrezska and Drahanska uplands from the west and
the Nizky Jesenik mountain range from the northeast, while
the elevation ranges from 200 to 800 m a.s.l. Te lowland areas
are highly urbanized, and include the major cities of Olomouc,
Prerov, Zabreh, Sumperk, and others. Because of favorable climate
and fertile soils, lowlands historically and currently represent the
substantive agricultural areas in the Czech Republic. Despite its
intensive development, the core of the Olomouc region consists
of the Litovelské Pomoraví Protected Landscape Area. Tis ex-
ceptional piece of natural landscape is formed by the naturally
meandering Morava River and its several permanent and periodi-
cal branches with wetlands, meadows, and unique complexes of
foodplain forests, some of the few remnants in central Europe.
Te other major forested habitats in the Olomouc region are
located in the northeastern upland areas, and are predominantly
composed of coniferous and mixed stands, which are used for
timber production.
Data Collection
Because the study area is located in central Europe, the images
from the high-resolution SPOT earth observation satellite would
be the appropriate data source for the intended study. However,
the SPOT data for the study site was not freely available when it
was needed. Terefore, the Landsat Multispectral Scanner (MMS)
and Enhanced Tematic Mapper (ETM) scenes were acquired for
change detection analysis (see Table 1). Te MSS data included
one scene (path 204, row 25) from May 8, 1976; the ETM+ data
included two scenes (path 190, row 25, and path 190, row 26)
from May 24, 2001. Described data sets were downloaded from
the Global Land Cover Facility (GLCF) (http://glcf.umiacs.umd.
edu/data/) through the Web interface and imported to IDRISI
geographic information system software using the GEOTIFF/
TIFF conversion module.
Ancillary sets of data were collected to support the land-
change analysis. Two sets of scanned and georeferenced black-
and-white aerial photographs from 1970s and 1990 and a set
of color orthophotographs from 2002 were obtained from the
Litovelské Pomoraví Protected Landscape Area Administration
to serve as reference ground-truth data during the map classifca-
tion process. Vector data of the Czech Republic boundary and
the Litovelské Pomoraví PLA area were obtained from the Czech
Environmental Information Agency (CENIA) ArcIMS server
(http://geoportal.cenia.cz).
Image Processing and Classifcation
Acquired data sets were processed and examined in the Clark
lab’s GIS software IDRISI 15.0, Andes edition. Figure 2 shows
the steps of image processing and classifcation needed to achieve
defned study objectives. After the satellite data were downloaded
from the Global Land Cover Facility and imported to IDRISI, it
was assessed for image quality. While both ETM+ images did not
exhibit any signifcant radiometric noise in the entire scene, the
MSS image contained a fair amount of haziness in the northeast-
ern portion of the scene and subtle striping throughout the entire
Figure 1. Study area
Table 1. Acquired satellite images
Scene Num-ber Path/
Row
Acquisition Date Sensor Format Spatial Resolution
(m)
Bands
044-131 204/25 1976-05-08 Landsat MSS GEOTIFF 57x57 1 - 4
036-343 190/25 2001-05-24 Landsat ETM+ GEOTIFF 30x30 1 – 5, 7
036-344 190/26 2001-05-24 Landsat ETM+ GEOTIFF 30x30 1 – 5, 7
URISA Journal • Václavík 47
area. As there were no meteorological data available for the time of
MSS image acquisition, an absolute atmospheric correction could
not be performed. Instead, the Principal Component Analysis
(PCA) was run, using standardized variance/covariance matrix
and all four MSS bands as inputs. PCA created four principal
component images in which the frst two explained more than 98
percent of the total variance and the remaining two components
contained most of the noise. Te original four MSS bands were
restored through an inverse PCA technique using the frst two
components.
The study area of the Olomouc region is located in the
overlap of the ETM images 036-343 and 036-344 from 2001.
A composite of the two overlapping images was created using
a mosaic technique by spatially orienting them and optionally
balancing the numeric characteristics of the image set based on
the overlapping areas. The average mosaic method was applied
to average the base image values with the adjusted overlap image
values. In addition, the WINDOW module, extracting subimages
from the set of original images, was utilized to isolate the desired
extent of the study area.
The last step before actual image classifcation was to syn-
chronize the spatial resolution of the images from both times. The
original resolution of the MSS image was 57x57 meters. For the
purpose of its comparison with the ETM image with resolution
of 30x30 meters, the MSS image needed to be resampled. The
resample module using parameters from the ETM image and map
corners as ground control points was applied, producing a total
root-mean-square error of 0.8 m, which is less than 0.5 pixels.
The MSS 1976 and ETM 2001 images were classifed using
the Maximum Likelihood supervised classifcation because most
of the land-cover mapping projects have applied either supervised
or unsupervised parametric classifcation algorithms to identify
spectrally distinct groups of pixels (Smits et al. 1999). With
supervised classifcation, the spectral signatures of the known
land-cover categories are frst developed, using digitized training
sites. The software then uses a specifc algorithm to assign all
pixels in the image data set into defned land-cover classes (Jensen
2004). The Maximum Likelihood classifcation is based on the
probability density function that is associated with a particular
training site signature. All pixels are assigned to the most likely
category based on an evaluation of the subsequent probability
that the pixel belongs to the signature (class) with the highest
probability of membership (Jensen 2004). Seven land-cover
categories were recognized in the Olomouc region: water, de-
ciduous forest, coniferous forest, mixed forest, developed (urban)
areas, areas of (intensive) agriculture, and meadows (grassland).
Training sites were digitized based on the personal knowledge of
the study area and ground-truth data of aerial photographs and
orthophotographs. Spectral signatures of individual land-cover
classes were developed and assessed for their separability. Spectral
Figure 2. Work-fow diagram
48 URISA Journal • Vol. 20, No. 1 • 2008
values of urban areas and agricultural areas with bare soil were
mixing, therefore their training sites had to be redefned, and also
the texture analysis using Dominance index and kernel window
of 5x5 pixels was conducted. Finally, the Maximum Likelihood
classifcation was run with original bands and the texture image
as inputs, producing two fnal land-cover maps of 1976 and 2001
that were compared.
An error matrix was constructed to estimate classifcation
accuracy of produced land-cover maps. The error matrix provides
a basis for characterizing types of errors by cross-tabulating the
classifed land-cover categories in sample locations against those
observed in ground reference data (Smits et al. 1999, Foody 2002).
A random sampling scheme was applied to defne ground-truth
locations (n = 100) and the aerial photographs from 1970s and
2002 were used to check for the “true” land-cover classes. The
overall accuracy was calculated for both maps. This represents
the probability that any point on the land-cover map is assigned
exactly the same category by the classifer, as the category that is
identifed in the ground-truth sites (Wulder and Franklin 2003).
In addition, the producer’s and user’s accuracies that measure
omission and commission errors were estimated for individual
land-cover classes.
A cross-classifcation procedure is one of the fundamental
pairwise comparison techniques used to compare two images
of qualitative data (Eastman 1995). It overlays two images and
calculates all their possible combinations of classes. In the case
where images represent the same land-cover categories at different
times, persistence occurs where areas fall in the same land-cover
categories, and change occurs where a new category is created
(Eastman 1995). IDRISI Andes offers an effcient and easy-to-use
tool for rapid assessment of land-cover change and its implications
based on cross-clasifcation principles. The Land Change Modeler
(LCM) for Ecological Sustainability allows a user to evaluate gains
and losses in land-cover classes, land-cover persistence, and specifc
transitions between selected categories. Using the classifed land-
cover maps from 1976 and 2001 as input parameters, this tool
was applied to identify the locations and magnitude of the major
land change, land persistence, and trends in transitions between
land-cover categories in the study area.
rEsults
Figures 3 and 4 represent the results of Maximum Likelihood
classifcation: land-cover maps depicting the situation in 1976
and 2001. Te change analysis tool provides efcient statisti-
cal assessment of changes in individual land-cover categories.
Its results in Figures 5 and 6 demonstrate that there have been
signifcant changes in all land-cover/land-use categories between
1976 and 2001 with the exception of water, where the subtle
change can be caused by location errors in land-cover classifca-
tion. Concerning the net change, which represents the earlier
area of a category with added gains and subtracted losses, three
land-cover categories experienced major transitions. Te total area
of meadows (grassland) increased by 942 km
2
, while the area of
Figure 3. Land-cover map of 1976
Figure 4. Land-cover map of 2001
Figure 5. Gains and losses between 1976 and 2001 in km
2
Figure 6. Net change between 1976 and 2001 in km
2
URISA Journal • Václavík 49
intensive agriculture decreased by 592 km
2
, as well as the area of
coniferous forests, which decreased by 603 km
2
. Te category of
developed (urban) area also was afected by distinct change, with
a net gain of 127 km
2
.
A simplifed cross-classifcation map (see Figure 7) represents
persistence in land-cover categories, areas where no change oc-
curred, and land-cover change, areas with any type of transitions
between categories (depicted in black). However, the land-change
and persistence map is diffcult to visually interpret if the areas
of individual land-cover classes are not clustered, and also the
type of change is not represented in this map. Therefore, the
contribution to net change, i.e., the transition between specifc
classes, was calculated to achieve the objectives of the study.
Data in Figure 8 represent the contribution to net change for
categories of meadows (grassland), developed (urban) area, and
mixed forest. They reveal that agricultural areas explain about
63 percent of the total increase in meadows, new development
occurred predominantly on former agricultural areas (more than
56 percent), and about 16 percent of previous coniferous forests
currently is identifed as mixed forest.
Analysis of the error matrices revealed that the proportion
of agreement between land-cover categories in the classifed map
and the ground-truth data was 77 percent for the 1976 period and
81 percent for the 2001 period. The producer’s accuracy was the
lowest for the agriculture class in both 1976 and 2001 maps (65
percent and 70 percent), as some of the agricultural areas were
classifed as meadows or developed. The user’s accuracy was the
lowest for the mixed forest class (60 percent) in the 1976 map
and for the developed class (65 percent) in the 2001 map. Some
sites with mixed forest were falsely identifed as the coniferous or
deciduous class in the reference data. Some developed sites were
falsely identifed as agriculture or coniferous forest categories.
dIscussIon And conclusIons
Tis study applied remote-sensing techniques to classify satellite
imagery of the Olomouc region, Czech Republic, from 1976
and 2001, from years before and after a major political change in
the country, and compared the two resulting land-cover maps to
identify the salient locations, types, and trends of the land-cover
change in the past 25 years. Te results support initial assump-
tions based on general knowledge of some of the land-use drivers
in diferent times. Tere have been signifcant losses in categories
of intensive agricultural areas and coniferous forest, and gains
in meadows and developed areas. From the former agricultural
areas, 23 percent became meadows and pastures, especially in the
northeastern hilly part of the study site, and 3 percent was devel-
oped in the lowlands around the Litovelské Pomoraví Protected
Landscape Area. About 16 percent from the previous coniferous
forest in the eastern hilly part of the region currently was identi-
fed as mixed forest.
This study provides no empirical evidence of direct causality
between discovered land-cover/land-use change in the study site
and political and cultural transformation of the country; however,
the location and trends of observed land change suggest there
might be distinct correlation. Concerning the transition from a
Figure 7. Land-cover persistence and change
Figure 8. Contribution to net change in selected categories (km
2
)
50 URISA Journal • Vol. 20, No. 1 • 2008
category of intensive agriculture to a category of meadows, the
major trend was observed in the northeastern uplands of the study
site. This observation is consistent with suggestions of Zemek et
al. (2005) that the marginalization of agricultural areas occurs
frst at locations with unfavorable natural conditions, especially in
uplands where the agricultural production was previously forced
by an extensive use of fertilizers and pesticides. Concerning newly
developed areas, the major trend occurred especially in the central
lowland area of the study site around the Litovelské Pomoraví
Protected Landscape Area. This observation is consistent with
the general suburbanization process in central Europe where
new residential areas tend to be developed in the form of “satel-
lite” towns in the vicinity of existing cities and recreational areas
(Ptáček 1998). Regarding transition from coniferous tree cover
to mixed forest, this change was observed predominantly in the
northeastern hilly part of the study site, where the elevation and
associated environmental conditions favor potential vegetation
of mixed and deciduous forest stands. This fact correlates with
the general diversion in forest management in the past 15 years
from clear-cut practices and spruce and pine plantations to the
alternative use of native deciduous species of trees in the lower
and middle altitudes of the country.
Classifcation of multispectral satellite data and comparison
of land-cover maps are essential tools for assessing large-scale land-
cover/land-use changes. However, this research left considerable
room for future improvement. Visual comparison of classifed
maps with training sites as well as accuracy statistics calculated
from error matrices showed inaccuracies in the classifcation
process. Spectral mixing was apparent between the classes of de-
veloped and agriculture areas where barren soil was present and
the texture analysis did not eliminate all of it. Also, all Landsat
scenes were acquired in the spring season, when certain types
of crops, such as cereals, are in a phenological stage that exhibit
similar spectral response as meadows and pastures. In addition,
the MSS imagery from 1976 suffered from large amount of hazi-
ness and radiometric noise, which were not entirely removed by
the principal component analysis and could distinctively affect
image classifcation. An effort to collect better-quality remote-
sensing data, such as the ones from the SPOT sensor, should be
made to improve the overall accuracy of land-change assessment.
Finally, the Maximum Likelihood classifcation results might have
been improved if unequal prior probabilities of the land-cover
classes had been assigned. Alternatively, decision-tree classifcation
techniques that derive probabilities of land-cover classes from
the distribution in the training data (Rogan et al. 2002) may be
considered for future improvement of the analysis.
Acknowledgments
Te author gratefully acknowledges Yelena Ogneva-Himmelberger
and John Rogan, professors from Clark University, for supervising
this work and the Fulbright program for enabling the author’s
training in the United States.
About the Author
Tomas Vaclavik was a Fulbright exchange student in the program
of Geographic Information Sciences for Development and
Environment at Clark University, Worcester, Massachusetts,
in the academic year 2006-2007. He earned both his
bachelor’s and master’s degrees in Ecology and Environmental
Sciences at Palacky University in the Czech Republic. He
recently has been working as a GIS analyst for the Agency for
Nature Conservation and Landscape Protection of the Czech
Republic. Currently, he is pursuing his Ph.D. in Geography
at the University of North Carolina at Charlotte, focusing
on applications of GIS in ecological research and working
as a research assistant in the Center for Applied GIScience
(CAGIS).
Correspondence Address:
Center for Applied Geographic Information Science (CAGIS)
Department of Geography and Earth Sciences
University of North Carolina at Charlotte
9201 University City Boulevard
Charlotte, NC 28223
Phone: (704) 687-5963
tova@email.cz
tvaclavi@uncc.edu
http://tova.euweb.cz
URISA Journal • Václavík 51
References
Eastman, J. R. 1995. Change and time series analysis. Worcester,
MA: United Nations Institute for Training and Research/
Clark Labs for Cartographic Technology and Geographic
Analysis, Clark University, 28.
Fanta, J., K. Prach, and F. Zemek. 2004. Status of marginalisation
in Czech Republic: Agriculture and land use. EUROLAN,
National Report—Czech Republic. Faculty of Biological
Sciences, University of South Bohemia, and Institute of
Landscape Ecology—Academy of Sciences of the Czech
Republic, 29.
Fanta, J., F. Zemek, K. Prach, M. Heřman, and E. Boucníková.
2005. Strengthening the multifunctional use of European
land: Coping with marginalization. EUROLAN, National
Report—Czech Republic. Faculty of Biological Sciences,
University of České Budějovice, Institute of System Biology
and Ecology, Academy of Sciences of the Czech Republic,
41.
Foody, G. M. 2002. Status of land cover classifcation accuracy
assessment. Remote Sensing of Environment 80: 185-201.
Jackson, J. 2002. Urban sprawl. Urbanismus a územní rozvoj
(Urbanism and urban development) 5, no. 6: 21-28.
Jensen, J. R. 2004. Introductory digital image processing. Upper
Saddle River, NJ: Pearson Prentice Hall, 544.
Neuhäuslová, Z., et al. 1998. Map of potential natural vegetation
of the Czech Republic. Prague: Academia, 341.
Ptáček, P. 1998. Suburbanizace—měnící se tvář zázemí velkoměst
(Suburbanization—the changing face of metropolitan
hinterland). Geografcké rozhledy (Geographical views) 7,
no. 5: 134-37.
Ptáček, P. 2000. Networking and local culture: Local community
transformation afer 1989 on the example of Olomouc, Czech
Republic. Acta Universitatis Palackianae—Geographica 36:
59-64.
Rogan, J., J. Franklin, and D. A. Roberts. 2002. A comparison
of methods for monitoring multitemporal vegetation
change using thematic Mapper imagery. Remote Sensing of
Environment 80, no. 1: 143-56.
Smits, P. C., S. G. Dellepiane, and R. A. Schowengerdt. 1999.
Quality assessment of image classification algorithms for
land-cover mapping: a review and proposal for a cost-based
approach. International Journal of Remote Sensing 20, no.
8: 1461-86.
ÚHÚL. 2006. Report on the state of forests and forestry in the
Czech Republic 2006. Ústav pro hospodářskou úpravu lesa
(Forest Management Institute), 128.
Václavík, T. 2004. Te use of GIS in ecological planning. (A case
study of Mount Desert Island). Master’s thesis, Department
of Ecology and Environmental Sciences, Palacky University,
Olomouc, 82.
Wulder, M. A., and S. E. Franklin. 2003. Remote sensing of forest
environments. Concepts and case studies. Norwell, MA:
Kluwer Academic Publisher, 519.
Zemek, F., M. Heřman, Z. Mašková, and J. Květ. 2005.
Multifunctional land use—a chance of resettling abandoned
landscapes? (A case study of the Zhůří territory, the Czech
Republic). Ecology 24, no. 1: 96-108.
URISA Journal • Goldstein 53
mAPPInG tHE futurE succEss
of PuBlIc EducAtIon
As we quickly move forward in the information age our nation’s
public school systems are placed precariously at a crossroad. While
most school districts strive to incorporate new technologies, many
teachers simply don’t have the luxury of time to teach material that
may not be directly aligned with state and federally mandated high
stakes tests, including new computer technology skills. Tere is
an urgent need to educate our students in technology, driven by
the business and world economies. We must become a premier
digital nation or face the consequences of not taking action. Te
ramifcation of these students falling through the cracks strikes
at the heart of public education and bankrupts all stakeholders:
parents, teachers, school administrators, students, and society at
large. Teaching GIS and geospatial technology may be just the
boost our public educational system needs to adequately prepare
students for entrance into the emerging global society.
WHy GIs
Many young students today have not known a world without
computers; they are naturally curious and gravitate to the tech-
nology. Te GIS system is a perfect vehicle to deliver necessary
content and contextualize the lesson so that students will not only
be engaged but will be motivated to gain the knowledge presented.
Using this application in the classroom promotes critical thinking
skills for students in addition to honing their communication
and presentation skills.
In 2005, the National Research Council published a report
entitled “Learning to Think Spatially: GIS as a Support System
in the K-12 Curriculum” that identifed the importance of
promoting spatial thinking skills across curriculum subjects. As
indicated by the report, GIS has the potential to successfully
cultivate those skills. The overarching goal of this educational
initiative is to create the next generation of students skilled in
mapping the future success of Public Education
Donna L. Goldstein M. Ed
ABSTRACT: For better or worse, computers have revolutionized every aspect of our lives. As we quickly make the transition from
an industrial to an information age, computer literacy skills have become a basic necessity. Technology skills are now referred to
as the “Fourth R” in education, as coined by Michael Goodchild. To successfully learn and use GIS (Geographical Information
Systems) technology, one must incorporate the skills of reading, writing, and arithmetic. Understanding and utilizing a GIS
system requires a holistic combination of reading instructions, data, and maps; writing hypotheses, reports, and presentations;
and using arithmetic to understand queries and spatial analysis. Thus the 4th R as it relates to GIS is a new elevated skill that
incorporates the three original R’s in education. Teaching GIS may be just the boost our public educational system needs to
adequately prepare students for entrance into the emerging global society.
thinking spatially and to provide them with the opportunity to
compete in an international society.
tHE sIGnIfIcAncE of GIs In
PuBlIc scHool currIculum
We live in a global society where competition for those skilled in
geospatial technologies will only increase. Governments interna-
tionally have paid close attention to the evolving technological
shifts and view the development of technology skills as the foun-
dation of the country’s future. So must we if we want our next
generation to be competitive and viable in the global marketplace.
For the sake of our future and the future of our youth, we have
a responsibility to teach them how to efectively utilize the new
and emerging geospatial technologies that are increasingly needed
in our new world.
Social and fnancial implications of GIS prompted the “U.S.
Department of Labor to identify GIS as one of the three most
important emerging and evolving felds, along with nanotechnol-
ogy and biotechnology, with over 900,000 additional jobs in the
U.S. in geospatial technology expected from 2002 to 2012” (U.S.
Department of Labor Employment & Training Administration,
2005). In addition “NASA says that 26% of their most highly
trained geotech staff is due to retire in the next decade, and the
National Imagery and Mapping Agency is expected to need
7,000 people trained in GIS in the next three years” (Gewin and
Virginia, 2004). With the aging of our many professionals in the
current geospatial workforce, there are a number of organizations
seeking to recruit the next generation.
What this means is that those of us in the geospatial industry
have an obligation to our chosen feld, to ensure its continued
growth and success. By not preparing the next generation to take
the helm of this industry we turn our backs on providing for the
future sustainability of GIS and spatial technology.
54 URISA Journal • Vol. 20, No. 1 • 2008
onE scHool dIstrIct’s
rEsPonsE
Since the 1980’s the Palm Beach County School District’s GIS
system was primarily used by staf on the operational or business
side of the house for facility planning, enrollment projections
and transportation. During 2006 the Planning Director, Kristin
Garrison and the GIS Coordinator recognized the tremendous
benefts GIS could ofer students including enhanced learning
techniques and potential career opportunities. Tis revelation
prompted communication with academic staf regarding imple-
menting GIS into the classroom curriculum. Enthusiasm for the
virtues of GIS in the educational environment quickly spread
and a GIS Steering committee was assembled. Te committee
involved participants across the spectrum of school district de-
partments including staf from Facilities, Transportation, Career
Academies, and Educational Technology, as well as social studies
and science curriculum planners. Te coordination of the business
and academic units of the organization provided a foundation to
methodically develop a plan.
A GIS Educational Model (GEM) Charter was developed, a
document that outlined the policy and structure for implement-
ing GIS into the classroom. The charter also included how GIS
directly addressed a number of school board-approved district
goals and policies. With a specifc timeline developed to reach
various milestones, work began on acquiring a district-wide GIS
site license from ESRI. The cost for this district-wide license was
less than the price for the individual seats purchased for a select
group of operational staff and was the springboard needed to
push GIS into the classrooms.
Recognizing that teachers hold the key that unlocks the po-
tential for implementation of any new program, the GIS Steering
Committee provided a professional development workshop for
social studies teachers designed to spark their interest. Stipends
and in-service points were included. A training room was confg-
ured with the GIS software loaded on PCs for 30 teachers. The
workshop included hands-on GIS lessons from the Mapping
Our World Lesson plan book (Malone et.al. 2005). The goals
of this exercise were (a) to expose teachers to the benefts of
incorporating GIS into their classroom curriculum (b) to have
them experience the ease of use, and (c) to prepare them to easily
transfer what they learned directly to the classroom by following
the GIS lesson plans.
A succEss story
Te barriers that Kerski points out in his article, “A National
Assessment of GIS in American High Schools, 2001,” are still
prevalent today, seven years later. However, slow progress is be-
ing made in a number of schools. One such success story is Boca
Raton Middle School in Palm Beach County Florida. During the
summer of 2007 a teacher from this school attended the one day
workshop on GIS and took on the bold task of implementing GIS
into several of her classes this past year. Armed with little more
than a one-day training session, a Mapping our World lesson plan
book, and assistance from administrative GIS staf, this individual
rose to the occasion and introduced her students to the future.
Teir enthusiasm for GIS is evident, as students look forward to
the lab sessions and pummel questions as to what types of jobs
are available in GIS and where is it used. Te program was so
successful that the principal, Jack Tompson, has requested that
this teacher incorporate GIS into all of her classes next year.
In addition to this recent success, plans are underway to
quickly expand GIS instruction throughout the school system.
Projects in the pipeline include the development of Criminal
Justice Career Academies and an in-house GIS Career Academy
at a high school. Another exciting project involves a GIS/GPS
science and social studies program for classroom instruction at
six schools. This endeavor is a joint effort with resources from
both the School District and South Florida Water Management
(SFWMD) and involves the development and delivery of a hands-
on workshop for 27 teachers.
PArtIcIPAtE In tHE cHAllEnGE
Skilled and experienced GIS professionals and those business
and government agencies who utilize GIS can greatly enhance
the eforts to expand GIS into our public school system. Tere
are many unique ways to “pay it forward.” Contact your local
schools and see if there is an interest (or extol the virtues of GIS
to curriculum administrators and generate the excitement). You
can ofer to present or demonstrate GIS to a classroom of students
or fnd a teacher interested in using GIS and provide technical
support and mentoring (GoTo meetings and web conferencing
tools make this easier than ever). If you institute a program of
internship for students, your ROI may be fnding the GIS Tech
or GIS web programmer you’ve been looking for. Donate your
replacement PCs or hardware to local schools. Your old computers
are probably superior to what many schools are currently using.
futurE PossIBIlItIEs
Our young people represent this nation’s greatest commodity and
our future rests in their ability to succeed. We will reap the seeds
we sow. If the U.S. in general and school districts in particular
are remiss in providing opportunities for our teachers and stu-
dents to learn new geospatial skills, we will be left behind in this
technological age. If we continue to focus on high stakes testing
at the cost of introducing new technologies, countries such as
China and India will surely pass us and the next generation will
be poorly equipped to compete (Friedman, 2005). Administrators,
parents, community leaders and businesses must recognize that
we need to introduce new and improved methods for education.
Our future is at stake.
It’s time to answer the call of progress and chart a course for
public education that elevates both the students and our industry.
Perhaps the greatest beneft of including GIS in public education
is societal: it is the legacy we leave for our next generation – the
opportunity to succeed and compete in the global society.
URISA Journal • Goldstein 55
urIsA’s rolE In k12 EducAtIon
Recognizing the imminent shortfall in trained GIS profession-
als to fll the upcoming void, URISA has embraced the topic of
introducing GIS in primary and secondary education. In fact
URISA now maintains a list of over 160 schools ofering related
programs.
In addition, at its 2007 Annual Conference the URISA
Program committee organized a special track for K-12 educators,
which included sessions to demonstrate the practical approaches
for K12 teachers to implement GIS into their classroom ac-
tivities. While attendance was good the unfortunate truth is it’s
diffcult for public school teachers to fnd the time and funding
to attend conferences, especially those workshops that are not
related to core academic topics covered in the high stakes testing
requirements.
The solution to this critical issue rests with GIS professionals
reaching out to the K12 school system. To address this URISA
also included a session to engage GIS practitioners to demonstrate
how they can volunteer their services to promote the use of GIS
in local school programs. One such example of integrating the
business community in the GIS education of K12 students is the
non-proft group Hopeworks, a non-proft organization from
Camden NJ whose mission is to educate intercity K12 youth.
This organization successfully promotes integration of business
and private sector resources to help fulfll the goal of educating
youth with marketable skills in the feld of GIS (Staff 2006).
As URISA becomes more involved with this very timely
topic we will provide additional information on how businesses
can reach out to the community and reap mutually benefcial
rewards. Check out the K12 track at URISA’s 46th Annual
Conference & Exposition October 7-10, 2008 in New Orleans.
And continue to look at URISA for ways to be part of the
part of the solution.
Acknowledgements
Te author would like to acknowledge the many contributions
provided by Kristin Garrison, AICP, over the past 17 years. In
her capacity as Executive Director of Planning Zoning and Build-
ing for Palm Beach County she supported and assisted with the
emerging GIS system and contributed greatly to the success this
countywide efort enjoys today. In addition as Planning Director
for the Palm Beach County School District her vision, support,
encouragement and unwavering belief in the value of GIS in the
K-12 educational system has paved the way for this program to
come to fruition. Te author would also like to extend apprecia-
tion to Dr. Lucy Guglielmino, without whom this article would
not have been developed. Many thanks for the editorial assistance
and perseverance in the production of this piece.
About the Author
Donna Goldstein received her Masters in Educational Leadership,
and is currently attending Florida Atlantic University, pursu-
ing a PhD in Educational Leadership, Adult and Community
education. Ms. Goldstein was employed by Palm Beach County
Planning Zoning & Building from 1990 thru 2001 as the GIS
Supervisor. In 2001 she accepted the position of GIS Coordina-
tor with the Palm Beach County School District and is currently
employed in that capacity.
References:
Friedman, Thomas L., 2005). The World is Flat, A Brief History
of the Twenty-frst Century, New York: Farrar, Straus and
Giroux, 257.
Gewin, Virginia (January 22 2004). Mapping opportunities.
Nature, International weekly journal of science opportunities.
427, 376. Retrieved June 26, 2007 http://www.nature.com/
nature/journal/v427/n6972/full/nj6972-376a.html
Goodchild, Michael (2006) The Fourth R? Rethinking GIS
Education. ESRI ArcNews. Retrieved April 24, 2008 from
http://www.esri.com/news/arcnews/fall06articles/the-
fourth-r.html
Kerski, Joseph (2001). A National Assessment of GIS in American
High Schools. International Research in Geographical and
Environmental Education. 10(1) Retrieved April 3, 2008
from http://www.multilingual-matters.net/irgee/010/0072/
irgee0100072.pdf
Malone, Lyn, Palmer, Anita M., Voight, Christine L., Eileen
Napoleon, and Feaster, Laura (2005). Mapping Our World:
ArcGIS Desktop Edition, GIS Lessons for Educators. 3-47,
215-239. Redlands Ca: ESRI Press
National Research Council’s Board on Earth Sciences and
Resources of the Division of Earth and Life Studies. Report
in Brief. (2005). Learning to Think Spatially: GIS as a
Support System in the K-12 Curriculum, Washington, D.C.:
The National Academies Press, Retrieved August 21, 2007
from http://dels.nas.edu/dels/rpt_briefs/learning_to_think_
spatially_fnal.pdf
Staff (2006) Hopeworks Founder Father Jeff Putthoff Encourages
Youth Development with Technology Training. ESRI ArcNews
Retrieved April 23, 2008 from http://www.esri.com/news/
arcnews/fall06articles/hopeworks-founder.html
U.S. Department of Labor Employment & Training Administration
(2005). High Growth Industry Profle. Retrieved July 5, 2007
from http://www.doleta.gov/BRG/Indprof/geospatial_
profle.cfm
Learn more about
GIS Professional Certifcation at
www.gisci.org
Copyright © 2007 ESRI. All rights reserved. The ESRI globe logo, ESRI, and www.esri.com are trademarks, registered trademarks, or service marks of ESRI in the United States, the European Community, or certain other jurisdictions. Other companies and products mentioned herein may be trademarks or registered trademarks of their respective trademark owners.
Share the value of your work throughout your organization with Server GIS.
Extend your GIS Across the Enterprise.
As a geographic information system
(GIS) professional, you receive countless
requests for maps and spatial informa-
tion. Answering special requests, while
important, takes time and reduces
your productivity.
Imagine being able to push your
maps, models, and tools out to the
rest of the organization via focused,
easy-to-use applications. Staf f in
other departments and out in the field
could query accurate, up-to-date data
without a lot of training. This increases
their productivity as well as yours.
By making your maps, data, and
analyses readily available, you can
help others reap the benefits of the GIS
work that you do. You already know
that spatial analysis and visualization
are important parts of daily decision
making. Use Server GIS from ESRI to
help others benefit from your work.
To learn how organizations are using ESRI server GIS
to improve productivity, visit www.esri.com/server.
In the United States: 1-866-447-3036
Outside the United States: +1-909-793-2853, extension 1-1235
On the Web: www.esri.com/international
“The server-based options
provided by ESRI allow the City of
Mesa to deliver more enterprise
GIS applications throughout
our organization without high
installation and training costs.”
Jason Bell
IT Services Leader,
City of Mesa, Arizona
P_Server-Products_L0307.indd 1 3/19/07 10:30:55 AM

Sign up to vote on this title
UsefulNot useful