You are on page 1of 2

We selected training sample points from individual land use types proportional to their respective

total parcel numbers instead of total parcel areas, since parcel areas were not the comprehensive ground
surface (road surfaces were excluded) and we are only interested in how many parcel records, instead of
how many parcel areas, are correctly classified based on parcel attributes. Our study area contains a total
of 33,025 parcels. To select training parcels proportionally from different land use, we first linked parcels
to the ground truth land use so that each parcel has information regarding its ground truth land use.
Afterwards, we randomly selected approximately half of the parcels as training parcels from each land use
type and the other half (test parcels) used for accuracy assessment. In total, we have 16,506 training parcels
and 16,525 test parcels. Then, we used training parcel centroids as training points to classify the entire
parcel image. Since many of the parcel attributes are not normally distributed, some commonly used
parametric classifiers are not appro- priate to use, such as the popularly used maximum likelihood classifier
(MLC). Nevertheless, we experimented with a variety of classification algorithms, including minimum
distance, parallelepiped, spectral angular mapper, Mahalanobis distance, binary encoding, neural network,
and decision tree, in addition to MLC. The decision tree classifier turned out to have the best classification
accuracy since the classifier does not require assumptions regarding the statistical properties of the input
data and also is capable of handling both numeric and categorical inputs.

3.2.3. Models for estimating population at census blocks


We tested the potential of building characteristics (count, area, volume, etc.) for producing
population estimates at census block level. Results indicated that there is a high correlation between
population and housing units regardless of its occupancy status. To this end, we developed a multiple linear
regression model that incorporated different building characteristics, which were area, perimeter, height,
volume, and compactness. Three statistics (sum, average, and range) were derived at block level for the
afore- mentioned five building characteristics, plus the count of buildings per block, yielding a total of 16
explanatory variables. Two cases were tested, respectively: (1) considering building data for all land cover
types and (2) considering building data for single-family and multifamily residential land uses alone. In the
latter case, blocks containing mixtures of residential with other land use types were excluded. The variables
that afford maximum explanatory power were also investigated using the path coefficient, the part
correlation, tolerance, and individual correlation.
Results indicated that (1) building properties did not correlate well with housing units even when
blocks containing single- family and/or multifamily residential buildings alone were used; (2) the average
building volume and the average building height amount nearly 90% of the total correlation that can be
achieved when using all 16 variables; (3) the incorporation of either the detailed land use information or
the volume information led to higher correlation coefficients; (4) although volume information is the most
appropriate building feature for modeling small-area population, the use of building count alone produces
the most accurate estimates due to errors introduced during automatic calculation of building volumes (and
areas).

4. Conclusions
Frequent and accurate intercensal population estimates are crucial for effective public- and private-
sector planning. Remote sensing provides unique opportunities to estimate the intercensal population. In
this study, the historical developments and recent advance- ments in population estimation with remote
sensing were introduced. The integrative use of LiDAR, high-spatial-resolution satellite imagery, and night
light satellite imagery, along with the advancement of dasymetric methods, has greatly facilitated the
generation of accurate population information. Specifically, a case study in Austin, Texas, was used to
illustrate how the integration of state-of- the-art remote sensors can help refine the small-area population
estimation. It is concluded that the image-based population esti- mates possess great potential for improving
the conventional demographic-based population estimates in noncensal years. The wealth of remotely
sensed information also provides a promising solution to population estimation in countries where reliable
census data are lacking or not available.
See also: 1.01. Volume 1 Overview. 1.12. Advances in UV Remote Sensing. 3.04. Inversion of Lumped
Parameters Using BRDF Kernels. 9.22. Urban Forests and Human Well-Being.

You might also like