Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11

Contents lists available at ScienceDirect

Int J Appl Earth Obs Geoinformation

journal homepage: www.elsevier.com/locate/jag

Research Paper

A comparative analysis of pixel- and object-based detection of landslides MARK

from very high-resolution images
⁎
Ren N. Keyporta, Thomas Oommena, , Tapas R. Marthab, K.S. Sajinkumara,c, John S. Gierkea
a
Department of Geological & Mining Engineering & Sciences, Michigan Technological University,1400 Townsend Drive, Houghton, MI 49931, USA
b
National Remote Sensing Centre (NRSC), Indian Space Research Organisation (ISRO), Hyderabad 500 037, India
c
Department of Geology, University of Kerala, Thiruvananthapuram 695 581, Kerala, India

A R T I C L E I N F O A B S T R A C T

Keywords: A comparative analysis of landslides detected by pixel-based and object-oriented analysis (OOA) methods was
Unsupervised classification performed using very high-resolution (VHR) remotely sensed aerial images for the San Juan La Laguna,
Object-Oriented analysis Guatemala, which witnessed widespread devastation during the 2005 Hurricane Stan. A 3-band orthophoto of
Segment optimization 0.5 m spatial resolution together with a 115 field-based landslide inventory were used for the analysis. A binary
Landslide mapping
reference was assigned with a zero value for landslide and unity for non-landslide pixels. The pixel-based
Lake atitlán
analysis was performed using unsupervised classification, which resulted in 11 different trial classes. Detection
of landslides using OOA includes 2-step K-means clustering to eliminate regions based on brightness; elimination
of false positives using object properties such as rectangular fit, compactness, length/width ratio, mean differ-
ence of objects, and slope angle. Both overall accuracy and F-score for OOA methods outperformed pixel-based
unsupervised classification methods in both landslide and non-landslide classes. The overall accuracy for OOA
and pixel-based unsupervised classification was 96.5% and 94.3%, respectively, whereas the best F-score for
landslide identification for OOA and pixel-based unsupervised methods: were 84.3% and 77.9%,
respectively.Results indicate that the OOA is able to identify the majority of landslides with a few false positive
when compared to pixel-based unsupervised classification.

1. Introduction of OOA methods for detection of landslides, particularly with medium

Historical collections of high-resolution remote sensing data pro-
vides us valuable opportunities to analyze natural hazard events to
advance our understanding of the hazard and to minimize its effects
(Oommen et al., 2013; Bialas et al., 2016), and help in creating an in-
ventory. Traditionally, remote sensing data have been analyzed using
pixel-based classification methods for studying natural hazards, espe-
cially landslides. The pixel-based methods rely solely on spectral
characteristics of the analyzed image which significantly limits the
potential for identification of spatially contiguous areas; often resulting
in a speckled appearance in the classification product with many small
regions or single pixels classified as hazard events (Sahoo et al., 2007;
Stumpf et al., 2011).
A new solution gaining popularity in the field of image processing
for landslide studies is Object-Oriented Analysis (OOA). Studies by
Barlow et al. (2012), Van Den Eeckhaut et al. (2012), Martha and Kerle
(2010), Lahousse et al. (2011), Martha (2011), Stumpf and Kerle
(2011), Chang et al. (2012), Blaschke et al. (2014), Moosavi et al.
(2014), Wu et al. (2014), and Dou et al. (2015) highlighted advantages
and low resolution remotely sensed data. The OOA begins with image
segmentation, which is a grouping of nearby pixels based on a homo-
geneity factor. These objects, or segments, can be analyzed further
using spatial, textural, contextual, geometric, and spectral character-
istics. The additional processing eliminates false positives that are
usually missed by pixel-based classification alone (O'Neil-Dunne, 2013;
Bialas et al., 2016).
Several studies have assessed the effectiveness of OOA and pixel-
based methods (Whiteside and Ahmad, 2005; Yan et al., 2006; Oruc
et al., 2004; Chang et al., 2012; Bialas et al., 2016). However, none of
these studies assessed the effectiveness of landslide detection with very
high-resolution (0.5 m spatial resolution) remotely sensed data.
2. Study area
San Juan La Laguna (14.695°N, 91.287°W) is a community of ap-
proximately 10,000 residents located on the shores of Lake Atitlán in
south central Guatemala (Fig. 1a–c). Lake Atitlán is a steep-sided cal-
dera lake (Fig. 1d). Three post-caldera active andesitic strato-volcanoes

⁎
Corresponding author.
E-mail address: toommen@mtu.edu (T. Oommen).

http://dx.doi.org/10.1016/j.jag.2017.08.015
Received 6 June 2017; Received in revised form 28 August 2017; Accepted 31 August 2017
Available online 06 September 2017
0303-2434/ © 2017 Elsevier B.V. All rights reserved.
R.N. Keyport et al.
Fig. 1. Location of San Juan La Laguna study area on western shore of Lake Atitlán, Guatemala (a) Map of Guatemala showing the location (white arrow) of Lake Atitlán (b) Lake Atitlán
and study area (c) Actual study area and reference image used for processing (d) Terrestrial photograph showing San Juan (Photograph faces the south).
() form the southern margin with San Pedro being the closest to San
Juan La Laguna (Newhall et al., 1987). A major eruption in the Atitlán
III complex occurred 84,000 years BP resulted in the creation of Lake
Atitlán. Quaternary tephra units cover most of the highlands with
thicknesses exceeding ∼15 m, and unconsolidated Quaternary ash de-
posits have filled most of the basins of western Guatemalan highlands
and in places are thicker than 200 m (Newhall et al., 1987).
The geomorphology of the Atitlán caldera is related to the volcanic
activity of the area and the different geologic units reflect various
eruptive stages of the volcanoes and landscape dynamics of the sedi-
ments (Luna, 2007). The topographic elevations increase from 300 m
on the surrounding coastal plain to 1562 m at Lake Atitlán and 3535 m
at the summit of Volcano Atitlán (Fig. 1b–d). The volcanic events have
produced a steep sided and often unstable crater rim with intermittent
deep canyons ranging in depth between 200 and 500 m. Downhill of the
rim are located several villages that are susceptible to landslides and
lahars. At San Juan La Laguna, slopes might be as steep as 80° and are
covered with a thin layer of clay and organic rich soil. This thin veneer,
of approximately one meter of soil, creates a very unstable setting,
which in the event of a heavy rainfall is likely to fail (Luna, 2007; Smith
et al., 2015; Kern et al., 2016).
On 4 October 2005, Hurricane Stan provided the heavy rainfall
(297.5 mm) needed to mobilize significant sections of the crater rim.
Thousands of landslides occurred along the steep crater walls
2
Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11
culminating in the destruction of communities and the loss of hundreds
of lives in nearby Panabaj. The destruction significantly affected the
fragile state of the predominantly subsistence farming community.
Despite the enormous toll on the local population, very little scientific
effort was expended in mapping landslides (Luna, 2007; Cobin et al.,
2017). Without a comprehensive record, the evidence may soon be
covered by vegetation and forgotten, leading to a future reoccurrence.
The steepness of the terrain and the larger number of landslides
make field-based mapping a daunting task. We propose and evaluate
herein the object-oriented approach to identify landslides. The perfor-
mance of this approach is compared to the more conventional pixel-
based approach.
3. Data and methodology
A detailed schematic of the method followed in this study is pre-
sented in Fig. 2. A field-based inventory of landslide initiation points
was created by the Instituto Nacional de Sismologia, Vulcanologia,
Meterologia, e Hidrologia (INSIVUMEH) after Hurricane Stan (Geoló-
gicos del Mundo (GM), 2009; Supplementary material-Appendix A).
This inventory together with the remotely sensed high-resolution or-
thophotos taken shortly after the event in early 2006 was highly ben-
eficial in identifying landslide locations. The multi-spectral (red, green,
and blue bands) orthophotos (GM, 2009) with 0.5 m spatial resolution
R.N. Keyport et al.
Fig. 2. Process flow chart of methods used in OOA and unsupervised classification. Key processing steps are grouped with examples of data result.
provide a detailed information of the region and were used in this study
to create a complete landslide inventory of the San Juan La Laguna
region. Additionally, 10 m elevation contours available from IN-
SIVUMEH were used to develop a Digital Elevation Model (DEM) to
represent the topographic features of the region. Slope, flow accumu-
lation, and curvature layers were derived from the DEM. These datasets
provided the basis to test OOA and unsupervised pixel-based classifi-
cation methods for automatic landslide detection and determine the
limitations of each method.
The extent of the region was chosen to represent a broad range of
physical features including developed land such as roads, buildings, and
farmland, as well as natural features such as rivers, lakeshore, and a
variety of other terrains. The selection of a broad range of physical
features was intentional to test the robustness of the method. The pixel-
based unsupervised classification was performed in Erdas Imagine
whereas OOA was performed in Trimble’s eCognition Developer soft-
ware.
The 0.5 m resolution data available for this region were processed
based on its spectral properties for pixel-based unsupervised classifi-
cation. Parameter selection is very limited in the unsupervised classi-
fication with the primary variable being the number of classes. The
reference image was classified using a range of class values from 4 to 14
unsupervised classes resulting in 11 output images. In each of these 11
unsupervised classification images, the brightest class best represented
the landslide because of the outcropping of the underlying felsic clay
deposited by ash fall from the nearby volcanoes. Thus, the fresh land-
slides generally appear in stark contrast to surrounding vegetation.
Each of these 11 images derived from unsupervised classification were
then compared to a reference dataset, which is discussed in the vali-
dation and accuracy assessment section. Fig. 3a–b shows the 4 and 14
class unsupervised classification results to demonstrate over and under-
3
Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11
classification, respectively.
OOA is a more complex method of image classification that allows
far greater flexibility in the parameterization of inputs. A variety of
approaches using OOA have been examined in previous studies
(Espindola et al. 2006; Boesch and Wang, 2008; Martha, 2011; Stumpf
and Kerle, 2011; Lu et al., 2011; Lahousse et al., 2011; Wu et al., 2014;
Shruthi et al., 2015). For this study, we utilized a trial-and-error ap-
proach to identify the best OOA. For the OOA, the red, green, and blue
spectral bands of the 0.5 m spatial resolution orthophotos were used.
Several segmentation algorithms available in eCognition were eval-
uated and the multi-resolution segmentation provided the best result.
The scale factor 44 for the multi-resolution segmentation was de-
termined through trial-and-error.
Potential landslide locations were then mapped using a brightness
threshold derived iteratively from the segmented objects. The iterative
comparison resulted in identifying the lower brightness threshold of
140 for the landslide objects, which was utilized for the eCognition
Feature Extraction Tool. K-means clustering was conducted using in-
creasing numbers of centers until a value close to 140 was attained,
which resulted in four cluster centers. All image objects with a mean
brightness of less than 140 were eliminated from the candidate pool.
Elimination of false positives was necessary since only the bright-
ness threshold did not fully eliminate objects other than landslides, such
as farmland, roads, buildings, beaches, and rivers. The elimination of
these false positives were conducted with the use of existing knowledge
of the region gained through field-work and spatial characteristics of
the objects derived by segmentation. It is much easier to identify spa-
tially consistent objects, such as those mentioned above, than to define
the characteristics of erratic and inconsistent landslide objects. For this
reason, false positive elimination is preferred over in-depth landslide
characterization (Martha, 2011).
R.N. Keyport et al. Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11

Fig. 3. Under-classified and over-classified results of landslide identification. Blue pixels identify potential landslide candidates (a) Under-classified result with four pixel classes; pixel
extent exceeds boundaries of landslide regions (b) Over-classified result with 14 pixel classes is clearly missing significant portions of landslide extent near landslide boundaries. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

The urban area of San Juan La Laguna exhibited the highest number
of false positives because of its many high-brightness features (e.g.,
roads and roofs). Fortunately, these features are urban structures
having geometric and contextual properties, which are different from
most landslides. Hence, geometric features, like rectangular fit, were
used in the identification of rooftops, parking lots, and roads whereas
compactness was used to delineate farmland. Compactness together
with length/width ratio are useful in the identification of individual
roads and road networks. The mean difference of an object to its
neighbors was also useful in the elimination of developed and barren
lands. A 2-step K-means clustering was used in the derivation of rec-
tangular fit, compactness, length/width, and mean difference false
positive elimination thresholds. A large portion of the remaining false
positives are comprised mostly of low-lying farmlands and open spaces
concentrated near and within the city and do not exist on the steep
slopes near San Juan La Laguna that are most susceptible to landslides.

4
R.N. Keyport et al.
Fig. 4. OOA result with segmentation scale factor 44. All colored objects indicate regions identified as potential landslides based on initial brightness threshold. False positives were then
eliminated with thresholds described in Table 1 and are colored based on this parameter.
To further eliminate false positive features, slope was selected as the
parameter. Fig. 4 displays all landslide candidates as colored objects
with each false positive colored differently based on its respective
parameter of elimination.
The elimination of false positives did not provide perfect identifi-
cation of all landslide candidates nor eliminate all false positives.
Ambiguity in the dataset and variability of the spectral characteristics
of the landslides made some identification by OOA and unsupervised
classification impossible. In particular, shadowed regions containing
landslides were not identified in either OOA or pixel-based un-
supervised classification because of the vast difference in brightness
value caused by the shadows. Fig. 5 shows three landslides hidden by a
shadow which resulted in missed identification. Table 1 outlines each
parameter that was used for false positive elimination.
4. Accuracy assessment and results
Assessment of the performance of unsupervised pixel-based classi-
fication and OOA methods of landslide detection required the creation
of a validation dataset to compare the results of each method. This
consisted of the ground truth dataset provided by INSIVUMEH and a
randomly generated group of 900 data points associated with non-
landslide cases (Supplementary material-Appendix A).
The ground truth dataset represents the initiation point of landslides
that occurred after Hurricane Stan in October 2005 collected through
field-based GPS survey. This approach represents each landslide to only
one verifiable point, which limits the validation potential and
5
Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11
epitomizes the challenges of working with historical data from a remote
region. To add to the challenges of this dataset, the accuracy of each
point was low enough to render some points outside of the actual
landslide boundary. This is likely due to the steep terrain and low po-
sitional accuracy of landslide data collected using hand-held GPS. A few
reference points that were visually outside the boundary of landslides
were adjusted to the closest landslide boundary. These instances were
easily distinguishable in the orthophotos. A total of 115 landslide in-
itiation points were identified within the study area, each composed of
a single pixel. Fig. 6 shows a small portion of the study area with
corrected initiation points.
The non-landslide validation dataset was created by the selection of
900 random points across the image. The random points were gener-
ated using ArcGIS random point generator tool with a condition of each
random point being at least 3 m away from the landslide point. The 3 m
buffer for the non-landslide point was used based on the average size of
the landslide feature. All the 900 random points were carefully verified
using high-resolution image to ensure that they are not within the
landslide boundary. Those points that were within the apparent land-
slide boundary were removed from the dataset to avoid false classifi-
cation. Since the full extent of the landslides could not be confirmed
due to lack of ground truth, these boundaries had to be inferred based
on spectral continuity of the region identified based on the initiation
points in the high-resolution image. This method proved successful for
all but one data point, which was located in a drainage channel beneath
several landslide initiation points. Since the exposed channel has the
same spectral characteristics as the associated landslides, it is
R.N. Keyport et al. Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11

Fig. 5. Landslide events missed because of concealment by shadows are denoted by white arrows. These events have no spectral variation from the shadow region and could not be
identified using only a single spectral image.

Table 1 negative data points to calculate the accuracy of each method.

Parameter selection for OOA of landslide cases at scale 44. Object character types re- From these data, an accuracy assessment was conducted in R soft-
mained the same, but values varied for other scale factors tested. The first target was the ware, which calculated overall accuracy, precision, recall, and F-score
selection of landslide candidates. The following five eliminated false positives based on
object character derived using the listed method.
of each classified dataset based on the standard equations (Oommen
et al., 2010; Supplementary material-Appendix B). Overall accuracy is
Target Property Object Character Method the arithmetic mean of the correctly identified landslide and non-
landslide location. Precision refers to the number of actual landslides
Landslides Spectral (RGB) Mean Brightness > 138 K-means
clustering
relative to all the points identified as landslides. Higher precision in-
Rooftops Object Rectangular fit > 0.87 Existing dicates that landslides were correctly mapped than non-landslides.
Geometry Knowledge Recall is the percentage of all landslide events that were properly
Barren Lands Spectral (Blue) Mean difference > 100 K-means identified. A high recall value indicates that most of the landslide
clustering
events were mapped. F-score is the harmonic mean of the landslide and
Farmland Object Compactness > 4.5 K-means
Geometry clustering non-landslide validation datasets and is important because the non-
Roads Object Length/width > 6 K-means landslide cases far outnumbers the landslide cases. By using precision
Geometry clustering and recall, no extra-weight is given to the higher number of non-
Developed land Slope Mean Slope < 9° Existing landslide events (Oommen et al., 2010). The results from the OOA and
Knowledge
the 11 different pixel-based unsupervised classification are presented in
Figs. 7 and 8.
impossible to interpret from the reference image the toe of the land- Overall accuracy maxima were obtained at unsupervised classifi-
slide, and the channel border. Therefore, this point was removed from cation levels of 9, 12 and 13 with a value of 94.3%, while the maximum
the validation dataset. This resulted in a data set of 894 non-landslide overall accuracy for OOA method was 96.5% (Figs. 7 and 8). Overall
reference pixels. accuracy does not provide a complete picture of the accuracy of the
The landslide and non-landslide validation datasets were combined methods, as it does not account for the fact that the number of non-
and a binary reference was assigned with null value for a landslide and landslide reference points outnumbers the number of landslide points
unity for non-landslide pixels. Extraction of values to these data was by nearly a factor of eight in the validation set. This leads to data
then done in ArcMap for each of the 11 unsupervised classifications and skewed toward non-landslide identification. It is important to evaluate
the landslide shape file created by OOA. Each produced a file of 1009 the precision, recall, and F-score of predicting landslide and non-land-
pixels consisting of true positive, true negative, false positive, and false slide classes to determine the strengths and weaknesses of each method

6
R.N. Keyport et al.
Fig. 6. Zoomed map indicating landslide initiation points that were corrected due to inaccuracy in data collection. Green indicates the original location of the initiation point and red is
the corrected location which falls definitively within the indicated landslide. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)
Fig. 7. Overall accuracy of OOA and unsupervised classification results. OOA scale factor
(44) is plotted in green square. Class 4–14 indicate unsupervised classification levels as
indicated by number. (For interpretation of the references to colour in this figure legend,
the reader is referred to the web version of this article.)
7
Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11
and to fully understand the quality of the result. When examining the
recall values of landslide identification, it appears that unsupervised
classification with four classes outperformed the best OOA by 14%. The
downfall of this is that four-class unsupervised classification had by far
the highest false positive identification, which resulted in having the
lowest overall accuracy and F-score. These results indicate that the
pixel-based unsupervised classification is identifying a much higher
percentage of the study area as landslides than actual.
The more preferable way of examining the result is with the use of
F-score. This provides a balanced view of the accuracy of how well each
test detects both landslide and non-landslide cases because it takes into
account the number of samples of each dataset using precision and
recall. F-score for OOA methods outperformed all pixel-based un-
supervised classification methods in both landslide and non-landslide
cases. The best F-score for landslide identification for OOA and un-
supervised methods were 84.3% and 77.9%, respectively. This differ-
ence highlights the effect of the low precision values achieved by un-
supervised classification. Precision values of 57% to 71% for pixel-
based and 87% for OOA demonstrate the poor false positive removal of
unsupervised classification.
Although methods and spatial resolution of data varied in previous
studies, the overall accuracy results for pixel-based and object-oriented
approaches varied from 38% to 4% improvement (Bialas et al., 2016).
The results from this study were most similar to those in Chang et al.
(2012) which also examined pixel- and object-based mapping ap-
proaches for landslides. The result of this study fit well within this range
validating the effectiveness of the methods performed.
When mapping the landslides, it is important to capture the
R.N. Keyport et al.
geometry besides the location. The geometry of a subset of the land-
slides was carefully mapped manually using the high-resolution image.
The manually mapped landslide geometry was then compared to the
landslides mapped using OOA (Fig. 9a–c). In Fig. 9, the manually
mapped landslides are in blue, and the OOA mapped landslides are in
the red. It is observed from Fig. 9 that in general the red lines overlay on
top of the blue lines and have a very good agreement between the
manual and OOA mapped landslides. Figs. 9b–c show zoomed locations
from the study area. It is evident from the zoomed image that though in
general there is good agreement between manual and OOA mapped
landslides, the OOA analysis seems to fail to map the entirety of the
landslides especially closer to the vegetated region. Besides, it is ob-
served in Fig. 9c that in some cases the OOA mapped landslides are
subdivided into smaller objects compared to a single landslide as as-
signed manually.
Also, we evaluated how well the OOA mapped landslides compare
to the landslide type. The landslide type was assigned based on a slope
threshold of 30° as observed in the field. Slopes greater than 30° being
rock fall and slopes between 15 and 30° being debris flow. Fig. 10 shows
the comparison OOA mapped landslide with rock fall and debris flow
within a subset of the study area. It is observed from Fig. 10 that in
general, the OOA mapped landslide boundaries have better agreement
with rock fall compared to the debris flow. The OOA mapped landslides
seem to be incomplete or entirely miss the boundary of debris flow
whereas the extent of the rock fall is mapped well by the OOA analysis.
8
Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11
Fig. 8. Precision recall, and F-score plots for
comparison of landslide and non-landslide
identification results. OOA scale factor (44)
is plotted in green square. Class 4–14 in-
dicate unsupervised classification levels as
indicated by number. (For interpretation of
the references to colour in this figure legend,
the reader is referred to the web version of
this article.)
5. Discussion
The mapping of landslides at San Juan La Laguna by pixel-based
unsupervised classification and OOA proved moderately successful with
very high-resolution images. Despite data limitations, overall ac-
curacies of 96.5% and 94.3% were obtained using OOA and un-
supervised classification methods, respectively. After confidence in-
terval testing, it was determined that the improvement in overall
accuracy was not statistically significant. The overall accuracy is
skewed toward non-landslide identification and does not accurately
represent the effectiveness of the methods. The highest F-score values
for landslide identification were 84.3% and 77.9% for OOA and pixel-
based unsupervised classification, respectively. This difference of 6.4%
more accurately represents the success of OOA methods than overall
accuracy.
The low precision values for landslide identification using pixel-
based unsupervised classification highlight the greatest weakness of this
method. False positive removal is limited to pixel values rather than
object characteristics, making the removal much more challenging
when little is known of the study area and landslide characteristics.
Large urban areas create a unique challenge in the processing of re-
motely sensed data and should be analyzed cautiously on a case-by-case
basis. This is especially true when dealing with high-resolution images
where individual buildings and small structures may be assigned their
own image objects. A potential solution would be the use of multiple
R.N. Keyport et al. Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11

Fig. 9. Comparison of the manually and OOA mapped landslide geometries. a) Comparison of the subset of the study area, b) and c) Zoomed areas from Fig. 9a shown as green and yellow
box. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

segmentation values optimized for the varying dominant regions of the
image.
The greatest limitation for both methods is the inability to confirm
the extent of landslides where no ground truth data are available. A
completed landslide inventory of this region would require significant
assumption of landslide extent, especially where landslides intersect
drainage channels, which are spectrally and texturally similar. Despite
these setbacks, OOA was able to identify the majority of the landslides
with a few false positive cases using a high-resolution orthophoto and
DEM. The comparison of the geometry of manually mapped landslides
with OOA mapped showed generally good agreement. The comparison
of the landslide type demonstrated that the OOA method mapped rock
falls better than the debris flows.
6. Conclusion
A comparative analysis of landslides detected by pixel-based and
OOA methods was performed using very high-resolution (VHR) re-
motely sensed aerial images for the San Juan La Laguna, Guatemala,
which witnessed widespread devastation during the 2005 Hurricane

9
R.N. Keyport et al.
Fig. 10. Comparison of the landslide type (rock fall and debris flow) with the OOA mapped landslides.
Stan. The OOA methods provided an improved mapping of landslides
with fewer false positives compared to the pixel-based approach. The
additional object characteristics that the OOA provide the ability for
improved false positive elimination compared to just the pixel char-
acteristics that unsupervised pixel-based method has to offer.
Comparing the geometry of the OOA mapped landslides to manually
mapped landslides demonstrates good agreement. The comparison of
the landslide type with OOA mapped landslide show that the OOA
approach has better success in mapping rock falls compared to debris
flows.
Acknowledgements
This work is partially supported by the National Science Foundation
(NSF) Partnership for International Research and Education (PIRE)
Grant No. 0530109. The work of Patrice Cobin during her Peace Corps
service and her collaboration with INSIVUMEH aided greatly in ob-
taining the data for this research and is herewith acknowledged.
Additionally, we are grateful to Rudiger Escobar Wolf, Aaron
Poznanovic, and Steven Gromatka for assistance and support.
Sajinkumar acknowledges University Grants Commission (UGC),
Government of India, for granting Raman Post-Doctoral Fellowship,
which enabled him to carry out the Post-Doctoral Research work at
Michigan Technological University.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in the
online version, at http://dx.doi.org/10.1016/j.jag.2017.08.015.
References
Barlow, J., Franklin, S., Martin, Y., 2012. High spatial resolution satellite imagery, DEM
derivatives, and image segmentation for the detection of mass wasting processes.
Photogramm. Eng. Remote Sens. 72 (6), 687–692.
10
Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11
Bialas, J., Oommen, T., Rebbapragada, U., Levin, E., 2016. Object-based classification of
earthquake damage from high-resolution optical imagery using machine learning. J.
Appl. Remote Sens. 10 (3). http://dx.doi.org/10.1117/1.jrs.10.036025.
Blaschke, T., Feizizadeh, B., Hlbling, D., 2014. Object-based image analysis and digital
terrain analysis for locating landslides in the Urmia Lake Basin, Iran. IEEE J. Sel. Top.
Appl. Earth Obs. Remote Sens. 7 (12), 4806–4817.
Boesch, R., Wang, Z., 2008. Segmentation Optimization for Aerial Images with Spatial
Constraints. The Int. Archives Photogramm. Remote Sens. Spat. Inf. Sci. XXXVII (Part
B4) (Beijing 2008).
Chang, K.T., Liu, J.K., Wang, C.I., 2012. An object-oriented analysis for characterizing the
rainfall-induced shallow landslide. J. Mar. Sci. Technol. 20 (6), 647–656.
Cobin, P., Oommen, T., Gierke, S., Sajinkumar, K.S., 2017. Significance of variable se-
lection and scaling issues for probabilistic modeling of rainfall-Induced landslide
susceptibility. Spat. Inf. Res (Under review).
Dou, J., Chang, K.T., Chen, S., Yunus, A.P., Liu, J.K., Xia, H., Zhu, Z., 2015. Automatic
case-based resoning approach for landslide detection: integration of object-oriented
image analysis and a genetic algorithm. Remote Sens. 7, 4318–4342.
Espindola, G.M., Camara, G., Reis, I.A., Bins, L.S., Monteiro, A.M., 2006. Parameter se-
lection for region-growing image segmentation algorithms using spatial auto-
correlation. Int. J. Remote Sens. 27 (14), 3035–3040.
GM (Geológicos del Mundo), 2009. Guidance for the development of susceptibility maps:
A hillside movement in the lagoatitlán basin, Guatemala.With the financial support of
the Spanish Agency for International Cooperation for Development (AECID).
Kern, A.N., Addison, P., Oommen, T., Salazar, S.E., Coffman, R.A., 2016. Machine
learning based predictive modeling of debris flow probability following wildfire in
the intermountain Western United States. Math. Geosci. 1–19.
Lahousse, T., Chang, K.T., Lin, Y.H., 2011. Landslide mapping with multi-scale object-
based image analysis − a case study in the Baichi Watershed, Taiwan. Nat. Hazards
Earth Syst. Sci. 11, 2715–2726.
Lu, P., Stumpf, A., Kerle, N., Casagli, N., 2011. Object-oriented change detection for
landslide rapid mapping. IEEE Geosci. Remote Sens. Lett. 8 (4), 701–705.
Luna, B.Q., 2007. Assessment and Modelling of Two Lahars Caused By Hurricane Stan At
Atitlan, Guatemala, October 2005. Thesis (unpublished). University of Oslo, Norway.
Martha, T.R., Kerle, N., Jetten, V., vanWesten, C.J., Kumar, V.K., 2010. Characterising
spectral, spatial and morphometric properties of landslides for semi-automatic de-
tection using object-oriented methods. Geomorphology 116, 24–36.
Martha, T.R., 2011. Detection of Landslides By Object-Oriented Image Analysis Thesis
(unpublished). University of Twente, The Netherlands.
Moosavi, V., Talebi, A., Shirmohammadi, B., 2014. Producing a landslide inventory map
using pixel-based and object-oriented approaches optimized by Taguchi method.
Geomorphology 2014, 646–656.
Newhall, C.G., Paull, C.K., Bradbury, J.P., Higuera-Gundy, A., Poppe, L.J., Self, S., Bonar
Sharpless, N., Ziagos, J., 1987. Recent geologic history of lake atitlan, a caldera lake
in Western Guatemala. J. Volcanol. Geotherm. Res. 33, 81–107.
O'Neil-Dunne, J., 2013. Letters from the SAL: ECognition. University of Vermont.
R.N. Keyport et al.
Oommen, T., Baise, L., Vogel, R., 2010. Validation and application of empirical lique-
faction models. J. Geotech. Geoenviron. Eng. 136 (12), 1618–1633.
Oommen, T., Baise, L.G., Gens, R., Prakash, A., Gupta, R.P., 2013. Documenting earth-
quake-induced liquefaction using satellite remote sensing image transformations.
Environ. Eng. Geosci. 19 (4), 303–318.
Oruc, M., Marangoz, A.M., Buyuksalih, G., 2004. Comparison of pixel-based and object-
oriented classification approaches using landsat-7 ETM spectral bands. In: Proc. XX
ISPRS Congress. Istanbul, July 19–23, 2004.
Sahoo, B.C., Oommen, T., Misra, D., Newby, G., 2007. Using the one-dimensional s-
transform as a discrimination tool in classification of hyperspectral images. Canadian
J. Remote Sens. 33 (6), 551–560.
Shruthi, R.B.V., Kerle, N., Jetten, V., Abdellah, L., Machmach, I., 2015. Quantifying
temporal changes in gully erosion areas with object-oriented analysis. Catena 128,
262–277.
Smith, D.M., Oommen, T., Bowman, L.J., Gierke, J.S., Vitton, S.J., 2015. Hazard assess-
ment of rainfall-induced landslides: a case study of San Vicente volcano in central El
11
Int J Appl  Earth Obs Geoinformation 64 (2018) 1–11
Salvador. Nat. Hazards 75 (3), 2291–2310.
Stumpf, A., Kerle, N., 2011. Combining random forests and object-oriented analysis for
landslide mapping from very high-resolution imagery. Procedia Environ. Sci. 3,
123–129.
Van Den Eeckhaut, M., Kerle, N., Hervas, J., 2012. Mapping Vegetated Landslides Using
LiDAR Derivatives and Object-Oriented Analysis Thesis (unpublished). the
Netherlands, University of Twente.
Whiteside, T., Ahmad, W., 2005. A comparison of object-oriented and pixel-based clas-
sification methods for mapping land cover in northern Australia. Proc. SSC 2005
Spatial Intelligence, Innovation and Praxis 1225–1231.
Wu, X., Ren, F., Niu, R., 2014. Landslide susceptibility assessment using object mapping
units, decision tree, and support vector machine models in the Three Gorges of China.
Environ. Earth Sci. 71, 4725–4738.
Yan, G., Mas, J.F., Maathuis, B.H.P., Xiangmin, Z., Van Dijk, P.M., 2006. Comparison of
pixel-based and object-oriented image classification approaches − a case study in a
coal fire area, wuda, inner Mongolia, China. Int. J. Remote Sens. 27 (18), 4039–4055.