Content uploaded by Sibel Canaz Sevgen
Author content
All content in this area was uploaded by Sibel Canaz Sevgen on Feb 18, 2019
Content may be subject to copyright.
45
International Journal of Engineering and Geosciences (IJEG),
Vol;4, Issue;1, pp. 045-051, February, 2019, ISSN 2548-0960, Turkey,
DOI: 10.26833/ijeg.440828
AIRBORNE LIDAR DATA CLASSIFICATION IN COMPLEX URBAN AREA
USING RANDOM FOREST: A CASE STUDY OF BERGAMA, TURKEY
Sibel Canaz Sevgen 1*
1Ankara University, Faculty of Applied Sciences, Department of Real Estate Development and Management, Ankara,
Turkey (ssevgen@ankara.edu.tr); ORCID 0000-0001-5552-6067
*Corresponding Author, Received: 05/07/2018, Accepted: 06/08/2018
ABSTRACT: Airborne Light Detection and Ranging (LiDAR) data have been increasingly used for classification ofurban
areas in the last decades. Classification of urban areas is especially crucial to separate the area into classes for urban
planning, mapping, and change detection monitoring purposes. In this study, an airborne LiDAR data of a complex urban
area from Bergama District, İzmir, Turkey were classified into four classes; buildings, trees, asphalt road, and ground.
Random Forest (RF) supervised classification method is selected as classification algorithm and pixel-wise classification
was performed. Ground truth of the area was generated by digitizing classes into features to select training data and to
validate the results. The selected study area from Bergama district is complex in urban planning of buildings, road, and
ground. The buildings are very close to each other, and trees are also very close to buildings and sometimes cover the
rooftops of buildings.The most challenging part of this study is to generate ground truth in such a complex area. According
to theobtained classification results,the overall accuracy of the results is found as 70,20%.The experimental results showed
that the algorithm promises reliable results to classify airborne LiDAR data into classes in a complex urban area.
Keywords: Random Forest, LiDAR, Classification, Complex Urban Area
International Journal of Engineering and Geosciences (IJEG),
Vol;4, Issue;1, pp. 045-051, February, 2019,
46
1. INTRODUCTION
Classification of objects in an urban area is a popular
subject in a variety of research areas, such as computer
vision, machine learning, pattern recognition,
photogrammetry, remote sensing, and urban planning. In
the literature, satelliteand aerial images have been widely
used in urban area classification. Especially, land cover
changes studiesover the years by classifying satellite data
are abundantly presencein the literature. For instance, Yu
et al. (2012) monitored land cover changes and urban
sprawl dynamics 1989, 1999, and 2009 of Yantai China
by classifying satellite images in five classes. Atlanta,
Georgia’s land cover changes 1973-1998 were
categorized into six different classes (Yang and Lo, 2002).
Canaz et al. (2017), classified Istanbul, Turkey, in four
different classes to monitor land cover change between
the years of 1986-2015. On the other hand, comparing
with optical sensor data, a new technology to collect
remotely sensed data is called as Light Detection and
Ranging (LiDAR) have also been subjected as a popular
data for classification studies. LiDAR technology is
capable of collecting 3 Dimensional (3D) point cloud data
in a short time day or night. Because of the direct 3D data
acquisition, LiDAR data also have been increasingly used
for classifying urban areas into classes.
Classical data-driven techniques have been
developed for urban area classification(Rottensteiner and
Briese, 2002, Charaniya et al. 2004), the recent trend is to
use machine learning techniques to classify LiDAR data
in urban area (Lodha et al., 2006). Supervised machine
learning techniques are based on selected features and
classifier algorithm. In the literature, a variety of
supervised classification techniques, support vector
machines, neural networks, exists (Richards, J.A., and
Jia), in this study one of the supervised classification
technique, called as Random Forest (RF) was selected
and used because of its stability and robustness to the
features.
RF classification for airborne LiDAR data has been
studied using different features in order to label different
classes. For instance, Niemeyer et al. (2012) classified
three different area from Vaihingen, Germany LiDAR
dataset named as ‘ISPRS Test Project on Urban
Classification and 3D Building Reconstruction’. The
authors classified data into five categories; building, low
vegetation, tree, terrain, and asphalt ground using
Conditional Random Field (CRF) approach. However,
they only showed and evaluated the result only for classes
building and tree. Their correction result for classification
for the 3 subset area of the datawas found in average 73%
and 92% for the tree and building classes, respectively.
Guo et al. (2010) use a combination of optical
multispectral and LiDAR data to classify LiDAR data in
urban area in four classes using the Random Forest (RF)
algorithm. Many other studies using RF algorithm to
classify LiDAR data can be found in the literature
(Immitzer, et al., 2012; Rodriguez-Galiano et al. 2012;
Guan et al., 2013).
Lodha et al. (2006) employed another LiDAR data
classification work. The authors used Support Vector
Machines (SVM) for classifying LiDAR data into
buildings, trees, roads, and grass using five features:
height, height variation, normal variation, LiDAR return
intensity, and image intensity. To evaluate result they
compare ground truth and the classification result and
observed 90% accuracy. Chen et al. (2013) classified
LiDAR data to detect landslides in Three Gorges, China
by using the mean aspect, Digital Terrain Model (DTM),
and slope textures based on four texture directions;
aspect, DTM, and slope textures based on aspect; and the
moving average and standard deviation (stdev) filter of
aspect, DTM, and slope and RF algorithm. By combine
feature selection method with RF algorithm, they found a
reliable result for classifying LiDAR data and detection
of landslides. Ma et al. (2017) studied a comparison
between SVM and RF algorithm to classify LiDAR data.
The authors classified data in four categories: trees,
buildings, farmland, and ground. According to their
findings, the RF algorithm gave a better result than the
SVM algorithm for the classification of the LiDAR data.
In this study, an area from the Bergama district of
İzmir province, Turkey was chosen as study area. The
study area is very complex in shape. The feature classes
in interest are located very close to each other and some
buildings and trees are embedded. Thus,the originality of
the study is that the selected study area is very complex
in shape. Therefore, digitization andgeneration of ground
truth for the study area were carried out very carefully.
After creating the ground truth and 12 features (which
were generated from LiDAR data such as intensity,
planarity, DSM etc.) were used to employ classification
of LiDAR data.
2. STUDY AREA AND DATA
The study area was chosen from Bergama District of
İzmir. İzmir is one of the biggest provinces in Turkeyand
located in western Turkey. Bergama is the biggest district
of İzmir in the size of the area. The area of Bergama is
1573 km2. The population of the district in 2017 is
102.961.
The study area is located in the center of Bergama
district (Fig. 1). The boundary of İzmir province is shown
with the blue line, and the boundary of Bergama district
is shown in red line in Fig. 1. The true orthophoto of the
study area is also shown in Fig. 1. Since the study area’s
land cover mainly consists of ground, roads, trees, and
buildings, the study area divided four groups for
classification: buildings, trees, ground, and asphalt road.
International Journal of Engineering and Geosciences (IJEG),
Vol;4, Issue;1, pp. 045-051, February, 2019,
47
Figure 1. Location of study area, İzmir province
boundaries (a), Bergama district boundaries (b) (source:
google maps), and the chosen study area (c)
True orthophotos of the study area were generated by
Directory of Geographic Information Systems. The
images were acquired in May 2016. The pixel size of the
images is 10 cm. LiDAR data of the study area was
collected by Optech Pegasus HA-500 technic by Turkish
General Command of Mapping on 20-21 October 2014
(Kayı et al. 2015). Detailed information about the Optech
Pegasus HA-500 is given in Table 1 (Optech, 2018).
Table 1. Technical information of Optech Pegasus HA-
500 (Kayı et al. 2015)
Feature
Value
Height
150-5000 m
Effective laser repetition rate
100-500kHz
Scanning Angle
0-75º Adjustable
Accuracy (KOH)
≤ 5-20 cm.
Scanning Mechanism
Oscillating
3. METHODOLOGY
RF algorithm is often used in remote sensing
applications to classify data such as multi and
hyperspectral images, radar, LiDAR and thermal data
sets. A literature review of these applications was
presented in Belgiu and Dragut article (2016). This study
is based on RF on one of the remote sensing data airborne
LiDAR for a complex urban area. The flowchart of the
methodology of this study is given in Figure 2.
Figure 2. Flowchart of the methodology
The classification of the LiDAR data involves pixel-
based classification; therefore, 12 features were generated
and rasterized to 50 cm images. Before generating
features images and classifying the study area, airborne
LiDAR data was cleaned from noisy and duplicate points.
After preprocessing, feature images were generated in
four groups (Chehata et al., 2009; Dittrich et al., 2017)
intensity, height, eigenvalue, and echo based. The
intensity-based feature relies on the reflected energy of
the objects in the LiDAR dataset. It helps to separate
different characteristics objects such as asphalt road and
ground classes. Intensity feature image is created using
ArcGIS “Las to Raster” tool in 50 cm pixel size. Height
based features, on the other hand, were generated from
height values of the points and they play a really
important role in separating ground and other non-ground
classes, such as buildings and trees. The lidardata set was
filtered to ground points, then from those points, a DTM
in 0.5 m pixel size was generated. In addition to that, a 50
cm Digital Surface Model (DSM) was generated from all
points in the LiDAR dataset. Normalized DSM (nDSM)
was obtained by subtracting DSM from DTM. Besides,
height features based on local neighborhood helps to
determine objects, which are also different levels of the
surface. minh, minimum height value in the
neighborhood, andHd, height difference from minh of the
interested point, were generated for each point in the
LiDAR dataset (Table 2). From those features, 0.5 m
feature raster were generated using Python programming
language (Python, version 2.7).
Table 2. Height based features
Feature
Description
nDSM
Normalized Digital Surface Model
minh
Minimum height in local neighborhood of
a point
hd
The difference between minimum height
in the local neighborhood of a point and
that point height
Turkey
International Journal of Engineering and Geosciences (IJEG),
Vol;4, Issue;1, pp. 045-051, February, 2019,
48
Eigen-value based features were obtained from
eigenvalues which were calculated from the local
neighborhood covariance matrix. Eigen-values describe
the shape of the object, thus they give valuable
information about the object, whether it is a plane, line or
sphere; therefore, those features are a good indicator of a
tree or building roofs, depending on the feature (Table 3).
Sphericity, S, planarity, P, linearity, L, anisotropy, A, the
sum of eigenvalues, Sum, and change of curvature, C,
were calculated and 0.5 m feature images for each feature
were generated using Python Programming Language.
Geometric features, sphericity, planarity, linearity, and
anisotropy describe the shape of the object and give
useful information about the object whether it is a line,
plane or sphere. All geometric features were created in 3
m neighborhood points per point and then rasterized into
1 m range of mean values.
Table 3. Eigen-value based features
Feature
Description
Anisotropy
λ1 − λ3
λ1
Planarity
λ2 − λ1
λ1
Sphericity
λ3
λ1
Linearity
λ1 − λ2
λ1
Change of curvature
∑
Sum of eigenvalues
Last feature set, echo based features, helps to differentiate
objects, which have multiple returns. Therefore, a total
number of return, n, and the ratio of a number of return
over a total number of returns, t/n, were calculated for
each point and rasterized to 0.5 m images (Table 4).
Table 4. Echo based features
Feature
Description
n
Total number of returns
t /n
Number of returns over a total number of
returns
A total number of twelve features was selected and
images were generated using Python programming
Language and its machine learning and geospatial
libraries, including scikit learn (Pedregosa et al., 2011)
and GDAL (GDAL, 2018). Some of the features and
orthophoto of a part of the study area areshown in Figure
3.
RF classification (Breimen, 2001) is an ensemble
method of decision trees, which relies on randomly
selecting a subset of features and creating multiple trees
in training. and predicting new unlabeled data by voting
each tree in the ensemble. Two parameters are required
by the user, a number of trees, that define how much a
tree can grow up andnumber of features, which determine
how many new nodes can be split from parent node in the
tree.
(a) (b) (c)
(d) (e) (f)
International Journal of Engineering and Geosciences (IJEG),
Vol;4, Issue;1, pp. 045-051, February, 2019,
49
(g) (h) (i)
Figure 3. (a) True orthophoto and example of generated feature images; (b) intensity, (c) nDSM, (d) sphericity, (e)
planarity, (f) linearity, (g) total number of returns, (h) anisotropy, (i) number of total returns over number of
return images.
Ground truth of the study area (red boundary) and the
training area (blue boundary) are shown in Figure 4. The
study area and training areaswere chosen froma different
area. According to the similar studies in the literature, the
size of the training area was chosen as no lower than the
following size: 0,3 x size of the study area. The study area
was fully digitized to use it for quality control of the
classification results. Pink, green, black, and yellow
colored features represent buildings, trees, asphalt road,
and ground, respectively.
Figure 4. Study area (red), training area (blue) and their manually digitized features
Using the manually digitized training area (blue
boundary Fig. 4) and the twelve features, the
classification results were acquired by the RF algorithm.
The results are described in the following section.
4. RESULTS
Ground truth of the area was created by digitizing the
features from orthophoto of the area. Buildings, trees,
asphalt road, and the ground were carefully digitized
(Figure 5a). A part of the ground truth is used for
classification as a training site, while the ground truth of
the study area is used for the quality control of the results.
The results were classified into four groups is shown
in Figure 5b. In the figure, red, green, gray and blue
represent the buildings, trees, asphalt road, and ground
classification results, respectively. As it can be seen in
figure 5, the classes are extracted with high accuracy by
comparing the proposed methodology classification
results and the orthophoto of the study area. The
qualitative analysis was employed by comparing the
ground truth and the classification result. For this
International Journal of Engineering and Geosciences (IJEG),
Vol;4, Issue;1, pp. 045-051, February, 2019,
50
purpose, the difference between the ground truth and
result of the classes were created andillustrated in the Fig.
5c.
Figure 5. (a) Ground truth, (b) Classification result, (c) Difference
By using the ground truth and the classification results,
the quality control was employed, and a confusion matrix
was calculated. According to the results, the accuracy is
found as, 77,90% , 58,37% , 72,90%, and 71,53%for the
buildings, trees, asphalt road, and the ground,
respectively. Overall accuracy for the results is 70,20%.
Although the area is very complex, the classification
results are reliable. Only trees class have lower results
than results of the other classes. Some errors occurred
since LiDAR data, and the orthophotos, which were used
to create ground truth, were acquired in different years
and seasons. Therefore, some of the trees might
misclassified just because in LiDAR data acquisition time
(October 2014), the trees might not have leaves on the
trees. On the other hand, orthophotos were collected and
created in May, when trees have leaves. Another reason
that affects the results is that, in orthophotos, some of the
buildings were demolished while they are present in the
LiDAR data. For instance, one of the case for this kind of
building is shown in Figure 5 with red circles. Finally,
there were cars that wereon the roads in the LiDAR data,
while they are not presence in the orthophoto. This
phoneme alsomismatch the classification of asphalt roads
5. CONCLUSION
In conclusion, in this study, LiDAR data ofa complex
urban area from Bergama district, İzmir, Turkey was
classified into four groups using the RF algorithm. The
classes are as following, buildings, trees, asphalt road,
and ground. The area is very complex in terms of city
planning for instance buildings’ shapes are irregular. The
most challenging part for this study was a generation of
the ground truth since the area is very complex in shape.
Digitization of roads and buildings was very difficult and
carried out very carefully. After digitization of the area,
twelve features were created from LiDAR data, and using
the features and ground truth together, the area is
classified by RF algorithm. According to the results, the
RF algorithmwas classified the area reliably with 70,20%
overall accuracy. However, some errors occur because
the LiDAR data was acquired in October 2014 and the
orthophoto used in this study was collected in May 2016.
Because of the seasonaleffect, some of the trees were not
classified by the proposed methodology. Moreover, in
some cases, some building and trees that are available in
the orthophoto images, is not found in the LiDAR data,
Finally, for the asphalt road, there are car on the roads,
which may not be on the LiDAR data or vice versa. These
affected the classification results. Even though, these
limitations, the proposed methodology is able to classify
the complex urban area with high accuracy.
ACKNOWLEDGEMENTS
The author is very thankful to Eray Sevgen, a Ph.D.
student at the Hacettepe University for sharing Python
scripts for the RF algorithm, helping the feature
extraction and digitizing of ground truth data. The author
is also thankful to the Turkish Directory of Geographic
Information Systems for providing true orthophoto
images of the study area and the Turkish General
Command of Mapping for providing the LiDAR data of
Bergama district.
REFERENCES
Belgiu, M. and DrǎguțL. (2016). Random forests in
remote sensing: a review of applications and future
directions. ISPRS J.Photogramm. Remote Sens., 114, pp.
24-31.
Breiman, L. (2001). “Random Forests.” Machine
Learning 45: 5–32.
Canaz S., Aliefendioğlu Y. and Tanrıvermiş H. (2017).
Change detection using Landsat images and an analysis
of the linkages between the change and property tax
values in the Istanbul Province of Turkey. Journal of
Environmental Management. Vol. 200:446-45.
Chehata, N., Li, G. and Mallet, C. (2009). Airborne
LIDAR feature selection for urban classification using
random forests. Geomat. Inform. Sci. Wuhan Univ. 38,
207–212.
Dittrich, A., Weinmann, M. and Hinz, S. (2017).
Analytical and numerical investigations on the accuracy
and robustness of geometric features extracted from 3D
point cloud data. ISPRS J. Photogramm. 126, 195–208.
International Journal of Engineering and Geosciences (IJEG),
Vol;4, Issue;1, pp. 045-051, February, 2019,
51
Charaniya, A.P., Manduchi, R. and Lodha, S.K. (2004).
Supervised parametric classification of aerial LiDAR
data. In Proceedings of 2004 Conference on Computer
Vision and Pattern Recognition Workshop(CVPRW’04),
Washington, DC.
Chen, W., Li, X., Wang, Y., Chen, G. and Liu S. (2014).
Forested landslide detection using LiDAR data and the
random forest algorithm: a case study of the ThreeGorges,
China. Remote Sens. Environ. 152 (2014), pp. 291-301.
GDAL/OGR contributors (2018). GDAL/OGR
Geospatial Data Abstraction software Library. The Open
Source Geospatial Foundation. URL http://gdal.org
Guan, H., Li, J., Chapman, M., Deng, F., Ji, Z. and Yang,
X. (2013). Integration of orthoimagery and lidar data for
object-based urban thematic mapping using random
forests. International Journal of Remote Sensing, vol. 34,
issue 14, pp. 5166-5186.
Guo, L., Chehata, N., Mallet, C. and Boukir, S. (2011).
Relevance of airborne LiDAR and multispectral image
data for urban scene classification using Random Forests.
ISPRS Journal of Photogrammetry and Remote Sensing,
66 (1) , pp. 56-66/
Kayı A. Erdoğan M. and Eker O. (2015). Results of
LİDAR test performed by OPTECH HA-500 and RIEGL
LMS-Q1560. Harita Dergisi, Volume 153, pp 42-46.
Lodha, S.K., Kreps, E.J., Helmbold, D.P. and Fitzpatrick,
D. (2006). Aerial LiDAR data classification using support
vector machines (SVM). The Third International
Symposium on 3D Data Processing, Visualization, and
Transmission pp. 567-574.
Ma L., Zhou M. and Li C. (2017). Land Covers
Classification Based On Random Forest Method Using
Features From Full-Waveform Lidar Data. The
International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences, Volume XLII-
2/W7, 2017 ISPRS Geospatial Week 2017, 18–22
September 2017, Wuhan, China
Niemeyer, J., Rottensteiner, F. and Soergel, U. (2012).
Conditional random fields for LiDAR point cloud
classification in complex urban areas. In: ISPRS Annals
of the Photogrammetry, Remote Sensing and Spatial
Information Sciences I-3, pp. 263-26.
Optech, (2018), http://www.Optech.Com/Specification
[Accessed on: 14 May 2018]
Richards, J.A. and Jia, X. (1999). Supervised
Classification Techniques Remote Sensing Digital Image,
Analysis, Springer-Verlag GmbH, Heidelberg (1999) pp.
193–247.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,
Thirion, B., Grisel, O., Blonde,l M., Prettenhofer, P.,
Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay
E. (2011). Scikit-learn: Machine Learning in Python.
Journal of Machine Learning Research 12:2825-2830.
Python Software Foundation. Python Language
Reference, version 2.7. available at
http://www.python.org. Retrieved on 15.04.2018.
Rodriguez-Galiano, V., Ghimire, B., Rogan, J., Chica-
Olmo, M. and Rigol-Sanchez, J. (2012). An assessment
of the effectiveness of a random forest classifier for land-
cover classification. ISPRS Journal of Photogrammetry
and Remote Sensing, 67 (0), pp. 93-104.
Rottensteiner, F. and Briese C.(2002). A new method for
building extraction in urban areas from high-resolution
LiDAR data. Int Arch Photogramm Remote Sens Spat Inf
Sci 34(3A):295–301
Yang, X. and Lo, C.P. (2002). Using a time series of
satellite imagery to detect land use and land cover
changes in the Atlanta, Georgia metropolitan area ,
International Journal of Remote Sensing, 23, pp. 1775—
1798.
Yu, X., Zhang, A., Hou, X., Li, M., and Xia, Y. (2013).
Multi-temporal remote sensing of land cover change and
urban sprawl in the coastal city of Yantai, China.
International Journal of Digital Earth. Vol. 6, Supplement
2, 137-154.
