DEVELOPING METHODOLOGY TO MAP TREE CANOPY IN URBAN AREAS
FROM LOW COST COMMERCIAL UAVS
Niluka Munasinghe1, U.I. Manatunga1, H.M.R. Premasiri1, N. Lakmal Deshapriya2, S.L. Madawalagama2 and Lal
1Department of Earth Resources Engineering, University of Moratuwa, Sri Lanka
2Geoinformatics Center, Asian Institute of Technology, Thailand
KEYWORDS: UAV, Urban Planning, Tree Canopy, DSM, Orthoimage
Urban Green Spaces are essential for the well-being of the urban society. Mapping and monitoring of tree canopy in
urban landscape is an integral part of the Urban Planning. Recent popularity and improvement of UAV technology
enables planners to assess available urban green spaces in high resolution with greater accuracy. One of most famous
UAV series is Phantom drone series. In this study, we have used Phantom 3 Drone (Professional Edition) which has
4K camera, approximately 20minutes flight time and range of maximum 5 km. But, most of the available commercial
Drones lacks Near Infrared (NIR) observation which is crucial to map and monitor vegetation. In this paper, we are
exploring an efficient methodology which is based on textural information from RGB image and elevation
information derived from a Drone, to extract tree canopies in Urban Area. A university environment which consists
of a desirable amount of tree canopy and buildings have been mapped to demonstrate this methodology in this paper.
Finally, results were compared to digitized tree canopy layer to validate the results.
Urban planning means dealing with the constant change as communities, cities, and towns which continue to
evolve. In response, professional planning, aside from requiring expertise, understanding, and political savvy,
requires the correct technology to collect topographical information/data to facilitate effective translation of
strategic visions into strategic action plans. However, collecting data is a challenging aspect of urban planning,
because, effective urban planning requires large amounts of accurate data, which is difficult to collect since most
often, locations, prove difficult to access. Further, most of the traditional technology for aerial surveys is highly
expensive. An ideal methodology that is cost effective, highly accurate, and able to navigate difficult locations is
required. One such methodology is UAV photogrammetry.
UAV technology has revolutionized the aerial photogrammetric mapping, and its applications are speedily
increasing. When compared with present day conventional methods of photogrammetry, UAV photogrammetry
has improved and facilitated the creation of an advanced tool for photographic measurement. The major
components of a UAV include an unmanned aircraft, a transmitter, a communication link, an image sensor such
as a digital or infrared camera and mission planning. UAV photogrammetry can also be integrated with a LiDAR
system. UAV photogrammetry facilitates both autonomous and semi-autonomous remote controlling and ensures
the accuracy of an approximate 1cm to 2cm. Further, it creates new opportunities for application in close range
domains by combining terrestrial and aerial photogrammetry, but its major advantage is that it is highly cost
effective, as it enhances problem-solving applications and offers an alternative to the traditionally manned and
high-cost aerial photogrammetry. UAV photogrammetry enables application in both small scale and large scale,
with varying or similar system costs, depending to an extent on the intricateness of the system. Nevertheless, this
study utilizes a UAV system approach and its cost effective applications to develop a methodology that facilitates
mapping tree canopies in urban areas.
Digital images can be directly captured by use of a digital camera, or they can be scanned from aerial photographs.
Being in digital form facilitates the easy storage and management of the photographic information as digital maps,
digital orthoimage and DEM in a computer. This method of using a computer to process digital images and aerial
photography is called the digital photogrammetric method.
A consumer grade UAV system was used to acquire the aerial photographs required for the study. Specifically,
the Professional Edition, DJI Phantom 3 with a Sony EXMOR camera was utilized as the system for data
acquisition of aerial photographs of the area of study.
2. STUDY AREA AND DATA USED
2.1 Study Area
The University environment (The Asian Institute of Technology, Thailand) was selected as the study area.
Because, it has land-cover type consists of combination of buildings and trees, which can be, considers as eco-
friendly urban setting.
Figure 1: Study Area (c WorldView-2 image 2012 DigitalGlobe)
2.2 Data from UAV System
Phantom 3 professional drone is selected as the platform to collect aerial images of the study area which are
processed to make the orthoimage and DSM. Phantom series drones by DJI are one of the most common civilian
drones among the community mainly used for photography and as hobby. In recent studies, accurate mapping
capacity of phantom 3 drone is exploited and it is used in this task to acquire aerial images. The drone is chosen
considering the affordability and mobility. According to DJI, the Phantom 3 professional weights 1280g and
falls to the micro UAV category. It is listed with a flight time of 23 minutes with a fully charged lithium polymer
battery. The included remote control unit operates on a 2.4GHz ISM frequency and has the ability to
communicate with the Phantom 3 up to 5 km range from the remote control. The 12 megapixel Sony EXMOR
camera is factory built and mounted in 3 axis stabilization gimbal which provides clear stabilized images.
Flight planning was done with a software named Map Pilot for DJI, which is used to execute full autonomous
flight. The advantage of modern flight planning software is requirement of minimal interference for operators
which allows to focus on field operation rather than calculation. The app automatically computes the exposure
location and flight paths according to the given flying altitude and image overlap. Flying altitude is set to 100m
above ground level which provides 4.2cm per pixel average resolution in unprocessed photographs. Both side
overlap and forward overlap were set to 80% to provide more key points as possible for accurate
photogrammetric processing. Total number of 1670 near vertical photos were collected by 7 successive flights
which were processed together to obtain DSM and orthoimage.
3.1 DSM & Orthoimage Generation
The DSM & orthoimage generation was carried out using Agisoft PhotoScan software. The below diagram shows
the basic processing steps of the software.
Figure 2: Processing steps of Agisoft PhotoScan
The first step is the feature matching across photos. Here the PhotoScan software identifies, in the source
photos, the stable points under variations of viewpoint and lighting. After identifying, the software produces
descriptors for the stable points based on their local neighborhood. The produced descriptors are then utilized to
discover any correlation, communication or consistency among the photos. Aspects of this process are similar to
the popular SIFT approach, but unlike the SIFT approach, this software employs different algorithms to acquire
an alignment of slightly higher quality. Subsequently, the PhotoScan software applies a greedy algorithm to
approximately locate the camera positions and then with the use of a bundle adjustment algorithm, proceeds to
refine them. A variety of algorithms are at hand at the following dense surface reconstruction phase. Here, the
methods dependent on the pair-wise depth map computation are the Exact, Height field and Smooth methods
while the one dependent on the multi-view approach is the Fast method. At the next phase, the software
develops a texture atlas by creating parameters of the surface possibly by cutting it (the surface) into smaller
pieces and blending source photos. Resulted DSM and orthoimage shown below.
Figure 2: Left: DSM; Right: Orthoimage
across the photos
reconstruction Texture Mapping
3.2 Tree Canopy Mapping
A decision tree was developed to map the tree canopy
(See Figure 2). Tree canopy mapping proves to be a
complex task without near infrared band. So we used the
height information and object parameters to map the
canopy. This task was carried out using eCognition
The most elementary step while using eCognition for
image analysis is the segmentation or separation of a
scene depicting an image turn, into image objects.
Therefore, segmentation at the initial stage involves the
separation of an image into isolated regions that are
represented by simple unclassified image objects, known
as ‘image object primitives’’. To get the most realistic
image objects, the necessary elevation information must
be represented, and it should not be too big or too small
for the analysis task.
The first step involved is creating image objects. To do
this, the multiresolution segmentation algorithm was
utilized. This is because it organizes the image into
objects, regions with similar pixel values. Therefore, the
areas that are homogenous produce objects that are
larger, and areas that are heterogeneous produce smaller
objects. How homogeneous/heterogeneous the objects
are allowed to get is operated by the ‘scale parameter’.
An assumption, that the buildings and trees were always
elevated was made. Further, the study area had not
experienced much change in terrain elevation, so simply
the applying elevation threshold is good enough to
classify elevated objects. In the second processing, step
trees had to be separated from buildings, as both were
elevated objects. Here again the DSM information was
used. It is apparent that Elevation (DSM) texture of the
buildings is far smoother than trees, which can be
represented by object’s standard deviation (Trees - high
Elevation Standard Deviation and Building - law
Elevation Standard Deviation). So, the thresholding on
Standard Deviation was used to separate Trees and Buildings, which are both elevated objects.
Therefore, the standard deviation of the DSM was applied to separate the trees and the buildings. Nonetheless,
some trees were still classified as ‘Buildings’ even after refining the classification using the standard deviation
of the DSM layer. It was obvious that the green layer contained significant information about vegetation. So
well know Green Ratio “green/(red+green+blue)”, which measures green component of the color was used for
further refinements. Some areas of the buildings had not thus far been classified or were de-classified again
because they accomplished one of the conditions before. Some of the not classified objects were highly
surrounded by ‘Buildings’ objects. If an unclassified object had a high common border to ‘Buildings’ objects it
ought to belong to the class ‘Buildings’. The neighborhood relationships of objects were then used to refine the
classification. There were still some misclassified objects, but they were very small. The compactness of the
‘object of interest’ was the separating condition here, based on the assumption, that the buildings have a regular
shape. Refining the classification depended on that characteristic.
Results showing the separation of Tree Canopy from the building is shown in Figure 4. Additionally, these results
were compared with respect to manual digitized results of the study area, which is shown in Table 1. 64% of Total
Accuracy was obtained through this methodology using consumer-grade UAV.
Figure 3: Decision Tree
Figure 4: Resulted tree canopy map
Table 1: Accuracy Assessment of the Classification
5. CONCLUSION AND DISCUSSION
Total Accuracy with respect to manual digitizing and visual inspection of the results; validate the approach that
we have used in this paper. Even without Near Infrared band, which is not available in consumer-grade UAV’s,
this approach successfully maps the Vegetation Canopy. It is prominent that this accuracy was good enough for
urban planners to assess accessibility to green spaced of urban population for sustainable planning of the urban
Unavailability of freely available Software for Object Oriented Classification, high cost of UAV’s and
legal barriers associated with UAV are the main obstacles to this approach. With the current advancement of
technology and the popularity of UAVs will help to overcome these obstacles in the future.
Uzar, Melis and Naci Yastikli. "Automatic Building Extraction Using Lidar and Aerial Photographs". Bol. Ciênc.
Geod. 19.2 (2013): 153-171. Web.
Attarzadeh, R. and Momeni, M. (2012) ‘Object-based building extraction from high resolution satellite imagery’,
ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences,
XXXIX-B4, pp. 57–60. doi: 10.5194/isprsarchives-xxxix-b4-57-2012.
Baatz, M., Benz, U., Dehghani, S. and Heynen, M. (2004) eCognition User Guide 4 , Definiens Imagine GmbH,
Baatz, M. and Schape, A. (2000) Multiresolution Segmentation: an optimization approach for high quality
multi-scale image segmentation, Proceedings of Angewandte Geogr. Informationsverarbeitung XII , Strobl, J.
and Blaschke, T. eds., Wichmann, Heidelberg, pp. 12-23.
Algorithms used in Photoscan (2015) Retrieved 17 September 2015, Available at
Hofmann, A.D., Mass, H., Streilein, A., 2002. Knowledge-based building detection based on laser scanner data
and topographic map information, IAPRS, Vol. 34, Part 3A+B, pp.163-169.
Trimble Navigation Limited, 2010. Getting started – Example: Simple building extraction. In eCognition 8.0:
Guided Tour Level 1. United States, pp. 32-55.
Total Accuracy: 64%
Lee, D.S., Shan J., Bethel, J.S., 2003. Class-guided building extraction from IKONOS imagery.
Photogrammetric Engineering and Remote Sensing 69 (2), 143-150.
Jiang, N., Zhang, J.X., 2008. Object-oriented building extraction by DSM and very high-resolution orthoimages.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol
XXXVII. Part B3b.
Song, W. and Haithcoat, T., 2005. Development of comprehensive accuracy assessment indexes for building
footprint extraction. IEEE Transactions on Geoscience and Remote Sensing 43(2), pp. 402–404.
Wei, Y.; Zhao, Z.; Song, J. Urban Building Extraction from High-Resolution Satellite
Panchromatic Image Using Clustering and Edge Detection. In Proceedings of IEEE International Geoscience
and Remote Sensing Symposium, Honolulu, HI, USA, 20–24 September 2004;Volume 7, pp.2008-2010.
Wei, Y.; Zhao, Z.; Song, J. Urban Building Extraction from High-Resolution Satellite Panchromatic Image Using
Clustering and Edge Detection. In Proceedings of IEEE International Geoscience and Remote Sensing
Symposium, Honolulu, HI, USA, 20–24 September 2004; Volume 7, pp. 2008-2010.
Agisoft, L. (2013). Agisoft PhotoScan user manual. Professional edition, version 0.9. 0. AgiSoft LLC (Pub),