Automated Tree Crown Detection and Delineation using Watershed

Segmentation
Muhammad Athar Haroon, Mirza Muhammad Waqar
Key Laboratory of Digital Earth Science
Institute of Remote Sensing and Digital Earth (RADI)
University of Chinese Academy of Science (UCAS),
Beijing 100094, China
Muhammad Atif Wazeer
Institute of Tibetan Plateau Research
University of Chinese Academy of Science, Beijing, China
Corresponding Author:

mirza.waqar@seecs.edu.pk

Abstract

In this study, a tree detectiondelineation algorithm designed

specifically for high-resolution digital imagery of different
age group trees is presented and rigorously evaluated.

Tree Crown delineation and species type identification

are the core issues in forest inventory mapping and carbon
stock estimation. With the availability of high resolution
satellite imagery, many methods have been developed for
accurate detection and delineation of tree crowns. However tree crown detection and delineation become very difficult when stand density is very high or if there is exist
mixed forest specifies in a particular area. Few researchers
accurately delineate tree crowns even the stand density is
very high by incorporating Lidar data but the availability
of lidar data is limited over certain regions. Hence tree
crown detection and delineation is still an issue in data
poor regions. The main objective of this research is to detect and delineate tree crowns using morphological image
segmentation techniques without incorporating high profile
(DEM, Lidear etc.). Image Morphological Segmentation
(IMS) is successfully implemented to detect and delineate
tree crowns in sparsely dense forest. Confusion matrix is
also constructed in order to check the accuracy of develop
algorithm.

Many efforts have been made to detect and delineate

tree crowns automatically / semi-automatically. Different
researchers used multi-scale filtering and segmentation to
extract from single band aerial imagery [4] [3]. Le Wang
(2008) delineated individual tree crowns (ITC) by implementing watershed segmentation method in combination
with orthogonal wavelets, edge enhancement filter and suppressing texture over tree crown resulted from tree branches
and twigs [9]. Francois A. Gougeon et al (2006). used valley following and ruls based isolation to detect and delineate
indiviusal tree crown using IKNOS images [2]. M. Kalapala
(2014) perform tree crowns delineation using morphological reconstruction and watershed transformation [5]. [8]
identified individual tree crowns as local extrema points in
the Laplacian-of-Gaussian scale-space pyramid that is constructed based on linear scale-space theory.
Most of previously developed algorithms are processing extensive and hence are not suitable for large size
datasets, [1] has developed an algorithm for tree crown
delineation using discriminative undirected probabilistic
graphical model based on Conditional Random Field (CRF)
based on Performance Computing Architecture and used
this algorithm to estimate tree crown for whole of Continental United States. Hence there is a need of an algorithm
that first discard those components which are of least importance before processing actual data. Object based image
analysis techniques are also frequently used in order to detect tree crowns accross the world [10] [6] [11].

1. Introduction
Tree Crown delineation and species composition estimation are cornerstones of forest inventory mapping and key
elements to forest management decision making. Improved
mapping techniques are constantly being sought in terms
of speed, consistency, accuracy, level of detail, and overall
effectiveness. Automated tree detection and delineation algorithms may be an effective means to accomplish this task.

There are some studies that focus on comparison of

1

different tree crown detection and delineation techniques.

Y. Ke (2012) et al. compared Marker Controlled Watershed Segmentation, Region Growing and Valley following
techniques and found all of these algorithms delineate tree
crowns effectively however region growing method perofm
well in terms of overall accuracy [7].

1.1. Goal and Objectives

The key idea of this research work is to implement Image morphological segmentation to detect and delineate tree
crowns automatically. This is achieved based on following
objectives:
1. Apply different image morphological segmentation
(IMS) techniques to identify suitable IMS technique
to detect and delineate tree crown.
2. Implementing selected IMS on testing samples to detect and delineate tree crowns.
Figure 1. Selected Portion of Chichawatni Reserved Forest, Punjab
Pakistan

3. Developing ground truth data by using object based

image analysis region growing technique.
4. Estimating confusion matrix by using ground truth
data and delineated tree crowns with the help of IMS.

image segmentation and image denoising. In the context of

remote sensing, it has been applied successfully to generate morphological features useful for the discrimination of
objects in high spatial images. The two fundamental oper-

2. Study Area
The study area chosen for this study is forest areas
is the Chichawatni Irrigated Plantation, also known as
Chichawatni Reserved Forest, in the Sahiwal District of
Punjab Province, Pakistan. This irrigated plantation is managed by the Punjab Forest Research Institute (PFRI), Forest,
Wildlife and Fisheries Department, Govt. of Punjab, Faisalabad, Pakistan. Map of study area is shown in Figure 1.

3. Dataset and Preprocessing

In order to develop tree crown detection and delineation,
WorldView-3 satellite data of June 15, 2015 was acquired
from Digital Globe. Although WorldView-3 data of whole
Chichawatni Irrigated Plantation was available but a small
subset of whole forest was chosen to develop and test algorithm. A representative subset of Chichawatni Irrigated
Plantation was selected that have different age of trees with
sparse to high density. As image morpohlogical operations
normally required gray scale images; for this purpose RGB
image is converted to gray scale using gray scale transformation. In order to enhance features in selected image,
background was estimated and removed from satellite image.

Figure 2. Methodological Framework

ators in mathematical morphology are erosion and dilation.

These operators are applied to an image with a set of known
shapes, called the structuring elements (SEs). The erosion
consists of finding where the SE fits the objects in the image. The dilation, which is dual to the erosion, shows where
the SE hits the objects. From these basic operators, differ-

4. Methodological Framework
Mathematical morphology is a tool widely used in very
diverse image processing tasks such as feature detection,
2

ent combined operators can be constructed. The most common are opening and closing. An opening result in a new
image where small bright objects (compared to their surroundings) are deleted. This means that they get assigned
the gray-scale value of their surroundings. Dark objects are
left unchanged. The dual operation of the opening is the
closing. This operator reverses the interpretation of grayscale values and thus removes small dark objects and leaves
bright objects unchanged. Usually, these operators are applied using structural elements (SE) with different shapes
and sizes such as disks and lines SEs. As the size of SE
increases, objects are removed from the image. A detailed
adopted methodologyical chart is presented in (Figure: 2)

5. Results and Discussion

As Image morphological based segementation (Watershed Segmentation) is used in this research. Normally all
image segmentation techniques inputs are gray scale images instead of RGB. For this purpose, 8-bands WorldView3 image is transformed to gray scale using gray scale transformation. After analyzing the resultant image, it is identified that there are few bright pixels that may lead to oversegmentation. To avoid oversegmentation, low pass filter is applied to gray scale image until these speckles are removed,
resultant image is shown in (Figure: 1) Furthermore Image gradient was computed as shown in (Figure: 3). After

Figure 4. Initial Watershed Segmentation

niques called opening-by-reconstruction and closing-byreconstruction to clean up the image are used. Result of
image morphological opening are shown in (Figure: 5).

Figure 5. Image Morphological Opening

Similarly results from image morphological opening by

reconstruction and combine results of image morphological
opening closing are shown in (Figure: 6 and 7).
Beside applying combined image opening and closing,
image opening and closing by reconstruction was also applied. These operations created flat maxima inside each
object. Opening is erosion followed by dilation, while
opening-by-reconstruction is erosion followed by a morphological reconstruction. First compute opening followed
by opening-by-reconstruction. Following the opening with
a closing can remove the dark spots and stem marks.
Reconstruction-based opening and closing are more effec-

Figure 3. Image Gradient

computing image gradient, watershed segmentation was applied on gradient image which leads to oversegmentation as
shown in (Figure: 4). The reason of this oversegmentation
is that the local tonal variations play major role in segmentation.
Next step is to mark the foreground objects, which
must be connected blobs of pixels inside each of the foreground objects. For this purpose, morphological tech3

Figure 8. Image Morphological Opening Closing by Reconstruction

Figure 6. Image Morphological Opening by Reconstruction

then watershed segmentation is used.

Opening-closing by reconstruction removes features
smaller than the structuring element, without altering
the shape; reconstruct connected components from the
preserved features; Opening by reconstruction removes
unconnected light features; Closing by reconstruction
removes unconnected dark features. After applying image
morpological opening closing by reconstruction, regional
maxima was estimated as shown in (Figure: 9). Initial

Figure 7. Image Morphological Opening Closing

tive than standard opening and closing at removing small

blemishes without affecting the overall shapes of the objects. It was identified that image opening closing by reconstruction gives much better results as compared to simple
image opening closing as shown in (Figure: 8).
To obtain background markers threshold openingclosing by reconstruction. Separating touching objects in
an image is one of the more difficult image processing
operations. The watershed transform is often applied to
this problem. The watershed transform finds catchment
basins and watershed ridge lines in an image by treating
it as a surface where light pixels are high and dark pixels
are low. Watershed transform without image pre-processing
steps like morphological operations leads to over segmentation as shown in (Figure: 4), so in this algorithm first
morphological operations are applied to the input image

Figure 9. Regional Maxima of opening Closing by Reconstruction

estimated regional maxima was not accurate, to make

it more accurate it was modified and modified regional
maxima was estimated as shown in (Figure: 11).
Next step is to threshold opening-closing by reconstruction, then fill holes as shown in (Figure: 12).
4

Figure 10. Regional Maxima superimposed over original image

Figure 12. Threshold Opening Closing by Reconstruction

Figure 11. Modified Regional Maxima superimposed over original

image

Figure 13. Watershed Ridge Lines

is identified that watershed segmentation based tree crowns

are detecting and delineating tree crowns with acceptable
accuracy.

Using image morpological opening closing along with

modified regional maxima and threshold opening closing,
watershed reidge lines are estimated as shown in (Figure:
13). There ridge lines are the separating lines between different watersheds or in other words, these lines will act as
separators between different tree clusters.
Finally watershed segmentation is applied using watershed ridge lines and modified regional maxima. The resultant object markers well separate different trees from the
background as shown in (Figure:14). As markers color
is mixing with gray scale raw image, just to mark markers more clear, colors are assigned to different segments as
seen in (Figure: 15)
In order to access the quality of segmentation results, obtained results are cross checked with ground truth data. It

6. Conclusion
In this research, watershed based image segmentation
technique has been implemented successfully to detect and
delineate. It has been identify that without applying image morphological opening and closing by reconstruction,
watershed based image segmentation leads to over segmentation. While estimating regional maxima in order to apply
watershed segmentation, major variations in the segments
should be considers otherwise it may leads to misleading
results or under/over segmentation. Results obtained from
segmentations are separating trees from other land covers
5

[3]

[4]

[5]

[6]

[7]

Figure 14. Watershed Segmentation based Object Boundaries superimposed on original image

[8]

[9]

[10]

[11]

Figure 15. Markers Superimposed Transparently on Original Image

accurately however in order to detect and delineate individual tree crown level 02, level 03, level 04... segmentation
should be performed until individual tree crowns are identified.

References
[1] S. Basu, S. Ganguly, R. Nemani, S. Mukhopadhyay,
C. Milesi, P. Votava, A. Michaelis, G. Zhang, B. Cook,
S. Saatchi, et al. A semi-automated machine learning algorithm for tree cover delineation from 1-m naip imagery
using a high performance computing architecture. In AGU
Fall Meeting Abstracts, volume 1, page 3698, 2014.
[2] F. A. Gougeon and D. G. Leckie. The individual tree crown
approach applied to ikonos images of a coniferous planta-

tion area. Photogrammetric Engineering & Remote Sensing,

72(11):12871297, 2006.
M. K. Heenkenda, K. E. Joyce, and S. W. Maier. Mangrove
tree crown delineation from high-resolution imagery. Photogrammetric Engineering & Remote Sensing, 81(6):471
479, 2015.
L. Jing, B. Hu, T. Noland, and J. Li. An individual tree crown
delineation method based on multi-scale segmentation of imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 70:8898, 2012.
M. Kalapala. Estimation of tree count from satellite imagery
through mathematical morphology. International Journal,
4(1), 2014.
M. Kamal, S. Phinn, and K. Johansen. Object-based approach for multi-scale mangrove composition mapping using
multi-resolution image datasets. Remote Sensing, 7(4):4753
4783, 2015.
Y. Ke and L. J. Quackenbush. Comparison of individual tree
crown detection and delineation methods. In Proceedings of
2008 ASPRS annual conference, 2008.
A. N. Skurikhin, S. R. Garrity, N. G. McDowell, and D. M.
Cai. Automated tree crown detection and size estimation using multi-scale analysis of high-resolution satellite imagery.
Remote Sensing Letters, 4(5):465474, 2013.
L. Wang. A multi-scale approach for delineating individual tree crowns with very high resolution imagery. Photogrammetric Engineering & Remote Sensing, 76(4):371
378, 2010.
W. Yin, J. Yang, H. Yamamoto, and C. Li. Object-based larch
tree-crown delineation using high-resolution satellite imagery. International Journal of Remote Sensing, 36(3):822
844, 2015.
N. A. M. Zaki, Z. Abd Latif, M. Z. Zainal, and K. Zainuddin. Individual tree crown (itc) delineation using watershed
transformation algorithm for tropical lowland dipterocarp. In
Space Science and Communication (IconSpace), 2015 International Conference on, pages 237242. IEEE, 2015.