Tree counting analysis using eCognition

Juan C. Surez
Forest Research
Roslin EH25 9SY
[email protected]

1. Data collection
High resolution aerial photography was taken in Aberfoyle in September 2002.
We selected 12 plots representing mature Sitka spruce stands between 30 and 40
years in age that were surveyed at the beginning of 2003. The aim of this field
data collection was the validation of the results of the image processing analysis
in the tree counting and LiDAR analysis projects. The proposed method
contemplated a survey of both stand and individual tree characteristics covering
thinned and unthinned mature Sitka spruce stands (see Figure 1).

Twelve plots of 50 by 50 metres were chosen in each forest district, where stand
parameters such as top height, tree diameters and dominance were surveyed.
Each individual tree within the plots was located by means of differential GPS
and laser relascope. Additionally, three small plots of 10 by 10 metres were
located within each 50 by 50 metres plot in order to measure all the individual
tree parameters such as tree height, canopy dimensions in a N-S and E-W axis
and height to the first live whorl.

All the field surveys were undertaken in parallel to the aerial photography and
LiDAR surveys before the start of the growing season to avoid a temporal
decorrelation in the predictions of the analysis.
Figure 1. Configuration of the sample plots in the field. Plots were oriented to the magnetic
north. Each corner was labelled with a c and a number starting from the Southwest corner in
sequence and finishing in the Southeast one. Each quadrant was labelled in the same sequence
as the corresponding corners. Small plots of 10 x 10 metres were labelled with an a and the
number of the quadrant they lay in.

In parallel, the digital aerial photographs were orthorectified and geo-referenced

between MGIU and Forest Research using the LiDAR data layers as the
reference points. This technique ensured a good match between the two data
layers and a superior geo-correction than the one normally obtained using OS
1:10,000 maps.

The working packages involved in the field data capture are outlined in the table
below.
Task ID Task Development Milestones Delivery
time
(man/days)
1.1 Training in Criterion 2 2nd week in Good understanding of
Laser and DGPS at May 2003 DGPS and Criterion Laser
Positioning Resource and their use within the
in Aberdeen. Husky hand held computer.
The system will allow an
easy integration of field
collected data to ArcView.
1.2 Location of reference 4 1st week in Location of the position of
points using a GPS May 2003 for every starting point for the
Aberfoyle survey of each plot in
F.D. Aberfoyle F.D. and Kielder
3rd Week in F.D. Estimated horizontal
June 2003 for accuracy is 1 m. Vertical
Kielder F.D. accuracy 2.5 m.

1.3 Location of four 4 Aberfoyle end Accurate location of the

corners in each plot. of May 2003. boundaries of each sample
Kielder end of plot and their overlay with
June. Lidar and A.P. data.
1.4 Video Recording of 2 End of June Video recording with digital
the two surveyed areas 2003 video camera will allow us
to keep a record of each plot.
A virtual revisit revisit to
each plot will help us to
understand the results of the
subsequent analysis.
1.5 Location of individual 24 End of June This is the most fundamental
trees in each Plot 2003 piece of information to relate
field data with the
subsequent analysis in
eCognition. Expected
horizontal accuracy is 1.5
m.
Work undertaken by
contractors using a laser
relascope and a total station.
1.6 Post-processing Laser 4 1st week in Increase the horizontal
survey July2003 accuracy of the points and
their efficient location in
ArcView.
1.7 Post-processing of 2 End of Increase the horizontal
GPS data June2003 accuracy of the points
located with DGPS using OS
reference stations.
TOTAL 42
2. eCognition analysis
The aims of the Tree Counting project were:
1. Evaluation of the eCognition software for counting trees in mature Sitka
spruce plantations
2. The identification of sensible grounds for the implementation of a fully
automated method for tree counting using the software and the three
visible channels (RGB) from commercial digital aerial photography.
3. The exploration of the main issues involved in the development of a
fully automated method for tree counting such as image characteristics,
image acquisition, image segmentation methods and image
classification. The aim was the development of a rule-based system
transferable across a wide range of situations knowing the limitations of
the method and data inputs.

The different working packages are outlined in the table below. Milestones
have been altered mainly due to complications in the data collection process.

Task Task Development Milestones Delivery

ID time
(man/days)
2.1 Select plots in 2 1st week in July Selection of the A.P,s. that cover
ERDAS Imagine each surveyed plot
2.2 Orthorectification 5 2nd week in July Provide the necessary
of Images information to MGUI to geo-
reference and ortho-rectify the
A.P,s. Refine the ortho-
rectification method to reduce
horizontal errors to less than 1 m.
2.3 Develop a rule- 10 End of July Develop a method to quantify the
based system for homogeneity of the spatial
image objects we want to detect (tree
segmentation tops) and relate them to the scale
parameter in eCognition.
2.4 Develop a rule- 10 2nd week in Analyse the topological
based system for August relationships between objects,
image and the individual object
classification reflectance to develop a fuzzy
logic classification system
2.5 Relate estimation 5 End of August Develop a method to relate
of individual tree location of individual trees in
identification to eCognition and the field surveys.
location in X and
Y
2.6 Estimate scale 5 End of August Quantify the effect of image
effects resolution in the segmentation
and classification processes.
2.7 Estimate 10 2st week in Quantify the effect of temporal
temporal September decorrelation in time series of
decorrelation images by testing the consistency
effects of the segmentation and
classification processes. In
particular, the method will try to
estimate the thresholds for
consistent predictions with varied
levels of noise.
TOTAL 42
3. Outline of the main conclusions

3.1 Main conclusions.

eCognition is a very flexible and easy to operate package that suits the purpose
of tree counting. The initial validation of the software has produced promising
results in most of the validation plots. However, the process still relies quite
heavily on the interactions of the operator with the program. This may be one
of the most critical bottlenecks towards a fully operational use of eCognition for
tree counting, unless the whole process can be automated and seamless
integrated to our corporate GIS system.

3.2 The segmentation method previous to the classification process identifies

geographical features using scale and homogeneity parameters. Scale relates to
the minimum size required to identify a particular object, which depends on the
resolution of the images. So, higher image resolutions will require larger scale
parameters to identify a particular object.

Homogeneity is described by a mutually exclusive interaction between colour

and shape. Colour refers to the spectral response of the objects; whereas shape
conveys information about the semantic consistency of the objects. Shape is
divided into two equally exclusive properties: smoothness and compactness that
respectively define the boundaries of the objects and their transition to others.
See Figure 2.

Figure 2. Segmentation method in eCognition.

The segmentation process selects groups of pixels with similar characteristics

into object primitives. These objects are the target of the subsequent
classification process, not the individual pixels in the images. Therefore it is
essential the optimisation of the segmentation process for a sensible
classification of the images. As it is at the moment, the parameters are vaguely
defined by the operator as a trial error process.

In order to make this process more effective, we have tried two geostatistical
methods, which basically quantify the degree of spatial autocorrelation (i.e. its
similarity) of the groups of pixels in the images. These groups are based on the
transition from areas of high reflectance (normally tree tops) to low values
(lower parts of the canopy). These methods quantify this transition as an
attenuation distance in metres that can be related to their size in the image. This
information is strongly correlated to the scale parameter in eCognition.

The technical details of each method will be fully described in the report.
However, it is worth noting here that the first method, the semivariogram
analysis, seems more suitable to model the size of the individual canopies in
unthinned stands or that were thinned long time ago (Figure 3). This technique
was already evidenced in the analysis of a six plots in Kielder forest (see
AGILE paper, 2003).

y = 1.7946x semivariogram analysis

R2 = 0.8195
10
9
8
7
lag distance

6
5
4
3
2
1
0
0 1 2 3 4 5
s pacing

Figure 3. The use of semivariogram analysis to quantify the transition between

individual tree canopies. The figure relates the relationship between actual
spacing and the transition distance in metres.

The Fourier Correlogram seems to model better the transition between canopies
in recently thinned stands. This technique seems to cope more efficiently with
the abrupt transition from one canopy top to the other when areas of the ground
become more visible (see Figure 4). Canopy structures seem to be smaller than
in the first case. However, the important factor here is the prospect of clearly
visible underlying ground in the transition between canopies. The reflective
characteristics of the ground are markedly different from the information
obtained from the tree canopies.
 Fourier correlograms y = 0.3871x
R2 = 0.6024
2.2

1.8

1.6

1.4

1.2

1
3 3.5 4 4.5 5 5.5

Figure 4. Use of Fourier Correlograms.

Another crucial factor that may confuse the choice of one model or the other is
the fact that some plots are viewed at an angle, which creates anomalously large
canopies (as in Figure 5). This situation produces confusing criteria in the
application of one method or the other. However, large slant view angles can be
described with the Fourier correlograms when the ratio between the longest and
the shortest axis in a central area (defined at a distance of 1/e from the centre)
reaches a critical value. When the ratio approaches a value of 1 means that the
area of interest has been capture in the image at an angle close to nadir (Figure
6).

Figure 5. Aberfoyle Plot 11 and Correlogram. Notice that the trees are observed at a
slant angle that is also reflected on the shape of the Correlogram in the middle of the
image that is depicted as an ellipse. The 1/e distance is shown in red.
 Figure 6. Aberfoyle Plot 2 and Correlogram. The trees are observed at nadir.
Therefore, the Correlogram does not show any anisotropy effects and its shape
approaches a circle.

At the present, we do not have enough evidence to provide a valid definition of

this critical value. Nevertheless, there are important implications for a
continuing research into this area, as this ratio can effectively determine the
exclusion of portions of the image situated on its edges.

3.3 The weights assigned to colour and shape in the segmentation process are
determined using a trial error process. This situation is far from ideal and we
still do not understand what determines the right balance between these two
parameters. At the present, the evidence seems to indicate that a weight of 0.8
to colour seems to offer the best solutions in the majority of the plots.
However, larger weights to the shape parameter appear to work better when the
images appear highly texture. This is the normal situation in low resolution
images or with a poor quality (i.e. affected by noise).

3.4 The weights assigned to all the visual bands have to be empirically
determined. At the moment, there is no evidence that supports the importance
of any particular band above the others. In theory, a largest weight should be
assigned to bands presenting wider reflectance values. This is the normally the
case of the green band. As the reflectance in this band is directly correlated to
chlorophyll content, it appears to provide a better discrimination between
vegetation types (Figure7). The histogram that shows the distribution of values
in this band appears to be slightly flatter in all the plots when compared to the
red and blue bands.

However, this information is not relevant since it can be easily biased by

shadows, sun angle and above all, by colour enhancement processes. The
literature review shows this problem and suggests the use of near infrared
reflectance. This is not a clear option available by most of the data providers
despite this is not an impossible option and it is worth to consider in the future.
Meanwhile, the use of descriptive statistics can generate qualitative information
about the central tendency and variability of the data (e.g. mean and standard
distribution), that can be used to automate the importance of each band in the
future. This analysis is very simple to perform in an automated pre-processing
procedure. Alternatively, a principal component analysis (PCA) can also
weight the relative importance of each band when compared to each other.
However, these two processes will have to be performed with selected portions
of the images that represent the best concentrations of trees in them, not the
whole images.

Figure 7. Histograms showing an example of the reflectance values in the three

visual bands. The red band is display above, the green in the middle and the blue at
the bottom.

In conclusion, a weight of 1 has been assigned to all the bands in all the
images in the absence of relevant information that indicates otherwise.

3.5 Image classification in eCognition.

There are two methods for image classification available in eCognition: the
Nearest Neighbour classifier and the use of membership functions. At the
moment, both methods seem to require considerable input from the user, which
constrains the possibility of a fully automation of the process.

The easiest and more effective method appears to be the Nearest Neighbour. It
requires fewer interactions from the user and the parameters are more
transparent. Currently, this method has produced the most encouraging results
in the majority of the plots.

The use of membership functions requires a methodical analysis of each one of

a long list of properties that include reflectance characteristics, topological
relationships of the objects and logical operators. The program offers the
possibility of selecting the most relevant properties in order to construct a rule-
based system that quantifies the degree of membership of the objects, obtained
in the segmentation process, to a class defined as tree top. This method, despite
being more complete and powerful than the Nearest Neighbour, needs a deep
understanding of both the image characteristics and the nature of the objects
from the user. Besides, this process limits the transferability of the rules
between images as they seem to be scene dependant.

3.5.1 Nearest-Neighbour classifier (NN)

The Nearest-Neighbour (NN) classification system is a quick method that
requires the selection of sample objects within a defined feature space. This
method is quite dependant upon a careful definition of classes and the selection
of the most representative samples by the operator. Therefore, the method is
very interactive and needs a lot of trial error attempts, which leaves very little
room for automation.

In addition, results may be seriously affected by the size of the objects (defined
by the scale parameter) and the location of the training samples. The use of
thematic layers (i.e. sub-compartment boundaries) improves substantially this
classification process because both the selection of samples and the
classification become restricted to the boundaries of the objects declared in the
thematic layers (i.e. stand borders).

The process of analysis has required a combined use of eCogniton and

ArcView. This process is outlined as:
Segmentation of the original images defines scale by the number of trees
in each plot. The rest of the parameters are equal in all the 12 areas:
Colour 0.8, Shape 0.2, Smoothness and Compactness 0.5
Polygons are exported to ArcView as a shape file.
Selection of polygons classified as tree tops in ArcView.
Location of the centroid within each polygon. Normally, this centroid
will correspond with the location of the tree top in the field. The results
evidence this statement.

The table below provides a summary of the best results achieved after an
exhaustive process of trial-error combinations. These results evidence a strong
correlation with the scale parameter that shows consistent in all the plots.

Plot ID No of trees Colour Scale Estimated Percentage

weight parameter no of trees of success
1 142 0.8 8 126 89
2 75 0.8 10 65 87
3 235 0.8 5 162 69*
4 109 0.8 10 98 90
5 270 0.8 5 223 83
6 123 0.8 9 121 98
7 96 0.8 10 90 94
8 103 0.8 10 111 108**
9 134 0.8 9 131 98
10 133 0.8 8 161 1.22***
11 167 0.8 8 152 91
12 126 0.8 9 123 98
 The next images show, as an orange circle, the estimation of the location of the
tree top from eCognition compared with the field observations. Dominance is
represented by a triangle of different colours, being red for dominant, green for
co-dominant and yellow for sub-dominant.

Figure 8
Plot 1
Scale 8
Colour 0.8
Shape 0.2

Figure 9
Plot 2
Scale 10
Colour 0.8
Shape 0.2
Figure 10
Plot 3
Scale 5
Colour 0.8
Shape 0.2

Figure 11
Plot 4
Scale 10
Colour 0.8
Shape 0.2
Figure 12
Plot 5
Scale 4
Colour 0.8
Shape 0.2

Figure 13
Plot 6
Scale 9
Colour 0.8
Shape 0.2
Figure 14
Plot 7
Scale 10
Colour 0.8
Shape 0.2

Figure 15
Plot 8
Scale 10
Colour 0.8
Shape 0.2
Figure 16
Plot 9
Scale 9
Colour 0.8
Shape 0.2

Figure 17
Plot 10
Scale 8
Colour 0.8
Shape 0.2
Figure 18
Plot 11
Scale 8
Colour 0.8
Shape 0.2

Figure 19
Plot 12
Scale 9
Colour 0.8
Shape 0.2
Dominance is an important parameter that will define the visibility of a tree top
from the air. It is assumed that tree tops will show higher reflectance in the
three visible bands. Likewise, reflectance will decrease towards the bottom of
the canopy. Dominance is assessed in the field by comparing the height of a
particular tree with the surrounding ones. Dominant trees present a tree top
clearly above the others. Co-dominants have similar heights than most of the
neighbours. Sub-dominants have a canopy below the majority of the adjoining
trees.

The results indicate that dominance is a decisive factor that will determine the
probability of detecting a tree top. Normally, dominant trees will present
clearer canopy structures with tree tops clearly visible from above. On the
contrary, sub-dominants are not so evident and the apices present reflectance
characteristics that can be easily confused by the lower parts of the canopy in
the dominant and co-dominant types.

However, as the largest tree tops present largest areas of high reflectance, the
process of analysis may identify more than one apex. This is illustrated in
Figure 20.

Figure 20. Common problem with trees presenting large canopies. The
canopies present large surfaces with several areas with high reflectance.
Therefore, the process of classification often misinterpret this information
and shows several tree tops.

Another common problem is the misclassification of areas of high reflectance as

tree tops. This is a common problem that normally happens with saturated
images. When this is the case, the images have to be pre-processed to mask
these areas. Alternatively, a new class can be defined in eCognition in order to
redirect the assignment towards them and not to the tree top class (Figure 21).
 Figure 21. Common problem areas of high reflectance that sometimes
may be confused with tree canopies.

3.5.2 Membership functions

As already stated, the membership functions are a highly complex process that
requires a skilful interpretation of the characteristics of the image and a no less
careful definition of the properties of the classes. This method is equally
dependant on a good segmentation prior to the classification process. Even if a
low number of characteristics are selected to define a tree top class, like mean
value in each band or brightness, the transferability of these rules across the
different plots do not seem to be very evident at the moment. At the present, we
have obtained some good results with different combinations of class properties
in each plot. However, the results are not necessarily the same in other plots
even if they share similar characteristics such actual spacing or planting age.

We continue working a quantitative method that may allow us to estimate

membership functions in each image for a selected number of properties.
Especially, those object characteristics that provide information about the
texture, the shape and the mean reflective values of the images.