The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1091
and the multispectral imagery (in 2004). These were the only
available images for this research project; therefore we had to
compromise on this issue. The study area is a slow growing
forest and within this time frame it had not seen any abrupt
changes such as tree damage by bush fire or logging. However,
some temporal effects were found, due mainly to the natural
growth of trees, which is always a challenge to address in high
resolution data fusion.
3.2 Watershed segmentation
The lidar derived nDSM represents the tree canopies of the
forest. Single and disjoint tree canopies can easily be delineated
in this process. However, a segmentation procedure is needed to
isolate trees which are grouped. This study uses marker-
controlled watershed segmentation for tree canopy isolation.
Watershed segmentation, first proposed by Beucher and
Lantuejoul (1979), is a well known image segmentation method
that incorporates region growing and edge detection techniques
(Soille,2003). To avoid the over segmentation problem, Meyer
and Beucher (1990) introduced marker-controlled watershed
segmentation. The idea is to perform watershed segmentation
around user-specific markers rather than the local maxima in
the input image.
In the watershed segmentation of the nDSM data, the tree
crown model was treated as a 3D surface, with lateral
dimensions representing the image plane, and the vertical
dimension representing the grey values (Figure 2a). Internal
markers were used to locate the local minima, which were
associated with high grey values (i.e. selected tree crowns) and
external markers were pointed to the local maxima, which were
associated with the background. Through flooding from the
local minima, the watershed segmentation was performed:
neighbouring watersheds were merged unless boundaries were
built to isolate individual tree features (Figure 2c). The process
of merging regions and building boundaries continued until no
more region growing could take place.
(a) (b) (c)
Figure 2. An illustration of watershed segmentation, (a) A
canopy model derived from nDSM, (b) 3D view of
the canopies, and (c) Segmentation results with
dams (in red) built at the divide line.
3.3 Data processing
After segmentation, the resulting crown polygons were overlaid
on the lidar and multispectral imagery to extract the spectral
signatures and texture information of the tree crowns for tree
species discrimination. Firstly, the extracted signatures from
four of the original multispectral bands were processed with a
directional convolution filter using a 3x3 window. This filtering
procedure allowed the suppression of shadow effects within the
sunlit area of the tree crown. The weighting factors and the
dimensionality of the filter are primarily dependent on the solar
direction at the time of over flight, the tree size, and the
illumination conditions within the tree crown.
Secondly, image enhancement by principal components
transformation was applied to the filtered four-band data set.
The objective being the replacement of the highly correlated
original bands with those of reduced correlation. The
transformation resulted in four new components: the brightness,
the redness, the greenness, and the blue-yellowness, for each of
the tree types.
In addition, two more lidar derived layers were included in the
fusion procure. A ninth layer was generated by a texture
analysis of the first return lidar intensity and the tenth layer
from lidar derived nDSM layer.
3.4 Supervised classification
A supervised classification of the 10 layer datasets into three
different categories as listed in Table 1 was carried out. Much
of the success of the maximum likelihood classifiers depends on
the choice of training areas. Extensive field survey
measurements were conducted to collect the training data. The
processed datasets were also used to redefine the training area
in order to maximize the classification results. These datasets
allowed a much better class-specific delineation of the training
areas involving a reduced sample size for the different tree
categories. However, the selected training areas still met the
minimal requirement of 5 x k (no. of layers) pixels from a
statistical point of view (Kalayeh and Landgrebe,1983).
Class
Tree type
Description
1
Black Box
Rough bark
2
Grey Box
Fine, pale, fibrous bark
3
River Red Gums
Smooth bark
Table 1. Selected tree classes and associated degree of disease
3.4.1 Filling the tree polygon: In high spatial resolution
data fusion, the class variability within tree crown is caused
mainly by the variability in crown structure (shadow effects),
crown density (background material) and different tree
components (bark, needles/leaves) (see figure 3a). In addition,
the class variability is also affected by the categorisation of the
tree types with respect to the leave and bark patterns (Table 1).
( a ) «Black Box (b)
« River Red Gum
■ Grey Box
Figure 3. Refining the tree classification; (a) the classified tree
crowns; (b) Filling the tree crown area with majority
species.
In order to increase the significance of the classification results,
the entire tree polygon was filled with the most frequent class
(Figure 3a). In this way, only one class occupied the entire