The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1092
polygon despite the classification of only a fraction of a tree CrownSize = 12 080-25 ^2e^°' m6{TreeHeighl)
crown (Figure 3b).
3.5 Accuracy assessment
To evaluate accuracy, ground truth tree types and crown maps
for six plots (Appendix A), each with an area of 50m by 50m,
were acquired in the field. Existing aerial photography was used
to design the sample plots and tree information was collected by
the field survey. There are a total of 61 trees in the six plots. An
error matrix was generated from the field sample data
corresponding to the fused results.
4. RESULTS AND DISCUSSION
Figure 4 illustrates the close correspondence {r 2 = 0.87, 95%
Confidence level, standard error = 0.67m) between tree mean
heights derived from both field measurements and lidar data
was observed. The comparison suggested, however, that the
mean height was more reliably estimated for trees with large
and relatively flat crown areas than those were small and
pointed crowns.
Figure 4. Relationship between individual tree mean heights as
estimated in the field and from lidar.
Results from previous studies have shown that isolating
deciduous tree species in lidar data is difficult due to their
complex structure (Chen et al.,2006). However, the use of the
marker-controlled watershed segmentation algorithm with the
lidar data achieved a satisfactory result for eucalypt trees. The
success of the tree crown extraction algorithm in old growth
areas was higher than in more juvenile areas where the crowns
are more scattered. It was also observed that large crowns were
better delineated than small ones.
To study the correlation between tree height and crown size,
tree height and crown size were measured from the crown
segments. Crown size is the average crown diameter and was
derived from the shape file generated area and algorithm for the
relationship of the crown area and radius. From the tabular
dataset, the trees were randomly sampled over the whole study
area and the sample size was 100 trees. It was found that crown
size has larger variability when a tree height is higher, which
will contradict the assumption of homoscadasticity if a linear
model is fitted. To avoid this issue, a parameterised non-linear
model was fitted:
Using Equation (1) a fitted line was generated through the
scatter plot as illustrated in Figure 5. In the regression analysis,
the relationship between crown size and tree height followed a
non-linear curve and mean squared error is 0.52, implied a low
level correlation.
Figure 5. The relationship between crown size and tree height.
The application of the maximum likelihood classification
technique involving the original four-band data set led to low
classification accuracy. The main reason was the confusion
within classes due to the noise effects such as shadows,
background vegetation and lack of information. An
improvement of the classification was achieved with the
integration of lidar derived height and intensity data.
Additionally, eight new layers generated from multispectral
data substantially reduced interclass confusion compared to the
original four-band data. The class separability was also
improved for particular tree species by increasing the gap
between the class means and reducing the class variability. The
use of the tree polygons derived from watershed segmentation
markedly improved the classification results through the
assignment of the most frequent pixel to the particular tree
polygon as shown in Figure 3b. In this way, only one class
occupied the entire polygon despite the classification of only a
fraction of a tree crown.
The results of the accuracy assessment are summarised in Table
2. The accuracy was assessed by comparing the classified tree
crown with the true tree information derived from field survey.
An average classification accuracy of 86 percent was achieved
and this procedure outperformed on average the original four-
band maximum likelihood classification by 23 percent. The
separation between the classes Black Box and Grey Box was
improved with the fused 10 layer datasets. This is mainly due to
the incorporation of the lidar data and the four principal
components of the mutispectral imagery, into the classification.
Classification types
Classification accuracy (%)
Only multi spectral (4 layer)
63
Fused multi spectral and lidar
data (10 layer)
86
Table 2 Comparative accuracy assessments.