5. Results and Discussion
Fig. 2 shows the characteristics of backscattering from different land cover types. Among land
cover types, bare land represents the lowest backscattering signals. Next, the second smallest
mean NRCS is observed for water. The bare land of the study area was composed of various
land covers such as paved roads, ground with low grass, open land and bare soil. During the
field data acquisition, there was a heavy rainfall, which continued for four days. Peculiar
backscattering occurred presumably due to the water accumulation on some flat land surfaces.
Fig. 2 Changes in the mean backscatter of each land cover type for
different polarizations
The standard deviation of mean backscattering of these (bare land and water) land cover types
ranges from 1.17 to 3.25, and the dynamics of the mean backscattered signal ranges from -2 dB
to -5 dB for all polarizations between water and bare land. Another peculiar backscattering was
observed from the urban area. In particular, the standard deviation of mean backscattering is the
largest among all land cover types (5.6). However, the result confirmed that the quality of
PI-SAR is fairly good, and therefore PI-SAR is useful for land cover classification.
Fig. 3 Composite image based on PI-SAR images of the study area.
HH, HV and VV polarization shown in RGB (Acquisition date: 14 Jul. 1999).
In the unsupervised classification, land cover was classified into seven classes. The resultant
classes were compared with the ground truth and other previously published data. Each cluster
was then defined as an appreciable land cover type. Figure 4 and Table 1 show the unsupervised
classification result. The unsupervised classification procedure produced an adequate
classification result for bare land, forest and bush land but poor results for urban area, small
structures, and wetland. Figure 4 shows that forest pixels were misclassified into the urban
category. Similarly, urban forest pixels were misclassified into the urban category. Presumably,
these errors are due to the complexity of land cover type of the urban area. For the accuracy
assessment of the unsupervised classification result, the overall accuracy is 61.7 and the overall
Kappa statistics, 0.55.
Fig. 4 Unsupervised classification of the study area.
Table 1 Signature separability for unsupervised classification
(Euclidean distance)
The maximum likelihood classification was performed for the composite image using the
evaluated signatures. Figure 5 and Table 2 show the land cover type distribution and the error
matrix derived from the supervised classification. A resultant thematic layer containing seven
land cover types in the study area was thus generated. However, water, small structures and
wetland agreed well with the actual conditions. As in the unsupervised classification, the
supervised training signatures were not representative enough to identify some classes such as
urban areas. A textural image was generated in order to overcome misclassification of the urban
area and improve the classification accuracy. An unsupervised ISODATA algorithm was then
used to classify the textural image. As a result, the urban area itself was correctly classified as
urban and others (two classes only), this classification was consistent with the actual urban area.
Fig. 5 Supervised classification of the study area.
Table 2 Error matrix derived from the supervised signatures.
