The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B6b. Beijing 2008 
Figure 6. C-band HH image 
Figure 7. Forest Mask with threshold 0.65 
Figure 8. Forest Mask with threshold 0.7 
From the accuracy assessment, the C and L bands intensity 
classification in which the forest mask was used has the highest 
accuracy for ‘Forest’, which is 100% in PA, 92% in UA. The 
result is much better than using the whole coherency matrix of 
which the best PA accuracy is only 76.5%. H( 1 — A) is indeed 
a good indicator for forest. And the threshold we used is 
suitable for this study. 
The bigger the threshold is, the less the forest is recognized. 
Further experiment is needed to figure out whether it can also 
be used to classify the forest according to the forest volume, and 
what is the threshold. However it should be pointed out that 
roads in the forest is mixed with forest in both mask images. 
INTENSITY 
Combining with other polarimetric features: H(l-A) and 
“entropy” texture of ¿-band HV intensity image, the overall 
accuracy of the intensity classification is better than that of the 
single-look coherency matrix. The OA is 81%, and Kappa 
achieved 0.70. It also recognized the most land-cover types. Part 
of the reason is that we use three masks in the classification for 
the intensity and the data were filtered two times: multi-look 
and texture analysis. The accuracies for the crops are good. PA 
is around 80% for four crop types. ‘Crop5’ was best classified 
with the PA 99.0%. But ‘Crop4’ has a relative low accuracy, 
with the PA 52% only. 
The accuracy of the ‘entropy’ texture of HV polarization of L 
band intensity image achieved PA 71.70% for ‘Road’ and 
94.00% for ‘Crop6’. Since ‘Road’ was not recognized in 
coherency matrix, this result was relative good. Although the 
coherency matrix has a higher accuracy for ‘Crop6’, it should 
be noted that this class is combined with other land-cover types 
as we described before. 
COHERENCY MATRIX 
In this study, the land-cover classification of coherency matrix 
is not as effective as that of intensity. Both C-band and L-band 
have a lower OA accuracy than intensity. However, C-band data 
is better than L-band data in crop classification. 
The main reason for the low classification accuracy was the 
speckle level in the image. The speckle in the image decreased 
the classification accuracy. Higher accuracy was produced by 
using intensity because two filterings were performed, while 
in coherency matrix, only one filtering was carried out. More 
filtering will be tested in the further study, using, for example, 
MAP filter (H. Skriver, 2005). 
5. CONCLUSION 
This study evaluated the performance of different polarimetric 
features for land-cover classification in order to develop an 
effective classification procedure. Two polarimetric features: 
coherency matrix and intensity were investigated by 
classification of the whole image. Other two polarimetric 
indicators: H(\ — A) of L-band and “entropy” texture of 
¿-band HV intensity image were evaluated as a classifier for one 
or two specific land-cover types. 
The results indicate that the supervised classification of the 
intensity of both C- and L- bands has the potential for 
land-cover mapping in this study area. The results also 
revealed that both classification results of coherency matrix and 
the intensity can be improved. It is very difficult to find one 
polarimetric feature that will be effective for all land-cover 
types. A hierarchical classification approach is highly desirable. 
The second classification method in this study is a good attempt 
and the result is also promising. More polarimetric features need