The authors introduce unsupervised Wishart classification technique for fully polarimetric SAR data based on H-alpha
decomposition of POLSAR images, and applied this technique to AIRSAR data of Flevoland, Netherlands. By applying
the Cloude H-alpha decomposition to the original L-band image, we segment the image to 9 classes. While take this as
the initial input, Wishart classification is followed. The most valuable in this paper is the section of application analysis.
We found H-alpha classification has lower classification accuracy than Wishart iteration which use coherence matrix, but
why? By analyzing the classification results for each type of land cover, this paper concluded the reason is that
parameters of entry and alpha angle lose the original polarimetric information. While coherence matrix does not lose the
original polarimetric information, we suggest that directly use coherence matrixes could derive much higher
classification accuracy. There is also another found. Middle entropy scattering such as low vegetation often does not a
single target while high or low entropy scattering, such as the deep forest and water, the coverage relatively much denser,
often has single component; thus, the classification accuracy of high of low entropy land cover will be much higher than
middle entropy scattering.
The authors introduce unsupervised wishart classification technique for fully polarimetric SAR data using H/α
decomposition of POLSAR images. This paper we applied this technique to AIRSAR data of Flevoland, Netherlands.
The most valuable in this paper is our evaluation. From the following tree aspects we evaluate the algorithm mentioned
in this paper and the results it produced. (i) By calculating the Jeffries-Matusit Distance (J-M Distance) J<sub>mn</sub> between two
classes, which represents the separation between classes, the property of this classifier is measured. J-M Distance is a
measurement of average difference between Probability Distribution Function (PDF) of two classes. Usually J-M
Distance is between 0 and 2, and the bigger J-M Distance represents that two classes has a good separation. This paper
we have most J-M Distances 1.8-2.0, thus indicates good separation; (ii) According to the average entropy and alpha of
each final class, the classification results are analyzed; (iii) by comparing the classification results with the ground truth,
the classification algorithm is evaluated. The results have a good simulation of ground truth. Experiment in this paper,
according to the measurement criterion, analysis and evaluation, demonstrates that the region of Flevoland is well
classification and the method has the advantage of edge holding that in the case of non-smooth borders this advantage is
helpful. Also this paper gives a better repeat time.