| dc.description.abstract | This paper will present a method based on a fuzzy neural network that will use fully polarimetric information for SAR image classification. The approach makes use of the statistical properties of the polarimetric data while taking advantage of a fuzzy neural network that requires no a priori information about the data. A distance measure based on the complex Gaussian distribution was applied to the fuzzy clustering algorithm and then subsequently incorporated into the neural network. Instead of pre-selecting the polarization channels as has usually been done before, the inputs to the neural network are now all elements of the covariance matrix which serve as the target feature vector. It is thus expected that the neural network will be able to take full power of the fully polarimetric information for the purposes of image classification. With the generalization, adaptation, and other capabilities of the neural network, general information contained in the covariance matrix, such as the amplitude, phase difference, degree of polarization, etc. are well preserved and thus are fully explored. One of the essential features in this setup lies in that the chosen neural network must be able to handle such high dimensional and yet diverse input feature vectors, while maintaining a sufficiently fast learning speed in order drive itself as a practical tool. To demonstrate the advantages of the proposed method, we compare four different configurations, which are categorized by their uses of feature vectors, classifier, distance measures, and whether fuzzy c-means are applied or are applicable. The validity and effectiveness of the proposed scheme support the utilization of this polarimetric information. It is shown that with fully polarimetric data, the fuzzy neural network can substantially reduce the learning time and improve the classification accuracy as well. It must be noted that a Lee polarimetric filter, that reduces the speckle noise while preserving the polarimetric properties has proven to be useful in improving the classification accuracy. It is also demonstrated that the proposed approach gains adaptability and flexibility for high dimensional feature vectors, such as the complete polarimetric data. | en_US |