A SUPERVISED DYNAMIC LEARNING BACK-PROPAGATION (DLBP) NEURAL NETWORK APPROACH FOR AIRBORNE FULLY POLARIMETRIC SAR TARGET CLASSIFICATION

摘要

In this paper, we propose a supervised dynamic learning back-propagation (DLBP) classifier for target classification using airborne fully polarimetric SAR data. This approach is composed of a speckle noise filtering mechanism, fuzzy c-means approach, a non-linear scaling process of digital number for each dimension of feature space, and a dynamic learning back-propagation algorithm. The fuzzy approach is utilized to take mixed pixels or regions into account. A distance measure based on the complex Gaussian distribution has been applied to represent the distance between any two classes in the feature space. The proposed non-linear scaling operation could provide feature space data with both as higher signal-to-noise ratio as possible and with better separability between two classes. We use the NASA JPL fully polarimetric SAR (POLSAR) data in Taiwan for testing. The test results of the proposed DLBP are analyzed and compared with the minimum distance method. They show that separability measure might be better than SNR for defining the data scale for target classification. After changing the data scale, separability measures could raise 1.01～1.34 times than Lee filtered data. The accuracy of DLBP image classification is 89.37～94.40%.