Land cover classification for polarimetric SAR image using convolutional neural network and superpixel

Abstract

The classification algorithms of polarimetric synthetic aperture radar (PolSAR) images are generally composed of the feature extractors that transform the raw data into discriminative representations, followed by trainable classifiers. Traditional approaches always suffer from the handdesigned features and misclassification of boundary pixels. Following the great success of convolutional neural network (CNN), a novel data-driven classification framework based on the fusion of CNN and superpixel algorithm is presented in this paper. First, the region-based complex-valued network utilizes both the intensity and phase information to predict the label of each pixel and constructs the label map based on spatial relations. Second, superpixel generating algorithm is adopted to produce the superpixel representation of the Pauli decomposition image, and the contour information which reflects the boundary of each category is preserved. Finally, the original label map and contour information are fused to make the decision of each pixel, outputting the final label map. Experimental results on public datasets illustrate that the proposed method can automatically learn the intrinsic features from the PolSAR image for classification purpose. Besides, the fusion of the superpixel features can effectively correct the misclassification of the boundary and singular pixels, thus achieving superior performance.