Adversarial Network With Higher Order Potential Conditional Random Field for PolSAR Image Classification

Effective and efficient pixel-level classification of polarimetric synthetic aperture radar (PolSAR) images represents an important step toward their interpretation and knowledge discovery. After the extraction and representation of discriminative features, the related important issues are how to enforce consistency and coherency of labeling using contextual information, and how to make the process computationally efficient. In this article, we propose a method to approach these two issues. The first issue is dealt with from three different levels, namely, first, to combine features at coarse resolution with pixel-wise information for pixel-level classification, we adopt the idea of Unet; second, to reduce labeling inhomogeneity among similar pixels, we relate the PolSAR image with graph theory through the conditional random field (CRF), and develop third-order potentials for a fully connected CRF to account for both higher order effects and long-range interactions; and, third, to utilize adversarial network to improve learning of the label distribution. The efficiency issue is handled by, first, adopting fully convolutional network for the contracting arm of the Unet to speed up hierarchical feature extraction; and, second, for the fully connected third-order potential, by making connection with the pairwise potential, to reduce the computational complexity of the most expensive message passing procedure from quadratic to linear in the number of variables. The effectiveness of the proposed method has been demonstrated by its application to the pixel-level classification of ALOS-2, RADARSAT-2, and ESAR PolSAR images, where its performance has qualitatively and quantitatively shown superiority over several other state-of-the-art models.