PolSAR Image Classification by Multi-objective optimization-based High-Order Discriminative Random Fields

To effectively improve PolSAR image classification accuracy, which is adversely affected by feature redundancy and the curse of dimensionality, this paper proposes a multi-objective optimization-based high-order discriminative random fields (MoODRF) model. The model integrates NSGA-II and ReliefF to achieve optimal feature selection for high-dimensional data,the optimal feature set is then input into using Support Vector Machines (SVMs) to construct the unary potential, while a binary potential is modeled using a generalized Ising model with an additional edge penalty function. The high-order potential energy is captured using high-order cliques generated by superpixels,enabling the model to simultaneously incorporate texture and polarimetric features, as well as spatial neighborhood information, and higher-order structural information.Finally, the classification result of the PolSAR image is obtained using the Iterated Conditional Modes(ICM) method.The efficacy of the suggested strategy is demonstrated by experiments using actual PolSAR images.