Land Cover Classification for Polarimetric SAR Images Based on Global Membership Tensor

The increasing resolution of PolSAR images makes it challenging to achieve satisfactory land cover classification performance with a single feature. Existing research considers combining multiple features to integrate polarimetric and multi-modal information from images, thereby better representing the diversity of land cover. Nevertheless, in the process of extracting multimodal information, the challenges of dimensionality disaster and dynamic range discrepancies persist, and how to effectively eliminate redundant information while preserving higher-order correlations among features remains a significant research challenge. This paper proposes a PolSAR land cover classification framework based on a global membership tensor, which avoids the dimensionality disaster and dynamic range difference issues by concatenating the pre-classified membership matrices, and utilizes Tucker decomposition to extract higher-order correlation information embedded in the tensor structure of the global membership tensor for final land cover classification. Experiments are carried out on the multi-frequency Hainan dataset acquired by the Aerospace Information Research Institute Chinese Academy of Sciences (AIRCAS). The experimental results indicate that the proposed classification framework based on global membership tensors successfully integrates polarimetric and texture features for effective land cover classification.