Interpretability of Transformer Structures Based on Attribution Inference for Polarimetric SAR Image Classification

Transformer-based deep networks have gained prominence in Polarimetric Synthetic Aperture Radar (PolSAR) image classification due to their strong capability for modeling long-range dependencies. However, existing Transformer-based methods fail to reveal the polarimetric scattering characteristics and their spatial interactions in classification decisions, which leads to limited interpretability and reduced reliability of the models. To address this challenge, this paper proposes an interpretability analysis framework based on attribution inference for Transformer-based PolSAR image classification. Specifically, an attribution inference module is first constructed by integrating structural dependency relationships with gradient-based saliency information, enabling explicit simulation of polarimetric scattering features and their interactions within the model. To enhance the interpretability of PolSAR classification models, we further conduct a full-chain analysis of the Transformer’s internal structure based on the attribution inference matrix, rather than relying on qualitative visualization from a single attention map. This integrated approach enables precise localization of critical attention heads and information flows, while simultaneously identifying discriminative regions that exert a decisive influence on the prediction results. To validate the proposed method, quantitative and qualitative experiments are performed on three PolSAR image classification datasets, which clearly demonstrate the effectiveness of the proposed interpretability framework.