SW-Former: Refined Transformer Models for Multi-Temporal PolSAR

Multi-temporal PolSAR data contains dynamic change information of surface targets over time. The increase in multi-temporal features adds pressure to represent temporal-polarimetric relationships. Therefore, some researchers prefer using Transformer model to handle these complex relationships. However, the native Transformer typically treats all Polarimetric features at a single time step as a token, and part of its structure is not suitable for multi-temporal PolSAR data tasks. To address this, we introduce an improved Transformer model called the Sliding Window Temporal-Polarimetric Transformer (SW-Former). In this model, sequences with the same polarimetric features in all time steps are treated as a token. Sliding window embedding is used to divide each token into equal-length patches. This extends the data volume and captures the inter-temporal semantic information that is not available in the native structure. Additionally, the native self-attention is replaced with a Temporal-Polarimetric Attention, which focuses on effective information from both the temporal and polarimetric dimensions. To prove the effectiveness of the proposed method, two datasets are considered in the experiments. Compared with state-of-the-art (SOTA) methodologies, the classification performance in terms of overall accuracy (OA) can be improved by about 1.59% and 1.41% respectively, which proves the superiority of the proposed methodology.