SAR Image Time Series for Land Cover Mapping via Sparse Local–Global Temporal Transformer Network

Due to the effective embedding of temporal information, time-series synthetic aperture radar (SAR) images can acquire more abundant land cover information, thereby effectively enhancing the accuracy of land cover mapping. However, the significant introduction of temporal features greatly increases the redundancy of deep features. Simultaneously, capturing the global–local features of time-series SAR images effectively while considering temporal features has a significant influence on time-series SAR land cover mapping. Motivated by these, a sparse 3-D Transformer network based on joint learning of temporal–spatial features, named sparse local–global temporal transformer network (SL-GT2Net), is proposed for SAR land cover mapping, which has the ability to learn sparse global–local temporal–spatial features of SAR land covers. To emphasize crucial spatiotemporal features and suppress noise features, we propose a spatial–temporal channel refinement block to mitigate feature redundancy. In addition, we develop novel and effective sparse multiscale local window multihead self-attention and sparse multiscale global window multihead self-attention blocks. Both incorporate parallel multiscale window partitioning and top-$K$ sparsification strategies to robustly model discriminative spatiotemporal features of land covers with vast scale variations, while maintaining linear complexity with respect to the number of tokens. The experimental results on three challenge time-series SAR datasets demonstrate that our SL-GT2Net exhibits outstanding competitiveness compared to other related networks.