STARNet: selective token attention with rectangular context network for hyperspectral image classification

Hyperspectral images are of substantial research significance in remote sensing land cover classification due to their rich spectral-spatial information. However, existing methodologies face dual challenges: global attention redundancy and local feature confusion. We present selective token attention with the rectangular context network to achieve breakthrough improvements. The selective token attention module introduces a multi-granularity token screening mechanism that dynamically integrates strongly correlated features through four branches, constructing a sparse attention map to reduce computational redundancy. The central self-tuning module leverages axial pooling and strip convolution to establish a rectangular receptive field, dynamically calibrating the central pixel’s feature response to enhance the local feature extraction capability. Complemented by a dynamic context classification head for final classification, the STARNet framework achieves a precise and efficient land cover classification. Experimental results demonstrate that STARNet performs exceptionally well in three benchmark datasets.