MCESAN: A Modality Complementary Enhancement and Sparse Attention Network for Joint Hyperspectral and LiDAR Image Classification

Abstract

Hyperspectral images (HSIs), with their hundreds of contiguous narrow bands, capture detailed physicochemical properties of land covers; LiDAR provides high-precision 3-D spatial and geometric structure, and the two modalities are highly complementary. However, differences in imaging mechanisms lead to insufficient cross-modal feature consistency, and conventional self-attention mechanisms tend to introduce numerous redundant dependencies in long-range modeling, impairing fusion accuracy. To address this, we propose a modality complementary enhancement and sparse attention network (MCESAN) for joint classification. First, we construct a modality complementary enhancement and correction module, composed of an adaptive discrepancy perception unit and a global consistency correction unit, to achieve precise feature alignment and semantic calibration. Then, we design a selective key sparse attention (SKSA) mechanism that uses LiDAR spatial features as queries and HSI spectral features as keys and values to establish a directional interaction framework, suppressing redundant associations and focusing on critical cross-modal dependencies. Experiments on the Houston2013, Trento, and MUUFL datasets demonstrate overall accuracy (OA) superior to mainstream methods, validating its effectiveness and robustness in complex scenes.