Cascaded Dual-directional Cross-attention Transformer Network for hyperspectral imaging classification

Hyperspectral images possess rich spectral information and spatial characteristics, and traditional classification methods often rely excessively on high-reflectance bands while ignoring crucial spectral information in low-reflectance bands. Under small-sample conditions, this information is often not fully utilized. To solve this problem, this paper proposes a Cascaded Dual-directional Cross-attention Transformer Network (CDCATNet). First, it utilizes 3D convolution to maintain spatial-spectral continuity and designs a cascaded dual cross-attention module to dynamically strengthen the interaction between spatial and spectral features; secondly, it introduces a multi-scale fusion module to capture land cover information at different scales; finally, through tokenized global aggregation and Transformer cross-interaction, it breaks the limitations of local patches to achieve global context modeling. Experiments on the xuzhou, Houston2013, Salinas, and WHU_Hi_HanChuan datasets demonstrate that CDCATNet outperforms existing mainstream methods across multiple evaluation metrics.