New AI Reads Invisible Light to Spot Crops, Minerals and Pollution

Hyperspectral images carry rich spectral-spatial information, but pixel-level classification remains challenging because of spectral-spatial heterogeneity and complex spatial structures. Existing vision state space models (Mamba), according to the paper, typically build sequences from predefined spatial neighborhoods without explicitly accounting for semantic similarity or spatial non-stationarity. The proposed fix is Token Clustering and Semantic Sequence Mamba, or STMamba. At the macro level, a hierarchical encoder-decoder progressively selects semantic tokens with a Token Clustering Module and restores dense features using a parameter-free Cross-scale Neighborhood Attention Upsampler. At the micro level, the clustering module identifies representative cluster centers through density-aware clustering, estimates soft memberships based on feature similarity, then uses a quadtree-based dynamic selection strategy to keep sparse, spatially distributed tokens from each semantic cluster — forming coherent semantic-token sequences while reducing redundant pixel-wise representations. Parallel Spatial and Spectral Semantic-wise Sequencing Mamba modules capture complementary long-range spatial and spectral dependencies within homogeneous token sequences while suppressing irrelevant interactions across heterogeneous regions. On three large-scale benchmark datasets, results show STMamba outperforms state-of-the-art methods in both quantitative and qualitative terms.