GeoSMAViT: geographical-context spectral-invariant multi-scale adaptive vision-language transformer

Official URL: https://dx.doi.org/10.1109/WHISPERS69515.2025.11501561

Cross-domain hyperspectral image classification faces challenges from deficiencies in geographical context modeling, spectral instabilities, and suboptimal multimodal fusion. We present GeoSMAViT, a geography-aware vision-language transformer with: (1) climate-aware spectral processing using region-specific band weighting and atmospheric correction, (2) multi-scale regional pattern recognition for different geographic land-use types, (3) single-stage cross-modal fusion with multi-head attention between vision and geographic text features, (4) progressive domain adaptation using source features for invariant representation, and comprehensive anti-overfitting measures including dynamic regularization and adaptive text dropout. On Houston 2013-to-2018 and PaviaU-to-PaviaC, GeoSMAViT achieved 98.37% and 96.29% OA, respectively, surpassing state-of-the-art baselines. Ablations confirmed that geographic awareness provided the largest contribution (+38.33 %) to cross-domain generalization.