Hypothesis-Conditioned Spectral Reallocation and Confidence-Coupled Spatial Routing for Lightweight Hyperspectral Image Classification

Hyperspectral image (HSI) classification imposes stringent requirements on computational efficiency, especially for onboard edge platforms with limited computing resources such as small satellites and unmanned aerial vehicles (UAVs). Existing deep learning methods commonly process all spectral and spatial information with nearly uniform computation, although class-discriminative wavelengths and informative regions vary from sample to sample. This uniform processing introduces redundant information extraction and makes lightweight deployment difficult. To address the aforementioned dual redundancy of class-irrelevant spectral bands and overlapping spatial regions, we propose ReRouteFormer, a sample-adaptive computational allocation framework for efficient HSI classification. The proposed network contains three task-oriented modules: a Hypothesis-Conditioned Spectral Reallocation (HCSR) module that suppresses weakly informative spectral channels before expensive feature extraction, a Confidence-Coupled Spatial Evidence Routing (CSER) module that performs content-aware token selection with a fixed retention budget and supports confidence-aware early exit, and a Sample-Adaptive Spectral–Spatial Meta-Fusion (SS-MetaFusion) module that fuses spatial, spectral, and mixed features with lightweight adaptive weights. These modules are jointly optimized and replace parameter-heavy operations in the baseline rather than adding extra complexity. Extensive experiments on four benchmark HSI classification datasets demonstrate that ReRouteFormer achieves highly competitive classification accuracy while maintaining substantially lower parameter counts and the lowest FLOPs among all compared methods, yielding a superior accuracy–efficiency trade-off.