Spectral-Structural Collaborative Learning for Fine-Grained Hyperspectral Mineral Classification

Fine-grained hyperspectral mineral classification remains challenging due to spectral homogeneity among minerals with different morphologies, severe spectral mixing from intergrowth, and high dimensionality. Existing methods rely on spectral separability assumptions, which become insufficient when spectral differences are subtle and spatial-structural ambiguity is high. To address these limitations, we propose S³AM-ECA-3DCNN, a spectral-structural collaborative feature learning framework. It uses a 3DCNN backbone to jointly model spectral-spatial features with progressive spectral downsampling. The spectral-similarity-based spatial attention module (S³AM) performs spatial purification by suppressing interference from spectrally mixed neighboring regions, and the efficient channel attention (ECA) module adaptively recalibrates discriminative spectral bands to enhance fine-grained representation. This establishes a spatial-first, channel-second collaborative optimization paradigm. To improve generalization under limited training samples, adaptive global pooling and a lightweight classification head are employed to reduce model complexity and mitigate overfitting. Experiments on a 146-class hyperspectral mineral dataset (covering silicates, carbonates, and sulfates) show that the framework achieves 93.424% overall accuracy, 91.099% average accuracy, and a Kappa coefficient (×100) of 93.368, outperforming mainstream methods. It significantly reduces misclassification among spectrally similar but morphologically distinct minerals, demonstrating strong robustness and discriminative capability for large-scale fine-grained tasks.