Spectral Tube Masking for Robust Hyperspectral Image Classification via Hybrid 3D-2D Convolutional Neural Networks

Hyperspectral image (HSI) classification with deep learning is constrained by scarce labeled samples, making generalization a persistent challenge. While data augmentation is well-established for natural images, its adaptation to the three-dimensional spectral–spatial structure of hyperspectral cubes remains largely unexplored. We introduce Spectral Tube Masking (STM-lite), a physically motivated augmentation strategy inspired by the tube masking paradigm of VideoMAE. Rather than applying independent spatial or spectral corruption, STM-lite masks entire spatial patches consistently across all spectral bands simultaneously, forming coherent 3D spectral tubes that preserve the physical coherence of hyperspectral measurements while forcing the network to learn more robust spectral–spatial representations. STM-lite is integrated as a lightweight, training-only module into an optimized hybrid 3D-2D CNN combining 3D convolutions for joint spectral–spatial feature extraction with 2D convolutions for spatial refinement, evaluated on the Indian Pines benchmark using a single stratified 20%/80% train/test split, with 4-fold stratified cross-validation performed only within the 20% training partition for model selection while the held-out test set remained untouched throughout. Across five masking ratios, STM-lite consistently improves over the unaugmented baseline (OA 99.06%, AA 98.09%, Kappa 98.93%), peaking at 50% masking (OA 99.32%, AA 99.40%, Kappa 99.22%) and recovering at 90% (OA 99.17%, AA 99.32%, Kappa 99.05%) after a transient dip at 75%; we tentatively attribute this recovery to an increased reliance on global spectral signatures once local spatial texture is degraded, a hypothesis that echoes the reconstruction regime of masked autoencoders and warrants further analysis. The most challenging minority class (Oats) reaches 100% recall at both 50% and 90% masking, confirming the regularization effect on heavily class-imbalanced data. STM-lite constitutes an effective, physically grounded augmentation for HSI classification, improving accuracy and per-class robustness with zero architectural overhead during inference; its main limitation is that both the masking ratio/tube geometry and the pixel-wise random split were validated on a single benchmark and dataset, leaving cross-dataset generalization and spatially disjoint validation for future work.