Few-Shot Hyperspectral Image Classification: A Review of Deep Architectures, Paradigms, and Benchmarks

Abstract

The critical bottleneck in hyperspectral image (HSI) classification lies in the inherent conflict between the extreme scarcity of labeled samples and the massive data requirements of deep models, making few-shot classification a vital research frontier. This article presents a comprehensive review of recent advances in deep learning for few-shot HSI classification, systematically structured around core architectures, learning paradigms, and rigorous benchmarking. At the architectural level, we trace the evolutionary trajectory from traditional convolutional neural networks to global-modeling Transformers, and further to emerging Mamba models that achieve linear complexity with global receptive fields. Concurrently, from a learning paradigm perspective, we analyze the mechanisms of transfer learning, meta-learning, and contrastive learning in mitigating sample constraints. Crucially, to eliminate pervasive cross-study evaluation biases, we establish a unified benchmark across four public datasets, conducting standardized quantitative cross-validation of three representative architectures and three advanced paradigms. Extensive experimental results demonstrate that long-range dependence models (Transformer and Mamba) exhibit superior generalization under extremely limited samples. Furthermore, contrastive and transfer learning paradigms significantly expand interclass feature margins, enhancing overall model robustness. Finally, we outline promising future directions, including physics-informed deep learning, frequency-aware reconstruction, and multilevel semantic enhancement for complex cross-domain scenarios.