Rethinking AI in Hyperspectral Image Classification: Paradigms, representations, and interpretability

The past decades have witnessed remarkable accuracy gains in hyperspectral image (HSI) classification, driven by rapid advances in artificial intelligence (AI). Although several surveys have reviewed AI-based HSI classification from specific methodologies or subdomains, a comprehensive understanding of its evolutionary trajectory remains lacking. In particular, the intrinsic relationship between the rise of deep learning and the growing emphasis on interpretability has not been systematically explored. To address these gaps, this paper rethinks and comprehensively reviews the field through a unified hierarchical framework that integrates three interconnected perspectives: classification paradigm, feature representation, and interpretable learning. First, we trace the evolution of HSI classification from model-driven to data-driven and toward interpretability. Then, we reveal that feature representation mechanisms shift from handcrafted representations to self-learned latent representations and move to traceable internal representations. Furthermore, we systematically investigate three mainstream interpretable learning approaches in HSI classification: prior knowledge embedding, deep unfolding, and post-hoc analysis. Based on this unified framework, three key findings are highlighted: (1) the shift in feature representation mechanisms is a major driver of classification paradigm evolution; (2) achieving interpretability essentially requires revealing the intrinsic mechanisms of feature representation; (3) interpretable learning is not a rejection of deep learning, but rather an enhancement of its transparency and trustworthiness. Finally, three promising research directions are outlined. These include developing quantitative interpretability evaluation frameworks, exploring interpretable mechanisms for multimodal data fusion, and constructing trustworthy AI agents. We hope this review serves as a navigational compass for researchers and catalyzes further advances in this field.