Advances in Hyperspectral Imaging for Remote Sensing, Biomedical Imaging, and Gastroenterology: A Review of Deep Learning and Hybrid AI Models

Abstract

Hyperspectral imaging (HSI) has emerged as a transformative sensing technology capable of capturing high-dimensional spectral–spatial information across hundreds of contiguous bands, enabling advanced analysis in remote sensing, biomedical imaging, and gastroenterology. In recent years, the field has undergone a significant transition from traditional signal processing and statistical learning methods toward deep learning, transformer-based architectures, and hybrid AI–physics models. Despite these advancements, challenges such as high dimensionality, limited labeled datasets, domain shift, computational complexity, and lack of standardization continue to limit large-scale real-world deployment.
This study presents a comprehensive systematic review of recent developments in hyperspectral imaging, focusing on AI-driven methodologies, multimodal fusion strategies, and physics-informed modeling approaches. The review synthesizes findings from high-impact studies published between 2024 and 2026, covering key tasks including classification, super-resolution, anomaly detection, change detection, and cross-domain generalization. Comparative analysis reveals that transformer-based and hybrid AI–physics models consistently outperform traditional machine learning and CNN-based approaches, achieving classification accuracies of up to 95.2%. Similarly, advanced reconstruction and anomaly detection frameworks demonstrate improved spectral fidelity, robustness, and noise resistance.
The findings further indicate that multimodal few-shot learning and hybrid architectures significantly enhance generalization across different sensors and domains, reducing performance degradation under cross-scene conditions. In biomedical applications, particularly gastroenterology, hyperspectral imaging shows strong potential for non-invasive tissue characterization and early disease detection through hyperspectral endoscopy. However, persistent limitations remain in interpretability, computational efficiency, and dataset availability.