Deep Learning-Enhanced Hyperspectral Imaging for Forensic Evidence Analysis: Methodological Frameworks, Applications, and Critical Perspectives

Hyperspectral imaging (HSI), characterized by integrated spatial-spectral acquisition, offers a powerful platform for nondestructive and fine-grained forensic evidence analysis. However, high data dimensionality, spectral mixing complexity, and limited forensic sample availability constrain traditional methods reliant on handcrafted features, necessitating more robust data-driven approaches. This review systematically examines the integration of deep learning and HSI in forensic science. We first summarize three representative paradigms: convolutional neural networks (CNNs) for spatial-spectral feature extraction, attention mechanisms for discriminative enhancement and interpretability, and generative adversarial networks (GANs) for data augmentation and nonlinear spectral unmixing. We then analyze multimodal integration strategies at the input, feature, and decision levels, elucidating their methodological rationale. Representative forensic applications-including trace fiber identification, mixed ink discrimination, soil analysis, and questioned document examination-are critically evaluated to compare the strengths and limitations of different approaches. Furthermore, key challenges are discussed, including model interpretability and judicial admissibility, computational efficiency and deployability, and barriers to data standardization and sharing. Finally, future directions, including lightweight architectures, physics-informed learning, and multimodal fusion, are proposed to advance intelligent and legally defensible forensic analytics.