Few-Shot Hyperspectral Representation Learning for Multi-Task Forensic Document Analysis Using Episodic Prototypical Networks

Hyperspectral imaging has emerged as a promising modality for forensic document analysis due to its ability to capture spectral and material information beyond conventional RGB representations. However, existing hyperspectral forensic studies typically focus on individual tasks and provide limited insight into the transferability of learned representations across different forensic applications. This study investigates episodic hyperspectral representation learning for forensic document analysis using a shared spectral–spatial embedding space learned through prototypical optimization. The proposed framework is evaluated on the complete UWA WIHSI and iVision HHID hyperspectral datasets across multiple forensic applications, namely writer identification, writer verification, ink mismatch detection, forgery detection, gender prediction, and age prediction.The proposed method achieved high performance on the controlled UWA WIHSI dataset across multiple forensic tasks. On the more challenging iVision HHID dataset, it attained a writer identification accuracy of 90.74%, together with writer verification and ink mismatch detection ROC–AUC scores of 0.9936 and 0.9915, respectively. Cross-dataset experiments further highlighted both the potential and the challenges of transferring learned hyperspectral representations across different acquisition conditions and writer populations, while extensive ablation studies confirmed the importance of spectral information, embedding design, and optimization strategy in learning effective hyperspectral representations. Overall, the results show that shared hyperspectral representations can support diverse forensic document analysis applications within a common embedding space while providing insight into the opportunities and challenges of developing transferable hyperspectral representations for forensic settings.