Hyperspectral Remote Sensing and Deep Graph Embedding Fusion for Spatiotemporal Crime Pattern Analysis in Forensic Archaeological Site Protection

Spatiotemporal crime pattern analysis in forensic archaeology requires precise modeling of subtle environmental perturbations and complex multisource event dependencies. Traditional geographic information system (GIS) mapping and statistical frameworks rely on coarse spatial resolution and fixed temporal windows, failing to capture nonlinear spectral signatures or implicit topological interactions, which limits both predictive accuracy and forensic interpretability. This study presents the hyperspectral remote sensing and deep graph convolutional embedding (HRS‐DGCE) framework, a unified architecture designed to detect illegal site excavations, looting, and site disturbances. The framework processes high‐resolution UAV‐acquired hyperspectral data across 224 spectral bands (400–2500 nm), spatially aggregated into a 10 m × 10 m grid topology. Graph nodes represent these spatial grid units, while edges capture spatial proximity, temporal co‐occurrence, and spectral similarity via spectral angle mapper (SAM) metrics. Digital metadata and forensic logs are integrated directly into node feature vectors without spatial distortion. Low‐dimensional spatiotemporal embeddings are learned using a dual‐layer graph convolutional network coupled with temporal kernel smoothing and supervised graph regularizes. Evaluated on the benchmark CrimeSpatioTemporal2025 dataset using five‐fold temporal cross‐validation, the proposed HRS‐DGCE framework achieves an overall accuracy of 92.40% ± 0.85%, a macro F1‐score of 0.90 ± 0.02, a temporal anomaly detection rate of 91.00% ± 1.20%, and a spatial correlation index of 0.87 ± 0.01, outperforming state‐of‐the‐art spatio‐temporal graph convolutional networks and deep baseline architectures. The system establishes a verifiable, mathematically rigorous framework for forensic site monitoring and risk mitigation.