GMoE-AD: Generalized Hyperspectral Anomaly Detection via Mixture-of-Experts and Domain-Invariant Learning

Hyperspectral (HS) sensing provides detailed high-dimensional spectral data to identify subtle anomalies and material variations in complex scenes. HS anomaly detection (HS-AD) aims to identify small, spectrally distinct objects or materials in HS imagery without prior target information. However, most HS-AD methods assume that the training and test data are drawn from the same distribution (single-domain). This assumption is often violated in operational sensing because of differences in sensor characteristics, scene composition, atmospheric conditions, and acquisition geometry. To address this challenge, we propose Generalized MoE-AD (GMoE-AD), a neural Mixture-of-Experts (MoE) architecture for robust cross-domain hyperspectral anomaly detection. The framework fuses the outputs of six unsupervised base detectors with representations from a pretrained HS foundation model, combines four neural experts through learned top-2 routing, and applies gradient-reversal-based domain-adversarial training to improve robustness under distribution shifts. A drift-aware test-time adaptation (DTA) variant is evaluated separately. We evaluate GMoE-AD on six public real-world HS benchmark datasets-San-Diego, Salinas, HYDICE-Urban, ABU-Airport, ABU-Beach, and ABU-Urban-and one private Arizona dataset comprising 22 images acquired by four different sensors. Under all-domain training, the model uses the training portions of all seven datasets and is evaluated without access to domain identity or dataset-specific information. GMoE-AD achieves an average ROC-AUC of 0.943, PR-AUC of 0.623, and F1-macro of 0.821 in this setting. In a leave-one-dataset-out (LODO) evaluation, where the target dataset is completely unseen during training, the model maintains an average ROC-AUC of 0.910 and F1-macro of 0.773. The optional DTA variant increases mean ROC-AUC from 0.938 to 0.953 but reduces F1-macro from 0.822 to 0.815, indicating a metric- and dataset-dependent adaptation trade-off. These results suggest that combining neural expert routing, transfer learning, and domain-adversarial representation learning improves robustness across heterogeneous HS datasets and provides a unified approach to anomaly detection in high-dimensional sensor data, supporting deployment in real-world hyperspectral applications where training and deployment conditions differ.