Cross-Scene Spectral Instance Discovery: From pixel-level anomaly detection to cross-scene instance discovery in hyperspectral imagery

Decades of hyperspectral anomaly detection (HAD) research culminate in pixel-level score maps, yet operational missions require instance extraction and cross-scene instance matching that lie beyond what these maps provide. Although prior studies have addressed individual stages of the detect–extract–match progression, the manual workflows analysts still depend on are breaking down as hyperspectral archives expand in scale across scenes. To formalize and automate this detect–extract–match chain, we introduce cross-scene spectral instance discovery (CrossSID). To the best of our knowledge, it is the first unified framework that organizes pixel-level detection, instance extraction, and cross-scene instance matching under a shared three-level task definition and hierarchical evaluation protocol. Because detection and matching require spectral relationships that remain comparable across scenes, CrossSID employs a self-supervised scene-robust relational encoder (SRRE). The resulting per-pixel embeddings, the raw spectra, and a training-free relational descriptor are each scored by Mahalanobis distance and combined through size-conditioned fusion. A quantile-adaptive seed-and-grow procedure then extracts instances from the fused score map, and score-weighted averages of these SRRE embeddings form instance descriptors for cross-scene retrieval. Experiments on 117 scenes from seven sensors validate all three levels of the hierarchy, with leading pixel-level detection on in-distribution and out-of-distribution scenes. They also show that instance extraction does not require per-scene threshold tuning, and they assess cross-scene matching quantitatively under both cross-flight-line and cross-sensor protocols.