Skip to content
Open access

SRBNet: Selective Reconstruction Bypass for Edge-Robust Hyperspectral Anomaly Detection

2026 · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · Vol 19, pp. 33500-33516 · 0 citations · 48 references

Abstract

Reconstruction-basedhyperspectral anomaly detection (HAD) suffers from high false-alarm rates at background edges, where edge residuals are comparable in magnitude to those of true anomalies, while increasing reconstruction capacity risks reconstructing the anomalies themselves. Edge and anomaly residuals, however, differ sharply in spatial topology: the former are elongated and continuous, the latter compact and isolated. Exploiting this distinction, we propose the Selective Reconstruction Bypass Network (SRBNet). Rather than reconstructing edges more faithfully, SRBNet learns a spatial gate that bypasses reconstruction at edge regions, restoring the original spectra there to remove edge-induced false alarms while preserving anomaly residuals. The gate is trained in a self-supervised, topology-aware manner from residual geometry alone, requiring no annotation, and is confined to edge structures. Because it discriminates residuals by shape rather than by how the reconstruction is produced, the gate also transfers as a plug-and-play refinement to existing reconstruction-based detectors. Experiments on 33 scenes from three benchmarks show that SRBNet outperforms eight representative methods in detection accuracy and background suppressibility (average AUC$_{(D,F)}$ and AUC$_{\text{BS}}$ of 0.9922 and 0.9824), and its gate consistently improves four other detectors.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.