A Practical Fixed-Configuration, Training-Free Baseline for Within-Sensor Hyperspectral Anomaly Detection

Abstract

Hyperspectral anomaly detection (HAD) aims to identify rare targets without labeled anomaly samples, but many recent deep HAD approaches require scene-specific neural optimization, increasing deployment cost. This paper presents MASCR-UAF, a fixed-configuration baseline designed for training-free within-sensor HAD. Global hyperparameters are calibrated on a designated calibration split and frozen before evaluation. No target-scene labels, neural-network retraining, gradient-based adaptation, or target-label-based hyperparameter selection are used, although unlabeled statistical quantities are estimated independently from each test scene following classical HAD conventions. MASCR-UAF integrates complementary spectral, density, and spatial anomaly cues through an unsupervised adaptive fusion rule. On ten held-out AVIRIS ABU scenes, the framework achieves competitive AUC-ROC performance while providing improved AUC-PR and low-false-alarm operating characteristics compared with the evaluated individual statistical detectors. Additional evaluations on Indian Pines, San Diego, and Pavia University reveal regime-dependent behavior across agricultural, airport, and sensor conditions. These results suggest that MASCR-UAF can serve as a practical reference for within-sensor cross-scene HAD, while its applicability beyond this setting requires further investigation.