Spectral matched filter ensembles for hyperspectral target detection

Hyperspectral target detection (HTD) faces significant challenges, including the dimensionality and substantial spectral variability. The objective of this work is to develop robust and computationally efficient matched filter-based detectors that overcome these limitations through ensemble learning. Although conventional matched filters (MF) are widely used as a primary and straightforward method in this field, they suffer from covariance matrix instability in high-dimensional spaces and spectral confusion in complex target-background environments. Therefore, this paper presents three improved ensemble-based variants of MF: the random subspace matched filter (RS-MF), its enhanced version with an adaptive target signature (Adaptive RS-MF), and the selective subspace matched filter (Selective RS-MF). Experiments conducted on three benchmark hyperspectral target detection datasets demonstrate that the proposed variants not only improve detection performance over the conventional MF but also outperform other existing HTD methods. The final results show that Selective RS-MF achieves an area under the receiver operating characteristic curve (AUC-ROC) of approximately 0.99, providing robust and effective solutions for HTD in diverse environments, while maintaining computational efficiency.