10 April 2026 in IEEE Signal Processing Letters
Low-rank representation with sparse outliers is a powerful paradigm for hyperspectral anomaly detection, yet its performance in complex scenes is often limited by weak spatial modeling and the unreliable use of coarse background priors. Existing low-rank/tensor detectors typically rely on globally tuned regularization and ignore the spatially varying reliability of prior estimates. Although recent learning-based HAD methods are promising, lightweight model-driven methods remain attractive for interpretable and training-free deployment. This letter proposes a confidence-weighted prior-guided RPCA (CWP-RPCA) model that decomposes an HSI into a subspace low-rank background and sparse anomalies while incorporating a superpixel-based collaborative background prior in a pixel-wise gated manner. CSRD reconstruction residuals are used to derive a confidence map for adaptively weighting prior strength, and a mild TV penalty on background coefficient maps enhances spatial continuity and suppresses superpixel-induced mosaic artifacts. The resulting convex formulation is solved by an ADMM algorithm with a closed-form FFT-based update of complexity $O(NL\log L)$. Experiments on three remote-sensing datasets and one intraoperative brain HSI dataset demonstrate higher AUCs, cleaner backgrounds, and fewer false alarms than recent representative detectors, together with favorable robustness and efficiency.