BBO-MARNet: Bayesian Band Optimization and Mask-Aware Reconstruction Network for Efficient Hyperspectral Anomaly Detection
Abstract
Hyperspectral anomaly detection is essential for blind target discovery, yet its effectiveness is hindered by high spectral redundancy and target–background entanglement. Existing reconstruction-based methods rely on task-agnostic full-spectrum modeling, which lacks discriminative subspace alignment and often leads to anomaly memorization. To address this limitation, we propose the Bayesian band optimization and mask-aware reconstruction network (BBO-MARNet). The framework first employs Bayesian optimization (BO) to identify a compact, detection-oriented band subset, reducing spectral redundancy prior to feature learning. A lightweight spatial–spectral masked autoencoder is then applied to reconstruct missing information from contextual cues, mitigating anomaly memorization. Additionally, a tripartite composite loss function, integrating multiscale reconstruction, gradient consistency, and internal regularization, is designed to enhance background representation and preserve structural boundaries. Experiments on five benchmark datasets, including a newly curated maritime dataset, demonstrate that BBO-MARNet consistently outperforms state-of-the-art methods while maintaining high efficiency with only 1.69M parameters. The source code is available at https://github.com/Sonel-YXL/BBO-MARNet