Robust asymmetric encoder-decoder nonnegative matrix factorization for hyperspectral anomaly detection

Hyperspectral anomaly detection remains a challenging task due to the limitations of traditional low-rank matrix factorization methods, which often fail to effectively isolate anomalies. Most approaches model hyperspectral images as a combination of low-rank background and high-rank anomaly components, but do not explicitly learn the anomaly, leading to contamination of the background representation and suboptimal detection. Furthermore, existing factorization-based models often lack an encoder, restricting their ability to enforce mutual consistency between latent features and reconstruction, especially in noisy environments. To overcome these challenges, we propose Robust Asymmetric Encoder–Decoder Nonnegative Matrix Factorization (REDNMF), a novel optimization framework that jointly learns a sparse anomaly matrix and low-rank background components. REDNMF incorporates an explicitly modeled anomaly matrix, improving robustness and detection accuracy. Unlike traditional decoder-only schemes, our asymmetric encoder–decoder architecture enables flexible and discriminative feature learning, enhancing both reconstruction fidelity and anomaly-background separation. Additionally, REDNMF enforces pixel-wise sparsity on the anomaly matrix using the L2,1 norm, promoting structured sparsity that aligns with the hyperspectral image geometry. This design helps localize anomalies more precisely and reduces sensitivity to noise and modeling errors. All components are integrated into a unified framework and optimized jointly. Extensive experiments on both real-world and benchmark hyperspectral datasets demonstrate that REDNMF consistently outperforms state-of-the-art methods in detection accuracy and robustness.