Non-Negative Latent Factorization Autoencoder for Interpretable Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) requires effective background modeling while limiting the reconstruction of rare anomalous spectra. Conventional non-negative matrix factorization provides an interpretable endmember–abundance representation but offers limited flexibility for estimating background representations under complex spectral variability, whereas reconstruction-oriented autoencoders provide flexible nonlinear inference but often lack an explicit mixing-model structure. To bridge these two paradigms, we propose a Non-negative Latent Factorization Autoencoder (NLFA) for anomaly-oriented background modeling. NLFA employs a nonlinear encoder to estimate simplex-constrained abundance-like latent coefficients and a linear mixing model (LMM)-consistent linear decoder to represent the dominant background through endmember-like spectral bases. A column-sparse residual is jointly optimized with the latent factorization so that spectral components insufficiently explained by the learned background are retained for anomaly detection, while local spatial smoothness and endmember-diversity constraints regularize the background representation. Experiments on eight benchmark hyperspectral datasets show that NLFA achieves competitive detection performance.