Normalizing Flow-Based Probability Distribution Representation Detector for Hyperspectral Anomaly Detection

Due to the powerful reconstruction ability, deep learning based hyperspectral anomaly detection methods have been prevalent in recent years. However, the capability of neural networks and the meaning of latent space remains unexplainable to some extent. To address the issue, we propose a normalizing flow-based probability distribution representation detector (NF-PDRD) for hyperspectral anomaly detection in this article, which clarifies the capability of the model from a probabilistic perspective. The framework first utilizes the variational autoencoder to acquire the probability distribution representation with the mean vector and standard deviation vector for the original data. Subsequently, we introduce a normalizing flow to transform the Gaussian approximate posterior to a more complex distribution, making the model generative and expressive. We finally accomplish the detection process with the extracted probabilistic representation data using the strategy of Gaussian mixture model estimation to fully leverage the spatial information. Experimental results on both synthetic and real data sets demonstrate the outstanding performance of the proposed NF-PDRD.