Performance of hyperspectral target detection algorithms is determined by the spectral variability and separability of target and background materials within the scene. Practical matched filter detectors typically utilize only background statistics due to the assumed rarity of target materials. Background materials are additionally modeled scene-wide as Gaussian, which allows for straightforward estimation of statistics but oversimplifies the complex manifold on which spectra are typically distributed. These simplifications can lead to detection errors in the form of both missed detections and false alarms. The variational autoencoder (VAE) is a general neural network architecture that implements a generative probabilistic model for data. This is accomplished via a deep latent variable model in which data generation is modeled by the mapping of a lower dimensional isotropic Gaussian latent variable through a neural network to a conditional distribution on the observation. VAEs are thus capable of learning distributions over complex, high dimensional manifolds. We propose utilizing the VAE as a probabilistic model for hyperspectral data to aid in target detection, discrimination, and false alarm mitigation. We fit a VAE to a hyperspectral cube and make use of the learned latent space. The VAE encoder is trained to map each pixel to a posterior Gaussian distribution, which we compare to encoded library and scene background posteriors to determine the presence or absence of target materials. Comparative analysis of posteriors is placed in an information theoretic framework, and we establish a connection to standard detection statistics. We demonstrate detection and discrimination performance using two real hyperspectral datasets.
|