Path-Based Background Model Augmentation for Hyperspectral Anomaly Detection

摘要

We consider the detection of submerged targets in a hyperspectral image comprising a difficult maritime scene with littoral, open water, and land regions, characterized by the presence of false alarms arising from highly variable depths and sandbars. We employ a baseline detection scheme that uses kernel principal component analysis (kPCA) to learn a background model from a random small sub-sample of the scene. Detection statistics for test pixels are formed from the reconstruction error between their Nystr?m projection into the kPCA feature space and their synthesis in the feature space using the learned background principal components. We show that false alarms associated with mixed littoral pixels are reduced by augmenting the background model with spectral samples drawn from either the linear geodesic constructed between sand and water end-members or an alternative nonlinear geodesic constructed using an optimal transport technique. We show that background augmentation using the the optimal transport geodesic yields significant improvements in detection performance.