Hyperspectral Algorithms for Anomaly Detection

We propose to develop methods for enhancing computational efficiency and effectiveness of HSI anomaly detection. Many existing algorithms typically suffer from significant false alarms due to the assumptions that the local background is Gaussian and homogeneous. These approaches also require significant computational resources providing a significant impediment to near-real-time implementation. We will first investigate the optimization of a subset of existing algorithms that are known to perform reasonably well, to provide feasible near real time implementation for anomaly detection. We further propose the use of a new non-parametric model tuned specifically to the data at hand to both enhance computational efficiency as well as performance. The proposed method, based on the application of kernel-based methods, has distinct computational advantages as it is linear relative to spectral dimensionality. Furthermore, the non-parametric modeling employed provides enhanced separation between target and background with increased detection and lower false alarm (FA) rates. Computational and algorithm performance will be analyzed and quantified. Feasibility of the proposed approach will be demonstrated in Phase I, with the development of a fully evaluated prototype in Phase II.