Anomalies detection approach using projection pursuit in hyperspectral image based on parallel genetic algorithm

Abstract

This anomalies detection approach seeks the directions that maximize the projection index, so as to gain the anomalies structure information. Using genetic algorithm in this approach can search accurate optimal projection directions, but it's a computation-intensive task. So, a parallel algorithm under distributed memory system was presented. The projection directions were searched efficiently by parallel genetic algorithm model, and the projection directions' precision was guaranteed by using a strengthened terminal qualification. Then, the detected anomaly components were wiped off by projecting the data onto the subspace orthogonal to the previous projection directions, and the other anomalies were searched in the residual space. The final task of projection and objects segmentation was also completed in parallel. Using an OMIS hyperspectral data to test the parallel algorithm's performance under an eight-node cluster, the process time reduced from 15 minutes to 2.8 minutes. The results show the validity and comparative good parallel efficiency.