Parallel algorithm of anomalies detection in hyperspectral image with projection pursuit

Projection pursuit can extract non-Gaussian structure in hyperspectral data to reveal the anomalies information, but searching the best projection directions is a computational intensive task. To improve the process efficiency, a parallel process algorithm under cluster system was presented. The combination of skewness and kurtosis was selected as projection index (PI). Using all the pixels' spectral as a special projection dierection set and searching the best projection directions in it. While parallel computing the PI value for each candidate dierection, the hyperspectral data was distributed to each computing node after partitioned evenly. After projecting each data subblock to a candidate direction in parallel, the index computation was transformed and decomposed. This makes all the data needed during index computation be in local memory for each node and decreases the communication. Furthermore, a "be host in turn" method was put forward to improve the degree of load balance. Using an operative modular imaging spectrometer data to test the efficiency on cluster, the results show that the parallel algorithm achieves good parallel performance.