Multilevel parallel algorithm of PCA dimensionality reduction for hyperspectral image on GPU

Based on the CPU/GPU heterogeneous system, the classical principal component analysis (PCA) method was researched. A G-PCA algorithm for single GPU and a Gs-PCA algorithm for multiple GPUs were implemented and validated on a mini-supercomputer system. Experimental results showed that the performance can be remarkably enhanced using the G-PCA algorithm, but the simulation is constricted by the limited memory; in comparison, the problem can be overcome by using the Gs-PCA algorithm, and can reach a maximum speed-up of 128X in test. ©, 2014, Northeastern University. All right reserved.