Study of Multilevel Parallel Algorithm of KPCA for Hyperspectral Images

Hyperspectral remote sensing image data has been widely used in a variety of applications due to its continuous spectrum and high spectral resolution. However, reducing huge dimensions with high data relevance is time-consuming, and parallel processing is required to accelerate this process. In the previous work, the KPCA (Kernel Principal Component Analysis), a nonlinear dimensionality reduction method was studied, and a parallel KPCA algorithm was proposed based on heterogeneous system with a single GPU, and achieved the desired experimental results. However, as data scale grows, the proposed solution would consume all the available memory on a single node and encounter performance bottleneck. Therefore, to tackle the limitation of insufficient memory caused by the reduction of large-scale hyperspectral data dimension, in this paper the intra-node parallelization using multi-core CPUs and many-core GPUs are exploited to improve the parallel hierarchy of distributed-storage KPCA. Finally, we designed and implemented a multilevel hybrid parallel KPCA algorithm that achieves 2.75–9.27 times speedup compared to the traditional coarse-grained parallel KPCA method on MPI.