Parallel Algorithm Design And Optimization Of Nonlinear Dimensionality Reduction For Hyperspectral Image On GPU

Dimensionality reduction of hyperspectral remote sensing image(HSI)is a key step in hyperspectral image processing.By extracting main features from the continuous spectrums,it reduces the huge computation overhead and avoids weakening the classification accuracy.With high computational complexity,traditional serial method is time-consuming,so it is unable to meet the demand in many real-time applications,such as military and geology.With outstanding logic control ability and powerful computing capacity,heterogeneous system base on CPU and GPU has turned into one of the mainstreams in the field of high performance computing.Equipped with numerous computing cores and effective memory bandwidth,GPU eliminates the bottleneck in data calculating and memory accessing,thus improving task throughput greatly.In this research,we bring in parallel computing on GPU to solve the performance problems in nonlinear dimensionality reduction of HSI.Linear algorithm has great limitations for hyperspectral data due to its nonlinear characteristics.However nonlinear models are difficult to be applied practically for time-consuming procedures with high computational complexity.In this paper,we study the parallel schemes and optimization methods of nonlinear dimensionality reduction algorithm of HSI on multicore and many-core heterogeneous platform.Our main works and innovations are listed in the following:1)Taking KPCA algorithm for example,the research is focused on parallel schemes and optimization strategies of kernel-based nonlinear dimensionality reduction algorithms on GPU.By analyzing and excavating the parallel potentials of accelerating hotspots including gaussian kernel matrix computation,bilateral jacobi iteration and KPCA transformation,we discussed various mapping schemes and memory accessing modes.In the experiment,we validated the performance enhancement of improved methods and compared the results of different designs and optimizations.The experiment proves that the speedup is proportional to the amount of data and the best performance improvement of parallel KPCA algorithm based on GPU is up to 173 x.2)Manifold learning is one of the research hotspots in the area of HSI nonlinear dimensionality reduction.Based on CPU/GPU heterogeneous system,we proposed its parallel implementation.During parallel realization and optimization,neighborhood graph,minimum K values and all pairs' shortest paths were studied successively.In further research,the designing schemes were improved by making better mapping method,employing shared memory and reducing extra overhead.Moreover,we analyzed and quantified performance diversity of different parallel implementations.The experiment shows that GPU-based parallel ISOMAP can obtain speedups between 9.06 and 91.15.3)When the amount of hyperspectral data increases to a huge size,nonlinear dimensionality reduction of HSI cannot run on a single node due to execution space limitation,so distributed storage is introduced.In this paper,we proposed KPCA algorithm based on MPI.Combining with shared storage and many-core GPU architecture,we came up with two kinds of parallel plans which are respectively based on MPI+CUDA and MPI+OpenMP+CUDA.Furthermore,the research made detailed discussion about communication scheme,task occupancy,partition granularity,storage consistency and memory accessing efficiency.The experiment shows that the parallel KPCA based on MPI+OpenMP+CUDA obtains 2.75~9.27 x speedups over the MPI KPCA method.