Abstract:

Hyperspectral image classification is a hot issue of hyperspectral remote sensing information processing. Under the structure of kernel sparse representation classification, SpatialSpectral Kernel Sparse Representation Classification (SSKSRC) of hyperspectral images can achieve better performance by joint spectral features and information of spatially adjacent pixels. However, it is impossible to utilize it in timecritical condition because of the large scale of data and calculation. A parallel optimization method of SSKSRC is proposed based on GPU/CUDA. A memory access optimization strategy is designed to optimize the data exchange between the host and the device. The parallel computing ability of GPU is fully used to accelerate the calculation of the kernel matrix in the process of classification. The matrix operation that is realized according to the parallel feature of GPU is used to optimize the solving process of the classification model based on the alternating direction multiplier method. The experiments with real hyperspectral image data validate the effectiveness and efficiency of the proposed method.

Key words: remote sensing;hyperspectral;GPU;parallel;classification