Research And Application For Hyperspectral Imagery Classification Based On GPU

Hyperspectral imagery has an important prospect in classification with the rich spectral information. Due to the huge dataset of hyperspectral imagery and high complexity of classification algorithm, the computational performance of existing classification algorithms in serial mode is often low, it is difficult to be popularized in the time-critical scenarios. Recently, GPU has opened new avenues to accelerate hyperspectral image classification algorithm due to its high level computational performance with low cost.The spectral information of neighboring spectral bands may be highly correlated in the hyperspectral image, and the redundant information will bring big computational challenge, reducing dimension is an effective method to discard redundant features, but the additional algorithm will increase the complexity of the program. The fusion of spectral and spatial information will improve the classification accuracy, but it also will further increase the complexity. How to increase computational efficiency while not degrading classification accuracy is an important aspect of hyperspectral image processing. In this article, a faster classification framework is proposed by implementing parallel dimensionality reduction, spatial feature extraction and classification algorithms based on GPU. The distributions of this article are as follows:Firstly, this article illustrates the research status of hyperspectral image classification and processing technology based on GPU according to the characteristics of hyperspectral and GPU, it provides technical support for the parallel implementation of the algorithm.Secondly, parallel spectral and spatial feature extraction algorithms based on GPU are proposed. The LPE algorithm is used to select the most distinctive but informative bands. The LBP algorithm is used to extract local texture information. All the algorithms are parallel implemented and get a better speedup ratio in real hyperspectral image datasets.Thirdly, a parallel CRC with distance-weighted Tikhonov regularization based on GPU is proposed. It has obtained a higher speedup ratio compared with the serial implementation when keeping high classification accuracy. And then the parallel algorithms are applied to achieve the faster classification of cell.