Rapid Processing Of Hyperspectral Remote Sensing Data Based On GPU For Mineral Resources

Hyperspectral remote sensing technology is widely applied in the earth sciences, and it has achieved great success. The mineral resource exploration is one of its important applications in the field of geological survey. But hyperspectral data and the computational complexity are large-scale so it is a great challenge to be efficient in practical application. In recent years, with the rapidly development of GPU general computing technology, due to its characteristics of multi-GPU processing cores and strong processing power and high memory bandwidth, GPU can improve the efficiency of large data processing effectively. At the same time, GPU hardware, small size, light weight and low cost advantage, has greater potential for high-performance computing technology.Based on the analysis of heterogeneous CPU+GPU programming model, this paper study the theory of hyperspectral remote sensing information extraction. It also proposed a method of mineral resources data processing quickly for hyperspectral remote sensing based on GPU/CUDA, and the experiments could obtain to meet the high-efficiency application results in real hyperspectral data. The paper work is as follows:Firstly, a fast GPU-based mineral information extraction method was proposed based on researching in the theory of hyperspectral remote sensing mineral information extraction. This paper analyzed the spectral characteristics, spectral reflectance and absorption mechanism of mineral surface features in detail. And combining the characteristics of hyperspectral data for information processing in the process of mineral on GPU/CUDA platform, it designed the process of hyperspectral mineral recognition as the key steps that included to feature extraction, endmember extraction, removal of the envelope and spectral matching.Secondly, the design and the implementation of parallel optimization on the GPU were carried out for the key steps of mineral information quickly extraction process. In addition, in the process of the design of parallel algorithms, this paper put forward the corresponding performance optimization strategies, including process optimization, improving efficiency and reducing memory access data access conflicts. (1)The principal component analysis(PCA) algorithm was used to do feature extraction, and it was improved to defend poor matrix calculation and reasonably allocated the computing tasks for the PCA algorithm on GPU+CPU; (2) The pixel purity index(PPI) algorithm was used to do endmember extraction, and it paralleled into the traditional vector projection algorithm for matrix multiplication problem in the PPI; (3) The removing envelopes and the spectral angle match(SAM) algorithm were used to do removal of the envelope and spectral matchin, and it were calculated for pixel set concurrency and paralleled the algorithms by thread mapping, storage optimization, etc. in envelope removal algorithms and the spectral matching algorithm. Then it used the real hyperspectral data to mineral mapping of information extraction in the GPU/CUDA platform, and the experimental results showed that the parallel design model and optimization methods proposed could not only guarantee the accuracy, but also able to quickly and efficiently mineral information extraction, and the maximum speedup achieved 81 times.Finally, the parallel and fast classification of support vector machines based on combination kernel on hyperspectral remote sensing image was studied. Mineral identification classification is an important part of the mineral resource exploration. Traditional methods are low computational complexity, but accuracys are not high. Combined with spectral information and spatial information as a kernel function of SVM, the support vector machine classification method based on a combination of kernel (SVMCK), could achieve better classification accuracy. But the serial algorithm is time consuming long time and cannot meet the high requirements of real-time remote sensing applications. This paper analyzed the process of SVMCK serial algorithm to identify performance bottlenecks for SVMCK serial algorithm, and proposed parallel optimization strategies to the most time-consuming part of the design based on GPU. For the most time-consuming part of the design of the nuclear precomputed, this paper proposed some parallel optimization strategies that including launching kernel function of the spectrum on the device, a reasonable allocation of airspace kernel computing tasks, improving data transmission efficiency. The practical experiments show that the effectiveness of the optimization algorithm can meet the for higher real-time requirements of hyperspectral remote sensing applications.