Research On Application Of Cloud Computing In Hyperspectral Image Classification

About Remote sensing technology, hyperspectral remote sensing technology is very conducive to deep excavation of the physical and chemical properties of different objects and the good identification of different objects,so for image classification and identification it has a unique advantage.However, due to the different hyperspectral techniques in different time,space and spectral bands collected by the remote sensing images of different,resulting in a large remote sensing image data, and the collected data with multi-band, large dimension characteristics, so remote sensing into large data era. Based on the distributed cloud computing architecture Hadoop, this paper designs the task assignment strategy of pixels, the distributed interclass distance solution and the class calculation method, and designs the nearest normal subspace algorithm for hyperspectral image data classification.Experiments show that for the hyperspectral image classification algorithm data and computationally intensive problems, 8 compute nodes can achieve about 6 times faster. Compared with Hadoop, the algorithm implemented on the Spark architecture is faster and more than twice as fast, and the fusion of Hadoop and high performance computing GPU is studied to further improve the parallel efficiency. The main contents of this paper are as follows?Firstly, through a large number of literature researches, to understand the remote sensing image classification, cloud computing platform and cloud computing technology in remote sensing image processing applications. Based on the study of several hyperspectral image classification algorithms, the nearest regularized subspace classification algorithm is selected and programmed.Secondly, based on the analysis of the characteristics of hyperspectral image data, the Hadoop architecture and the related technology and workflow of Spark framework based on memory computing are studied in cloud computing technology.Thirdly, the Nearest Regularized Subspace classification algorithm is applied to Hadoop platform, and the hyperspectral image data is quickly classified by MapReduce programming using parallel programming framework. Considering the massive growth of hyperspectral image data and the real-time requirement of image classification, this paper analyzes the advantages of using the distributed parallel architecture Hadoop to realize the classification algorithm, and divides the hyperspectral image data into multiple nodes Process, in order to achieve a good acceleration effect.Fourthly, based on the memory computation of open source cluster computing system Spark, combined with the nearest regularized subspace classification algorithm, we compare the running time and classification accuracy of parallel program based on Hadoop under the condition of using the same hyperspectral image, and discuss the effect of different nodes on the efficiency of classification at run time.Finally, the fusion of distributed architecture Hadoop and high-performance computing GPU to achieve matrix addition to parallel acceleration, Test multiple sets of data, in the Hadoop and Hadoop - GPU platform, the program running time.