Study On Application Of Parallel Computing To Seismic Damage Analysis From Remote Sensing Images

Earthquake is one of sudden natural disasters, and brings huge disaster to human society and civilization. As earthquake prediction is a worldwide science problem, pre-earthquake prevention and post-earthquake rescue are the most effective methods to reduce earthquake disaster losses.Remote sensing is an integrated earth observation technology, which was developed in the 1960s. Compared with the traditional observation methods, it has unique advantages: large range, comprehensive and macroscopic observation; large amount of information, multi-means, multidimensional and all-weather observation capabilities; quick access to information and shorter update cycle. For a long time, the earthquake disaster information was obtained through on-site investigation with shortages of a heavy workload, low efficiency, high cost and so on. The remote sensing technology can be applied to obtain quickly the comprehensive and macroscopic disaster information on the earthquake-hit areas and provide timely the key basis for disaster investigation, loss assessment, emergency rescue, recovery and reconstruction.With the development of remote sensing technology, remote sensing images with high resolution were gradually increased, and the amount of data was also rapidly grown. Single computer can't meet the fast processing requirement of the massive remote sensing image data. The speed of remote sensing image procession becomes a bottleneck of the application of remote sensing technology to a large extent. This brings a new challenge to remote sensing image processing technology. Thus the study on the application of parallel computing technology to the massive remote sensing image procession and seismic damage extraction was selected as the main topic in the thesis in order to improve substantially the processing speed and precision of emergency remote sensing image.This thesis introduces firstly the application status of seismic damage extraction from remote sensing images, the related concepts of parallel computing, the development of parallel computer and the mainstream parallel programming environments, as well as the development trend of parallel computer—CPU + GPU heterogeneous cluster and the hybrid parallel programming environment of MPI, OpenMP, CUDA and OpenCL. Then the application status of parallel computing technology to the remote sensing image procession was introduced. The construction of Beowulf cluster based on MPI and Windows is described, The correlation coefficient method and ratio method, as the change detection methods commonly used in seismic damage extraction from remote sensing images are selected. The parallel computing algorithm and program are implemented in the cluster computers based-on master/slave mode. The tests of parallel computing of change detection are performed by using the pre-earthquake and post-earthquake remote sensing images in the city area of Beichuan County. Finally, the speedup and efficiency of parallel computing are analyzed.The Conclusions from the above study can be obtained as following:(1) Good correspondence between the correlation coefficient and seismic damage level can be obtained through the correlation coefficient method. The lower the correlation coefficient is, the serious the seismic damage will be, and vice versa. The above results show reliability and feasibility of the parallel computing methods.(2) The analysis of the speedup and efficiency of correlation coefficient method under window 9×9 demonstrates that, when the number of computing processes generated in each computing node is not greater than the number of cores in CPU, a good performance of the parallel algorithm of correlation coefficient method can be obtained; the speedup was generally 0.9 times the number of processes and the efficiency was maintained above 90% in the case of the non calculated processes are excluded, In principle, the speedup and efficiency will be most optimal when the number of computing processes generated in each computing node is equal to the number of cores in CPU. Otherwise they will be descended as the number of computing processes generated in each computing node greater than the number of cores in CPU.(3) The analysis indicates that there are still good speedup and efficiency for parallel computing of ratio method if only the time of computing processes is considered. While, because of low computation time of ratio method, the time of reading or writing data and communicating spent higher proportion of time. As a result, the slave process has to wait the master process to receive computed result or to send the data block to be processed. This was associated with the design model available for the large computing issues. Therefore the speedup and efficiency were unable to reflect performance advantages of parallel computing in remote sensing image processing.(4) Tests of parallel computing of correlation coefficient method with different size of window indicate that, the speedup and efficiency will be better for larger size of window. The phenomena are caused by the small changes in time spent by the master process for the non computing functions such as the data access, partition and communication, while there is a significant increase proportion of time spent by computing processes in the total run time of program. The result implicates that the master/slave design computing model is suitable for the large computing case in remote sensing image procession.In the parallel process of remote sensing damage extraction, there are some problems and shortcomings:(1) There were large changes in surface features of remote sensing image after the earthquake. It affected the image registration accuracy of pre-earthquake and post-earthquake remote sensing images. Since the accuracy of registration did not reach sub-pixel level, it had an impact on the processing result, and needed to improve the accuracy of registration. While, from the point of performance analysis of parallel algorithms, there was no practical impact.(2) Using the data partitioning and fixed data buffer size in this paper, it did not take computing processes into account. And the result in load balancing between each node was not very satisfactory. Thus it affected the improvements of parallel computing performance.(3) The master/slave design model and the method of distribution of processes between the nodes adopted in the thesis might cause a relatively higher time rate spent by data access and communication, which will cause the reduction of the performance of parallel computing.(4) Although a preliminary optimization to serial algorithm on the correlation coefficients method in this paper, there was room for optimizing parallel program. It needed for further to research and analyze the algorithm and program to achieve better optimization.Overall, the results of this research work in this paper achieved the desired goals, and expected to apply future study.