Research On Fast Implementation Of SAR Algorithm Based On Memory Distribution Computing

With the increasing number of synthetic aperture radar(SAR)imaging tasks,the traditional single computer serial imaging has the problem of insufficient computing power and scalability,and it is difficult to meet the needs of SAR rapid imaging.Distributed parallel computing is to combine computers with common configuration into a cluster,which greatly further the computing ability and scalability of computers.Besides,compared with the MapReduce framework,the Spark framework does not require disk drop operations during map conversion,and can be directly cached in memory,improving the execution efficiency,which provides an efficient and feasible technical approach for SAR imaging in large scenes.Therefore,in this paper,Spark is applied to SAR imaging algorithm,and SAR memory distributed parallel imaging method based on Spark is studied.The main work of this paper is as follows:1.Introduce the basic algorithm of SAR imaging,these include Back Projection(BP)algorithms and Compressed Sensing(CS)algorithms,besides,the principle of memory distribution calculation is also introduced.Including resource scheduling management module,distributed file system and computational framework,which lays a conceptual foundation for the subsequent realization of distributed parallel computing.2.Aiming at the problem that BP algorithm uses a single computer to image for a long time and cannot quickly image,a Spark memory distributed parallel implementation method for BP-SAR imaging is proposed.The imaging method first partitions the range-compressed data,packages the partitions into independent tasks,and sends them to the Spark cluster for parallel computing,and then coherently accumulates the calculation results of each node to complete the BP imaging of SAR.Due to the Spark computational framework,this method has the benefit of distributed parallel computing,which can improve the imaging speed of the algorithm.In the experiment,the method runs in a4-node Spark cluster,and the imaging speed is 4.9 times that of the single computer serial computing BP-SAR imaging method,and 1.4 times that of the MapReduce SAR back-projection imaging method.3.Aiming at the problems of many iterations,large amount of computation and long imaging time in CS imaging algorithm,a Spark memory distributed parallel implementation method of CS-SAR imaging is proposed.The approach benefits from Spark’s design pattern for resilient distributed datasets.In the process of range imaging,the range imaging results are cached in Spark’s elastic distributed data set;in azimuth imaging,data is directly read from the elastic distributed data set to complete azimuth imaging.This method runs in the Spark cluster of four nodes,and its imaging speed is 1.9 times faster than the CS-SAR imaging method of single computer serial calculation,and 1.4 times faster than the SAR compressive sensing imaging method of MapReduce.