面向多任务粒度的遥感影像镶嵌并行技术研究

发布时间：2018-03-27 11:04

本文选题：遥感影像并行处理　切入点：多任务粒度　出处：《南京大学》2015年博士论文

【摘要】：遥感影像是地理信息系统(GIS)中的重要地理数据类型。随着遥感影像在资源调查、环境监测、气象天气预报、灾害监测和评估、军事目标识别等方面广泛应用,对于实时或接近实时的处理大数据量遥感影像的需求日益迫切。并行技术是高效快速处理大数据量遥感影像的有效途径。针对多幅遥感影像的并行处理方法具有多任务粒度的特征,解决在不同任务粒度下存在的任务依赖和数据依赖是实现遥感影像并行处理方法的关键问题。遥感影像镶嵌是遥感应用技术的重要内容,遥感影像镶嵌并行方法设计涵盖了多种任务粒度,本研究以遥感影像并行镶嵌技术研究为例,研究面向多任务粒度的遥感影像处理方法并行技术,构建多粒度的遥感影像并行镶嵌框架,提高遥感影像镶嵌的效率,为实现面向多任务粒度遥感影像并行处理提供实例和方法借鉴。本论文的主要研究内容和结论如下:(1)遥感影像地图投影及坐标变换方法并行技术。遥感影像地图投影及坐标变换是进行遥感影像镶嵌的基础。遥感影像并行地图投影及坐标变换的任务粒度为像元和单幅影像。研究分析了遥感影像地图投影及坐标变换方法原理,从并行算法设计的角度分析了以像元和单幅图像为任务粒度的并行方法的特点,提出了遥感影像地图投影及坐标转换并行方法设计的关键在于多幅遥感影像并行地图投影及坐标转换时影像数据的划分策略。研究综合考虑了数据存储格式、并行环境等对并行方法效率的影响,提出了基于遥感影像数据存储格式、处理器性能的数据划分策略,结合静态任务调度和对等模式实现了并行方法,对并行方法的性能进行了测试。测试结果表明,采用提出的数据划分策略设计的遥感影并行方法有效减少了遥感影像地图投影及坐标变换的时间,具有较高的加速比,在处理大数据量遥感影像时有着较好的表现。(2)遥感影像相对辐射校正并行技术。遥感影像相对辐射校正是保证遥感影像间色彩均衡,形成色彩一致的镶嵌影像的重要步骤。遥感影像并行辐射校正的任务粒度为部分影像。研究选择了精度和稳定性较好的迭代加权多元变换检测和正交回归方法实现并行相对辐射校正。通过对遥感影像相对辐射校正原理的分析,提出了以部分影像为任务粒度的相对辐射校正并行方法设计的难点为:选取参考影像、确定影像的校正路径、解决在校正路径上存在的任务依赖。研究提出了基于最短路径的参考影像选取和校正路径生成方法使得整体的辐射校正误差最小。研究解析了校正路径对校正系数的影响,建立了校正路径上各个任务执行结果的关系式,使得任务可以独立执行,任务依赖通过计算量很小的串行计算解决。采用了动态任务调度和主从模式实现了并行方法。并行方法在高性能集群上的实验结果表明,基于以上并行策略实现的并行方法中串行部分的执行时间占总执行时间的比例非常小,并行方法具有较高的加速比,且当处理数据量较大时,加速效果明显。(3)遥感影像拼接并行技术。遥感影像的拼接是在影像的重叠区内找到一条接缝线,作为影像范围线,将影像拼接起来形成镶嵌影像的过程。遥感影像并行拼接的任务粒度为全部影像。研究选择了形态学拼接方法实现影像并行拼接。以全部影像为任务粒度的并行拼接方法的设计需要解决以下问题:计算需要将全部影像读入内存,对内存的存储容量要求高;输出大量的中间结果,占用大量外存;方法为全局分层次计算,存在任务依赖和数据依赖。通过对形态学拼接方法原理的解析,提出了将重叠区划分为独立的相交区的数据划分方案来减少内存负担;提出了基于矢量-栅格互转的数据存储优化方案以减轻读写负担和外存占用;根据数据划分和数据存储优化结果,研究了任务依赖的特性,重新定义了任务的粒度,结合主从模式,实现了动态任务调度策略。在高性能计算集群上对并行方法性能的实验表明,采用以上并行策略,有效的减少了并行算法中的串行计算,使得遥感影像拼接执行时间大幅减少,遥感影像并行拼接方法具有较好的可扩展性,负载均衡效果较好。本研究的创新之处在于构建了多粒度的遥感影像并行镶嵌框架,提出了以部分影像为任务粒度的遥感影像并行辐射校正任务层次依赖的解决策略及提出了基于数据存储优化和动态任务调度的遥感影像并行拼接方案。
[Abstract]:The remote sensing image is the geographic information system (GIS) is an important geographic data type. With remote sensing monitoring of environmental resources in the survey, weather forecast, disaster monitoring and assessment, widely used in military target recognition. To deal with large numbers of real-time or near real-time remote sensing image according to the amount of parallel technology is increasingly urgent demand. Efficient and effective way of rapid processing of massive remote sensing image. The characteristics of the parallel processing method for remote sensing images with multi task granularity, solve the problems in different task granularity task dependence and data dependence is the key problem in the parallel processing method of remote sensing image. Remote sensing image mosaic is an important content of the application of remote sensing technology. Remote sensing image mosaic method of parallel design covers a variety of task granularity, based on the research of remote sensing image parallel mosaic technology as an example, research on multi task oriented particle Remote sensing image processing method of parallel technology, remote sensing images to construct multi granularity parallel embedded framework, improve the efficiency of remote sensing image mosaic, for parallel processing to provide examples and methods for reference implementation for multi task granularity of remote sensing image. The main research contents and conclusions are as follows: (1) remote sensing image map projection and coordinate transformation method in parallel the technology of remote sensing image. Map projection and coordinate transformation is the basis of remote sensing image mosaic. Remote sensing image parallel map projection and coordinate transformation of the task granularity is single image pixels and. Research and analysis of the remote sensing image map projection principle and the method of coordinate transformation, the parallel algorithm is designed to analyze the characteristics of parallel task granularity by pixel method and a single image for the proposed remote sensing image map projection and coordinate conversion method of parallel key lies in the design of several remote sensing images and The partitioning strategy of image data for map projection and coordinate transformation. The research considers the data storage format, influence the parallel environment on the parallel efficiency of the proposed method, remote sensing image data storage format based on data partitioning strategy of processor performance, combined with the static task scheduling and peer-to-peer mode realizes parallel method, the performance of parallel method the tests were carried out. Test results show that the proposed data partitioning strategy of remote sensing image parallel method reduces image map projection and coordinate transformation time, with a relatively high speed, has a good performance in dealing with large amount of data of remote sensing images. (2) the relative radiometric correction of remote sensing image parallel technology the remote sensing image. Relative radiometric correction is the guarantee of the remote sensing image color balance, an important step in the formation of color mosaic image consistent with the remote sensing image and radiation for school. Is the task granularity is part of the image. The research chose the better precision and stability of iterative weighted multivariate change detection and orthogonal regression method to realize parallel relative radiometric correction. Through the analysis of the principle of relative radiometric correction of remote sensing image, put forward to the part of the image for task granularity relative radiometric calibration method for the design of parallel difficulties: selection of reference image determination of image correction, path, solve the problems in the correction on the path dependence. The research task put forward the reference image of the shortest path selection and correction path generation method makes the overall radiometric correction based on minimum error correction. Study analysis the influence on path correction coefficient, a correction on the path of each task execution result of the relationship that makes the task can be executed independently, depending on the tasks by a small amount of computation. The serial computation to solve the dynamic task scheduling and Slave mode to realize parallel method in high performance. The experimental results show that the above method of parallel, parallel serial parallel method strategy to achieve part of the execution time of the total execution time based on the proportion of very small, the parallel speedup method has high, and when processing large amount of data, (the acceleration effect is obvious. 3) remote sensing image mosaic mosaic parallel technology. Remote sensing image is the overlapping area in the image of the interior to find a joint line as the range of image line, the image mosaic process will be formed. The mosaic image mosaic task granularity parallel remote sensing images. All images were chosen to achieve parallel image mosaic method of morphology design of parallel stitching. Splicing method of task granularity to all images for the need to solve the following problems: computing all images needs to be read into memory, the memory requirements for high storage capacity; output The amount of intermediate results, take up a lot of memory; hierarchical method for global computing, task dependence and the data dependence. By analyzing the principle of morphological mosaic method, proposed scheme divides the data overlapping area is divided into independent intersection area to reduce the burden of memory; a vector raster conversion data storage optimization a way to reduce the burden and to read and write memory occupancy based on data partitioning and data storage; according to the optimization results, study the characteristics of task dependence, redefined the task granularity, combined with master-slave mode, realizes the dynamic scheduling strategy. In high performance computing cluster of parallel method of performance experiments show that the parallel strategy. Effectively reduce the parallel algorithm in serial computing, the remote sensing image mosaic significantly reduce the execution time of parallel remote sensing image mosaic method has good scalability, negative Load balancing is better. The innovation of this study is to construct the multi granularity parallel remote sensing image mosaic framework, put forward to the part of the image of remote sensing task granularity parallel radiation correction task level dependent solving strategy and proposes a parallel connection scheme for remote sensing image data storage and dynamic optimization based on task scheduling.

【学位授予单位】：南京大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：P237

【相似文献】