Research On Multiple Task Granularities Oriented Parallel Computing Technology For Remote Sensing Image Mosaics

Remote sensing images are important data source of Geographic information system.Remote sensing images have been widely used in resource investigation,environmental monitoring,weather forecast,disaster monitoring and evaluation,and military target recognition and so on.There is an increasing need for real-time or near real-time processing of massive remote sensing images.Parallel computing technology brings an opportunity to process a large number of remote sensing images efficiently.In the research of remote sensing images parallel processing,one of the essential content is how to deal with the task dependence and data dependence under multiple task granularities.Remote sensing mosaicking is an important remote sensing processing method,and there are multiple task granularities in parallel remote sensing processing.In this paper,parallel computing technology for remote sensing images mosaics are studied with multiple task granularities to improve the processing performance of generating mosaics with massive remote sensing images,and to provide technology support for the parallelization of other remote sensing processing methods.The contents and conclusion are as follows:(1)Parallel map projection and coordinate operation of remote sensing images.Map projection and coordinate operation is the mathematical foundation for remote sensing mosaicking.The task granularity of parallel map projection and coordinate operation is a pixel and an image.The key issue in the parallelization is how to determining the size of blocks which these remote sensing images are partitioned into,so that the parallel method can be executed efficiently.Considering the environment of parallel computing and the storage format of remote sensing images,a partition scheme with the limit of memory has been propose.According to the scheme,the remote sensing images are partitioned into blocks with almost equal size and the parallel algorithm is designed based on peer-to-peer mode of static load balance.Experiments on a high performance computing(HPC)cluster with different number of processes and different data volumes show that the proposed parallel methods have significantly reduced the execution time of map projection and coordinate operation of massive remote sensing images.The parallel methods also show better performance when dealing with images of larger data volumes.(2)Parallel relative radiometric normalization of remote sensing images.The relative radiometric normalization is a vital step in remote sensing mosaics for radiometric consistence.In this paper,the method of iteratively re-weighted multivariate alteration detection and orthogonal linear regression is adopted for parallel relative radiometric normalization.As the task granularity of the parallel method is a pair of overlapping remote sensing images,there is task dependence when dealing with all remote sensing images.To parallelize the method of IR-MAD and orthogonal regression,there are two key problems:the normalization path determination and the task dependence on normalization coefficients calculation.In this paper,the reference image and normalization paths are determined based on the shortest distance algorithm to reduce normalization error.Formulas of orthogonal regression are acquired considering the effect of the normalization path to reduce the task dependence on the calculation of coefficients.A master-slave parallel mode is proposed to implement the parallel method,and a task queue and a process queue are used for dynamic task scheduling.Experiments on a HPC cluster show that the parallel RRN method provides good normalization results and favourable parallel speed-up,efficiency and scalability,which indicate that the solutions proposed for the above problems are effective.(3)Parallel image compositing of remote sensing images.Image compositing gives the rule to output a unique value for all those pixels that are presented in more than one image.In this paper,a compositing procedure based on mathematical morphology is selected as the rule of image compositing.The task granularity of the parallel image compositing method is all remote sensing images for mosaics,which brings task independence and data independence in parallelization.Also with massive remote sensing images for mosaicking,the parallel method will take lot of memory and disk space.Solutions to these problems includes:taking the overlapping areas as the subject to dealing with instead of the whole remote sensing images to decrease the occupation of memory,storing the results and essential information of remote sensing images in vector form via conversion between raster and vector to save disk space.A real-time task scheduling strategy is proposed and new tasks are added timely according to the state of task execution to reducing task dependence.Experiments on a HPC cluster show that with these solutions,there is less task dependence with more processes,which indicates a good extendibility of the parallel method.And the parallel method can greatly reduce the execution time of images compositing with good load balance between processes.In this paper,a parallel framework for remote sensing image mosaics is established.New solutions are proposed to the problem that there is task-dependence in the parallelization of relative radiometric normalization taking a pair of images as task granularity.New schemes based on data storage optimization and real-time task scheduling are proposed in the parallelization of remote sending compositing...