| With the development of remote sensing technology,the remote sensing images obtained by human beings have higher spatial resolution and spectral resolution,and the data volume of remote sensing image also increases exponentially.However,a single sensor which reflects the same scene can’t simultaneously obtain remote sensing images with high spatial resolution and hyperspectral resolution.Therefore,it is of great scientific value to apply fusion of multi-source remote sensing images to achieve this purpose.At the same time,due to the large volume of remote sensing image and the complicated task flow of the fusion algorithm,the traditional single-machine processing mode is not applicable to the processing of remote sensing data with continued accretion of the size of the data.Nevertheless,cloud computing platform can effectively solve the bottleneck problem of single machine,with its powerful storage capacity and efficient distributed parallel computing ability.In this paper,the Pan-Sharpening fusion method based on BP neural network for processing multispectral remote sensing images is deeply studied,and the distributed parallel algorithm of Pan-Sharpening method is designed and implemented on the platform of TensorFlowOnSpark,In view of the complexity among tasks of the fusion algorithm,a task scheduling strategy for multi-spectral image fusion method on cloud computing platform is proposed,which uses the strategy of particle swarm optimization.This method improved the computational efficiency of fusion algorithm by optimizing the assignment of tasks on each node.The main contents of this paper are as follows:1.A Pan-Sharpening method of multi-spectral remote sensing image based on BP neural network is deeply studied.This method utilizes BP neural network to model the complex relationship between high-resolution remote sensing images and low-resolution remote sensing images.Then,we use a trained network to forward low-resolution multispectral images to generate high-resolution multispectral images.The experimental result shows that the fusion image generated by this method is of high quality which ensures the high spatial resolution of the multispectral image as well as avoids the spectral distortion.2.Aiming at the operation efficiency of multi-spectral remote sensing image Pan-Sharpening method,a parallel BP neural network optimization method based on cloud platform and TensorFlowOnSpark is designed and implemented in this paper.The method combined with Spark platform and the platform of TensorFlowOnSpark to optimize BP neural network in parallel.Large-scale remote sensing data is stored effectively through HDFS distributed file system.It also makes full use of Spark memory-based features to improve the efficiency of the algorithm,which greatly reduces the training time of the neural network while ensuring the accuracy of the algorithm3.In order to further improve the performance of the distributed multi-spectral image fusion algorithm,this paper also proposes a task scheduling optimization method based on PSO(particle swarm optimization),which optimizes the task scheduling for Pan-Sharpening method on cloud platform.On the basis of the parallel optimization of Pan-Sharpening mentioned above,this method analyzes the task flow of the algorithm firstly.Then,the mutual restraint relationship among tasks is described by a directed acyclic graph.Finally,each task is assigned to appropriate virtual machine according to the PSO algorithm.The running time of the fusion algorithm is significantly reduced by the optimization of task allocation with the task scheduling method.4.Based on the research above,a multi-spectral remote sensing image fusion system is designed and implemented.The system framework and function module design are proposed,including user management,image input,parameter input,display of fused images,etc.The system also performs the corresponding fusion accuracy and the performance analysis of task scheduling algorithm. |
PDF Full Text Request