分块变换和GPU并行的遥感影像快速正射校正方法

A Fast Remote Sensing Image Ortho-rectification Method Based on Block-wise Transformation and GPU Parallel Processing

正射校正是整个遥感数据处理过程中计算量最大、耗时最长的步骤之一,已经成为制约整个遥感数据处理快速完成的瓶颈。为了提高正射校正处理效率,文章系统地探讨了基于分块三维直接线性变换和图形处理单元(GPU)并行的遥感影像快速正射校正方法。首先针对正射校正坐标转换计算量过大的问题,提出了分块三维直接线性变换策略,有效地降低了坐标转换的计算量;在此基础上,采用"渐进式"策略开展GPU并行处理,首先通过GPU并行映射(核函数任务映射、基本设置),使方法在GPU上可执行,然后通过"两层次"性能优化(核函数性能优化、整体流程性能优化),进一步提高了方法的执行效率。在CPU和GPU组成的实验环境中,使用文中方法对"高分二号"卫星全色标准景影像进行实验,GPU执行时间仅为5.13s,与CPU相比,相应加速比达到142.42倍,可以满足对大数据遥感影像的快速正射校正需求。

Ortho-rectification is one of the most computationally and time-consuming steps in remote sensing data processing,and has become a bottleneck restricting the rapid completion of remote sensing data processing.Therefore,a fast ortho-rectification method,based on block-wise three-dimensional linear transformation and GPU parallel processing,is proposed in this paper.First,a three-dimensional linear transformation is introduced to reduce the computational amount of coordinate projection conversion.Then,based on progressive strategy,the GPU parallel mapping(including kernel mapping and basic settings)is conducted.Furthermore,two performance optimization methods,i.e.,kernel performance optimization and overall performance optimization,are used to further improve the efficiency.The approach is tested with GF-2 panchromatic images with CPU and GPU.The experimental results show that the processing time is only5.13s,with a speedup ratio of 142.42,which can meet the requirement of fast ortho-rectification for big-data-volume remote sensing images.