面向汛旱情监测的遥感影像GPU并行处理算法

GPU-based parallel image processing algorithm for flood and drought monitoring

针对面向汛旱情监测应用中遥感影像处理耗时过长的问题,包括辐射校正、几何纠正、遥感指数计算等过程,对其业务化工作流程进行了分解分析。结合统一计算架构(compute unified device architecture,CUDA)的存储结构和程序设计模型,将数据处理过程划分为数据读取、直方图统计、栅格分割、波段计算、重采样和数据输出等模块,对波段计算及重采样等模块设计了并行处理方案,并通过实验确定了栅格划分的最佳尺度,基于栅格数组图形处理器(graphics processing unit,GPU)映射方法加速了数据传输效率,最终提出了基于CUDA架构CPU-GPU协同的并行处理算法。实验结果表明,辐射校正及遥感指数计算的波段计算模块可节约58.9%的时间;几何纠正效果最为显著,最邻近像元重采样和双线性内插重采样模块的最终加速比分别能够达到9倍和7倍以上。

Aiming at the time-consuming problems in the remote sensing image processing for flood and drought monitoring,the authors analyzed related workflows and algorithms including radiometric correction,geometric correction,and the calculation of remote sensing indices.Based on the storage structure and program design model of compute unified device architecture(CUDA),the remote sensing image processing was divided into several modules,including data reading,histogram statistics,grid partition,band calculation,resampling,and data output.Among them,parallel processing schemes were designed for the modules of band calculation and resampling,and the optimal cell sizes were determined for the module of grid partition.Meanwhile,the data transfer efficiency was increased through the grid data mapping based on a graphics processing unit(GPU).Finally,a parallel processing algorithm based on CPU-GPU collaboration in CUDA was proposed.The experiment results are as follows.The modules of radiometric correction and band calculation of remote sensing indices showed a 58.9%saving in time.Meanwhile,the geometric correction module enjoyed the most significant time-saving effects,and the final speedup ratios of the resampling methods of nearest neighbor and bilinear interpolation reached up to nine and seven times,respectively.