基于CPU和GPU协同处理的光学卫星遥感影像正射校正方法

本文系统地探讨了基于CPU和GPU协同处理的光学卫星遥感影像正射校正方法。首先使用“层次性分块”策略设计了基于CPU和GPU协同处理的正射校正方法,然后通过配置选择优化和存储层次性访问等手段进一步提高了方法执行效率。在Tesla M2050 GPU上对资源三号卫星下视全色影像进行正射校正的实验结果表明,本文方法大幅提高了光学卫星遥感影像正射校正效率,与传统串行正射校正算法相比,加速比最高达到110倍以上,相应的处理时间压缩至5s以内,可满足对大数据量光学卫星遥感影像进行快速正射校正的要求。     

对吉林一号光学卫星的应急快速处理方法研究

在CPU/GPU协同处理框架下,系统地探讨了光学遥感影像高性能处理方法。首先,实现光学卫星数据处理算法的GPU高效映射,将MTF补偿、传感器校正(包括波段配准和影像拼接)、系统几何校正等算法高效映射至GPU并行执行;在此基础上,为充分利用CPU高频化优势和加速I/O运算,利用Ramdisk技术在内存盘处理程序数据与结果输出,利用Intel C++Compiler编译优化框架。在GeForce GT 755M(GPU)、Mobile Quad Core Intel Core i7-4700MQ(CPU)的Windows系统环境下,对吉林一号卫星全色影像和多光谱影像原始数据进行0~2级产品的光学遥感影像预处理。实验结果表明,与传统预处理算法相比,此高性能算法的处理时间缩短到了40s以下,最高加速比达到11.216,可满足对海量光学卫星遥感影像数据的应急快速处理需求。     

CPU和GPU协同的多光谱影像快速波段配准方法

随着遥感影像数据量的飞速增长,传统的串行波段配准方法已无法满足大数据多光谱影像的实时配准需求。针对该问题,提出了一种CPU和GPU协同的多光谱影像快速波段配准方法。首先进行计算量和并行度分析,将同名点匹配和微分纠正映射至GPU执行,仿射变换系数拟合仍驻留在CPU执行。其次通过核函数任务映射和基本设置,使算法步骤在GPU上可执行,并设计了3种性能优化方法(访存优化、指令优化、传输计算堆叠),进一步提高了波段配准的执行效率。在NVIDIA Tesla M2050 GPU和Intel Xeon E5650 CPU组成的实验平台上,对遥感26号卫星多光谱影像的实验表明,使用该方法加速后的波段配准执行时间仅为3.25 s,与传统串行方法相比,加速比达到了32.32倍,可以满足大数据多光谱影像的近实时配准需求。     

GPU-CPU协同航空影像快速正射纠正方法

为满足应急测绘中大序列航空影像快速正射纠正的要求,提出了一种GPU-CPU协同快速正射纠正方法。实验结果表明,通过对GPU程序进行配置选择优化和存储层次性优化,该方法较传统的基于CPU正射纠正方法,其平均加速比达到52倍。     

基于GPGPU的并行影像匹配算法

提出一种基于GPGPU的CUDA架构快速影像匹配并行算法,它能够在SIMT模式下完成高性能并行计算。并行算法根据GPU的并行结构和硬件特点,采用执行配置技术、高速存储技术和全局存储技术三种加速技术,优化数据存储结构,提高数据访问效率。实验结果表明,并行算法充分利用GPU的并行处理能力,在处理1280×1024分辨率的8位灰度图像时可达到最高多处理器warp占有率,速度是基于CPU实现的7倍。CUDA在高运算强度数据处理中呈现出的实时处理能力和计算能力,为进一步加速影像匹配性能和GPU通用计算提供了新的方法和思路。     

一种面向CPU/GPU异构环境的协同并行空间插值算法

CPU/GPU异构混合系统是一种新型高性能计算平台,但现有并行空间插值算法仅依赖CPU或GPU进行加速,迫切需要研究协同并行空间插值算法以充分利用异构计算资源,进一步提升插值效率。以薄板样条函数插值为例,提出一种CPU/GPU协同并行插值算法以加速海量激光雷达（light detector & ranger,LiDAR）点云生成数字高程模型（DEM）。通过插值任务的分解与抽象封装以屏蔽底层硬件执行模式的差异性,同时在多级协同并行框架基础上设计了Greedy-SET动态调度策略,策略顾及底层硬件能力的差异性,以实现异构并行资源的充分利用和良好负载均衡。实验表明,协同并行插值算法在高性能工作站上取得19.6倍的加速比,相比单一CPU或GPU并行算法,其效率提升分别达到54%和44%,实现了高效的协同并行处理。     

基于PMVS算法的大规模数据细粒度并行优化方法

三维多视角立体视觉算法（patch-based multi-view stereo,PMVS）以其良好的三维重建效果广泛应用于数字城市等领域,但用于大规模计算时算法的执行效率低下。针对此,提出了一种细粒度并行优化方法,从任务划分和负载均衡、主系统存储和GPU存储、通信开销等3方面加以优化;同时,设计了基于面片的PMVS算法特征提取的GPU和多线程并行改造方法,实现了CPUs_GPUs多粒度协同并行。实验结果表明,基于CPU多线程策略能实现4倍加速比,基于统一计算设备架构（compute unified device architecture,CUDA）并行策略能实现最高34倍加速比,而提出的策略在CUDA并行策略的基础上实现了30%的性能提升,可以用于其他领域大数据处理中快速调度计算资源。     

自适应分块加权Wallis并行匀色

针对区域范围内多幅待镶嵌影像之间的色彩差异问题,提出一种基于GPU的分块加权Wallis并行匀色算法。首先,根据变异系数对影像自适应分块并利用双线性插值确定每一个像素的变换参数,利用加权Wallis变换消除影像间的色彩差异。然后,为了控制区域整体的匀色质量,利用Voronoi图和Dijkstra算法确定影像间的处理顺序。最后,利用GPU技术进行并行任务设计并从配置划分、存储器访问和指令吞吐量等方面进行优化,提高算法运算效率。实验结果表明,本文方法既能有效地消除影像间色彩差异,又能消除影像间的对比度差异。与CPU串行算法相比,GPU并行算法显著减少了计算时间,加速比最高达到60倍以上。     

遥感影像正射纠正的GPU-CPU协同处理研究

提出了一种基于CUDA的遥感影像正射纠正GPU-CPU协同处理方法,以实现重采样操作的GPU细粒度并行化。根据GPU的并行结构和硬件特点,采用执行配置优化技术提高warp占有率,利用共享存储器优化减少对效率低下的全局存储器中坐标变换系数的重复访问,通过纹理存储器代替全局存储器优化对原始影像数据的访问。实验结果表明,并行算法能够充分发挥GPU的并行处理能力,利用GeForce 9500 GT显卡,对大小为6 000像素×6 000像素的全色影像进行多项式纠正对比实验,最邻近灰度内插重采样和双线性灰度内插重采样的最终加速比分别能够达到8倍和10倍以上。     

基于GPU的快速影像匹配

随着图形处理器(GPU)的飞速发展和计算同一设备架构(CUDA)的推出,GPU的并行性和可编程性不断增强,本文提出了一种基于CUDA的Harris算子影像匹配并行处理方法,在GPU中完成对影像的灰度化、Harris角点提取、重采样、灰度相关匹配,并从线程分配、内存使用、共享存储器等方面进行优化。实验结果表明,该方法与CPU串行处理方法相比,其速度得到了明显提升。     

CPU/GPU near real-time preprocessing for ZY-3 satellite images: Relative radiometric correction,MTF compensation,and geocorrection

ZY-3 is the first high-accuracy civil stereo-mapping optical satellite of China. It greatly improves China’s optical satellite image resolution with a boom in data volume, calling for new challenges in processing real-time applications. On the other hand, using central processing unit (CPU)/graphic processing unit (GPU) to resolve data-intensive remote sensing problems becomes a hot issue. In this paper, we present an approach for CPU/GPU near real-time preprocessing of ZY-3 satellite images, focusing on three key processors: relative radiometric correction (RRC), modulation transfer function compensation (MTFC), and geocorrection (GC). First, basic GPU implementation issues are addressed to make the processors capable of processing with GPU. Second, three effective GPU specific optimizations are applied for further improvement of the GPU performance. Furthermore, to fully exploit the CPU’s computing horsepower within the system, a CPU/GPU workload distribution scheme is proposed, in which CPU undertakes partial computation to share the workloads of GPU. The experimental result shows that our approach achieved an overall 48.84-fold speedup ratio in ZY-3 nadir image preprocessing (the corresponding run time is 11.60 s for one image), which is capable of meeting the requirement of near real-time response to the applications that follow. In addition, with the supportability of IEEE 754–2008 floating-point standard in the Fermi type GPU, preprocessing ZY-3 images with our CPU/GPU processors could maintain the quality of image preprocess as done traditionally with CPU processors.     

A Parallel Scheme for Large‐scale Polygon Rasterization on CUDA‐enabled GPUs

This research develops a parallel scheme to adopt multiple graphics processing units (GPUs) to accelerate large‐scale polygon rasterization. Three new parallel strategies are proposed. First, a decomposition strategy considering the calculation complexity of polygons and limited GPU memory is developed to achieve balanced workloads among multiple GPUs. Second, a parallel CPU/GPU scheduling strategy is proposed to conceal the data read/write times. The CPU is engaged with data reads/writes while the GPU rasterizes the polygons in parallel. This strategy can save considerable time spent in reading and writing, further improving the parallel efficiency. Third, a strategy for utilizing the GPU's internal memory and cache is proposed to reduce the time required to access the data. The parallel boundary algebra filling (BAF) algorithm is implemented using the programming models of compute unified device architecture (CUDA), message passing interface (MPI), and open multi‐processing (OpenMP). Experimental results confirm that the implemented parallel algorithm delivers apparent acceleration when a massive dataset is addressed (50.32 GB with approximately 1.3 × 10⁸ polygons), reducing conversion time from 25.43 to 0.69 h, and obtaining a speedup ratio of 36.91. The proposed parallel strategies outperform the conventional method and can be effectively extended to a CPU‐based environment.     

CBERS-02B卫星WFI成像在轨MTF估算与图像MTF补偿

基于高分辨率图像对比法, 利用同一卫星平台上空间分辨率19.5m的CCD相机图像对CBERS-02B卫星上空间分辨率为258m的 WFI成像仪图像进行在轨MTF(modulate transfer function)测量, 获得WFI(wide field imager)相机沿轨、跨轨与45o方向的MTF曲线, 并计算出3个方向的线扩展函数LSF(line spread function), 获得3个方向的有效半带宽。结果表明WFI相机红波段跨轨、沿轨与45o方向的有效半带宽, 即有效瞬时视场, 分别为1.188, 1.165与1.281个像元, 近红外波段为1.258, 1.195与1.326个像元。基于获得的MTF, 利用维纳滤波法对WFI图像进行补偿, 部分恢复了WFI图像的细部信息。     

渤海近岸水体漫衰减系数Kd(490)遥感反演模型

Irrigated land is one of the most important parts in land cover classification system, but until now it has seldom been reported about extracting it from remote sensing data and monitoring its dynamic change. The existing researches are focused on the land use-cover change (LUCC) and the irrigated area mapping, so the study on extracting irrigated land is much more important. On the basis of close relationship between water deficit index (WDI) and soil water content, this paper firstly combines two neighbor NDVI and the surface-air temperature (Ts-Ta) and conforms one Vegetation Index Temperature (VIT). It makes the timely incomparable WDI comparable and then computes the change of WDI during the monitored period. Secondly it infers the change of soil water. Thirdly it removes the rainfall’s influence under some assumption and extracts the irrigated land of the survey area. Results show that comparing with census data all the deviation is below 7% in the survey area except Shanxi province, which means that the extracted results are comparable with census data. Extracted irrigated land mainly distributes around rivers, lakes, and reservoirs or on the irrigated regions and oasises. The results are consistent with the centralized region which has abundant irrigated land known before. The results are checked elementarily using TM images. All the precisions are above 70% except in Shanxi province. The precision in Xinjiang municipality is the highest.     

FY-3微波成像仪遥感图像地理定位方法研究

MWRI (MicroWave Radiation Imager) is one of the payloads on our next generation polar meteorological satellite FY-3. MWRI conically scans with a fixed incident angle on the earth surface. It is the first time for Chinese remote sensor to use this scan mode. In this work, we present a geolocation method for FY-3 MWRI’s remote sensing image based on its special scan geometry. The integrated coordinate systems and the specific relationships with these coordinate systems are defined. A spatial relationship model between the remote sensing data and the earth-based coordinate system is established. This method also includes an algorithm of satellite orbit computation, which is used to get the satellite’s instantaneous velocity vector from its position. This method has been applied to MWRI’s remote sensing image geolocation. The results show that the accuracy of this method can achieve 1 pixel. The 33 GCPs (Ground Control Points) which are in the regiones of FY-3 MWRI’s observation have been collected and used to analyze the precision of the geolocation. By statistical analysis, the error along-track is about 1.5km, and the error along-scan is about 3.0km. It is obvious that this method fulfills the requirement of precision for FY-3 MWRI whose space resolution exceeds 5km.     

An efficient data organization and scheduling strategy for accelerating large vector data rendering

Rendering large volumes of vector data is computationally intensive and therefore time consuming, leading to lower efficiency and poorer interactive experience. Graphics processing units (GPUs) are powerful tools in data parallel processing but lie idle most of the time. In this study, we propose an approach to improve the performance of vector data rendering by using the parallel computing capability of many‐core GPUs. Vertex transformation, largely a mathematical calculation that does not require communication with the host storage device, is a time‐consuming procedure because all coordinates of each vector feature need to be transformed to screen vertices. Use of a GPU enables optimization of a general‐purpose mathematical calculation, enabling the procedure to be executed in parallel on a many‐core GPU and optimized effectively. This study mainly focuses on: (1) an organization and storage strategy for vector data based on equal pitch alignment, which can adapt to the GPU's calculating characteristics; (2) a paging‐coalescing transfer and memory access strategy for vector data between the CPU and the GPU; and (3) a balancing allocation strategy to take full advantage of all processing cores of the GPU. Experimental results demonstrate that the approach proposed can significantly improve the efficiency of vector data rendering.