摘要：
近年来随着遥感数据的爆炸性增长,快速、稳定的自动化影像正射校正成为遥感大数据处理的重要环节.该文在分析GF-1遥感大数据组织方式与元数据特征的基础上,将有理多项式模型正反变换与数字高程数据提取结合,设计实现自动化正射校正系统,并以提高正射校正计算效率与稳定性为目标,研究待校正影像对应数字高程数据快速提取方法,待校正影像分块读取策略等关键问题.在此基础上针对20景覆盖不同地形区域GF-18 m多光谱正射校正影像选择均匀分布的检查点,以Google Earth影像中同名点坐标为真值,分析校正误差及收敛情况,试验结果X(纬线方向)方向和Y(经线方向)方向最大误差均小于16.863 m,距离误差小于23 m,并且92.25%的检查点误差小于16 m(2个像元).该文提出的自动化正射校正方案在山地地形与平原地形均表现出良好的校正精度与稳定性.%Remote-sensing imaging is a complicated process. It is influenced by many factors like optical distortion, sensor attitude change, satellite platform movement, curvature change of the earth, terrain fluctuation, and so on, which cause geometric distortion such as excursion, extension, and shrink compared with location of real ground objects, and the extent of distortion turns severe with the increase of distance between pixel and sub-satellite point. Therefore in practical application, correcting geometric distortion caused by terrain fluctuation or satellite platform, becomes one of the basic works. We analyzed data organizing mode and metadata structure of GF-1 satellite image data, and on that basis RPC (rational polynomial coefficient) model-based forward and inverse transformation was combined with the DEM (digital elevation model) data extraction; the process of RPC model-based images orthorectification was elaborately calculated; automatic orthorectification system (GF1AMORS) were designed and implemented, which could fit 2, 8 and 16 m resolution images. There are the critical questions: 1) Method for DEM data rapid extraction; 2) Strategy for image blocking. Firstly, DEM data were reorganized and coded based on 0.5°×0.5° geographic grid system in order that DEM data could be read to system memory rapidly according to coordinate range of image being rectified, and relevant test showed that our grid-based DEM data dynamical extraction method could achieve good efficiency with different image range, while the system memory might overflow when the image range turned larger than 3.5°×3.5°. Secondly, comparative experiments were done to study the relation between image block size and orthographic correction efficiency of GF-1 WFV (wide field of view, 16 m resolution) multi-spectral images. Experiments showed that the computational efficiency of single image converged to 98 s when the block size was set as from 384×384 to 480×480. To test the conversion accuracy of our automatic orthorectification process, 20 GF-1 PMS (Pansharpen/Multispectral Sensor, 8 m resolution) multi-spectral images that covered area with different terrain features (mountainous and plain terrain) in Heilongjiang Province with less cloud were extracted, and on that basis 400 control points were selected and compared to their homonymy points selected in Google Earth (by ENVI "SPEAR Google Earth Bridge") to analyze error and convergence. The experiment showed that our automatic orthorectification process exhibited a nice accuracy and stability in both mountainous terrain and plain terrain: For mountainous terrain, the maximum error inXorientation was less than 16.863 m and inY orientation was less than 16.811 m, and the standard deviation inX orientation was less than 5.514 m while that inY orientation was between 2.872 and 4.336 m; for plain terrain, the maximum error inX orientation was less than 10.959 m and in Y orientation was less than 13.546 m, and the standard deviation inX orientation was less than 3.051 m while that inY orientation was less than 3.761 m. The maximum distance error was 23 m, and the distance error of 92.25% control points was less than 16 m (namely 2 pixels), and that of 38.75% control points was less than 8 m (namely 1 pixel). At last, we presented the limitation and our future work about our automatic orthorectification method. Considering that there is no physical significance for each parameter of RPC model, the calibration precision of our system needs to be improved (e.g. by integrating control points to weaken the system error) before it is used in some applications with higher accuracy requirement; furthermore, there is still a large optimization space of computational efficiency for our system, and high performance computing method (e.g. graphics processing unit) will be integrated based on data block feature to improve the calculating speed.
