基于有理多项式模型区域网平差的GF-1影像几何校正

2013年4月成功发射的GF-1卫星是中国高分系列卫星的首发星,影像在中国农情遥感监测业务中得到了广泛应用,已成为大宗农作物种植面积遥感监测的主要数据源之一。高精度几何位置的配准是卫星农情定量化应用的基础与前提,该文提出了一种基于区域网平差方法修正GF-1卫星WFV(wide field view,WFV)影像RPC(rational polynomial coefficients,RPC)参数,获取更高几何定位精度的校正方法,形成了模式化的业务处理流程,为该影像在农情遥感监测中的应用奠定了基础。算法流程包括2个部分,首先是基于像面间仿射变换关系及有理多项式RFM(rational function model,RFM)模型构建轨道间的区域网平差数学模型,其次是根据影像连接点及少量控制点输入求解所有参与平差的卫星影像定向参数,获取亚像元级的校正结果。平差参数的解算是通过两步求解完成的,初始平差参数是根据连接点及对应的DEM高程值进行平差迭代至收敛,结果平差参数是将初始平差参数作为初始值代入区域网平差模型,并以逐点消元方式约化法方程,解算出各影像的仿射变换参数。该文在求解平差参数过程中,直接使用DEM(digital elevation model)上获取的高程值作为约束条件,消除了平面坐标与高程的相关性,保证了区域网平差模型能够解算。混合地形、平原、山区3种情况下区域网平差结果表明,全连接点平差结果具有较高的相对定位精度,其行方向的中误差分别为0.3046、0.4674、0.3365像元,列方向的中误差分别为0.3677、0.2849、0.2889像元;而结合少量控制点的区域网平差则同时具有很高的绝对定位精度,其行方向的中误差分别为0.3648、0.5041、0.3605像元,列方向的中误差分别为0.4954、0.4039、0.6323像元,整体达到了亚像素级。最后,在农业应用基础控制底图的支持下,分别对原始影像、RPC校正影像、区域网平差后的影像进行几何配准,分析不同输入影像条件下的几何校正精度,仅有区域网平差后的影像达到了亚像元的校正精度,混合地形、平原、山区3种情况下行方向的中误差分别为0.6857、0.6664、1.0646像元,列方向的均方差分别为0.4342、0.4696、0.5609像元,但与几何校正前精度相比没有明显改善,说明本文提出的研究方法可以实现少量控制点条件下的几何精校正。不同DEM校正结果表明,对于山区,更高分辨率的DEM可以获得更好的定位精度。上述研究充分表明,该方法对GF-1/WFV数据的处理有效且可行,并在农业部中国农情遥感业务工作中得到了初步应用。

Geometric correction of GF-1 satellite images based on block adjustment of rational polynomial model

Abstract: GF-1 satellite was launched successfully in April 2013, and its wide field view (WFV) images have been widely applied in agricultural remote sensing monitoring practice in China. To obtain high precision of image positioning, the paper proposes a correction method for acquiring higher geometric positioning precision based on block adjustment method to correct rational polynomial coefficients (RPC) of high resolution WFV. A business process including adjustment model construction, adjustment parameter calculation and geometric registration based on reference images has been formed. Firstly, affine transformation relation among images, namely, a block adjustment model, has been built based on the relationship between the RPC parameter image points and ground points; secondly, initial value of connection points is identified, and combined with a few ground control points, the affine transformation coefficients of various images are calculated so as to conduct the image block adjustment. Meanwhile, orthoimage is produced based on adjustment results and DEM (digital elevation model), and the correction result with sub-pixel accuracy is achieved. The calculation of adjustment parameters is achieved through 2 steps. The first step is to identify the initial value of the connection point. By utilizing one degree term of RPC model, the plane coordinates of the connection points are iterated and updated, till they converge to a stable state. The elevation values are extracted from DEM; the second is to identify affine transformation parameters. The updated connection points with a few ground control points are entered into block adjustment model to establish error equation. The elevation values acquired from DEM are taken as a constraint condition, and the relevance between plane coordinates and elevations is eliminated to ensure that the block adjustment model works. Meanwhile, the unknown variable is calculated by using the point-by-point elimination method. The block adjustment results under 3 different conditions of mixed terrain, plain area and mountainous area show that, the adjustment results of whole connection points have a relative higher positioning precision, with the errors of 0.3046, 0.4674 and 0.3365 pixels respectively at the row direction, and with the errors of 0.3677, 0.2849 and 0.2889 pixels respectively at the column direction; block adjustments with a few control points have very high absolute positioning precision, with the errors of 0.3648, 0.5041 and 0.3605 pixels respectively at the row direction, and with the errors of 0.4954, 0.4039 and 0.6323 pixels respectively at the column direction. Finally, under the support of base control map of agriculture application, the geometric registration of raw images, RPC correction images, and images after block adjustment is conducted, and the geometric correction precision under different input image conditions is analyzed. Only images after block adjustment have reached correction precision of sub-pixel. The errors at the row direction under 3 conditions of mixed terrain, plain area and mountainous area are 0.6857, 0.6664 and 1.0646 pixels respectively, and those at the column direction are 0.4342, 0.4696 and 0.5609 pixels respectively, indicating that the research method proposed by this paper can achieve accurate geometric correction under the condition of a few control points, though there is no significant improvement compared with the precision before geometric correction. After comparing the DEM with different resolutions in the model, we find that the precision of DEM affects the correction result. Applying higher resolution in mountainous areas can achieve better positioning precision. The above results show that this method can effectively improve the geometric correction precision of WFV images of CF-1 satellite, and it has been preliminarily applied in the operation of agriculture remote sensing monitoring.