一种光学卫星无控定位误差智能建模方法

An Intelligent Modeling Method for Uncontrolled Positioning Error of Optical Satellite

光学卫星的无控定位精度是决定影像应用效果的重要因素;研究表明,影响光学卫星无控定位精度的主要因素包括姿态测量随机误差、时间同步误差、结构变形等引起的姿态低频误差等,影响因素多,难解耦,传统通过控制点评估来建立无控定位误差模型的方法难以客观全面地揭示误差规律;为了更准确地建立无控定位误差变化规律模型,该文将卷积神经网络引入无控定位精度建模,以卫星成像参数和业务系统全自动几何质检结果作为学习样本,利用网络训练无控定位精度与成像参数的关系,更全面揭示无控定位误差规律,并通过预测定位误差来提升无控定位精度;试验中选取了10019景珞珈一号01星数据,采用7514景影像作为无控定位误差变化规律的训练样本集,剩余的数据开展无控定位误差预测补偿精度验证;结果表明,模型预测精度小于1个像素,验证了该文方案的有效性和可行性。

Uncontrolled positioning accuracy of optical satellite is an important factor to determine the effect of image application.The research shows that the main factors of image optical satellite uncontrolled positioning accuracy include attitude measurement random error,time synchronization error,attitude low-frequency error caused by structure deformation,etc.There are many image factors which is difficult to decouple.The traditional method of establishing uncontrolled positioning error model through control point evaluation is difficult to objectively and comprehensively reveal the error law.In order to establish the uncontrolled positioning error model of variety rule accurately.The convolutional neural network is introduced into the modeling of uncontrolled positioning accuracy.The satellite image parameters more comprehensively reveal the law of the uncontrolled positioning error,and improve the uncontrolled positioning accuracy by predicting positioning error.Finally,the effectiveness and feasibility of this method are verified by using the 10019 images data of Luo-Jia01 satellite.By using the 7514 scene images,the training sample set is taken as the uncontrolled positioning error,the compensation accuracy for the uncontrolled positioning error is verified through the remaining data.The results show that the prediction accuracy for the model is less than 1 pixel,the effectiveness and feasibility of this paper are verified.