高分辨率遥感影像位置精度建模的机器学习方法  

Machine Learning Methods for Positional Accuracy Modeling in High Resolution Remote-Sensing Images

作　　者：杨蕊宁 张景雄[1] 贺超 辛亮 王斯健 YANG Ruining;ZHANG Jingxiong;HE Chao;XIN Liang;WANG S ijian(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China;Shanghai Municipal Institute of Surveying and Mapping,Shanghai 200063,China)

机构地区：[1]武汉大学测绘学院,湖北武汉430079 [2]上海市测绘院,上海200063

出　　处：《测绘地理信息》2023年第4期11-15,共5页Journal of Geomatics

基　　金：国家自然科学基金(41471375);上海多源卫星遥感数据高精度几何校正研究项目(Z20-6003)。

摘　　要：提出了利用机器学习对高分辨率影像的位置精度进行建模分析的方法。实验提取上海市资源三号遥感影像中心城区与郊区空间纹理统计特征,用作解释变量,构建位置误差回归模型,对比分析不同机器学习模型之间拟合差异,并进行实验验证。结果表明,支持向量回归模型整体拟合精度最优,其在两块实验区域拟合均方根误差(rootmean squares error,RMSE)均在0.9 m左右,随机森林(random forest,RF)模型在郊区也有较好拟合结果。与常规方法相比,机器学习方法更有效,平均绝对误差(mean absolute error,MAE)相差较小,预测结果对异常值较小的样本可信度更高。所提出方法可通过模型预测影像逐像素位置误差,为精细地物信息提取和精度控制提供技术支持。This paper proposes exploring machine learning methods for modeling and analyzing positional accuracy of high-resolution images.In this research,certain spatial texture statistical characteristics of the central urban and suburban areas in Resources satellite-3 satellite image subsets over Shanghai are extracted,and used as explanatory variables for regression modeling of positional error.We compare and analyze model fitting by different machine learning models,based on experimental verification.The results show that the SVR model performs the best,with RMSEs in both experimental areas being about 0.9 m.The RF model performs better in the suburbs.Compared with conventional methods,machine learning methods are more effective,with smaller MAE and more reliable results for samples containing smaller outliers.The proposed method can predict per-pixel position errors and will provide technical support for the needs of fine-resolution information and quality control.

关 键 词：高分辨率遥感卫星影像 位置精度 多元线性回归 支持向量回归 随机森林 

分 类 号：P237[天文地球—摄影测量与遥感]