基于随机森林和多层感知机集成算法的大气可降水量建模方法  

Atmospheric Precipitable Water Modeling Method Based on IntegrationAlgorithm of Random Forest and Multilayer Perceptron

作　　者：李锴 李黎 余航 董州楠 何琦敏 LI Kai;LI Li;YU Hang;DONG Zhounan;HE Qimin(School of Geography Science and Geomatics Engineering,Suzhou University of Science and Technology,Suzhou 215009,China;Research Center of Beidou Navigation and Environmental Remote Sensing,Suzhou University of Science and Technology,Suzhou 215009,China)

机构地区：[1]苏州科技大学地理科学与测绘工程学院,江苏省苏州市215009 [2]苏州科技大学北斗导航与环境感知研究中心,江苏省苏州市215009

出　　处：《大地测量与地球动力学》2025年第5期518-525,共8页Journal of Geodesy and Geodynamics

基　　金：江苏省研究生科研与实践创新计划(SJCX23_1718,SJCX24_1901);江苏省科技计划(BK20230660);国家自然科学基金(42204014)。

摘　　要：基于2009-2019年中国地区56个探空站资料,分析大气可降水量(PWV)与天顶对流层延迟(ZTD)、水汽压(E_s)、大气压(P_s)、地面温度(T_s)、加权平均温度(T_m)之间的相关性;然后基于多元线性回归法(MLR)、随机森林(RF)、多层感知机(MLP)和MLP-RF集成算法分别构建中国地区的PWV模型,并评估各类PWV模型精度。结果表明,PWV与T_s、P_s、E_s、T_m和ZTD之间的相关系数分别为0.78、-0.71、0.77、0.74和0.99;基于MLR、RF、MLP与MLP-RF的全年PWV模型的平均RMSE分别为1.25 mm、0.44 mm、0.28 mm和0.21 mm,平均MAE分别为0.98 mm、0.31 mm、0.18 mm和0.14 mm;基于MLR、RF、MLP与MLP-RF的分季度PWV模型的平均RMSE分别为0.60 mm、0.45 mm、0.29 mm和0.21 mm,平均绝对偏差分别为0.49 mm、0.33 mm、0.19 mm和0.14 mm;基于MLP-RF的全年PWV模型精度最高,可用于中国地区PWV的获取及研究。Based on the data of 56 sounding stations in China from 2009 to 2019,we analyze the correlation between precipitable water vapor(PWV)and zenith tropospheric delay(ZTD),water vapor pressure(E s),atmospheric pressure(P s),surface temperature(T s),and weighted mean temperature(T m),and then construct the China PWV models using multiple linear regression(MLR),random forest(RF),multilayer perceptron(MLP),and MLP-RF integration algorithm.The accuracy of these models are evaluated.The results show that the correlation coefficients between PWV and T s,P s,E s,T m and ZTD are 0.78,-0.71,0.77,0.74 and 0.99,respectively.The average RMSE of yearly PWV models based on MLR,RF,MLP and MLP-RF are 1.25 mm,0.44 mm,0.28 mm,and 0.21 mm,and the average MAE are 0.98 mm,0.31 mm,0.18 mm,and 0.14 mm,respectively.The average RMSE of seasonal MLR,RF,MLP,and MLP-RF PWV models are 0.60 mm,0.45 mm,0.29 mm,and 0.21 mm,and the MAE are 0.49 mm,0.33 mm,0.19 mm,and 0.14 mm,respectively.The yearly MLP-RF PWV model is with the highest accuracy,which is suitable for real time PWV acquisition and research in China region.

关 键 词：大气可降水量 随机森林 多层感知机 多元线性回归 

分 类 号：P228[天文地球—大地测量学与测量工程]