基于集成学习算法的全国降水量分布制图研究  

Study on Predictive Mapping of Spatial Distribution of Precipitation in China Using Ensemble Learning Algorithm

作　　者：王妍[1] 赵文旻[1] 虞亚楠 李林 任杰 赵玲妹 李珈慧 WANG Yan;ZHAO Wen-min;YU Yan-nan;LI Lin;REN Jie;ZHAO Ling-mei;LI Jia-hui(Shanghai College of Agriculture and Forestry Vocational Technology,Shanghai 201699,China)

机构地区：[1]上海农林职业技术学院,上海201699

出　　处：《中国农村水利水电》2025年第3期55-60,70,共7页China Rural Water and Hydropower

基　　金：上海市科委“科技创新行动计划”农业科技领域项目(20392001300)。

摘　　要：针对传统空间插值模型对样本数据的依赖性及其预测偏差的缺陷,提出一种新的降水量空间分布数字制图方法——集成学习(Ensemble Learning, EL)算法。基于降水量分布随多元地理环境因素(地理位置、地表覆盖、地形特征)变化的假设,以中国地区2019年的618个气象站点年降水量观测资料为样本数据,建立基于EL降水量空间制图模型,该EL模型以广义线性回归(Generalized Linear Regression, GLM)模型为元学习机来整合样条函数(Anusplin)、地理加权(Geographically Weighted Regression, GWR)和高斯过程(Gaussian Process Regression,GPR)模型产生的初级预测,最后制取全国1 km空间分辨率的降水量栅格面。结果显示,EL模型取得较可靠的预测结果,模型验证精度决定系数(Determination Coefficient, R2)达0.96,均方根误差(Root Mean Square Error,RMSE)仅为55.17 mm;EL模型性能优于其他模型,比传统的GPR、GWR和Anusplin模型的RMSE分别降低了10.95%、16.54%、18.02%。本文提出的EL模型在大尺度范围的站点式气候要素空间制图领域中显示出良好的应用潜力。In view of the defects of traditional spatial interpolation model sample’s dependence on sample data and prediction uncertainty,this paper proposes a new digital mapping method for spatial distribution of precipitation,ensemble learning(EL)algorithm.Based on the relationship between precipitation and multiple geospatial factors(geographic location,vegetation cover,topographic features),an EL-based precipitation spatial mapping model is established with the sample data of 618 meteorological station observations in 2019 in China.The EL model uses the generalized linear regression(GLM)model as a meta-learning machine to integrate the primary predictions generated by Anusplin,geographically weighted regression(GWR),and Gaussian process regression(GPR)models,and then generates precipitation grid point data at 1 km resolution for the whole country.Results show that the EL model achieves the best prediction with an determination coefficient(R2)of 0.96 and an Root mean square error(RMSE)of 55.17 mm,which is 10.95%,16.54%,and 18.02%lower than the RMSEs of the single model of GPR,GWR,and Anusplin,respectively.The EL model proposed in this paper shows potential applications in large-scale site-based spatial distribution mapping of precipitation.

关 键 词：集成学习 单一模型 降水量 空间制图 

分 类 号：TV125[水利工程—水文学及水资源] P426.6[天文地球—大气科学及气象学]