面向复杂地形区气温场拟合的回归学习方法精度比较  

Accuracy Comparison of Regression Learning Methods for Temperature Field Fitting in Complex Terrain Area

作　　者：孙行 黄泽纯 SUN Hang;HUANG Zechun(Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学地球科学与环境工程学院,四川成都611756

出　　处：《测绘与空间地理信息》2022年第5期18-23,共6页Geomatics & Spatial Information Technology

基　　金：中央高校基本科研业务费专项资金(2682014CX017)资助。

摘　　要：针对复杂地形区域气温具有非线性变化特征,常规函数模型难以准确构建气温场进行气温预测的问题,本文利用机器学习处理非线性问题的优势,比较分析支持向量回归、径向基神经网络回归和k近邻回归3种机器学习方法气温场拟合的模型精度。首先,以中国西部六省气温监测资料为基础,运用普通克里金插值增强样本容量,并划分训练数据集和测试数据集。然后,利用训练数据集训练3种气温场回归学习模型,根据测试数据集训练得到的模型获得气温预测值。最后,从气温场三维表面、气温偏差统计特征、误差指标3个方面比较分析了气温场拟合的模型精度。实验结果表明,3种回归学习方法的精度都非常高;模型精度从高到低的方法依次为加权k近邻回归、支持向量回归和径向基神经网络回归;在气温变化细节精细建模方面径向基神经网络回归更具优势。研究结果可为复杂地形区高精度气温预测提供参考。Due to the non-linear characteristics of temperature in complex terrain,the conventional function models are difficult to accurately construct the temperature field for temperature prediction.In view of the advantages of machine learning in dealing with non-linear problems,the purpose of this paper is to compare and analyze the model accuracy of three machine learning methods,namely support vector regression(SVR),radial basis function neural network regression(RBFNNR)and k-nearest neighbor regression(kNNR).First,based on the temperature monitoring data of six provinces in Western China,the sample size was enhanced by ordinary kriging interpolation,and these samples were divided into the training data set and test data set.Then,three kinds of regression learning models of temperature field were trained using training data set,and temperature prediction values were obtained according to the trained model using test data set.Finally,the accuracy of three regression models was compared and analyzed from three aspects:three-dimensional surface of temperature fields,statistical characteristics of temperature deviation and error indexes.The experimental results show that the accuracy of the three regression learning methods is very high;the methods of model accuracy from high to low are weighted kNNR,SVR and RBFNNR;the RBFNNR method has more advantages in fine modeling of temperature change details.The results could provide a reference for the selection of methods for high-accuracy temperature prediction.

关 键 词：机器学习 气温场拟合 精度比较 气温预测 非线性回归 

分 类 号：P237[天文地球—摄影测量与遥感]