融合多种机器学习模型的2 m气温空间降尺度方法  

Spatial downscaling methods for the 2-meter air temperature grid data based on multiple machine learning models

作　　者：李显风[1] 袁正国[1] 邓卫华[1] 杨立苑 周雪莹[1] 胡丽丽 LI Xianfeng;YUAN Zhengguo;DENG Weihua;YANG Liyuan;ZHOU Xueying;HU Lili(Jiangxi Meteorological Information Center,Nanchang 330096,China)

机构地区：[1]江西省气象信息中心,南昌330096

出　　处：《自然资源遥感》2023年第1期57-65,共9页Remote Sensing for Natural Resources

基　　金：江西省重点研发计划项目“江西省高时空分辨率多源降水融合技术研究与网格化产品研制”(编号:S2020ZPYFB0099);江西省03专项及5G项目“江西省大数据智慧气象服务示范应用平台”(编号:20212ABC03W02)共同资助。

摘　　要：高分辨率气象资料是精细化气象业务服务的重要数据基础,文章利用2020年1月—2021年3月逐小时的2 m气温网格数据,选取海拔、经度、纬度等地形因子,综合应用LightGBM(LGB)、XGBoost(XGB)、梯度提升树(gradient boosting tree,GBT)和随机森林(random forest,RF)4种机器学习方法,实现1 km分辨率的2 m气温网格数据降尺度至100 m,并对4种机器学习降尺度结果进行加权融合。将不同模型降尺度结果与双线性插值结果对比,结果表明:各降尺度模型结果与站点观测值较为一致,LGB,XGB和RF模型与双线性插值降尺度结果空间结构相似,但更为精细;各降尺度模型具有相同的时空误差分布特征,与双线性插值结果相比,LGB,XGB和GBT的数据精度均有明显提高,均方根误差(root mean square error,RMSE)分别降低了5.2%,4.1%和4.6%,而加权融合后的RMSE降低了5.9%,优于单一机器学习模型;LGB,XGB和GBT模型对不同地形条件下的降尺度结果均具有一定改善,尤其对高海拔地区(海拔在600 m以上)的改进效果更为显著,LGB,XGB和BGT和融合模型的相关系数分别提高了0.45%,0.40%,0.63%和0.66%,RMSE分别降低了9.1%,8.0%,12.7%和13.1%。研究显示,多种机器学习加权融合的降尺度模型兼顾了提升空间分辨率和保持数据精度两方面的要求,适用于研究区2 m气温数据的降尺度研究,为研制高分辨率数据产品具有一定参考意义。High-resolution meteorological data serve as an important data basis for fine-scale meteorological services.Using the hourly 2-meter air temperature grid data from January 2020 to March 2021 and the terrain factors such as altitude,longitude,and latitude,this study aimed to enhance the resolution of 2-meter air temperature grid data with a resolution of 1 km to 100 m through downscaling based on four machine learning methods,namely LightGBM(LGB),XGBoost(XGB),gradient boosting tree(GBT),and random forest(RF).Then,this study conducted the weighted fusion of downscaling results of different models.Finally,the downscaling results of different models were compared with the bilinear interpolation results,and the results are as follows.The results of each downscaling model were relatively consistent with the observational data.Compared with the bilinear interpolation results,the results of the LGB,XGB,and RF models had similar spatial structures but were more detailed.All downscaling models yielded the same spatio-temporal distribution characteristics of errors.Compared with the bilinear interpolation results,the data of the LGB,XGB,and GBT models showed significantly higher precision,and their root mean square errors(RMSEs)decreased by 5.2%,4.1%,and 4.6%,respectively.Meanwhile,the RMSE after weighted fusion decreased by 5.9%,which was higher than that of any single machine learning model.The downscaling results of the LGB,XGB,and GBT models were improved to a certain degree compared with the bilinear interpolation results under different terrain conditions,especially in high-altitude areas(above 600 m).The correlation coefficients of results of the LGB,XGB,and BGT models and model based on weighted fusion increased by 0.45%,0.40%,0.63%,and 0.66%,respectively,and their RMSEs decreased by 9.1%,8.0%,12.7%,and 13.1%,respectively.These results indicate that the downscaling model based on the weighted fusion of different machine learning methods can both improve spatial resolution and maintain data precision and,thus,is suit

关 键 词：实况网格产品 2 m气温 降尺度 机器学习 加权融合 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]