多种特征因子结合GBDT的降水数据降尺度方法研究

Downscaling method of precipitation data based on GBDT combined with multiple eigenfactors

结合增强型植被指数(EVI)、大气可降水量(PWV)、地表温度(LST)、坡度(Slope)、坡向(Aspect)、海拔(Altitude)及经纬度(LON,LAT)提出了一种基于机器学习算法结合站点数据校正的降尺度框架.分析了前馈神经网络(FNN)、随机森林(RF)、梯度提升决策树(GBDT)在长江流域GPM IMERG数据空间降尺度过程中的鲁棒性差异;考虑了2001—2019年不同季节植被对降水降尺度结果的时滞效应;探究了降尺度数据对GPM的校正效果与降水量的关系.通过站点实测数据对模拟结果进行精度分析,结果表明:3种降尺度结果在得到1km空间分辨率降水数据的同时,也不同程度提高了数据的精度.其中,GBDT模拟结果对降水细节特征表达更明显,可以很好地反应降水的空间异质性.并且它在年、季时间尺度上模拟的结果在各研究区内的精度均为最优(年R^(2)=0.748~0.958,季R^(2)=0.518~0.909),有更好的降水捕获能力和更强的模型鲁棒性;春冬季植被的滞后性对GPM降尺度结果的响应最为敏感,最佳滞后期分别为1个月和2个月,夏秋季时滞效应不明显,基本无滞后期;降尺度数据相比原始数据提高的R^(2)与降水量呈正相关,相关系数达到0.630~0.844.

A spatial downscaling framework based on machine learning algorithms integrated with station data correction is proposed by combining enhanced vegetation index(EVI),precipitable water vapor(PWV),Land Surface Temperature(LST),Slope,Aspect,Altitude,LON,and LAT.Then,the robustness differences among three machine learning algorithms,namely Random Forest(RF),Feedforward Neural Network(FNN),and Gradient Boosting Decision Tree(DBGT),were evaluated for spatially downscaling the GPM IMERG over the Yangtze River basin.The time-lag effect of different seasonal vegetation on the downscaling results of precipitation from 2001~2019 was taken into account to explore how the correction effect of the downscaling framework on GPM is related to the precipitation.The results show that the three downscaling approaches can improve the accuracy of the data to different degrees while obtaining the GPM IMERG data with 1km spatial resolution.Specifically,the simulations of GBDT can express the detailed features of precipitation more obviously in terms of the spatial heterogeneity with the best simulation accuracy and stronger model robustness at annual and seasonal time scales in each study area(annual R^(2)=0.748~0.958,seasonal R^(2)=0.518~0.909).The hysteresis of vegetation in spring and winter were most sensitive to the response of GPM downscaling results,with the best hysteresis of 1and 2months in order.Yet the time lag effect was not significant in summer and autumn,with no lag period essentially.Compared to the original data,the increased R^(2)of the downscaled data was positively correlated with precipitation at R^(2) of 0.630~0.844.