顾及时空特征的参考作物蒸散量集成学习估算

Ensemble learning estimation of reference crop evapotranspirationtaking into account temporal and spatial characteristics

为提升参考作物蒸散量(ET_(0))的估算精度,以四川省为研究区域,发现全省ET_(0)的数据变化具有明显的时间和空间自相关性,继而在气象特征基础上引入时空特征构建以XGBoost, LightGBM,GBDT、随机森林和极限树为基模型的Stacking模型.将顾及时空特征的Stacking模型与其各个基模型以及经验模型彭曼公式(FAO 56 Penman-Monteith)的决定系数、平均绝对值误差和均方误差等多项指标进行了全面的精度对比验证.试验结果表明:在顾及空间特征的情况下,Stacking模型在测试集上决定系数精度提升了3%,平均绝对值误差和均方误差分别降低了51%和76%;在顾及时序特征的情况下,Stacking模型在测试集上的决定系数精度提升了4%,均方误差和平均绝对值误差分别降低了92%和72%.这表明时空特征的引入可有效提升模型估算ET_(0)性能.在同时顾及时空特征的情况下,Stacking模型相较于彭曼公式,决定系数提升了39%,均方误差、平均绝对值误差分别降低了95%和77%,并且,在2006—2010年逐年精度验证中,Stacking模型精度始终优于其每年最优基模型精度.因此,顾及时空特征的Stacking模型可有效提升四川省ET_(0)估算精度.

In order to improve the estimation accuracy of reference crop evapotranspiration(ET_(0)),Sichuan Province was taken as the research area to reveal that the data changes of ET_(0) in Sichuan Pro-vince had obvious temporal and spatial autocorrelation.Then,based on meteorological characteristics,temporal and spatial features were introduced to construct XGBoost and LightGBM,GBDT,random forest and extreme tree-based Stacking model.The Stacking model with its base model and empirical model(FAO 56 Penman-Monteith)was comprehensively verified from multiple indicators such as coefficient of determination,mean absolute error and mean square error.The experimental results show that considering the spatial characteristics,the Stacking model improves the coefficient of determination on the test set by 3%,and the mean absolute error and mean square error are reduced by 51%and 76%,respectively.Taking into account the temporal characteristics,the Stacking model improves the accuracy of the upper coefficient of determination on the test set by 4%,and the mean absolute error and mean square error are reduced by 92%and 72%,respectively.This shows that the introduction of temporal and spatial characteristics can effectively improve the ET_(0) performance of the model.With both spatial and temporal characteristics in mind,compared with Penman-Monteith′s formula,the Stacking model′s coefficient of determination increases by 39%,and the mean absolute error and mean square error reduced by 95%and 77%,respectively.In the year-by-year accuracy verification in 2006—2010,the accuracy of the Stacking model is always better than its annual accuracy optimal base model.Therefore,the Stacking model considering the spatial and temporal characteristics can effectively improve the ET_(0) estimation accuracy in Sichuan Province.