Ensemble learning estimation of reference crop evapotranspiration taking into account temporal and spatial characteristics
In order to improve the estimation accuracy of reference crop evapotranspiration(ET0),Si-chuan Province was taken as the research area to reveal that the data changes of ET0 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 co-efficient 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 ET0 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 ET0 estimation accuracy in Sichuan Province.
reference crop evapotranspirationStacking modeltemporal and spatial characteristicsensemble learning
国家科技期刊平台
NSTL
万方数据
维普
