Evaluation of SVM, ELM and four tree-based ensemble models for predicting daily reference evapotranspiration using limited meteorological data in different climates of China

Abstract

Accurate estimation of reference evapotranspiration (ET0) is of great importance for the regional water resources planning and irrigation scheduling design. The FAO-56 Penman-Monteith model is recommended as the reference model to predict ET0, but its application is commonly restricted by lack of complete meteorological data at many worldwide locations. This study evaluated the potential of machine learning models, particularly four relatively simple tree-based assemble algorithms (i.e. random forest (RF), M5 model tree (M5Tree), gradient boosting decision tree (GBDT) and extreme gradient boosting (XGBoost)), for estimating daily ET0 with limited meteorological data using a K-fold cross-validation method. For assessment of the tree-based models in terms of prediction accuracy, stability and computational costs, these models were further compared with their corresponding support vector machine (SVM) and extreme learning machine (ELM) models. Four input combinations of daily maximum and maximum temperature (Tmax and Tmin), relative humidity (Hr), wind speed (U2), global and extra-terrestrial solar radiation (Rs and Ra) with Tmax, Tmin and Ra as the base dataset were considered using meteorological data during 1961–2010 from eight representative weather stations in different climates of China. The results showed that, when lack of complete meteorological data, the machine learning models using Tmax, Tmin, Hr, U2 and Ra obtained satisfactory ET0 estimates in the temperate continental, mountain plateau and temperate monsoon zones of China (RMSE < 0.5 mm d−1). However, models with three input parameters of Tmax, Tmin and Rs were superior for daily ET0 prediction in the tropical and subtropical zones. The ELM and SVM models offered the best combination of prediction accuracy and stability. The simple tree-based XGBoost and GBDT models showed comparable accuracy and stability to the SVM and ELM models, but exhibited much less computational costs. Considering the complexity level, prediction accuracy, stability and computational costs of the studied models, the XGBoost and GBDT models have been recommended for daily ET0 estimation in different climatic zones of China and maybe elsewhere with similar climates around the world.