Accurate estimates of evapotranspiration (ET) over croplands on a regional scale can provide useful information for agricultural management. The hybrid ET model that combines the physical framework, namely the Penman-Monteith equation and machine learning (ML) algorithms, have proven to be effective in ET estimates. However, few studies compared the performances in estimating ET between multiple hybrid model versions using different ML algorithms. In this study, we constructed six different hybrid ET models based on six classical ML algorithms, namely the K nearest neighbor algorithm, random forest, support vector machine, extreme gradient boosting algorithm, artificial neural network (ANN) and long short-term memory (LSTM), using observed data of 17 eddy covariance flux sites of cropland over the globe. Each hybrid model was assessed to estimate ET with ten different input data combinations. In each hybrid model, the ML algorithm was used to model the stomatal conductance (G_s), and then ET was estimated using the Penman-Monteith equation, along with the ML-based G_s. The results showed that all hybrid models can reasonably reproduce ET of cropland with the models using two or more remote sensing (RS) factors. The results also showed that although including RS factors can remarkably contribute to improving ET estimates, hybrid models except for LSTM using three or more RS factors were only marginally better than those using two RS factors. We also evidenced that the ANN-based model exhibits the optimal performance among all ML-based models in modeling daily ET, as indicated by the lower root-mean-square error (RMSE, 18.67–21.23 W m⁻²) and higher correlations coefficient (r, 0.90–0.94). ANN are more suitable for modeling G_s as compared to other ML algorithms under investigation, being able to provide methodological support for accurate estimation of cropland ET on a regional scale.

Accurate estimation of evapotranspiration (ET) can provide useful information for water management and sustainable agricultural development. However, most of the existing studies used physical models, which are not accurate enough due to our limited ability to represent the ET process accurately or rarely focused on cropland. In this study, we trained two models of estimating croplands ET. The first is Medlyn-Penman-Monteith (Medlyn-PM) model. It uses artificial neural network (ANN)-derived gross primary production along with Medlyn’s stomatal conductance to compute surface conductance (Gs), and the computed Gs is used to estimate ET using the PM equation. The second model, termed ANN-PM, directly uses ANN to construct Gs and simulate ET using the PM equation. The results showed that the two models can reasonably reproduce ET with ANN-PM showing a better performance, as indicated by the lower error and higher determination coefficients. The results also showed that the performances of ANN-PM without the facilitation of any remote sensing (RS) factors degraded significantly compared to the versions that used RS factors. We also evidenced that ANN-PM can reasonably characterize the time-series changes of ET at sites having a dry climate. The ANN-PM method can reasonably estimate the ET of croplands under different environmental conditions.