Advanced long-term actual evapotranspiration estimation in humid climates for 1958–2021 based on machine learning models enhanced by the RReliefF algorithm

Abstract

Study region

Chengdu, Wuhan, Chongqing, and Kunming regions in China.

Study focus

Accurate estimation of crop water use or actual evapotranspiration (AET) remains a key obstacle in the effective design of irrigation schedules, plans, and design. This is due to the non-linear nature of this phenomenon. To address this issue and guarantee more accurate ET predictions, this study attempts the following: i) to assess the performance of five machine learning (ML) models optimized by the RReliefF algorithm in estimating actual ET values for each month in four Chinese provinces under various agroclimatic conditions; and ii) to select the optimal model based on statistical metrics while minimizing discrepancies between the estimated and actual ET values. AET was estimated using support vector machine (SVM), ensemble bagged and boosted trees, robust linear regression (RLR), and Matern 5/2 Gaussian process regression (M-GPR) models.

New hydrological insights for the region

The M-GPR model outperformed the other models and generated the best values for all statistical measures for training and testing stages: R² (0.979, 0.982), RMSE (5.56, 5.09), MAE (3.29,3.16). In comparison, the RLR model exhibited the lowest training and testing performances metrics. The results of this study demonstrate the capacity of the M-GPR model to accurately predict long-term AET values. This model is best suited for further research on AET prediction at the stations under investigation, which could improve irrigation and boost agricultural productivity.