Improving Evaporative Loss Forecasts in Arid Climates by Integrating Machine Learning Models With Feature Selection Algorithms

Abstract

Evaporation is a major water-loss process that significantly disrupts the hydrological cycle; therefore, reliable and continuous evaporation monitoring is essential for decision-makers in water resource management. However, hyper-arid climates exhibit accelerated evaporation rates, complicating hydrological modeling. This study represents the first attempt to integrate the RReliefF algorithm for meteorological feature selection with machine learning models for pan evaporation prediction in hyper-arid climates. This approach overcomes the arbitrary selection of features for ML model input. Daily average pan evaporation rates at the examined stations exceed 8 mm/day. Such extremely high evaporative losses have been shown to hinder ML model performance. Extreme gradient boosting (XGBoost), random forest model, and k-nearest neighbors were used. Meteorological datasets were preprocessed using the RReliefF algorithm to rank their influence on pan evaporation variability. Depending on the weather station, shortwave radiation, wind speed, and average diurnal temperature emerged as the best predictors of pan evaporation rates. During the validation period, the Nash–Sutcliffe efficiency coefficient (NS), root mean squared error (RMSE), and mean absolute error (MAE) were 0.85–0.94, 1.152–1.833, and 0.863–1.147 mm/day, respectively. The findings of this study offer a robust and efficient computational approach for forecasting evaporative losses in hyper-arid environments.