Hybrid decomposition-based deep learning and boosting models for accurate prediction of evaporation potential in multi-climate regions

Accurate prediction of evaporation potential an vital role in water resources management, drought monitoring, and the hydrological cycle. In this research, the performance of diverse models, including LSTM, VMD + LSTM, EMD + LSTM, HistGBRT, VMD + HistGBRT, EMD + HistGBRT, CatBoost, VMD + CatBoost, and EMD + CatBoost for predicting evaporation potential in eight different climates was investigated. Also, in this research, two satellite products, NASA-POWER and ERA5, were used on a daily scale to examine the performance of different products. Studies showed that the proposed models performed better in predicting evaporation using NASA POWER satellite data. Comparative results based on correlation coefficien (CC) demonstrated that the stand-alone LSTM model performed poorly in warm seasons, especially summer, and CC values were less than 0.30 and even negative (−0.10) in some regions. In contrast, the combined decomposition + Catboost models performed much better, and in most regions and seasons, their CC values were higher than 0.97 and in many regions. The results also showed that the CatBoost models, and especially EMD + CatBoost, were able to reproduce the distribution of observational data with high accuracy and were most similar to observations in both mean and range. Additionally, it can be stated that EMD combined with CatBoost demonstrated a robust capability to capture both long-term trends and short-term fluctuations in evaporation, effectively reproducing extreme values. Based on the results, it can be stated that data preprocessing can provide an efficient framework for predicting evapotranspiration in different regions.