Machine Learning-Based Estimation of Monthly Potential Evapotranspiration for Sustainable Water Resources Management using Random Forest and Support Vector Machine

Accurate estimation of potential evapotranspiration (PET) is essential for irrigation planning, agricultural water allocation, drought assessment, and sustainable water resources management. Conventional PET methods require several meteorological variables that may be incomplete or unavailable in data-scarce regions. This study evaluates Random Forest (RF) and Support Vector Machine (SVM) for monthly PET estimation in Supaul District, Bihar, India, using monthly meteorological data from 2001–2022. Seven predictors—specific humidity at 2 m (QV2M), relative humidity at 2 m (RH2M), wind speed at 2 m (WS2M), maximum and minimum air temperature (T2M_MAX and T2M_MIN), ultraviolet radiation (UVB), and sunshine duration (SDDN)—were used, while Penman-derived PET was the target. Model performance was evaluated using R², RMSE, MAE, MBE, NSE, KGE, Willmott’s index of agreement, residual diagnostics, Taylor diagrams, and Bland–Altman analysis. RF achieved R² = 0.976, RMSE = 6.30 mm month⁻¹, MAE = 4.30 mm month⁻¹, MBE = −0.21 mm month⁻¹, NSE = 0.976, KGE = 0.897, and d = 0.994. SVM achieved R² = 0.828, RMSE = 16.74 mm month⁻¹, MAE = 12.24 mm month⁻¹, and MBE = −1.94 mm month⁻¹. RF consistently showed closer agreement with the Penman-derived reference series and smaller residual dispersion. The results indicate that RF provides an efficient data-driven approach for reproducing monthly Penman-derived PET in the study region and may support irrigation and water-resource planning where complete operational calculations are inconvenient.