Modeling High Pan Evaporation Losses Using Support Vector Machine, Gaussian Processes, and Regression Tree Models

Evaporation is considered to be one of the most influential hydrological processes, contributing significantly to water loss within the hydrological cycle. This study aimed to address the challenge of modeling daily pan evaporation in arid climates, where harsh hydroclimatic conditions hinder modeling efficacy. In such climates, annual pan evaporation rates exceed 3,500 mm, exacerbating water scarcity in agricultural basins. Three machine-learning techniques: regression trees, Gaussian processes, and support vector machine regression were employed to model daily pan evaporation rates at two meteorological stations in Kuwait. Various meteorological variables, including average diurnal temperature, average wind speed, and average relative humidity, were utilized to formulate different modeling scenarios. The three modeling methods demonstrated robust efficiency in simulating historical pan evaporation under varied input formulations. In addition, the data-driven models were shown to outperform physically and statistically based conventional evaporation modeling methods. The mean absolute error (MAE) and coefficient of determination ($R^2$) ranged from 2.04 to $2.84\mathrm{mm}/\mathrm{day}$ and 0.73–0.85, respectively. Notably, a bias in model predictions was observed for daily pan evaporation rates exceeding $25\mathrm{mm}/\mathrm{day}$. A probabilistic assessment of model skill for operational forecasts on a weekly time scale affirmed the suitability of the selected data-driven models for operational and water management decision-making. This study sought to equip water managers in arid regions with powerful tools to formulate resilient water strategies mitigating the detrimental effects of water scarcity.