Abstract. We analyze the relationship between potential evapotranspiration (ET_o), actual evapotranspiration (ET_a), and surface water evaporation (E_sw) in the semi-arid south-central Texas, using hourly climate data, daily lake evaporation measurements, and daily actual evapotranspiration measurements from an eddy covariance (EC) tower. The deterministic analysis reveals that ET_o set the upper bound for ET_a, but the lower bound for E_sw in the study area. Unprecedentedly, we demonstrate that a newly developed probabilistic machine learning (ML) model, using a hybridized NGBoost-XGBoost framework, can accurately predict the daily ET_o, E_sw, & ET_a from local climate data. The probabilistic approach exhibits great potential in overcoming data uncertainties, in which 99 % of the ET_o, 90 % of the E_sw, and 91 % of the ET_a test data at three watersheds were within the model's 95 % prediction interval. The probabilistic ML model results suggest that the proposed framework can serve as a robust and computationally more efficient tool than the hourly Penman-Monteith equation to predict the ET_o while avoiding computationally-involved net solar radiation calculations. Additionally, the performance analysis of the probabilistic ML model indicates that it can be successfully implemented in practice to overcome the uncertainties associated with pan evaporation & pan coefficients in E_sw estimates, and to offset the high capital & operational costs of EC towers used for E_a measurements. Finally, we demonstrate, for the first time, a coalition game theory approach to identify the order of importance, dependencies & interactions of climatic variables on the ML-based ET_o, E_sw, and ET_a predictions. New knowledge gained through the game theory approach is beneficial to strategically locate weather stations for enhanced evapo(transpi)ration predictions, and plan out sustainability and resilience efforts, as part of water management and habitat conservation plans.