Optimizing actual evapotranspiration simulation to identify evapotranspiration partitioning variations: A fusion of physical processes and machine learning techniques

Evapotranspiration (ET) serves as a pivotal metric for studying ecohydrological processes. Its dynamics are intricately linked to the interplay among water balance, energy balance, and land use. However, the precise estimation of ET at a regional scale, coupled with a comprehensive understanding of its response to changes in water, heat, and vegetation factors, presents an ongoing challenge. In this study, we assessed the necessity to refine soil evaporation simulations in arid and semiarid regions. Subsequently, we introduced the random forest algorithm to optimize the soil evaporation module of the Priestley-Taylor Jet Propulsion Laboratory (PT-JPL) model to enable a more accurate prediction of ET. Then, we quantitatively explored the correlation between ET and pertinent factors, such as soil moisture, solar radiation, and the normalized difference vegetation index (NDVI) in the Inner Mongolia section of the Yellow River Basin (IMSYRB), which is a typical arid and semiarid region. The findings underscore that soil evaporation is the predominant constituent of ET in the IMSYRB. Moreover, soil moisture, as opposed to relative humidity, better represent the near-surface moisture conditions within the study area. The refined model (PT-JPL-RF) exhibited enhanced simulation performance at both individual stations and in the overall region for ET. Furthermore, the application of the PT-JPL-RF model for scrutinizing the spatiotemporal ET variations in the study area revealed an average ET of 250 mm over the past four decades. Since 1982, a fluctuating upward trend of 0.99 mm/year has been observed. Spatially, ET exhibits an eastward distribution, with elevated values in the eastern sectors compared to those in the western regions, particularly in the Big Black River Basin (296 mm). Vegetation was the primary factor influencing ET variations in the IMSYRB, contributing to an annual increase by 1.58 mm. These findings demonstrate the benefits of integrating machine learning algorithms with physical models and provide a practical example for accurately simulating and predicting ET across diverse regions.