Key Points: • Ecological and evapotranspiration characteristics of ten typical vegetation communities in semi-arid steppe were refined and decomposed. • Sensitive parameters of dynamic evapotranspiration improve the regional simulation effect. • Deep learning was used to downscale regional evapotranspiration at the 3-hour scale. Abstract Reports on ecohydrological models for semi-arid steppe basins with scarce historical data are rare. To fully understand the ecohydrological processes in such areas and accurately describe the coupling and mutual feedback between ecological and hydrological processes, a distributed ecohydrological model was constructed , which integrates multi-source information into the MY Ecohydrology (MYEH) model. This paper mainly describes the evapotranspiration module (Eva module) based on sensitive parameters and deep learning. Based on multi-source meteorological, soil, vegetation, and remote sensing data, the historical dynamic characteristics of ten typical vegetation communities in the semi-arid steppe are refined in this study and seven evaporation (ET) components in the Xilin River Basin (XRB) from 1980 to 2018 are simulated. The results show that the Naive Bayesian model constructed based on the temperature and three types of surface reflectance can clearly distinguish between snow-covered or-free conditions. Based on the refinement of typical vegetation communities, the ET process characteristics of different vegetation communities in response to climate change can be determined. Dynamic sensitive parameters significantly improve the regional ET simulation. Based on the validation with the Global Land Evaporation Amsterdam Model product and multiple models in multiple time scales (year, quarter, day, 3 h), a relatively consistent and reliable ET process 1 was obtained for the XRB at the 3-hour scale. The uncertainties of adding and dynamizing more ET process parameters and adjusting the algorithm structure must be further studied.