A novel evapotranspiration downscaling approach based on dynamic sensitive parameters and deep learning in the grassland lacking historical measurements

Abstract

To address the challenges of ecohydrological modeling in data-scarce semi-arid steppe basins, this study developed the Distributed Dynamic Process Model (DDPM), integrating multi-source data and a novel evapotranspiration (ET) module enhanced by dynamic sensitivity parameters and deep learning. By leveraging meteorological, soil, vegetation, and remote sensing data, we analyzed the ET dynamics of ten representative vegetation communities in the Xilin River Basin (XRB) from 1980 to 2018. Key findings include: (1) A Naive Bayesian model, utilizing temperature and surface reflectance, achieved high snow cover detection accuracy (>0.95); (2) Incorporating dynamic sensitivity parameters significantly improved ET simulation accuracy (R2 increased from 0.816 to 0.941, NSE from 0.583 to 0.769); (3) Multi-scale validation (annual, seasonal, daily, and 3-hour) against the Penman-Monteith model and Bowen Ratio-Energy Balance method confirmed the model’s robustness, particularly at the 3-hour scale. These results highlight vegetation-specific ET responses to climate change and underscore the need for refined dynamic parameterization and algorithmic structures. The DDPM framework provides a robust approach for ecohydrological modeling in data-scarce regions, offering insights for adaptive water resource management and ecological restoration.