Abstract

Emerging evidence suggests that climate-induced changes in vegetation profoundly affect the water cycle and hydrological processes. However, state-of-the-art hydrological models typically overlook this dynamic aspect, relying instead on static vegetation parameters when analysing the hydrological impacts of climate change. Although land surface models within earth system models account for dynamic vegetation, their computational demands render them impractical for integration into hydrological frameworks. In the current study focusing on India, we demonstrate the pivotal role of vegetation variability in influencing evapotranspiration, particularly during the latter half of the Indian summer monsoon season and the subsequent post-monsoon period characterized by high vegetation activity. Incorporating spatial and temporal vegetation variability into the widely used Variable Infiltration Capacity (VIC) model revealed an 18% increase in total annual evapotranspiration compared to conventional simulations. To efficiently integrate varying vegetation dynamics into the VIC framework, we developed a grid-scale Machine Learning model based on Long Short-Term Memory (LSTM). This model represents vegetation changes as a function of hydrometeorological variables, such as precipitation and temperature. The LSTM-derived vegetation properties—fraction of vegetation cover, Leaf Area Index, and albedo—are then incorporated into the hydrological model. Using this novel approach, we simulated future hydrometeorological scenarios in India based on downscaled NEX-GDDP-CMIP6 dataset projections from the MIROC6 General Circulation Model (GCM). Our findings highlight the significant impact of changing vegetation properties on future ET fluxes, a factor often overlooked when employing constant vegetation parameters. We restricted the analysis to a single GCM to demonstrate the importance of our approach, which considers changing vegetation dynamics—a critical yet neglected aspect in current hydrological impact assessments of climate change. This study not only addresses the inadequacies of prevailing methodologies but also presents an advanced framework for integrating machine learning into physics-based hydrological models to account for evolving vegetation properties in future simulations.