Deep Learning-Enhanced Global Hydrological Cycle Simulation and Its Response to Climate Change

Abstract:

The global hydrological cycle is a core component of the Earth system, and its response to climate change has profound impacts on ecological security and human society. Traditional hydrological models are limited by simplified parameterization schemes, leading to uncertainties in simulating complex hydrological processes. This study proposes a deep learning-enhanced global hydrological simulation framework that integrates long short-term memory (LSTM) networks with a physically based hydrological model (Variable Infiltration Capacity, VIC). By assimilating multi-source data (e.g., satellite-derived precipitation, evapotranspiration, and soil moisture), the framework optimizes parameterization of key processes such as infiltration and runoff generation. Validation using in-situ observation data from 500 global river basins shows that the proposed framework improves the simulation accuracy of runoff and soil moisture by 15% and 22% respectively compared to the traditional VIC model. Further scenario simulations indicate that under the SSP5-8.5 scenario, the global average runoff will increase by 8.3% by the end of the 21st century, with significant spatial heterogeneity—runoff in high-latitude regions may increase by more than 20%, while arid and semi-arid regions in the mid-latitudes may face a 10-15% decrease in runoff. This study provides a new approach for improving the accuracy of global hydrological cycle simulations and offers scientific support for formulating climate change adaptation strategies.