A multi-resolution deep-learning surrogate framework for global hydrological models

Global hydrological models (GHMs) are an important tool for sustainable development making in today’s water-scarce world. These models enable assessment of water scarcity by estimating both the natural water cycle and human activities around the world. Moreover, their process-based structure allows for projections under diverse climate change and socio-economic scenarios; information that is essential to support sustainable water management. Nevertheless, the need for better, higher resolution and larger ensemble simulations is reaching the limit of what is computationally feasible.
Recently, the deep-learning community has shown the potential of neural networks in providing highly accurate and computationally cheap hydrological predictions. This development has let to the emergence of deep-learning model surrogates that mimic process-based hydrological simulations using neural networks. Yet, the majority of these surrogates are restricted to assessing land-surface water fluxes at a singular spatial resolution, thereby limiting their application for global hydrological models.
We present a novel framework to create deep-learning global hydrological surrogates, with two salient features. First, our surrogate framework integrates spatially-distributed runoff routing that is essential to estimate water availability and human water withdrawals. Second, our surrogate framework offers scalability across various spatial resolutions and can match the wide variety of resolutions at which global hydrological models are applied.
To test our framework, we developed a deep-learning surrogate of the PCRaster Global Water Balance (PCR-GLOBWB) global hydrological model. The surrogate encompasses all water-balance components, including the impact of human activities on the water system. The PCR-GLOBWB surrogate runs faster than its process-based counterpart and performs well when compared to the original model’s output at different spatial resolutions. Interestingly, the multi-resolution surrogate actually outperforms model surrogates trained for a single resolution, even on their target resolution.
Deep-learning surrogates are a useful tool for the global hydrological modeling community, enabling enhanced model calibration (through parameter learning and flux matching) and more detailed model simulations. Our framework provides an excellent foundation for the community to create their own multi-scale deep-learning model surrogates.