An accurate and efficient deep learning emulator for hydrological modeling

Hydrologic data comes in many formats: point observations (such as stream gages), remote sensing products and networks of instruments (such as FluxNet or SNOTEL). Hydrologic models are often used to translate these data into operational outputs, thus informing stakeholders. Hydrologic models use mathematical equations to represent natural processes such as river flow, infiltration and exfiltration. Models that include those natural processes and the interaction between them are often referred to as integrated models. However, the more processes included in these models, the more computational demand is required for model simulations. Deep learning (DL) tools are increasingly used in Earth Science to link between components due to the tools' ability to capture nonlinear interactions. Recent advances in DL for modeling video dynamics have resulted in several solutions for spatiotemporal prediction. By applying a recent DL technique, namely PredRNN++, we are able to build an emulator version of an integrated hydrologic model, ParFlow, to spatially and temporally predict different components of the water cycle. We test the simulation performance of PredRNN++ model for different rainfall-runoff scenarios in a headwaters watershed in the Rocky Mountains, the Taylor River basin. Streamflow, water table depth and total water storage predicted by PredRNN++ and ParFlow agree well with the average bias of 0.085, 0.122 and 0.030 for streamflow, water table depth, and total water storage, respectively. PredRNN++ can speedup simulating rainfall-runoff scenarios from 5 to 40-fold, compared with the original ParFlow simulation. Given this promising proof of concept, we will train the PredRNN++ model with observed meteorology to efficiently and accurately predict the terrestrial water cycle.