Combining machine leaning and physics-based modeling to predict the hydrological response to an unprecedented climate
Abstract
Information on the availability and distribution of water in the subsurface is essential for a wide range of applications, like water management or natural hazard prediction. This information is often only available from observations or models. Observational data is often limited to point data that lack spatial coverage and remote sensing products that lack point-scale accuracy. Among models, physics-based hydrological models provide promising estimates of states and fluxes particularly when representing connections between components. They can accurately simulate water distribution in the subsurface, but are very computationally expensive, especially at high spatial resolution. The work presented here is part of the NSF convergence accelerator project HydroGEN that combines physics-based modelling with machine learning to improve accessibility to hydrological tools and data. Here, we use the state-of-the-art model ParFlow-CLM and historical meteorological data (NLDAS) to train a Convolutional Neural Network (CNN). The advantages of such a setup are the computational efficiency of the machine learning component and the accuracy of the estimates provided by ParFlow-CLM, which can be used to create not only historical timeseries, but also run climate scenarios for a no-analog future. This is important because the climate is changing and many locations only have limited (<50 year) historical records, that often do not contain sufficient variability to reflect more extreme events. We design this system to predict daily soil moisture using static (permeability, porosity, and slopes) and dynamic variables (soil moisture, infiltration, and evapotranspiration of the previous day). We use ParFlow-CLM to provide soil moisture fields using historical observed forcing, but also additional drought scenarios where the precipitation is homogeneously reduced by a scaling factor and/or temperature is increased by a fraction of a degree. The comparisons between performances obtained with CNNs trained only on observed forcing or also on drought scenarios confirms that generating additional simulations with ParFlow-CLM can help train computationally efficient ML models and better predict an unprecedented future.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H32B..01L