A Deep Learning Emulator for a Groundwater Model: Mapping Steady-State Water Table Depth
Abstract
Groundwater is an increasingly important water resource, especially as drought and climate change make other sources of water more scarce. Hence, mapping water table depth (WTD) and understanding the sensitivities of input parameters to WTD are of great use for decision making, as well as hydrological modeling. Developing a spatially continuous map of WTD solely with observations is challenging. Physically-based models that simulate WTD are computationally expensive, limiting the ability to address uncertainty. Here, a deep learning emulator is developed to predict steady-state WTD over the contiguous US (CONUS), with a focus on exploring the relationships between WTD, hydraulic conductivity (K), and precipitation minus evapotranspiration (PME). A U-Net architecture is trained on thousands of 32km*32km simulated data from the physically-based hydrologic model ParFlow. The emulator converged in 200 epochs and its test performance resulted in MAE values of 4.25m. Sensitivity analyses are conducted to evaluate model robustness. Uncertainty distributions for WTD, K, and PME are developed by injecting gaussian noise into the emulator; this enables an assessment of importance among various hydrological quantities in steady-state WTD modeling. The effects of varying input and output parameters are also studied: the same U-Net architecture is used to create mappings of K and PME from the other remaining parameters. This was done to study the possibility of calibrating input parameters to the emulator against observational WTD data. In particular, the calibration of K using Markov chain Monte Carlo methods is explored. The aforementioned simulation-based inference (SBI) strategy opens up possibilities for improving the true accuracy of steady-state WTD mapping over the CONUS, and refining K data for use in other hydrological models.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H52P0670P