Mapping the groundwater depth in Africa at high resolution (1 km²) based on the Parflow model and machine learning

Groundwater depth is the result of a balance between climatic conditions (rainfall, temperature, radiation, etc.), topography (slope and proximity of a river), land use, soil and subsoil characteristics (hydrodynamic parameters), and potential exploitation by man.

In order to better understand the evolution of water resources in Africa, the ParFlow-CLM model (Maxwell et al. 2015) was used to provide at high resolution simulation (1 km²) over West Africa (3.5 million km²). This simulation was obtained after a long period of groundwater equilibration using a simplified version of the Parflow model with a monthly time step, and forced by the variable P-ETR (rainfall minus evapotranspiration). Simulation started with a 30 m water table depth everywhere and equilibrium was reached after a few tens of hydrological years in the Sudanian part and 1700 years in the Sahelian part close to the Sahara. The depths of the water table obtained range from 1-2 m to 85 m below the surface. This preliminary operation has a high numerical cost and required just over 1,000,000 hCPUs over 2,560 cores.

In order to reduce computing time and allow modelling of water table depths over the whole of Africa (30.5 million km²), a method based on artificial intelligence (AI) has been applied, following the work of Tran et al. (2021) and Bennett et al. (2024), to reproduce the operation of ParFlow and allow computing times of the order of 1000x less than physical modelling. The approach consists of conducting the automatic learning of the AI on sub-regions of the Parflow simulation, and then evaluating the relevance of the results on other regions.

In this presentation, we show how AI can be used to estimate the equilibrium groundwater depths simulated by the Parflow model over West Africa, while drastically reducing the numerical cost. We also show the performance of this methodology for mapping groundwater depths over the whole of Africa using networks of piezometers and village wells.

Bennett, A., Tran, H., De la Fuente, L., Triplett, A., Ma, Y., Melchior, P., et al. (2024). Spatio‐temporal machine learning for regional to continental scale terrestrial hydrology. Journal of Advances in Modeling Earth Systems, 16, e2023MS004095. https://doi.org/10.1029/2023MS004095

Maxwell R.M., L. E. Condon, and S. J. Kollet (2015). A high-resolution simulation of groundwater and surface water over most of the continental US with the integrated hydrologic model ParFlow v3. Geosci. Model Dev., 8, 923–937, https://doi.org/10.5194/gmd-8-923-2015

Tran, H., Leonarduzzi, E., De la Fuente, L., Hull, R. B., Bansal, V., Chennault, C., et al. (2021). Development of a deep learning emulator for a distributed groundwater–surface water model: ParFlow‐ML. Water, 13(23), 3393. https://doi.org/10.3390/w13233393