Comparison of dimensionality reduction approaches of spatio-temporal inputs and outputs of an integrated surface-subsurface hydrological model
Résumé
Over the past decades, uncontrolled urbanization, poor land management, and rising groundwater withdrawals for drinking and irrigation, alongside agricultural pollution, have severely reduced groundwater recharge, threatening aquifer sustainability. The Water4All AQUIGROW project addresses this by: 1. Quantifying current recharge rates in key aquifers, 2. Identifying sustainable management challenges, and 3. Developing tools to enhance recharge quality and quantity. However, integrated surface-subsurface hydrological models (ISSHM), essential for simulating water and pollutant movement, are often too complex for practical decision-making. To bridge this gap, we use metamodeling, creating simplified surrogate models (specifically Gaussian Process (GP) metamodels) that offer uncertainty estimates and can be refined with additional data. This adaptive approach balances accuracy, computational cost, and environmental impact. A major challenge is the high dimensionality of ISSHM inputs (e.g., soil property maps) and outputs (e.g., spatio-temporal recharge fluxes), as traditional GPs handle only low-dimensional inputs and scalar outputs. Our work uses the CATHY model, a physics-based tool simulating coupled water flow and solute transport. We compare dimensionality reduction methods, from Principal Component Analysis to advanced autoencoder-based feature learning, to simplify the model while preserving accuracy.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |
