A comparative analysis of two machine learning models for forecasting salinity intrusion in a coastal aquifer in Marbella (Spain)

The project of thesis is based on the research carried out in the frame of the European project MAR2PROTECT, a 4-year (2022-2026) project that aims to develop a new holistic approach to Managed Aquifer Recharge (MAR). Among its key innovations are machine learning tools designed to assess climate change impacts on groundwater quality: DRONE, developed by the University of Bologna, and REACH, by CETAND (Andalusia Water Technology Center). This work explores the theoretical and practical foundations of their modeling approaches: Gaussian Processes (GPs) for DRONE and autoregressive models for REACH. Four GP configurations were designed, combining stationary and non-stationary covariance functions with two linear mean functions. Their performance in forecasting salinity intrusion in the Marbella coastal aquifer (Spain) was evaluated and compared to REACH using multiple metrics (e.g., MAE, SMAPE, MDA). Results indicate that REACH and the stationary GP with a simpler mean function outperformed other models in terms of accuracy, while a non-stationary GP showed slightly better MDA results.