Modeling Hydrological and Atmospheric Interactions in the Mar Menor Lagoon Through Machine Learning: Feature Selection, Time Association, and Predictive Modeling

The Mar Menor Lagoon (Spain), recently granted legal personality after severe eutrophication crises, is monitored by sensors that generate multi-source, high-frequency data, yet hydrological and atmospheric variables are usually analyzed in isolation. This study examines how these variables interact over time. Using 2023 data from the lagoon buoy, AEMET meteorological stations (San Javier Airport, Cabo Palos), the La Pedrera reservoir and the MURC2 tide gauge. We follow the CRISP-DM methodology and combine Lasso and Elastic Net feature selection, K-means and Gaussian Mixture clustering, time-rolling-window Spearman correlation, and Random Forest prediction of chlorophyll-a and turbidity. Results indicate that (1) air and lagoon water temperature are strongly coupled, and the disruption of this coupling flagged a sensor anomaly in June 2023; (2) the positive association between conductivity and temperature holds only during the first six months, suggesting a shift in the dominant drivers of ion mobility; and (3) wind power is negatively associated with temperature and conductivity, consistent with evaporative cooling and surface mixing. In this single-year record, associations are therefore time-varying, which challenges static, compartmentalized monitoring. The workflow is reproducible and can support sensor quality control and evidence-based management of legally protected water bodies.