How do meteorological insights boost groundwater level prediction in former sewage-irrigated urban areas?

The study applies linear regression and artificial neural networks to downscale groundwater level (GWL) to daily resolution in the former sewage-irrigated fields using past monthly GWLs and daily meteorological data. The best predictors, selected using F-test, were maximum air temperature, sunshine duration, and relative humidity from the previous day. Including prior-month GWLs provides high model accuracy (R2 = 0.945–0.965; root mean squared error (RMSE) = 0.095–0.118). Clustering yielded modest overall gains, with the most notable improvements observed for specific weather types. In the stable conditions RMSE was reduced by 13% (cluster 5: cool, humid, cloudy days) and the mean average percentage error (MAPE) by 19% (cluster 4: hot, dry and sunny weather). The comparison of algorithms showed that the linear regression (LR) provides high stability of forecasts by taking into account the system inertia, while the artificial neural networks (ANNs) show an advantage in representing nonlinear responses of the groundwater table after extracting specific meteorological patterns. Accurate GWL prediction supports climate-resilient planning and the protection of urban biodiversity.