Aquifer-specific flood forecasting using machine learning: A comparative analysis for three distinct sedimentary aquifers

Accurate flood prediction is critical for avoiding catastrophic impacts, but its difficulty varies by geological location. This study evaluates four machine learning models - TFT, Informer, LSTM, and XGBoost - for multi-horizon flood forecasting (1-4 days), across Limestone, Chalk, and Greensand located in the Thames Basin, UK. Stations were carefully chosen using the UK government flood risk maps, geological mapping, and Environment Agency hydrological data to guarantee a complete portrayal of aquifer-specific groundwater-river interactions. The results show that the model accuracy varies significantly depending on aquifer features. Rapid GWL-river interactions allowed Limestone aquifers to achieve very high precision (R² = 0.98-0.99), with transformers and LSTM clearly surpassing XGBoost. The accuracy of Chalk aquifers was moderate (R² = 0.77-0.80), indicating delayed reactions and intermediate permeability. Greensand aquifers were difficult to model due to delayed and complex reactions, resulting in low or negative R² values. Correlation study confirmed these findings: Limestone showed a significant groundwater-river linkage (r = 0.84), Chalk moderate (r = 0.26), and Greensand had a small negative association (r = -0.14). The novelty of this study highlights the significant impact of subsurface hydrology on predicted reliability, revealing aquifer-specific geological restrictions in ML-based forecasting. This research offers a more physically consistent early warning method by fusing GWL data with developed transformer architectures. The results highlight the significance of adjusting forecasting frameworks to geological environments, which has direct implications for resilience planning and flood risk management at the watershed scale.