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

Area of Science:

• Hydrology and Hydrogeology
• Geological Sciences
• Machine Learning Applications

Background:

• Accurate flood prediction is crucial for mitigating catastrophic impacts, yet its accuracy is influenced by geological conditions.
• Understanding aquifer-specific groundwater-river dynamics is essential for effective flood forecasting.

Purpose of the Study:

• To evaluate the performance of four machine learning models (TFT, Informer, LSTM, XGBoost) for multi-horizon flood forecasting (1-4 days).
• To assess the impact of different aquifer types (Limestone, Chalk, Greensand) in the Thames Basin, UK, on flood prediction accuracy.
• To investigate the relationship between subsurface hydrology and the reliability of machine learning-based flood forecasting.

Main Methods:

• Selected hydrological stations based on UK flood risk maps, geological data, and Environment Agency hydrological data.
• Employed four machine learning models: Transformer (TFT), Informer, Long Short-Term Memory (LSTM), and XGBoost.
• Analyzed model performance using R-squared values and correlation coefficients (r) to assess groundwater-river linkages.

Main Results:

• Model accuracy varied significantly across aquifer types: Limestone (R ² = 0.98-0.99), Chalk (R ² = 0.77-0.80), and Greensand (low or negative R ²).
• Transformer and LSTM models outperformed XGBoost, particularly in Limestone aquifers with rapid groundwater level (GWL)-river interactions.
• Correlation analysis confirmed strong GWL-river linkage in Limestone (r=0.84), moderate in Chalk (r=0.26), and weak/negative in Greensand (r=-0.14).

Conclusions:

• Subsurface geology significantly impacts the reliability of machine learning-based flood forecasting.
• Forecasting frameworks must be adapted to specific geological environments for improved flood risk management and resilience planning.
• Integrating groundwater level data with advanced transformer architectures offers a more physically consistent approach to early flood warning systems.