Neural Network Driven Early Warning System for Groundwater Flooding: A Comprehensive Approach in Lowland Karst Areas

Ruhhee Galal Tabbussum, Department of Civil, Structural and Environmental Engineering, Trinity College Dublin, Dublin , Ireland
Bidroha M. Basu, Department of Civil, Structural and Environmental Engineering, Munster Technological University, Cork, Ireland
Laurence Talas Mahammad Gill, Department of Civil, Structural and Environmental Engineering, Trinity College Dublin, Dublin , Ireland
Patrick Morrissey, Department of Civil, Structural and Environmental Engineering, Trinity College Dublin, Dublin , Ireland

Environmental Modelling & Software

Abstract

This study has developed forecast models for groundwater flooding in lowland karst region of south Galway (Ireland). It employed neural network models incorporating Bayesian regularization and Scaled Conjugate Gradient training algorithms for model training and optimization. Training datasets include years of field data and outputs from a calibrated hydraulic/hydrological karst model. The Bayesian model achieves Nash-Sutcliffe Efficiency (NSE) of 0.95 up to 45 days ahead, whilst the Scaled Conjugate Gradient models outperform this, maintaining NSE of 0.98 up to 20 days and 0.95 up to 60 days ahead, with reduced training time compared to Bayesian models. Both models exhibit high performance with a Coefficient of Correlation (r) value of 0.98 up to 60 days ahead and Kling Gupta Efficiency of 0.96 up to 15 days ahead. The research shows that integrating diverse data sources and using both daily and hourly models improve such a flood warning system's resilience and reliability.