Prediction of spatial-temporal flood water level in agricultural fields using advanced machine learning and deep learning approaches

Agricultural fields frequently experience flood disasters and significantly impacting food security, thus prompting the urgent need for efficient predictive flood mitigation mechanisms. This study presents an innovative approach for predicting spatial-temporal water levels in an agricultural field. Five ensemble machine-learning algorithms were developed to predict temporal channel water levels at four gauging points (GPs). Further, the ensemble Recurrent Neural Network - Long Short-Term Memory (RNN-LSTM), a deep learning-based model was employed for spatial prediction – surface water level. The models were trained and validated using observed rainfall and simulated water level data for both drainage and field surfaces derived from SWMM-based hydrological models. The Random Forest and Extra trees models achieved superior performance in temporal predictions at gauging points 1 and 4, achieving R² and KGE values greater than 0.800. For the spatial inundation predictions, the RNN-LSTM achieved R² and RMSE values of 0.999 and 0.094, respectively. This study underscores the critical influence of drainage network characteristics and design rainfall patterns in enhancing flood prediction accuracy. These results demonstrate the potential for precise flood prediction in agricultural fields and suggest the integration of machine learning and deep learning models into flood control and a decision support system, thereby enhancing flood management, decision-making, and preparedness against flood disasters.