Advancing flood warning in rural watersheds: A data-driven machine learning classification framework

Riverine flooding jeopardizes the health and livelihood of communities and environmental systems worldwide and is expected to become more frequent and intense as climate change conditions alter hydrological processes. Most regions worldwide have flood warning systems that incorporate streamflow thresholds to specify action for emergency management and public messaging. With prescribed thresholds, there is an opportunity to apply classification Machine Learning (ML) models for flood message forecasting to aid in flood warning systems. In this study, lagged surface and subsurface data were used to develop a framework for employing classification ML models to predict types of flood warning messages in a rural watershed in Ontario, Canada. Thirty-three classification ML algorithms were implemented on a five-year dataset to predict the flood warning message six hours in advance. Overall classification accuracy ranged from 89.5% to 99.7%, where the top performing models were fine K-Nearest Neighbour (KNN), ensemble subsurface KNN and cubic support vector machines. Feature sensitivity analysis of the top performing models showed that streamflow, hourly flow difference and groundwater elevation were important parameters for flood message prediction. Results from this study highlight that classification ML algorithms can be applied to small agricultural watersheds for predicting flood messages, and the importance of including subsurface parameters in flood forecasting models. The framework developed from this work can be applied to other regions to develop flood forecasting models with interpretable outputs for stakeholders including flood forecasting professionals, government officials, emergency managers and the public.