Machine Learning-Based Early Warning Systems for Urban Floods: A Case Study in Nilwala Basin

Abstract

This study pioneers the integration of Graph Neural Networks (GNNs) into flood forecasting systems, extending the predictive horizon from short-term forecasts to 7 days by effectively capturing spatial dependencies between rainfall stations. Focusing on the flood-prone regions of Matara and Galle districts within the Nilwala Basin, the research addresses the limitations of conventional forecasting methods by leveraging historical hydrological data, including daily rainfall records from six key stations and flow data from Pitabeddara. A hybrid machine learning framework combining Random Forest (RF) and K-Nearest Neighbors (KNN) models was developed to predict river discharge using rainfall data, overcoming challenges posed by limited water level data. The inclusion of GNNs introduces a novel approach to modeling complex spatial relationships, enabling improved accuracy in long-term flood prediction, particularly during extreme events. The proposed system demonstrates significant advancements in predictive reliability, offering a timely and accurate early warning tool to enhance disaster preparedness and risk management in the Nilwala Basin. This research underscores the transformative potential of datadriven methodologies in addressing the challenges of flood-prone regions.