Data-Driven Modeling for Urban Flood Warning Systems: A Case Study in the Guarará Basin, Brazil

Abstract

Urban flooding is a growing challenge in metropolitan areas, exacerbated by climate change and increasing urbanization. This study develops an innovative flood warning system for the Guarará Basin in Santo André, Brazil, leveraging both parametric and machine learning (ML) models. Rainfall data from the São Paulo State Flooding Alert System and historical flood records were processed using the dynamic Thiessen polygon method and advanced statistical techniques. A parametric model was calibrated to define alert thresholds, while a Random Forest (RF) classifier was trained to predict five alert levels: “No Rain,” “Raining,” “Vigilance,” “Warning,” and “Alert”. The models were validated against historical events from 2016 and 2019, demonstrating strong agreement in predicting alert levels and highlighting the benefits of combining physical interpretability with data-driven adaptability. The ML model achieved an overall weighted F1-score of 0.99, showcasing its effectiveness in classifying rainfall events and issuing timely warnings. This integrated methodology offers a robust framework for flood risk management in urban areas, contributing to the development of sustainable and resilient cities.