Flood prediction using machine learning

Abstract

Heavy rainfall events that cause floods are natural and unavoidable, yet result in devastating damage to society and, therefore, must be studied to mitigate their impacts. This study presents a predictive Artificial Intelli gence model capable of classifying the occurrence of flood events using Machine Learning techniques, including Multi-Layer Perceptron (MLP) neural networks and decision trees. This research was conducted in two separate case studies: one in the urban area of Hugo Cantergiani, the regional airport in Caxias do Sul, Brazil, and the second on the Faxinal microbasin, also located in the city. The model was trained using meteorological, such as rainfall, temperature, atmospheric pressure, and others, and hydrological data such as Antecedent Precipitation Index (API). The data collected from Instituto Nacional de Meteorologia (INMET) for the first case study, and provided by Serviço Autônomo Municipal de Água e Esgoto de Caxias do Sul (SAMAE) for the second. The first case study was treated as a classification problem, whereas the second was a regression problem. The performance of the model was evaluated using metrics coherent to each case study, such as accuracy and F1-Score for the first case and R2, Mean Absolute Error (MAE), and Root Mean Square Eeviation (RMSE) for the second case study. Ex plainability was also explored using the Shapley Additive Explanations (SHAP) method. The INMET case study obtained 93.3% accuracy in predicting one day in the future. In SAMAE case study, it was possible to reach R2 of 0.984 in the complete dataset, but only 0.779 for the outliers, that are the cases when floods happened. [resumo fornecido pelo autor]