Assessing Urban flood susceptibility using tree-based machine learning models

Código

I-EBHE0091

Autores

Marcos Roberto Benso, Maria Clara Fava, Marina Batalini de Macedo, Anaí Floriano Vasconcelos, Beliana Cavalcante Sawada de Carvalho, Maria Elisa Leite Costa, ALAN VAZ LOPES, Javier Tomasella

Tema

WG 1.03: Urban Water - Urbanization phenomenon & adequate water management

Resumo

Urban floods are among the worst kinds of hazards and one of the more challenging to predict because of their suddenness and small scale.. In this context, flood susceptibility mapping is a fundamental tool to help urban planners and decision-makers for planning flood risk management measures. Despite the growing interest in applying data-driven models, fewer examples have been applied in Brazil. Therefore, there is the need to further develop these techniques considering the Brazilian context. The aim of this paper is to test data-driven models for assessing urban flood susceptibility. In this article, classification algorithms were tested including Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Regularized Random Forest (RRF) to assess urban flood susceptibility in the Aricanduva catchment in the São Paulo Metropolitan Region. Urban flood events are categorized according to the disaster occurrence database of the city of São Paulo. Considering the distribution of events, we considered the following susceptibility classes: not susceptible (1,000 data points), waterlogging of drainage systems (196 data points), fluvial flood (207 datapoints) with a total of 1,403 data points collected from the years 2013 to 2024. The urban flood influential factors were derived from LiDAR resampled data to 5 meters horizontal resolution using Triangulated irregular network (TIN) and the following factors were calculated, Slope, Slope Length, Slope Steepness, Topographic Position Index (TPI), Topographic Index (TI), Height Above the Nearest Drainage (HAND), Profile Convergence Index (TCI), Stream power index (SPI), Aspect, and Drainage Direction. The dataset was split into training (80% of data) and testing (20 % of data), and all models were trained using 10-fold cross-validation with 3 repetitions. All models indicated HAND and TCI as the two most important factors for classifying urban floods. The highest Kappa and Accuracy was achieved by the RRF model (Kappa = 0.72, Accuracy = 0.87), followed by the RF model (Kappa = 0.70, Accuracy = 0.86) and XGBoost model (Kappa = 0.65, Accuracy = 0.83). The Specificity of all models for the three classes (not susceptible, waterlogging, and fluvial flood) were higher than 90%, indicating that the rate of false negatives is low. The Sensitivity of the classes was high for not susceptible class was 98%, indicating that the false positive rate is very low for this class. On the contrary, for waterlogging and fluvial flood the Sensitivity was 54% and 55%, respectively. This indicates that the models were successful for predicting the cases of regions susceptible and not susceptible to floods, but not to differentiate between the cases that there was waterlogging of the drainage system and fluvial floods. The results showed that tree-based models could successfully predict regions with urban flood susceptibility. The next step of this study is to further evaluate the factors that influence the classification of waterlogging of drainage systems and fluvial floods, this involves adding quality assurance of disaster report databases, including and testing more predicting variables and testing different machine learning models.