Content area
This study investigates the development of machine learning for monitoring flood-susceptible areas and applies the random forest (RF) and support vector machine (SVM) algorithms to predict flood susceptibility in the Cachoeira River watershed, located in the municipality of Joinville (Santa Catarina, Brazil). Comparative evaluations of the predictive efficacy between the RF and SVM models were performed, analyzing their performance in capturing non-linear relationships among the conditioning variables identified in the study area. Classification evaluation metrics and hyperparameter optimization techniques were utilized to establish a rigorous procedure for selecting the best models. For the final model, fourteen geo-environmental variables were integrated, encompassing topographic, geological, hydrological, and meteorological parameters, selected through a systematic review of flood conditioning factors. The computational implementation, developed in the R language, employed packages such as caret for hyperparameter optimization and k-fold cross-validation, raster for processing georeferenced data, and rgeos for geospatial analyses, ensuring methodological reproducibility and scalability. The models demonstrated elevated performance, with accuracies of 99.60% (RF) and 99.40% (SVM-R), attributed to the robustness in the selection of factors, the geomorphology of the basin, and the statistical representativeness of the sampling. The generated cartographic output, in addition to identifying critical zones with high resolution (10x10m), allows for dynamic updates through the inclusion of new environmental data or climatic scenarios, consolidating it as a strategic tool for risk management. The results show the high accuracy of supervised learning techniques in assimilating non-linear relationships in natural systems, reinforcing their potential for public policies on disaster mitigation in urban regions and with complex geomorphologies.