Uso de modelos de aprendizado de máquina para previsão de movimentos de massa na Cidade do Rio de Janeiro

Abstract

Landslides are recurring phenomena in mountainous regions, and their rate of occurrence intensifies during rainy periods. Correlation models of precipitation vs. landslides are based on the Theory of Effective Stress and have been playing a role as tools to mitigate the catastrophic effects of landslide occurrences. Such models are predominantly empirical, and, regardless of their typology, they are developed from a database. Currently, in the city of Rio de Janeiro, the Alerta Rio System is based on a graphic correlation model and uses the rate of precipitation as a reference to trigger sirens and evacuate risky areas. The main objective of this research is to propose a classifier as an alternative or as na complement to the present model in operation by adopting the databases of landslides and precipitation from the period 2009 to 2015 in the city of Rio de Janeiro. The best classifier model will be chosen among three distinct models built by three different algorithms: Random Forest (RF), Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP). Each model is trained, validated and tested considering a period of 96 hours, which allows the incorporation of the precipitation of the date (24 hours) of the landslide occurrence and also the preceding rain that precipitated within 72 hours prior to the date of the landslide occurrence. Regarding the models developed, the RF classifier presented better results than the SVM or MLP classifiers. Therefore, it can be said that the Random Forest (RF) shows better performance, although, in general, the validation and testing errors of the models indicate that they did not achieve the intended performance. Furthermore, the study carried out in this research highlights the importance of the database available for building models based on machine learning techniques, mainly due to the fact that the performance evolution of such models is related to the quantity and quality of data for carrying out training, validation and testing. In the case in question, the existence of more data related to landslides could result in better performance and, consequently, greater confidence in the model.