Use of machine learning classification algorithms for detecting landslide inducing rainfalls in Petrópolis/RJ, Brazil

Código

I-EBHE0145

Autores

Artur Nonato Vieira Cereto, Gean Paulo Michel, Franciele Zanandrea, Ivanovich Lache Salcedo

Tema

WG 1.07: Understanding drivers & feedbacks of soil moisture variability

Resumo

Due to its topography, climate, and the unplanned growth of its urban areas, Petrópolis has experienced many landslides, particularly during heavy rain events, with variable scales of impact. In 2011 and 2022, two catastrophic disasters resulted in significant death tolls, economic damage, and left hundreds of people displaced. As a way of mitigating the economic impacts and loss of human lives caused by these disasters, Landslide Early Warning Systems (LEWS) come as crucial tools, mainly by assisting decision-makers in taking timely actions to evacuate the populace to safer areas. LEWS work by gathering and processing data related to the occurrence of these disasters, such as recorded and predicted rainfall, pore water pressure in the soil of landslide prone slopes, soil displacements in slopes etc. This study tested two machine learning classification models, one based on Support Vector Machine (SVM) and the other on Multilayer Perceptron (MLP) , a feedforward Artificial Neural Network (ANN), with the objective of classifying a rainfall event as landslide-inducing or not. To train the models, we gathered rainfall data from 33 rain gauges installed and maintained by the Centro Nacional de Monitoramento e Alertas de Desastres Naturais (CEMADEN) throughout the city of Petrópolis over the past 10 years. After performing clustering analyses to understand rainfall patterns along the area, we investigated the relation of this data with landslide occurrence records from 2015 to 2024 provided by Petrópolis? Defesa Civil. The use of SVM and ANN techniques showed promising results and should be considered to be incorporated into the development of future Landslide Early Warning Systems (LEWS). One of the largest challenge for construction of reliable LEWS in Brazil remains on obtaining accurate information about local and time of landslide occurrences.