Proposta metodológica para a avaliação de suscetibilidade a escorregamentos: contribuições do aprendizado de máquina no município de Belo Horizonte, MG

Abstract

Quickly urbanization and the presence of numerous risk areas have made landslide susceptibility mapping a crucial aspect of urban planning in Belo Horizonte, southeastern Brazil. This study proposes novel machine learning (ML) models for landslide susceptibility classification in the region. An integrated database was developed, incorporating landslideconditioning factors and landslide inventories from URBEL and Belo Horizonte's Civil Defense. Four ML algorithms 3 Artificial Neural Networks (ANN), Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR) 3 were trained to classify landslide occurrence and non-occurrence locations. The ML task was treated as binary classification, with results expressed as probabilities to generate susceptibility maps. Twelve landslide susceptibility maps were produced for the Belo Horizonte region, derived from three evaluated scenarios: the entire municipality, the southern portion (anisotropic model), and the northern portion (isotropic model). Results indicated that RF outperformed other models with an accuracy of 81.17%, followed by SVM with 80.95%. ANNs achieved an accuracy of 79.03%, while LR exhibited the lowest performance with 76.45% in the overall scenario. In the anisotropic scenario, accuracy values for the test data were 82.87%, 81.82%, 81.29%, and 78.52% for RF, SVM, ANN, and LR, respectively. For the isotropic scenario, there was a general decline in performance for all models, with RF and SVM achieving approximately 77% accuracy, ANN showing 74.13%, and LR with an accuracy of 73.28% for the test data. RF and SVM emerged as the most effective and reliable techniques across the three evaluated scenarios, offering consistent and promising performance in identifying landslide-susceptible areas. The study presents novel approaches for utilizing artificial intelligence in generating susceptibility maps in Brazil, providing a promising perspective to reduce the subjectivity inherent in traditional processes. While implementing artificial intelligence techniques is complex, these approaches enable continuous data updates and revisions, resulting in more accurate and effective maps, along with the automation of tasks that can be optimized. This, in turn, has the potential to significantly transform cities' capacity to deal with geological risks, contributing to the development of safer and more resilient communities.