Use of artificial neural networks for shallow planar landslides mapping.

Author(s)

Caio da Silva Azevedo

• Document type: Master's Dissertation
• Press: São Paulo.
• Institution: Universidade de São Paulo (USP). Escola Politécnica (EP/BC)
• Defense date: 2022-03-24

Examining board members

Mauro de Mesquita Spinola; Alessandra Cristina Corsi; Vagner Luiz Gava

• Advisor: Mauro de Mesquita Spinola

Abstract

Landslides can cause serious environmental, economic and social consequences. One of the first steps in risk management and prevention of these events is the mapping of the territory\'s natural and anthropic susceptibilities, which provides subsidies for territorial planning and indications for urban expansion. In this sense, the mapping of susceptible areas acts as an important tool for local managers, in order to locate the areas most sensitive to these natural disasters and to think of strategies for land use. Among the techniques currently used for this mapping, artificial neural networks (ANN) stands out, considered one of the most accurate because of its high accuracy, ability to learn and generalization of results. Thus, this study aimed to propose a methodology for mapping areas susceptible to shallow planar landslides, through artificial neural networks, contributing to the context of smart cities. As input data for training the networks, seven attributes were used, most extracted from digital terrain models (DTM): slope, aspect, elevation, land use, lithology, topographic wetness index (twi) and curvature, from an inventory of shallow planar landslide scars, which occurred in the coastal municipality of Guarujá-SP in 2020, an atypical period, with extreme weather conditions, characterized by rainfall above 200 mm. The equivalent number of samples was created for the points of non-occurrence in different scenarios. Then, the importance of the attributes was evaluated, concluding that their relevance varies according to the data set used for ANN training. Finally, among the best configurations of the neural networks, trained with the backpropagation algorithm, the accuracy of 98.50% and 95.43% was obtained, a metric superior to the average of 86.74% identified in the works in the literature and, for a sample not used in the training of networks, the maps generated with RNA also performed better than the map of the IPT, the institute responsible for mitigating landslides in the state of São Paulo, whose accuracy was 68.3%, demonstrating that the proposed model meets with the objective of identifying areas susceptible to mass movements. (AU)

• FAPESP's process: 19/14011-0 - Machine learning for river flow modeling and flood prediction
• Grantee: Caio da Silva Azevedo
• Support Opportunities: Scholarships in Brazil - Master