Deep learning aplicada a segmentação semântica semiautomática de cicatrizes de deslizamentos de terra

Abstract

Each year, natural disasters are responsible for countless losses to society, both economically and socially. Landslides are disasters that can be caused by natural factors like heavy rains, earthquakes, characteristics of the soil and the topography of the region, as well as human actions like deforestation, constructions, ore mining and forest fires, and among others. The purpose of this work is to apply a methodology of landslide scars mapping, based on the application of fully convolutional neural networks, to detect and learn how to map landslide scars, and from that, to produce and make a publicly available and highly reliable landslide database for the states of Rio Grande do Sul, Santa Catarina and Paraná. In order to do so, the landslide points obtained from the previous step were manually sorted, and the best ones were submitted for expert analysis, to confirm those as definitive landslide scars. This database can make possible the development of landslide susceptibility maps, and be the first step of the development of alert and prevention system. This way, the negative consequences of this kind of disaster on society can be reduced. The F1-Score is used to measure overall performance in neural networks. For the U-Net architecture used in this work, the F1-Score was 0.7433, 0.7141 e 0.6825 for training, validation and test, respectively.