Aplicação de redes neurais convolucionais na detecção de assentamentos precários em João Pessoa.

Abstract

Due to population growth and the expansion of informal settlements, it is necessary to monitor and map these places so that public policies can be developed aimed at solving the precarious nature present in these spaces. Some of the current solutions involve image classification based on machine learning algorithms, however, those present in the state of the art require the extraction of many features, which takes a lot of time and generates a large amount of parameters that need to be processed by the algorithms. This work presents the use of a Convolutional Neural Network, a U-Net with Inception ResNet-V2, as a solution for the feature extraction automation and parameter reduction in satellite images, focusing on the city of João Pessoa, in Paraíba, together with the segmentation of the images aiming at the detection and classification of precarious settlements in urban spaces. The model was evaluated using the Jaccard and Dice coefficients, which presented respectively 53% and 69%, in the test dataset.