Método Ensemble de Modelos Híbridos para Detecção de Barragem de Rejeitos em Imagem de Satélite

Diego A. B. Cardoso; Matheus R. Guimarães; Roberta B. Oliveira

doi:10.20906/CBA2022/3183

Diego A. B. Cardoso Departamento de Ciência da Computação, Universidade de Brasília, DF
Matheus R. Guimarães Departamento de Ciência da Computação, Universidade de Brasília, DF
Roberta B. Oliveira Departamento de Ciência da Computação, Universidade de Brasília, DF

DOI: https://doi.org/10.20906/CBA2022/3183

Keywords: convolutional neural network, support vector machine, ensemble, hybrid model, tailings dam

Abstract

In Brazil, there is a high rate of irregular dams, including places that are not even cataloged. These dams have a high potential for socio-environmental damage and a possible rupture would be capable of destroying rivers and villages. This paper proposes an ensemble method composed of hybrid models, based on Convolutional Neural Network (CNN) and Support Vector Machine (SVM) for detecting tailings dams in satellite imagery. The database used was BrazilDAM, considering 1924 multispectral images, where some combinations were evaluated using 13 different bands. The proposed method achieved 96.1% accuracy, 96.8% F1- score, and 98.7% AUC. The application of the proposed method shows better results than the traditional methods proposed in the literature.