Segmentação Semântica de Imagens Aplicada à Detecção de Desmatamento na Amazônia

• Autor(a) principal: Andrade, Renan Bides de
• Data de Publicação: 2020
• Outros Autores: [e-mail]
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UERJ
• Texto Completo: http://www.bdtd.uerj.br/handle/1/20286
• topic: Deep Learning; Deforestation; Semantic Segmentation; Change Detection; Processamento de imagens – Segmentação; Linguagem de máquina; Desmatamento – Amazônia (Brasil); Desmatamento; Segmentação Semântica; Detecção de Mudanças; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO

Deforestation is a worldwide problem which is responsible for serious environmental issues, such as biodiversity loss and climate change. Containing approximately ten percent of all biomass in the planet and habitat of one tenth of the known species, the Amazon biome has, in the last decades, being submitted to important deforestation pressure. Therefore, devising e cient deforestation detection methods is key to combat illegal deforestation and to give support to public policies targeting sustainable development in the Amazon. Aiming at contributing to the use of state-of-the-art technologies in the environmental management, this work implements and evaluates a deforestation detection approach based on a Fully Convolutional Deep Learning (DL) model for semantic segmentation, the DeepLabV3+. The results obtained with the use of the proposed approach are compared to those obtained with previously proposed patch classi cation DL-based methods (Early Fusion and Siamese Convolutional Network). In the reported experiments, it is employed a database, consisting of Landsat OLI-8 images acquired on di erent dates covering regions of the Amazon forest. The aim is evaluating the sensibility of the compared methods to the amount of training data. Additionally, distinct values of the parameters of the loss function used during training were evaluated. Results suggest that the tested variants of the proposed method signi cantly outperformed the other DL-based methods in terms of overall accuracy, F1-score and precision, while performed similarly in terms of the recall metric. Performance gain, when present, tended to be more substantial when limited amounts of images was used in the training.