Redes neurais totalmente convolucionais aplicadas à identificação de focos de desmatamento

Abstract

Deforestation is one of the main environmental threats to the ecological balance worldwide, as well as one of the primary sources in terms of greenhouse gas emissions. Because of this, the development of methodologies with the potential to monitor deforestation increasingly accurate and scalable is essential for the attempt to combat and minimize this problem. Considering the rapid development of technologies such as data mining and machine learning and the characteristics of deep learning models, capable of performing the semantic segmentation of images, there are numerous potentialities in using these types of structures to create methods for mapping automated and real-time deforestation. In this work, therefore, a Fully Convolutional Neural Network (FCN) architecture called U-Net was studied and evaluated to be applied in a methodology to identify forest cover loss in two Brazilian biomes. This dissertation is presented in the form of three articles. The first presents the process of choosing the FCN to meet the objectives of the problem, as well as an initial assessment of the possibility of using the structure to identify deforestation in an Amazon region. Thus, it was found that U-Net presented the metrics improvements and, therefore, was selected to continue with the studies. The second article, on the other hand, evaluated the potential of U-Net more deeply, identifying the best configuration and evaluating different compositions of multispectral bands and their application to the Atlantic Forest and Amazon biomes. The study indicated that U-Net achieved the best results in the classification of forest and non-forest using 4 spectral bands (Red, Green, Blue and Infrared) and was able to map the forest regions satisfactorily, reaching, for the network trained with images of both biomes, accuracy, precision, recall and F1- Score values of 0.9880, 0.9871, 0.9882, and 0.9876, respectively. Finally, the third article presents a complete and automated methodology using the best U-Net configurations obtained in previous studies and applying them in a method capable of identifying deforested areas in the Atlantic Forest and Amazon biomes using Sentinel-2 images both temporally and spatially spaced. It was observed that the methodology is able to identify, with a high degree of accuracy, areas that have suffered loss of vegetation cover, which can be applied in monitoring systems or serve as a complementary method in existing systems.