Segmentação semântica de áreas desmatadas utilizando Redes Neurais Convolucionais no sul da Amazônia Legal

Abstract

In this work, we propose the evaluation of convolutional neural networks architectures to segment deforested areas in the southern region of the Brazilian Legal Amazon, using Landsat-8 satellite images. In order to carry out this work, a mosaic image data set was elaborated, consisting of samples of deforested areas and forest areas extracted from the Landsat-8 satellite images. The deforested areas were identified through the use of deforestation data from the PRODES project at INPE. The architectures of convolutional networks used in our research were those proposed in the work of Serrão et al. (2020) and Miyagawa et al. (2018). The training of the networks was carried out over 32 epochs using SGDM, RMSProp and ADAM optimization methods and the L2 and Dropout regularization methods. The combining of the three architectures with these methods, resulting in 36 simulations. To measure the performance of the architectures for segmented deforested region, allowing a comparison between the models, the accuracy metric was chosen. After evaluating the performance of the models in the validation set, six of them were selected to be evaluated with the test set. The model that presented the best result, with an accuracy of 99.97%, was the one that used the following combination: CNN2 + RMSProp + Dropout. The results of this work were compared with the results of the work of Ortega et al. (2019), Adarme et al. (2020) and De Bem et al. (2020). Our work obtained results best perfomance that the results obtained by these authors. The results showed that the convolutional neural networks are capable of performing the task of classifying deforested areas with high performance.