Evaluating the performance of convolutional neural networks to detect deforested regions in the Brazilian Legal Amazon using LandSat-8 satellite images

Abstract

In this study we used Convolutional Neural Network architectures to detect deforested regions in the Brazilian Legal Amazon, using LandSat-8 satellite images. To improve the network performance, some methods for improving generalization and different optimization methods were employed. Due to class imbalance, a new technique was used for training the networks called mosaic image training. From the satellite images, small rectangular samples of deforested and non-deforested areas were extracted. From these samples, a large image is created, with almost the same number of small deforested rectangles and small non-deforested rectangles. To evaluate the network performance the following metrics were used: accuracy, precision, sensitivity, specificity, and F1-Score. The best obtained accuracy in this study was 99.97%.