Detecção de áreas desmatadas na porção sul do estado do amazonas, utilizando técnicas de extração de características e redes neurais artificiais

Abstract

The monitoring of deforestation in the Amazon is a great challenge due to, among other factors, the enormous territorial extension. In this sense, images generated from remote sensing systems have been used to perform the detection of deforested regions. To facilitate this work, images generated from remote sensing systems have been used to perform the detection of deforested regions. In this sense, since 1988, the National Institute of Space Research has been monitoring the Amazon Forest through LANDSAT optical images acquired through remote sensing systems. Currently, the methodology adopted by INPE is based on the manual demarcation of deforested areas based on the knowledge of specialists. However, doing this work manually requires time and effort. In the literature, we can find several works that propose to perform an automatic classification of remote sensing images. This work proposes a methodology for the study of deforestation areas using LANDSAT-8/OLI optical images obtained by remote sensing. It is proposed the association of methods to improve the generalization of neural networks, such as early stopping and regularization, with characteristic extraction techniques, such as scalar selection, principal component analysis and linear discriminant analysis. The performance of the methodology is evaluated using measures such as precision, sensitivity, specificity, area under ROC curve. Finally, similarity metrics, Dice and Jaccard, were calculated. The best results were obtained with the early stop generalization method associated to the linear discriminant analysis technique. The study was carried out in three regions of the State of Amazonas, located in the municipalities of Apuí, Humaitá and Lábrea. In addition, the proposed methodology was compared with three papers in the literature that also used optical images from the Amazon region.