Modelo de classificação para monitoramento de desmatamento na Amazônia brasileira utilizando redes neurais artificiais.

Abstract

The Amazon rainforest boasts a rich abundance of fauna and flora, recognized as the world’s greatest biodiversity. Additionally, it plays a crucial role in maintaining global climate balance. Nevertheless, illegal deforestation in the Amazon has significant international repercussions, negatively affecting Brazilian biodiversity, climate change, among other aspects. Forest monitoring through satellites allows the visualization of deforested areas. Remote sensing images enable the classification of these areas and the monitoring of potential causes of deforestation. In this context, artificial neural networks play an important role and have been utilized in this research field. Therefore, the objective of this work is to develop a classifier through artificial neural networks capable of analyzing spatial images and identifying deforested or non-deforested areas. Geoprocessing techniques such as the Linear Spectral Mixture Model were studied, alongside advanced Artificial Neural Network techniques like Data Fusion. The results demonstrated that Multilayer Perceptron (MLP) and Convolutional Neural Networks (CNN) architectures were able to classify deforested areas with a ROC curve area above 95%. Furthermore, data fusion yielded even more significant results, matching the ideal ROC curve and correctly classifying, in three of the created models, all samples presented to the network. Thus, it can be said that the classifier provided accurate results, contributing to the scientific community working for the well-being of the Brazilian Amazon.