Previsão de desmatamentos ilegais baseado em dados temporais de sensoriamento remoto

Abstract

In order to study and understand the dynamics of illegal deforestation, it is necessary to measure information about deforested areas. One of the ways to do this is through remote sensing using satellite images. However, obtaining and processing these images is not easy, mainly due to their complexity of space and also the difficulty of obtaining validated points with the occurrence of the desired phenomenon, in addition to the creation of methods to automatically perform the classification of these areas. Therefore, it is useful to make available a database containing information that has already been processed and validated about spots that have suffered deforestation. This type of database makes it possible to reduce efforts in future research, besides to providing a common database to be used as a benchmark in machine learning solutions. In this situation, the Remote Sensing Deforestation Dataset (RSDD) was created containing tabulated data and images of areas affected by deforestation. The RSDD has data validated by a State agency, it has a size less than 1% of the size of the original remote sensing images, and it is publicly available. With such a base, it is possible to create machine learning models to be able to forecast deforested areas. In this work, a model of Artificial Neural Networks was built capable of detecting areas that will be deforested, obtaining an accuracy of 93.2% and an accuracy of 94.8% on the best tests.