Utilização de inteligência artificial para mapeamento de estradas não oficiais na Amazônia

Abstract

Mapping unofficial roads in the Amazon has a really important role in understanding the dynamics of access and utilization of the Biome. The Amazon Institute of People and the Environment (IMAZON), had been conducting this procedure in a manual way, making this precise process, longstading. With that in mind, this work propose a new way of making this mapping, using artificial intelligence and remote sensing technologies for accelerate the detecting of roads, making possible to tracking constantly activities in the Amazon. The model of neural net created to accomplish this task was made using a modified U-Net and using Google’s training and prediction services. The model was tested with five locations inside state of Pará, all of those with 1200 km² of area, and it produced promising results, which indicate the capacity of the network de make predictions over satellites images and identify the existence of roads, being capable of detect different formations - Dendritic, geometric and “Fishbone”. Besides that, another test to compare results between two types of images with different resolutions was conducted, showing improvement with the rising of the resolution, when made with the right input conditions. At last, this work accomplished its objective, presenting a new method to map unofficial roads in the Amazon, this being faster, less expensive and capable of keep up with the dynamic of the region.