Abstract
Amazon rainforest deforestation severely impacts the environment in many ways, including biodiversity reduction, climate change, and so on. A key indicator of deforestation is the sudden appearance of rural/unofficial roads, usually exploited to transport raw materials extracted from the forest. To early detect such roads and prevent deforestation, remote sensing images have been widely employed. Precisely, some researchers have focused on tackling this task by using low-resolution imagery, mainly due to their public availability and long time series. However, performing road extracting using low-resolution images poses several challenges, most of which are not addressed by existing works, including high inter-class similarity, complex structure, etc. Motivated by this, in this paper, we propose a novel approach to perform road extraction on low-resolution satellite images based on contextual and pixel-level decision fusion. We conducted a systematic evaluation of the proposed method using a new dataset proposed in this work. The experiments show that the proposed method outperforms state-of-the-art algorithms in terms of intersection over union and F1-score.