Road Multilabel Semantic Segmentation from Satellite Images Using Convolutional Neural Networks

Road semantic segmentation in satellite images is a very important and studied field in the state of the art since having road infrastructure is quite significant for decision making in various areas of a country. The recovery of this information occurs through time-consuming large processes, and they may involve a considerable logistical displacement. In this study, we propose two U-shaped architectures to the problem of road multilabel segmentation from high resolution satellite images. Since we do not have a proper dataset to optimize our model, a procedure is defined to create samples from a base dataset. Quantitative evaluations achieve values above 93.69% on the considered metrics. Likewise, qualitative evaluations exhibit an appropriate generalization of the segmentation models since inferences protrude over ground truths in the experiments. The full implementation of this study is available at https://github.com/edson2495/road-multiclass-segmentation.