Unsupervised Domain Adaptation For Automatic Road Extraction From Aerial Images

Automatic road extraction has been an active research topic for the past few decades, in which methods based on Deep Learning are state-of-the-art. However, to train those models, datasets containing a mass of dense annotations are a prerequisite and do not deal with the different scenarios that the road network has. To address these challenges, we propose an approach based on Unsupervised Domain Adaptation (UDA). Specifically, our UDA-based approach is evaluated on three strategies for Convolution Neural Networks: U-Net architecture, the use of multi-level and multi-scale information (DeepLabv3+), and the attention mechanism in U-Net. The experiments show that among the strategies, the attention mechanism enhances the connectivity of the road network over some shortcomings, and its contribution against the baseline reaches an increase of 2% and 3% in F1-Score and Intersection over Union metrics, respectively.