Road Extraction from High-Resolution Remote Sensing Images Based on Road Branch Enhanced U-Net

Extracting roads from high-resolution remote sensing images is pivotal in navigation, mapping and other fields. Compared with traditional machine learning methods to extract road, deep learning methods are better in expressing the feature of targets relying on the powerful feature extraction ability of neural network. However, existing models usually neglect the redundancy of information in feature transfer. Meanwhile, the extraction effect of tiny roads is poor. As a result, a RBEU-Net (Road Branch Enhanced U-Net) model for road extraction has been proposed based on the Road Branch Attention (RBA) Module and Road Branch Enhanced (RBE) Loss. First, we incorporate RBA mechanism at the skip connection structure, which not only resolves the problem that skip connection causes the redundancy, but also enhances the capacity of feature extraction of the network. On this basis, Road Branch Enhanced Loss has been proposed to enforce the model to enhance the learning of road branches and other tiny roads based on road information. Experiments on the Massachusetts roads and DeepGlobe road extraction datasets demonstrate that RBEU-Net outperforms other methods, especially in extracting road branches and small roads. RBEU-Net achieved a Recall of 86.09% and 81.02% respectively on the two datasets, with a notable improvement of 7.41% and 7.06% compared to the basic U-Net architecture.