A Global Context-aware and Batch-independent Network for road extraction from VHR satellite imagery

Road extraction is to automatically label the pixels of roads in satellite imagery with specific semantic categories based on the extraction of the topographical meaningful features. For governments, timely and accurate road mapping is crucial to plan infrastructure development and mobilize relief around the world. Recent advances in deep learning have shown their dominance on road extraction from very high-resolution (VHR) satellite imagery. However, previous road extraction based on deep learning mainly stacked the multiple convolution operators and failed to predict the contextual spatial relationship correctly. Besides, the precision of cross-domain road extraction is limited by an insufficient amount of labeled data and the transferability of the model. To remedy these issues, a Global Context-aware and Batch-independent Network (GCB-Net) is proposed, which is a novel road extraction framework extract complete and continuous road networks. In GCB-Net, the Global Context-Aware (GCA) block is added to the encoder-decoder structure to effectively integrate global context features. The Filter Response Normalization (FRN) layer is used to enhance the original basic network, which eliminates the batch dependency to accelerate learning and further improve the robustness of the model. Experimental results on two diverse road extraction data sets demonstrated that the proposed method outperformed the state-of-the-art methods both quantity and quality. Moreover, to test the robust generalizability of the proposed method, the proposed CHN6-CUG Roads Dataset was used for spatial transfer evaluation, and GCB-Net achieved significantly higher transferability than other methods.