Multiscale Global Attention Network With Edge Perceptron for Automatic Road Extraction From Remote Sensing Imagery

Automatic road interpretation using remote sensing images is crucial for intelligent city construction and is widely applied in various domains such as automatic driving navigation, cartography, and urban planning. Recently, deep learning algorithms, especially for convolutional neural networks (CNNs) and Transformers, have been utilized with large-scale remote sensing datasets to extract abundant semantic features, significantly improving the accuracy and efficiency of road extraction. However, these models ignore the correlation between multiscale local context and global semantics, which could cause fragmentary prediction in complex remote sensing environments. In addition, the edge features of roads often cannot be accurately constructed due to the lack of semantic guidance. To address the aforementioned issues, this study developed a hybrid deep neural network integrating CNN and Transformer structures. In the encoder, a multiscale global attention pyramid (MGAP) is constructed to enhance the overall semantic representation of the road with a local context. The road edge perceptron is designed in the decoder to improve edge prediction accuracy by establishing hierarchical spatial attention. Quantitative experiments and visual analysis on two public road datasets have confirmed that the proposed network architecture and modules can improve road extraction accuracy with high efficiency (achieving an average 71% IOU and 83% F1 score).