Multilevel distillation learning road extraction model with global feature awareness and fusion

Based on an end-to-end convolutional neural network (CNN), the paper proposed a road extraction model combining spatial attention mechanism, global information perception, and feature fusion module. It can extract road information with detailed spatial features and comprehensive semantic information and improve the inference speed of road information extraction. Firstly, the spatial attention mechanism and global information awareness module were used to obtain road contextual information and improve the spatial information representation of low-level features. Then the feature fusion module was built to consider the channel and semantic information to eliminate the semantic gap between low- and high-level features. Finally, the multilevel knowledge distillation learning strategy was used to reduce network parameters and computational complexity to obtain road information quickly and accurately. The training, validation, and evaluation were conducted on the Deep Globe, Massachusetts, and Beijing-Tianjin New Town remote sensing image road datasets. The results show that the proposed model provides high accuracy and a suitable extraction effect. It achieves excellent road information extraction ability both in satellite remote sensing data sources and UAV remote sensing data sources. The  scores reach 79.36%, 78.42%, and 84.27%, respectively, outperforming other road extraction models. Meanwhile, the multilevel knowledge distillation learning strategy significantly improves the accuracy and generalization ability of the model. The index IOU values were improved by 0.29%, 0.77%, and 0.46%, respectively. It achieves better results regarding model accuracy and network parameters and has broad application prospects.