Information extraction of roads from remote sensing images using CNN combined with Transformer

Deep learning-based methods for information extraction of roads from high-resolution remote sensing images face challenges in extracting information about both global context and edge details. This study proposed a cascaded neural network for road segmentation in remote sensing images, allowing both types of information to be simultaneously learned. First, the input feature images were sent to encoders CNN and Transformer. Then, the characteristics learned by both branch encoders were effectively combined using the shuffle attention dual branch fusion (SA-DBF) module, thus achieving the fusion of global and local information. Using the SA-DBF module, the model of the features learned from both branches was established through fine-grained interaction, during which channel and spatial information in the feature images were efficiently extracted and invalid noise was suppressed using multiple attention mechanisms. The proposed network was evaluated using the Massachusetts Road dataset, yielding an overall accuracy rate (OA) of 98.04%, an intersection over union (IoU) of 88.03%, and an F1 score of 65.13%. Compared to that of mainstream methods U-Net and TransRoadNet, the IoU of the proposed network increased by 2.01 and 1.42 percentage points, respectively. Experimental results indicate that the proposed method outperforms all the methods compared and can effectively improve the accuracy of road segmentation.