DF-DRUNet: A decoder fusion model for automatic road extraction leveraging remote sensing images and GPS trajectory data

Accurate road networks are of great importance to online food delivery (OFD) services. In recent years, various data sources have been used to extract road information. Remote sensing images and Global Positioning System (GPS) trajectories can provide complementary information about roads, and the fusion of these two data sources allows to enhance the accuracy of automatic road extraction. To make full use of the information available from these two data sources, we developed a decoder fusion model based on the dilated Res-U-Net (DF-DRUNet) which fuses the remote sensing images and GPS trajectories in an efficient way to extract the road network. The DF-DRUNet model is built on two components: First, two independent dilated Res-U-Net models are used, where one model uses remote sensing images as input whilst the other model takes GPS trajectories as input. Second, we fused the decoders from the modalities based on a gated fusion module, which can help to learn the selection from these two input modalities. Based on the road extraction from the DF-DRUNet model, we also developed various refinement strategies, i.e., noise removal, skeleton extraction, topology construction, and vectorization. Numerical experiments were conducted using the DF-DRUNet model and baseline models from the real dataset of remote sensing images and GPS trajectories. The quantitative evaluation shows that the DF-DRUNet model can integrate remote sensing images and GPS trajectories effectively and achieve the highest performance of F1-score (0.857) and IoU (0.746) among all baseline fusion models. Moreover, the proposed DF-DRUNet model needs relatively fewer parameters and takes shorter training time. •Automatic road extraction leveraging remote sensing images and GPS trajectory data.•A designed decoder fusion model is used for automatic road network extraction.•A novel gated fusion module is designed to help to learn the modality selection.•Results are validated by a proposed deep learning-based model yielding up to 85.7%.