CoA-DLinkNet: Connectivity-Enhanced Dual-Branch Road Extraction Network Based on D-LinkNet

Linghan Li, Heliu Chen, Renjie He, Yuchao Dai, Mingyi He

Abstract: Extracting roads from high-resolution remote sensing image data presents a challenging task in the field of remote sensing image processing, which is of great significance for urban planning, vehicle navigation, and geographic information system updates. In 2018, the champion of the DeepGlobe Road Extraction Challenge proposed a creative solution named as D-LinkNet, which employed a dilated convolution cascade module to expanded the network receptive field and achieved impressive results with a concise network structure. However, D-LinkNet still suffers from low connectivity caused by road breaks. Therefore, based on D-LinkNet, we have made three improvements in this paper: 1) incorporating a strip pooling module to capture long-range anisotropic contextual information, 2) employing a parallel upsampling structure in the decoder to supervise the shallow layers, and 3) introducing a connectivity branch with a connectivity attention module to model local connectivity of roads. We have named the upgraded network as CoA-DLinkNet. Our experimental results have demonstrated that the proposed network significantly improves the prediction accuracy and connectivity of the extracted road.