DAGMNet: A Dual-Branch Network for Connectivity-Preserving Road Extraction

Abstract

Accurate extraction of road networks from high-resolution remote sensing imagery is critical for autonomous driving, urban planning, and intelligent transportation systems. However, existing methods often struggle to preserve connectivity and topology due to the narrow and curvilinear nature of roads, complex background clutter, and frequent occlusions caused by buildings or vegetation. Although some approaches introduce auxiliary data or structural priors to improve connectivity, reliance on additional sensor modalities may reduce generalization and increase deployment cost. To address the loss of connectivity in thin, curvilinear, and partially occluded roads during deep semantic abstraction, we propose DAGMNet, a dual-branch network designed for connectivity-preserving road extraction. The network is organized into two coupled subsystems. The first subsystem performs gradient–morphology-aware structural encoding by deriving a gradient-magnitude structural prior from the input image and using dynamic snake convolution to model elongated and curvilinear road geometry. The second subsystem performs cross-branch feature coupling through anisotropic enhancement and multiscale gated cascaded fusion, allowing structural cues and semantic context to interact progressively across directions and scales. These two subsystems work jointly to recover continuous road structures rather than functioning as isolated plug-in modules. On CHN6-CUG, SpaceNet, and Massachusetts Roads, DAGMNet obtains F1 scores of 78.41%, 75.17%, and 78.18%, and intersection-over-union (IoU) scores of 64.48%, 60.78%, and 62.79%, respectively. On the two datasets evaluated with the Average Path Length Similarity (APLS) metric, it obtains scores of 67.20% and 76.40%, providing additional evidence of road-network connectivity under the evaluated protocols.