TSNet: A Two-stage Segmentation Network Guided by Uncertainty for Remote Sensing Image Road Extraction

Abstract

Road extraction from high‑resolution remote‑sensing images plays a vital practical role in multiple application scenarios including urban layout planning and autonomous driving systems. Nevertheless, current road‑extraction approaches still suffer from several bottlenecks under complex ground environments, including indistinct object boundaries, missing fine‑grained features and fragmented road outputs. Aiming to mitigate the above defects, this paper puts forward an uncertainty‑guided two‑stage segmentation network termed TSNet. The proposed framework divides road segmentation into two successive subtasks: global‑level semantic understanding and local boundary optimization. A cross‑layer residual feature enhancement module (CRM) is adopted to extract and aggregate multi‑scale feature information. On the basis of multi‑scale fused features, the first segmentation stage produces preliminary probability maps, and an uncertainty‑aware learning module (UALM) is introduced to measure the confidence degree for each pixel‑wise prediction. In the subsequent second stage, the local refinement module (LRM) takes the generated uncertainty map as guidance to dynamically concentrate on regions prone to prediction errors, and implements adaptive optimization for road edge areas. Quantitative experiments on the DeepGlobe dataset prove that the presented model achieves competitive performance in maintaining road connectivity, yielding Accuracy, Precision, Recall, F1‑score and Intersection over Union (IoU) of 98.68%, 83.72%, 82.36%, 82.85% and 71.07%, respectively. Further validation on the CHN6‑CUG dataset obtains an Accuracy of 96.10%, Precision of 80.48%, Recall of 79.56%, F1‑score of 79.37% and IoU of 66.20%. When compared with state‑of‑the‑art competing algorithms, our method achieves optimal results for all five evaluation metrics. In addition, qualitative visualization results reveal that the model exhibits prominent strengths in retaining connected road topology and recovering intricate edge contours, which effectively enhances the robustness and prediction accuracy of road extraction for complex real‑world scenes.