文献类型：期刊文献

英文题名：UGD-DLinkNet: An Enhanced Network for Occluded Road Extraction Using Attention Mechanisms and Uncertainty Estimation

作者：Yang, Peng[1,2];Xiao, Huanling[3];Lin, Cong[3];Xie, Xia[1]

机构：[1]Hainan Univ, Sch Comp Sci & Technol, Haikou 570228, Peoples R China;[2]Guangdong Ocean Univ, Sch Math & Comp, Zhanjiang 524088, Peoples R China;[3]Guangdong Ocean Univ, Sch Elect & Informat Engn, Zhanjiang 524088, Peoples R China

年份：2025

卷号：18

起止页码：24144

外文期刊名：IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

收录：SCI-EXPANDED(收录号：WOS:001585542000001)、、EI(收录号：20253719162677)、Scopus(收录号：2-s2.0-105015422319)、WOS

基金：This work was supported in part by the Haikou Science and Technology Plan Project under Grant 2023-053, in part by the Stable Supporting Fund of Acoustic Science and Technology Laboratory under Grant JCKYS2024604SSJS00301, in part by the Hainan Province Science and Technology Special Fund under Grant ATIC-202302001, and in part by the Guangdong Ocean University through the Program for Scientific Research Start-Up funds under Grant 060302112405.

语种：英文

外文关键词：Roads; Feature extraction; Remote sensing; Attention mechanisms; Uncertainty; Noise; Accuracy; Robustness; Convolution; Semantics; Attention mechanism; knowledge distillation; Monte Carlo (MC) dropout; road extraction

外文摘要：Road information extracted from high-resolution remote sensing images is crucial for urban planning and traffic navigation. However, challenges such as complex and variable geometric and contextual features of roads, occlusion by trees and buildings, and labeling errors in datasets present significant difficulties. To address these issues, we propose the UGD-DLinkNet model. Specifically, attention mechanisms are employed to enhance road feature extraction in occluded and complex regions by adaptively focusing on informative spatial and channelwise cues, while model uncertainty estimation helps reduce the impact of noisy annotations by guiding learning toward more reliable predictions. First, we introduce a hybrid attention module to strengthen the encoder's ability to capture essential features and structural information. Second, a self-attention unit is integrated into the bridging network to form a dilated convolution attention module (DCAM), improving the perception of semantic features across scales. Third, a channel attention module refines skip connections, bridging shallow and deep semantic information. Finally, we incorporate Monte Carlo dropout (MC dropout) and propose an uncertainty-guided knowledge distillation strategy to mitigate labeling noise. Experiments conducted on the Massachusetts, CHN6-CUG, and DeepGlobe benchmark datasets demonstrate the comprehensive superiority of UGD-DLinkNet. On the Massachusetts dataset, our model achieves the highest F1-score of 78.75%. On CHN6-CUG, it obtains the best overall performance with an F1-score of 76.50% and an IoU of 61.95%, exceeding the best results among recent mainstream methods by 4.68% and 2.33%, respectively. On DeepGlobe, UGD-DLinkNet also achieves top performance, with an F1-score of 81.49% and an IoU of 68.77%, highlighting its strong generalization and robustness across diverse scenarios.