High-precision road extraction is a critical technology for intelligent transportation, urban planning, and navigation applications. However, existing approaches face three primary limitations: (1) inadequate representation of complex scenes due to reliance on single-data sources, (2) low accuracy of traditional machine learning methods, and (3) a lack of systematic performance comparisons among deep learning models. To address these challenges, this study proposes a multi-source fusion approach, combining high-resolution remote sensing imagery with open-source geographic data to construct a refined, multi-city dataset. Quantitative evaluation metrics—including overall accuracy (OA), F1 score, and intersection-over-union (IoU)—were employed alongside qualitative visual analysis to systematically compare the semantic segmentation performance of deep learning models (DeepLabV3 and U-Net) against traditional maximum likelihood classification (MLC). The results demonstrate that DeepLabV3 achieves superior performance, surpassing UNet by 1.16% in OA, 8.53% in F1 score, and 13.05% in IoU. Compared to MLC, it exhibits even more significant improvements: 8.88% higher OA, 2.98% higher F1 score, and 17.83% higher IoU. Additionally, DeepLabV3 excels in boundary delineation, further validating the advantages of deep learning for high-precision road extraction.
Han et al. (Fri,) studied this question.