Road networks support navigation, urban planning and disaster response, yet extracting them from high-resolution imagery remains difficult. Roads vary in width, material and appearance and are often obscured by shadows, trees and buildings. Deep learning has improved accuracy, but reviews mainly summarize methods rather than provide unified comparisons. Existing datasets often emphasize primary urban roads, lack multi-level labels or underrepresent Chinese cities, limiting multitask learning and cross-region generalization. Models trained on small or idealized data often degrade in fragmented scenes. This gap constrains practical deployment. Given these challenges, in-depth research is needed on diverse, standardized and topology-aware road extraction benchmarks.
A team from the School of Geography and Information Engineering, China University of Geosciences (CUG, Wuhan), published (DOI: 10.34133/remotesensing.1080) the study on 14 October 2026 in Journal of Remote Sensing. It addresses a practical bottleneck: models can perform well on controlled benchmarks but often fail when transferred to new cities, imaging conditions or fragmented networks. CHN10-CUG and its unified evaluation framework help compare generic segmentation networks with road-specific designs and assess robustness, topology preservation and scalability before deployment. It supports reproducible comparison.
The study integrates a dataset, benchmark and review. CHN10-CUG covers ten Chinese cities, with 0.5 m imagery and 20 road subtypes. It combines pixel-level surfaces, refined centerlines and graph labels, enabling segmentation, topology inference and multitask learning. The benchmark reproduces 15 general segmentation models and 30 road-specific methods, including graph and foundation-model approaches. Cross-region tests in Beijing and Birmingham, UK, show competitive transfer, while temporal mapping in Wuhan demonstrates long-term monitoring potential. Road-specific models generally outperform generic ones, and graph methods better preserve connectivity; the authors describe this as the first comprehensive reference framework for fair cross-paradigm comparison, plus failure modes.
CHN10-CUG contains 8,000 image patches; roads span 10–80 pixels in width, and labels follow SpaceNet-style protocols. Experiments used MMSegmentation under PyTorch on four NVIDIA graphics processing units (GPUs). Generic models were trained under a unified protocol; road-specific and graph methods used original codebases adapted only for data reading. Mask2Former was the strongest generic model, reaching 83.70% mean intersection over union (mIoU) and 82.47% F1 score (the harmonic mean of precision and recall). Road-specific methods such as LG-RoadMapper, OARENet and CGCNet reached about 84% mIoU and F1 above 83%. SAM-Road (a Segment Anything Model (SAM)-based method) reached 92.43% F1 in road-graph extraction. MCMCNet performed well with limited labels. Cross-region transfer was competitive without fine-tuning. Long-term Wuhan mapping used morphological closing, connectivity filtering and an 8-pixel change tolerance to reduce temporal inconsistency. The dataset is split into training, validation and test sets, and all methods used consistent image size, hardware and metrics.
The authors write, “By integrating dataset, benchmark, and review, this study provides a reproducible resource for remote sensing and Earth observation.” They add that multi-level annotations and graph labels can guide more robust, topology-aware model design and support long-term infrastructure monitoring. Such labels help preserve connectivity under occlusion and fragmentation.
The dataset was built from Google Earth red, green, blue (RGB) imagery at 0.5 m resolution, cropped into 512×512 patches. Experts manually annotated road surfaces; OpenStreetMap (OSM) centerlines underwent geometric correction, offset repair and connectivity optimization, and graph labels followed SpaceNet protocols. Multistage quality control included cross-inspection and consistency checks. Benchmarking used unified splits and metrics. Generic models ran in MMSegmentation with standardized training, while road-specific and graph methods used original codebases adapted only for data reading, ensuring reproducible comparison.
Future work is expected to extend CHN10-CUG toward multimodal and multitemporal fusion, combining optical imagery with synthetic aperture radar (SAR), light detection and ranging (LiDAR) and time series. Such integration could improve robustness under occlusion and support function-aware road mapping for hierarchy, usage and service roles. Weakly supervised and foundation-model approaches may reduce annotation costs and improve cross-region transfer. The benchmark could support infrastructure monitoring, transportation diagnosis and climate-resilience planning, helping move road extraction from localized experiments toward scalable mapping.
###
References
DOI
10.34133/remotesensing.1080
Original Source URL
https://doi.org/10.34133/remotesensing.1080
Funding information
This work was supported by the National Natural Science Foundation of China under Grant Nos. 42622113 and 42271413.
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.