Towards Satellite Image Road Graph Extraction: A Global-Scale Dataset and A Novel Method

TL;DR

A global-scale satellite road graph extraction dataset is collected and a novel road graph extraction model is developed, i.e. SAM-Road++, which adopts a node-guided resampling method to alleviate the mismatch issue between training and inference in SAM-Road, a pioneering state-of-the-art road graph extraction model.

Abstract

Recently, road graph extraction has garnered increasing attention due to its crucial role in autonomous driving, navigation, etc. However, accurately and efficiently extracting road graphs remains a persistent challenge, primarily due to the severe scarcity of labeled data. To address this limitation, we collect a global-scale satellite road graph extraction dataset, i.e. Global-Scale dataset. Specifically, the Global-Scale dataset is $\sim20 \times$ larger than the largest existing public road extraction dataset and spans over 13,800 $km^2$ globally. Additionally, we develop a novel road graph extraction model, i.e. SAM-Road++, which adopts a node-guided resampling method to alleviate the mismatch issue between training and inference in SAM-Road, a pioneering state-of-the-art road graph extraction model. Furthermore, we propose a simple yet effective ``extended-line'' strategy in SAM-Road++ to mitigate the occlusion issue on the road. Extensive experiments demonstrate the validity of the collected Global-Scale dataset and the proposed SAM-Road++ method, particularly highlighting its superior predictive power in unseen regions. The dataset and code are available at \url{https://github.com/earth-insights/samroadplus}.

Figures (6)

• Figure 1: (a) The collection pipeline of our Global-Scale dataset; (b) The world map showing the region location of the collected images in the Global-Scale dataset.
• Figure 2: Overview of our proposed SAM-Road++. The red line indicates training only and the blue line indicates inference only. The satellite image is fed into SAM kirillov2023segment Encoder and Decoder to get the feature map and road mask. During training, the proposed node-guided resampling uses the ground truth and road mask to get the re-sampled nodes. In the inference stage, the nodes are obtained using only road masks through Non-Maximum Suppression (NMS). Finally, the connectivity classifier determines whether a road exists between nodes based on the node features and the extended line between the nodes. Both loss functions $\mathcal{L}_{mask}$ and $\mathcal{L}_{topo}$ are binary cross-entropy loss, where $\mathcal{L}_{mask}$ and $\mathcal{L}_{topo}$ are used to supervise the road segmentation and the topological connectivity of the road, respectively.
• Figure 3: Illustrating the process of node-guided resampling. The sampled nodes are obtained by sampling from the ground truth, and $R$ represents the maximum distance threshold between the source node and target node during the sampling process. Then for each target node, our node-guided resampling strategy will find the maximum probability point of the mask around the target node and save it as the new target node. $r$ represents the maximum distance threshold between the target node and new target node.
• Figure 4: Illustration of occlusion challenge in road nodes connectivity from satellite images. (a) is the raw satellite image, (b) is the prediction of SAM-Road hetang2024segment, and (c) is the prediction with the "extended-line" strategy.
• Figure 5: The visualized road network graph predictions of SAM-Road++ and two baseline methods. Better zoom-in and view in color. Overall, the prediction accuracy of SAM-Road++ is higher. In the crossroads regions (a and b), SAM-Road++ successfully predicted the complexity of multiple roundabouts and overpasses, and in predicting the tree-shaded region c, SAM-Road++'s prediction result is also more complete compared to the other two baselines.