Abstract
Road safety mapping using satellite images is a cost-effective but challenging problem for smart city planning. The scarcity of labeled data, misalignment, and ambiguity makes it hard for supervised deep networks to learn efficient embeddings to classify between safe and dangerous road segments. In this paper, we address these challenges using a region-guided attention network. Our model extracts global features from a base network and augments them with local features obtained via the region-guided attention network. Additionally, we perform domain adaptation for unlabeled target data. To bridge the gap between safe and dangerous samples from source and target domains, we propose a loss function based on within- and between-class covariance matrices. Experiments on a public dataset of London show the algorithm achieves a classification accuracy of 86.21%. We obtain a 4% increase in accuracy for NYC using the domain adaptation network. Furthermore, a user study demonstrates that our proposed algorithm achieves 23.12% better accuracy compared to subjective human analysis.
Introduction
Traditional road safety mapping is a logistical challenge, often requiring years of meticulously recorded traffic accident data. Such data is frequently scarce or poorly maintained, particularly in low- and middle-income regions. While satellite imagery offers a cost-effective alternative, it presents significant computer vision hurdles: extreme data imbalance (dangerous segments are rare), misalignment between satellite images and road grids, and domain shift, where models trained on one city (e.g., London) fail in another (e.g., New York City) due to differing urban layouts. This paper proposes a model capable of learning safety features in a source city and successfully adapting them to a target city without requiring local labels.
Methods
The authors introduce the Deep Attention Model (DAM) and its domain-adaptive counterpart, DAM-DA.
Global and Local Feature Extraction
The architecture uses a ResNet-50 backbone to extract global features from satellite images. To handle misalignment and capture fine-grained environmental cues, the system employs Region of Interest (ROI) pooling to divide images into subregions. An attention mechanism identifies which local features—such as specific intersections or road markings—are most predictive of danger, guiding the overall classification.
Domain Adaptation (DA)
To bridge the gap between a labeled source city and an unlabeled target city, the DAM-DA framework utilizes a covariance-based loss function. By minimizing the Frobenius norm between the within-class and between-class covariance matrices of both domains, the network learns to extract features that are consistent across different urban environments. The process is further supported by pseudo-labeling target data, which is iteratively refined.
Datasets
The research utilized accident records from London (99,516 records), New York City (1.25 million records), and Denver (143,776 records). Satellite imagery was obtained via the Google Static Maps API, with road segments processed in 30x30 meter grids.
Results
The DAM model achieved 86.21% accuracy on the London dataset, outperforming standard ResNet and VGG architectures. In the domain adaptation task from London to NYC, the DAM-DA model improved accuracy by 4% over non-adaptive baselines and significantly reduced the False Positive Rate to 18.94%.
A notable finding was the performance gap between AI and humans. A user study showed that humans achieved only ~63% accuracy in identifying road safety from satellite views, whereas the DAM model outperformed human subjects by 23.12%. Humans particularly struggled to identify dangerous segments, often missing subtle cues that the attention-guided network successfully captured.
Conclusion
This work presents a robust framework for road safety mapping that overcomes data scarcity and domain shift. By combining region-guided attention with covariance-based domain adaptation, the model provides a scalable solution for urban planning. Future research will focus on pre-training with urban-centric datasets and developing more granular safety metrics beyond binary classification.
Summary updated