City-Scale Road Extraction from Satellite Imagery v2: Road Speeds and Travel Times

Adam Van EttenWACV 2020

TL;DR

This paper introduces CRESIv2, a novel approach for automatically extracting road networks and estimating travel times from satellite imagery at a city scale. By combining advanced segmentation with graph-based analysis and incorporating semantic features like speed limits, CRESIv2 enables more accurate routing for applications like disaster response.

What Problem Does This Paper Address?

Automated road network extraction from satellite imagery is crucial for various applications, including urban planning, disaster response, and improving access to services, especially in developing countries. However, existing methods face challenges in accurately capturing road networks at scale, especially in areas with poor data quality or rapid changes. Furthermore, traditional road extraction focuses on geometric distance, which is insufficient for optimal routing when travel time is a critical factor. Existing datasets often suffer from misaligned labels, missing metadata (like speed limits), or are too small for comprehensive city-scale analysis.

Methods

CRESIv2 builds upon previous work by introducing several key advancements:

  • Multi-Class and Continuous Segmentation: The paper employs two segmentation strategies. The multi-class approach categorizes roads into different speed bins, while the continuous approach linearly scales pixel values with speed limits. Both use a ResNet34 encoder with a U-Net decoder and a custom loss function.
  • Graph Extraction and Refinement: The segmentation output is refined into a graph structure. This involves smoothing, morphological operations, skeletonization, and graph pruning to remove spurious connections and small gaps.
  • Speed and Travel Time Estimation: A novel aspect is the inference of road speeds and travel times. This is achieved by analyzing the prediction mask at segment midpoints and assigning speed limits based on metadata tags (road type, surface) or inferred from the segmentation mask itself. Travel time is then calculated as edge length divided by estimated speed.
  • Large-Scale Processing: To handle city-scale imagery, CRESIv2 integrates the Broad Area Satellite Imagery Semantic Segmentation (BASISS) methodology, allowing processing of arbitrarily large images by tiling and stitching predictions.

Datasets

  • SpaceNet 3 WorldView-3: This dataset, featuring 30 cm/pixel satellite imagery and detailed road centerline labels, is the primary source for training and evaluating CRESIv2. It covers 3000 km² with over 8000 km of labeled roads.
  • SpaceNet Test Regions: Four cities (Las Vegas, Khartoum, Paris, Shanghai) from the SpaceNet dataset were used for large-area testing, totaling 608 km² and 9065 km of roads.
  • Google Satellite Imagery with OSM Labels: This dataset, consisting of 60 cm/pixel imagery over 40 cities, was used to evaluate performance with OpenStreetMap (OSM) labels.

Key Insights and Ideas

  • Semantic Features for Routing: Incorporating semantic information like speed limits into road extraction significantly enhances the potential for optimized routing, moving beyond simple geometric distance.
  • Multi-Class Segmentation for Speed: The multi-class segmentation approach, which directly predicts speed bins, proves more robust for speed estimation compared to continuous scaling, especially when dealing with noisy or ambiguous predictions.
  • SpaceNet vs. OSM Labels: Models trained and tested on SpaceNet labels significantly outperform those trained on OSM labels, highlighting the importance of high-quality, well-registered, and detailed ground truth data.
  • Travel Time vs. Distance Routing: The study demonstrates that optimizing routing for travel time results in only a small decrease in performance (3-13%) compared to distance-based routing, making time-optimized routing a feasible and valuable goal.

Key Results

  • SpaceNet Performance: CRESIv2 achieved an APLSlength score of 0.68 ± 0.21 on SpaceNet test chips, with a slight decrease to 0.58 ± 0.21 when using travel time as the weight.
  • Comparison to Previous Work: On Google/OSM data, CRESIv2 achieved a TOPO score of 0.53, a significant improvement over existing methods like RoadTracer (0.43) and Orientation Learning (0.64, though this comparison is nuanced).
  • Large Area Performance: For large area SpaceNet testing, CRESIv2 achieved an APLSlength of 0.67 and APLStime of 0.64, showing minimal degradation when using travel time.
  • Inference Speed: The algorithm achieves an inference speed of 280 km²/hour on a single Titan X GPU, making it practical for large-scale mapping efforts.

Limitations and Future Work

  • Complex Intersections: The paper acknowledges that complex intersections, overpasses, and onramps remain a challenge for CRESIv2, requiring further refinement in gap connection and proximity-based logic.
  • OSM Data Quality: The variability and density of OSM road metadata present challenges for accurate speed inference, suggesting this as an area for future work.
  • Ensemble vs. Single Model: While a 4-fold ensemble approach showed marginal improvement in APLS scores, it significantly increased runtime. The paper opted for a single model for speed efficiency, but future work could explore optimized ensemble strategies.
  • Further Optimization: CRESIv2 has not been fully optimized for speed, indicating potential for further acceleration.

Summary updated

BibTeX
@inproceedings{etten2020city,
  author = {Adam Van Etten},
  title = {{City-Scale Road Extraction from Satellite Imagery v2: Road Speeds and Travel Times}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2020},
}
Figure