Sentinel-2 Imagery For Fine-Grained Road Map Extraction Using Deep Learning Techniques.pdf

### Summary

Title: Towards Fine-Grained Road Maps Extraction Using Sentinel-2 Imagery

Authors: C. Ayala, C. Aranda, M. Galar
Affiliations: Tracasa Instrumental and Institute of Smart Cities (ISC), Public University of Navarre

#### Key Concepts:
- Objective: Demonstrate the feasibility of accurately detecting road networks using high-resolution Sentinel-2 imagery (10 m resolution).
- Methodology: Propose a novel deep learning architecture that integrates semantic segmentation with super-resolution techniques.
- Advantages: Potential reduction in costs and increased frequency of cartography updates due to Sentinel-2's high revisit times.

#### Background:
- Traditional road network extraction relies on aerial images (<1 m) due to their higher resolution, which facilitates the detection of roads despite occlusions like trees.
- Deep learning methods have become standard for image processing tasks, with Convolutional Neural Networks (CNNs) being particularly effective in semantic segmentation tasks such as road network extraction.

#### Related Work:
- Early attempts at using deep learning for road extraction date back to 2010 with restricted Boltzmann machines and later CNNs.
- Recent studies have employed modified U-Nets and LinkNets, often utilizing aerial imagery due to its higher resolution.

#### Proposed Approach:
- Data Source: Sentinel-2 imagery combined with OpenStreetMap (OSM) annotations for training.
- Evaluation Metrics: Intersection over Union (IoU) and F-score metrics are used to assess performance.
- Experimental Setup: The study includes 20 cities across Spain, acknowledging potential labeling errors in OSM data but emphasizing its utility despite noise.

#### Conclusion:
The paper concludes by highlighting the effectiveness of using high-resolution satellite imagery for road detection and suggests future research directions. The approach leverages Sentinel-2's capabilities to potentially transform how frequently and cost-effectively road maps can be updated.