| Title |
Automatic road extraction from satellite imagery is a critical task in remote sensing and urban planning, with applications in transportation network analysis, infrastructure development, and smart city solutions. This paper proposes a novel methodology for road detection by integrating object-oriented deep learning algorithms, specifically combining the Faster R-CNN architecture with the Multi-Task Road Extractor model to enhance road identification accuracy. The study utilizes SpaceNet satellite imagery data, focusing on urban areas, to train and evaluate the models. The Faster R-CNN model is employed to detect candidate road regions, while the Multi-Task Road Extractor model refines these detections by leveraging a shared encoder to perform simultaneous road segmentation and classification tasks. Experimental results demonstrate the effectiveness of this integrated approach, achieving an average precision (AP) of 0.557 at a 0.6 intersection-over-union (IoU) threshold with Faster R-CNN and a 98% accuracy after refinement with the Multi-Task model. These results highlight the potential of combining multi-task learning and object detection for improved road extraction in complex urban environments.
|
| Type |
MSc |
| Supervisors |
Mahmoud Hamed; Mahmoud Salah |
| Year |
2025 |
| Abstract |
Automatic road extraction from satellite imagery plays a vital role in remote sensing and urban
planning. Its applications span transportation network analysis, infrastructure development, and
smart city solutions. This thesis introduces a detailed methodology for road detection by
combining object-oriented and pixel-based deep learning techniques, specifically integrating the
Faster R-CNN architecture with the Multi-Task Road Extractor model, to improve road
detection accuracy.
The study uses SpaceNet imagery data, focusing on urban environments, to train and evaluate
the models. The Faster R-CNN model is first utilized to identify candidate road regions within
the imagery. To enhance these results, the Multi-Task Road Extractor model is applied,
employing a shared encoder that performs road segmentation and classification simultaneously.
This dual approach enhances both pixel-level accuracy and the recognition of road attributes.
The experimental outcomes validate the effectiveness of the two models, showcasing notable
improvements in average precision (AP) across various intersection-over-union (IoU)
thresholds. Specifically, the Faster R-CNN model achieves an AP of 0.557 at a 0.6 detection
threshold, while the training metrics for the Multi-Task indicated high performance, with
accuracy reaching 0.986 and Dice coefficient achieving 0.861.
This study can support urban planning, smart city initiatives, and infrastructure development,
providing a solid foundation for future work. Future research will focus on utilizing multi-source
geospatial data, such as UAV imagery and laser scanning, to achieve the same objectives with
enhanced precision. |
| Keywords |
Road Extraction, Faster R-CNN, Deep Learning, SpaceNet Dataset, Environmental Sustainability |
| University |
Benha |
| Country |
Egypt |
| Full Paper |
download paper
|
