Graph residual U-Net: Automated deep learning-based road extraction using DeepGlobe dataset, cartosat-2 and sentinel-2 imagery

Avishkar Karnik (), Shilpa P Pimpalkar () and Manjusree Panchagnula ()

International Journal of Innovative Research and Scientific Studies, 2025, vol. 8, issue 2, 2047-2064

Abstract: Road detection is crucial for defensive strategies and disaster management since precise and timely mapping of road networks enables efficient operations during critical times. In defence, road detection facilitates strategic troop movement, the identification of secure pathways, and real-time monitoring of potential barriers or dangers, all of which contribute to tactical planning and operational safety. It helps disaster management by finding accessible roadways for first responders, assessing damaged infrastructure, and optimizing evacuation routes. Recent research work using deep learning techniques have significantly enhanced the processing capabilities of remote-sensing images for information extraction. Consequently, these deep-learning methodologies facilitate the automation of road extraction using very high-resolution imagery/satellite imagery. This study presents a novel deep learning approach, Graph Residual U-Net, designed for road extraction from satellite imagery. The Graph Residual U-Net model uses Deep Residual U-Net as the backbone U-Net architecture and incorporates graph neural network (GNN) layers into a residual U-Net framework, enhancing feature representation and improving road extraction accuracy. Findings are presented by implementing two models: i) Graph Residual U-Net Cartosat Model- Model trained using the DeepGlobe and Cartosat-2 Dataset and tested using DeepGlobe and Cartosat-2 Dataset and images acquired using Google Earth Engine, ii) Graph Residual U-Net Sentinel Model- the other model trained and tested using the Sentinel-2 imagery. This work highlights the potential of combining GNNs with convolutional neural networks (CNNs) for remote sensing applications. Graph Residual U-Net Cartosat Model showed an accuracy of 0.97702 and loss of 0.06150 and Graph Residual U-Net Sentinel Model demonstrated an accuracy of 0.99636, a loss of 0.00017, and . Graph Residual U-Net Cartosat Model showed superior performance considering overall performance metrics. Graph Residual U-Net Sentinel Model showed low precision and recall but the model demonstrated excellent findings by generating accurate predicted masks when tested on unseen images.

Keywords: Graph Neural Network; Graph Residual U-Net; Residual U-Net; Road extraction. (search for similar items in EconPapers)
Date: 2025
References: Add references at CitEc
Citations:

Downloads: (external link)
https://ijirss.com/index.php/ijirss/article/view/5634/1001 (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:aac:ijirss:v:8:y:2025:i:2:p:2047-2064:id:5634

Access Statistics for this article

International Journal of Innovative Research and Scientific Studies is currently edited by Natalie Jean

More articles in International Journal of Innovative Research and Scientific Studies from Innovative Research Publishing
Bibliographic data for series maintained by Natalie Jean ().