Utilizing GCN-Based Deep Learning for Road Extraction from Remote Sensing Images

Yu Jiang, Jinmei Zhao, Wei Luo, Bincheng Guo, Zhulin An, Yongjun XuSensors 2025 · DOI

TL;DR

This paper introduces FR-SGCN, a novel deep learning model that leverages Graph Convolutional Networks (GCNs) and attention mechanisms to improve road extraction from remote sensing images, especially in challenging scenarios like occlusions and similar background features. By integrating graph reasoning with enhanced feature extraction and separation, FR-SGCN achieves state-of-the-art performance, offering more accurate and robust road mapping for urban planning and smart city development.

What Problem Does This Paper Address?

Extracting roads from remote sensing imagery is crucial for urban planning, resource management, and developing smart cities. However, this task is notoriously difficult due to several challenges:

  • Ambiguous Visual Features: Roads often look very similar to other urban features like rooftops and parking lots, making it hard for models to distinguish them.
  • Occlusions: Roads can be hidden by trees, shadows, or buildings, leading to fragmented or missed detections.
  • Complex Backgrounds: The intricate details and variations in urban environments can confuse algorithms.

Existing deep learning methods, while powerful, often struggle with these issues, leading to incomplete road networks, false positives, and a need for complex post-processing. This paper aims to tackle these limitations by developing a more robust and accurate road extraction model.

Methods

The proposed FR-SGCN model is built upon several key components:

  • Hierarchical Feature Extraction (ResNeXt): The model starts by using ResNeXt, a powerful CNN architecture known for its efficient grouped convolutions, to extract rich, high-dimensional features from the input images. This captures a broad range of visual information.
  • Feature Separation and Enhancement (Attention Mechanisms): To better handle the ambiguity and interference, the extracted features are then separated into spatial and channel components. Attention mechanisms (spatial and channel attention) are applied to dynamically weight these features, emphasizing critical road-related information and suppressing background noise. This is achieved using depthwise separable convolutions.
  • Novel Graph Construction (Hybrid Adjacency Matrix): This is a core innovation. Instead of relying solely on traditional adjacency matrices, FR-SGCN constructs a new adjacency matrix by combining:
  • Gradient Operators: To capture local structural information (edges, orientation).
  • Graph Reasoning (Similarity Matrix): To model inter-regional feature similarity and road saliency. This hybrid approach allows the GCN to consider both structural and feature similarities between nodes, leading to a more comprehensive understanding of road topology.

  • Graph Convolutional Networks (GCNs): The GCNs then process this enriched graph structure to capture global contextual relationships among features, ensuring better connectivity and integrity of the extracted road network.

  • Hybrid Loss Function: To address the common class imbalance problem (far more background pixels than road pixels), a hybrid loss function combining Binary Cross-Entropy (for stability) and Dice Loss (for optimization) is used. This helps the model learn to detect road pixels more effectively.

Datasets

The authors used two datasets for their experiments:

  • Langfang City Road Dataset: A self-constructed dataset from UAV sensors in Langfang City, China. It comprises 11,710 images (1500x1500 pixels) with a Ground Sampling Distance (GSD) of 0.1m. The ground truth masks were generated semi-automatically with meticulous manual refinement.
  • DeepGlobe Dataset: A public dataset featuring high-resolution aerial imagery from Thailand, Indonesia, and India. It includes 6226 images (1024x1024 pixels) with a spatial resolution of 0.5m. The authors used a 75%/25% split for training and testing.

Key Insights and Ideas

  • Graph Structure is Key for Road Topology: The paper strongly advocates for explicitly modeling the graph-like structure of road networks, arguing that traditional CNNs struggle with non-local topological relationships.
  • Hybrid Adjacency Matrix is Crucial: Combining gradient information (local structure) with similarity information (regional context) in the adjacency matrix significantly improves the GCN's ability to understand road connectivity, especially in challenging areas.
  • Attention Mechanisms Enhance Feature Discrimination: Dynamically weighting spatial and channel features helps the model focus on relevant road cues and ignore distracting background elements.
  • Class Imbalance Needs a Smart Loss Function: The hybrid loss function is vital for achieving a good balance between precision and recall, preventing the model from being biased towards the majority background class.
  • Lightweight Attention for Efficiency: The proposed attention mechanism is significantly more parameter-efficient and faster than standard Transformer attention, making it suitable for high-resolution imagery and real-time applications.

Key Results

  • State-of-the-Art Performance: FR-SGCN achieved the highest F1-scores on both the Langfang City dataset (82.22%) and the DeepGlobe dataset (82.87%) compared to 12 other methods.
  • Superior Precision: On the Langfang dataset, FR-SGCN achieved the highest precision (86.69%).
  • Robustness to Occlusions and Similar Backgrounds: Visualizations clearly show FR-SGCN effectively extracting occluded roads and distinguishing them from visually similar backgrounds, outperforming other models in these challenging scenarios.
  • Efficiency: The proposed lightweight attention mechanism offers a significant speedup (32.5 FPS) with fewer parameters compared to Transformer-based methods.
  • Ablation Studies: Each component of FR-SGCN (ResNeXt backbone, graph construction, attention module, hybrid loss) was shown to contribute significantly to the overall performance improvement.

Limitations and Future Work

  • Dataset Privacy: Due to privacy concerns, the full Langfang City dataset cannot be made publicly available, though a subset is shared.
  • Generalization: While the model performs well, future work aims to explore more robust dynamic-inference mechanisms to bolster generalization in diverse urban environments.
  • Model Lightweighting: Further research will focus on model lightweighting to improve efficiency.
  • Integration with Sustainable Development: The authors highlight the potential for their method to support eco-friendly infrastructure and sustainable development, suggesting future work could explore deeper integration.

Summary updated

Abstract

The technology of road extraction serves as a crucial foundation for urban intelligent renewal and green sustainable development. Its outcomes can optimize transportation network planning, reduce resource waste, and enhance urban resilience. Deep learning-based approaches have demonstrated outstanding performance in road extraction, particularly excelling in complex scenarios. However, extracting roads from remote sensing data remains challenging due to several factors that limit accuracy: (1) Roads often share similar visual features with the background, such as rooftops and parking lots, leading to ambiguous inter-class distinctions; (2) Roads in complex environments, such as those occluded by shadows or trees, are difficult to detect. To address these issues, this paper proposes an improved model based on Graph Convolutional Networks (GCNs), named FR-SGCN (Hierarchical Depth-wise Separable Graph Convolutional Network Incorporating Graph Reasoning and Attention Mechanisms). The model is designed to enhance the precision and robustness of road extraction through intelligent techniques, thereby supporting precise planning of green infrastructure. First, high-dimensional features are extracted using ResNeXt, whose grouped convolution structure balances parameter efficiency and feature representation capability, significantly enhancing the expressiveness of the data. These high-dimensional features are then segmented, and enhanced channel and spatial features are obtained via attention mechanisms, effectively mitigating background interference and intra-class ambiguity. Subsequently, a hybrid adjacency matrix construction method is proposed, based on gradient operators and graph reasoning. This method integrates similarity and gradient information and employs graph convolution to capture the global contextual relationships among features. To validate the effectiveness of FR-SGCN, we conducted comparative experiments using 12 different methods on both a self-built dataset and a public dataset. The proposed model achieved the highest F1 score on both datasets. Visualization results from the experiments demonstrate that the model effectively extracts occluded roads and reduces the risk of redundant construction caused by data errors during urban renewal. This provides reliable technical support for smart cities and sustainable development.
BibTeX
@inproceedings{jiang2025utilizing,
  author = {Yu Jiang and Jinmei Zhao and Wei Luo and Bincheng Guo and Zhulin An and Yongjun Xu},
  title = {{Utilizing GCN-Based Deep Learning for Road Extraction from Remote Sensing Images}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25133915},
}
Figure