An Improved Method for Road Extraction from High-Resolution Remote-Sensing Images that Enhances Boundary Information

Shuai Wang, Hui Yang, Qiangqiang Wu, Zhiteng Zheng, Yanlan Wu, Junli LiSensors 2020 · DOI

TL;DR

This paper introduces the Coord-Dense-Global (CDG) network, a novel deep learning architecture for extracting road networks from high-resolution remote-sensing images. By integrating coordinate convolution and a global attention mechanism with a DenseNet backbone, CDG significantly improves road boundary delineation and global context awareness, outperforming existing state-of-the-art methods, especially in challenging scenarios like tree occlusion.

What Problem Does This Paper Address?

Extracting accurate road networks from high-resolution remote-sensing images is crucial for urban planning, navigation, and geographic information management. However, existing deep learning methods often struggle with preserving fine spatial details, leading to fragmented road extraction and loss of boundary information. Furthermore, global context is often lost during downsampling, and challenges like tree occlusion can severely degrade performance. This paper aims to address these limitations by developing a network that enhances both local spatial features and global contextual understanding for more robust and complete road extraction.

Methods

The proposed CDG model is built upon an encoder-decoder architecture leveraging a DenseNet backbone. The key innovations are:

  • CoordConv Module: This module enhances spatial information by explicitly incorporating coordinate information (horizontal and vertical) into the feature maps. This helps to reduce the loss of spatial details and strengthen road boundaries, which are often thin and linear.
  • Global Attention Module (GAM): This module is designed to capture global context information. It applies global average pooling to feature maps, processes them through deconvolution and pooling paths, and then uses a sigmoid classifier to generate an attention map. This map is multiplied with the original feature map to highlight high-level features and improve category classification, ensuring better continuity of extracted roads.
  • DenseNet Backbone: The network utilizes DenseNet's dense connectivity, where each layer is connected to all preceding layers. This promotes feature reuse, reduces parameter redundancy, and facilitates better information flow, making it well-suited for semantic segmentation tasks like road extraction.

The encoder progressively extracts features using dense blocks and global attention layers, while the decoder upsamples and concatenates features with corresponding encoder layers via skip connections to reconstruct the high-resolution segmentation map.

Datasets

The paper uses two primary datasets for training and evaluation:

  • Massachusetts Roads Dataset: This publicly available dataset consists of 1171 aerial images (1500x1500 pixels, 1.2m spatial resolution) with red, green, and blue channels. It includes 1108 training, 49 test, and 14 validation images, covering diverse urban, suburban, and rural contexts. The authors cropped these images to 512x512 pixels for training, creating 14,366 subset images.
  • Gaofen-2 Satellite Dataset: This dataset, with 1m spatial resolution, was used to test the model's generalization capabilities. The authors trained on a subset of this data (5892 samples of 512x512) and then tested on four larger images (4578x4442) from Hefei and Tianjin, China.

Key Insights and Ideas

  • Coordinate Information is Key for Thin Structures: Explicitly feeding coordinate information into convolutional layers (CoordConv) proves highly effective for delineating thin, linear features like roads, which are prone to losing spatial fidelity in standard CNNs.
  • Global Context Prevents Fragmentation: The global attention mechanism is crucial for understanding the broader scene, which helps maintain the continuity of roads and prevents them from being broken into disconnected segments, especially in complex environments.
  • Dense Connectivity Aids Feature Richness: The DenseNet architecture's inherent feature reuse and dense connections allow the model to leverage a rich set of features from various levels, contributing to more accurate and complete road extraction.
  • Tree Occlusion is a Significant Challenge: Even advanced models struggle with heavily occluded roads, highlighting the persistent difficulty of extracting roads obscured by dense foliage.

Key Results

  • Quantitative Performance: On the Massachusetts dataset, CDG achieved a mean IoU of 61.90% and a mean F1 score of 76.10%. This was 1.19% and 0.95% higher in mean IoU and F1 score, respectively, than D-LinkNet, the winner of a road-extraction contest. CDG also outperformed U-net and DeepLabV3+ in these key metrics.
  • Tree Occlusion: CDG demonstrated superior performance in extracting roads occluded by trees compared to D-LinkNet, U-net, and DeepLabV3+, successfully extracting a larger portion of these roads.
  • Generalization: On the Gaofen-2 satellite dataset, CDG achieved good results, extracting roads in Hefei and Tianjin, China, with a reported average precision of 68.38%, recall of 77.72%, F1 of 72.62%, and IoU of 57.11%.
  • Test Time: CDG's test time was 196s, which is higher than U-net (167s) but competitive and offers advantages over DeepLabV3+ (245s).

Limitations and Future Work

  • Severe Tree Occlusion: The paper acknowledges that while CDG performs better with tree occlusion, it still struggles with severely blocked roads where even human eyes find it difficult to discern the road.
  • Discontinuities in Complex Backgrounds: Despite improvements, discontinuities in road extraction can still occur due to complex backgrounds (forests, buildings, varying road types/widths).
  • Ground Truth Mismatches: The authors note that mismatches between ground truth labels and actual roads (e.g., unlabelled roads identified by the model) can limit performance.

For future work, the authors suggest incorporating road geometric information to further address incomplete or broken road extraction issues.

Summary updated

BibTeX
@article{wang2020improved,
  author = {Shuai Wang and Hui Yang and Qiangqiang Wu and Zhiteng Zheng and Yanlan Wu and Junli Li},
  title = {{An Improved Method for Road Extraction from High-Resolution Remote-Sensing Images that Enhances Boundary Information}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20072064},
}
Figure