Exploring the Potential of Deep Learning Segmentation for Mountain Roads Generalisation

Xiang Zhang (张翔), Azelle Courtial, Achraf El Ayedi, Guillaume Touya, Xiang ZhangIsprs Int. J. geo inf 2020 · DOI

TL;DR

This paper explores using deep learning, specifically a U-Net architecture, to automate the challenging task of generalizing mountain roads for cartographic purposes. By treating generalization as an image segmentation problem, the authors demonstrate that deep learning can learn to smooth roads, enlarge bends, and simplify complex road networks, though challenges remain in achieving human-level cartographic quality.

What Problem Does This Paper Address?

Cartographic generalization is the process of adapting detailed geographic data for visualization at smaller scales. A particularly difficult aspect of this is generalizing mountain roads, which often involves smoothing, enlarging bend summits, and simplifying series of bends. Traditional methods rely on complex geometric operators and manual intervention, making full automation a significant challenge. While machine learning has been explored, deep learning offers a promising avenue to capture the implicit spatial relationships and complex reasoning required for effective generalization. This paper investigates the potential of deep learning to tackle this specific, challenging problem.

Methods

The core of this research lies in framing mountain road generalization as an image segmentation task. The authors employed a U-Net architecture, a convolutional neural network known for its success in image segmentation, particularly in medical imaging.

Here's a breakdown of their approach:

  • Vector-to-Raster Conversion: Input vector road data (from a 1:25,000 scale map) was converted into raster images. This process involved careful consideration of resolution and symbolization to minimize information loss and represent road importance through width.
  • Image Tiling: To create training data, the study area was divided into tiles. Two tiling methods were explored:
  • Fixed-size tiles: A sliding window approach that creates images of a consistent scale.
  • Object-based tiles: Tiles are generated based on individual road objects, allowing for varying scales within tiles.
  • Data Augmentation: To increase the size and diversity of the training dataset, standard augmentation techniques like rotation and mirroring were applied.
  • U-Net Architecture: A U-Net model was trained to segment pixels into either "generalized road" or "background" classes. The network's contracting path captures context, while the expansive path reconstructs the segmentation map, aided by skip connections to preserve fine details.
  • Loss Function: A binary cross-entropy loss function was used, calculated for each pixel's probability of being a road.

Datasets

The study utilized two datasets from the French national mapping agency (IGN):

  • Input Data: A 1:25,000 scale extract of road data, representing high detail.
  • Target Data: A 1:250,000 scale extract of road data, representing the generalized version.

These datasets were processed to create paired input (ungeneralized) and output (generalized) raster images for training and evaluation. The authors note that the generalized dataset was created through a process combining automated algorithms and manual corrections over twenty years ago.

Key Insights and Ideas

  • Deep Learning for Cartographic Reasoning: The paper demonstrates that deep learning models can learn implicit cartographic rules for generalization, such as smoothing and bend enlargement, without explicit programming of geometric operators.
  • Image Segmentation as a Generalization Tool: Framing generalization as an image segmentation problem is a viable approach, allowing CNNs to learn pixel-wise classifications that represent the generalized output.
  • Tiling Strategy Matters: The choice of tiling method (fixed-size vs. object-based) significantly impacts the results, with fixed-size tiles showing better performance in this specific use case.
  • Sinuosity is a Challenge: Mountain roads, with their inherent sinuosity, pose a greater challenge than simpler road types, and the model struggles more with these complex features.
  • Evaluation Mismatch: Standard pixel-based evaluation metrics (like Dice and IOU) don't fully capture the quality of cartographic generalization, as minor pixel errors can lead to significant visual cartographic mistakes.

Key Results

  • Segmentation Accuracy: The model achieved a segmentation accuracy of around 65% on the evaluation dataset (Dice measure around 60%, IOU around 50%).
  • Qualitative Improvements: The U-Net model successfully learned to smooth roads and enlarge bends in many cases. The output images visually resemble generalized versions of the input roads.
  • Limitations in Complex Cases: The model struggled with tasks requiring displacement of roads to avoid overlaps with other features (which were not present in the input images) and sometimes produced disconnected road segments.
  • "Loop" Problem: A recurring issue was the creation of "loops" in bends, where road segments were added to close off bends, a problem attributed to data limitations and image readability.
  • Noise: White spot noise and the creation of fictitious road segments were observed, particularly with insufficient training data or epochs.

Limitations and Future Work

The authors acknowledge several limitations:

  • Evaluation Metrics: Existing quantitative measures are not ideal for assessing cartographic generalization quality, as they don't fully account for visual legibility and cartographic principles.
  • Dataset Size and Diversity: The dataset is limited in size and lacks sufficient examples of very narrow and sinuous bend series, hindering the model's ability to learn complex generalization.
  • Computational Constraints: Computation time and memory limitations restricted the exploration of deeper networks or larger image resolutions.
  • Road Displacement: The model's inability to learn road displacement due to the absence of contextual geographic features in the input images.

Future work suggested includes:

  • Integrating Contextual Features: Adding other geographic features (rivers, railways) to the input images to enable learning of road displacement.
  • Exploring Other Architectures: Investigating Generative Adversarial Networks (GANs) for their potential to generate more realistic and continuous road networks.
  • Improving Evaluation: Developing more suitable evaluation metrics and loss functions that better align with cartographic generalization principles.
  • Post-processing: Investigating techniques to re-establish connectivity and refine the output, potentially by converting raster outputs back to vectors.
  • User Studies: Evaluating generalization quality through user tests to better understand visual perception.

Summary updated

BibTeX
@article{2020exploring,
  author = {Xiang Zhang (张翔) and Azelle Courtial and Achraf El Ayedi and Guillaume Touya and Xiang Zhang},
  title = {{Exploring the Potential of Deep Learning Segmentation for Mountain Roads Generalisation}},
  journal = {Isprs Int. J. geo inf},
  year = {2020},
  doi = {10.3390/ijgi9050338},
}
Figure