An Enhanced GAN Model for Automatic Satellite-to-Map Image Conversion

Ying Zhang, Yifang Yin, Roger Zimmermann, Guanfeng Wang, Jagannadan Varadarajan, See-Kiong Ng

Introduction

Digital maps are essential for modern navigation, yet keeping them updated is a resource-intensive process. While automated satellite-to-map conversion using Generative Adversarial Networks (GANs) has shown promise, real-world conditions like cloud cover, shadows, and irregular road networks often lead to poor results. Researchers have introduced SG-GAN (Semantic-regulated Geographic GAN), a novel model that integrates satellite imagery with crowdsourced GPS data and high-level semantic masks to overcome these limitations.

The Challenge of Satellite-to-Map Conversion

Traditional map-making is labor-intensive, often resulting in significant delays between updates. Automated systems using standard GANs, such as Pix2Pix, frequently struggle with:

  • Irregular Road Networks: In cities with non-grid layouts, models often produce disconnected or "messy" patterns.
  • Visual Obstructions: Clouds, shadows, and vegetation can obscure ground features from satellite sensors.
  • Data Sparsity: In regions with limited information, GANs tend to generate visual artifacts rather than accurate road representations.
Satellite imagery serves as the primary input for the conversion process.

The SG-GAN Architecture

The SG-GAN model moves beyond simple image-to-image translation by incorporating two critical layers of geographic guidance:

  1. GPS-as-Image: Crowdsourced vehicle GPS traces are "rendered" into binary images. This allows geographic movement data to be concatenated with satellite imagery, providing a "hint" of road locations even when they are obscured by clouds.
  2. Semantic Estimator: A U-Net-based module predicts a coarse land-type mask (e.g., road vs. non-road). This acts as a high-level regulator, ensuring the generator respects regional structures.
  3. Hybrid Loss Function: The model is trained using a combination of Adversarial Loss (for realism), Content Loss (for pixel accuracy), and a new Semantic Loss (to ensure adherence to predicted land-type boundaries).

Key Insights

A significant finding of the research is that GPS density acts as an "X-ray" for clouds; high GPS ping density confirms the existence of a road regardless of visual clarity. Furthermore, the authors discovered that coarse guidance (simple road vs. non-road masks) often outperforms more granular classification for GAN regulation, providing better structural "guardrails."

The target output is a clean, accurate map representation of the urban environment.

Results and Performance

SG-GAN was evaluated against baselines like Pix2Pix and CycleGAN using the Grab-Posisi dataset and HERE API imagery. The model achieved a 20.76% to 27.91% improvement across key metrics, including Inception Score (IS) and Structural Similarity Index (SSIM). Qualitatively, SG-GAN excelled at reconstructing curved roads and maintaining connectivity in areas with partial cloud cover, where traditional models failed.

Conclusion and Future Work

While SG-GAN offers superior accuracy and robustness, its increased complexity leads to higher computational costs. Future work will focus on model compaction for improved efficiency and expanding the semantic regulation to include more diverse urban features, such as building types and water bodies.

Summary updated

Abstract

Location-based service significantly relies on accurate and up-to-date maps. The conventional map generation involves labor-intensive and time-consuming manual efforts, which restricts the map-update frequency to a few years or even longer. In recent years, satellite images become more ubiquitous, and converting them to map-style images has attracted attention due to its frequent-updating and cost-effective nature. Generative adversarial network (GAN) is a promising approach for automatic satellite-to-map image conversion. However, it is still challenging to process satellite images when the underlying road structure is complex and irregular, or when some objects are visually indistinguishable due to obstruction or bad weather. To address these issues, we propose an enhanced GAN model to generate improved quality map images by bringing in the external geographic data as implicit guidance. The textual geographic data is converted to an image so that it can work collaboratively and seamlessly with the satellite image during the conversion. A high-level semantic regulation is also introduced to further reduce the noisy patterns generated during the translation, which occur frequently for the regions with sparse geographic data. The proposed method is versatile to various backbone GAN structures with a 20% performance improvement on three popular metrics (Inception Score, Frechet Inception Distance Score and SSIM score). Our proposed Semantic-regulated Geographic GAN (SG-GAN) is anticipated to reduce the manual identification efforts in broad geospatial applications.
BibTeX
@article{zhang2020enhanced,
  author = {Ying Zhang and Yifang Yin and Roger Zimmermann and Guanfeng Wang and Jagannadan Varadarajan and See-Kiong Ng},
  title = {{An Enhanced GAN Model for Automatic Satellite-to-Map Image Conversion}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3025008},
}
Figure