Consistency-Aware Map Generation at Multiple Zoom Levels Using Aerial Image

Linwei Chen, Zheng Fang, Ying FuarXiv 2026 · DOI · arXiv

TL;DR

Researchers have developed a novel deep learning framework that automatically generates accurate, multi-scale maps directly from aerial images. By combining a "level-aware" generator with a "consistency-aware" refinement process, the system ensures that geographical features like roads and rivers remain stable and accurate as a user zooms in or out, outperforming existing image-to-image translation models.

What Problem Does This Paper Address?

Modern digital maps are the backbone of everything from logistics to emergency response. Traditionally, these maps are rendered from vector data—highly precise mathematical representations of geography. However, maintaining vector data is a slow, labor-intensive process requiring professional surveyors to physically verify points of interest. This makes maps expensive to update and slow to react to sudden changes, such as the aftermath of an earthquake or flood.

Aerial and satellite imagery, by contrast, is cheap and updated frequently. While we have "image-to-image translation" models (like GANs) that can turn a photo into a map-style image, they struggle with multi-level consistency. If you generate a map at zoom level 15 and another at level 16 independently, a river might shift positions, a road might disappear, or the colors might clash. This paper tackles the challenge of generating a "pyramid" of map tiles that are both visually realistic and structurally consistent across different scales.

Methods

The authors propose a two-stage framework to bridge the gap between raw pixels and cartographic consistency:

  1. Level-Aware Map Generator (LAMG): This stage creates the initial map tiles. It uses a Semantic Module (based on DeepLabv3+) to explicitly identify features like roads, water, and land. This prevents the model from just "hallucinating" textures and forces it to understand the topology. Crucially, they introduce Level Identification—a guidance channel that tells the generator exactly which zoom level it is working on, allowing a single model to handle multiple scales.
  2. Consistency-Aware Map Generator (CAMG): This is the "secret sauce." It treats the maps at different zoom levels as a sequence. Starting from the highest resolution (where details are clearest), it iteratively refines the lower-resolution maps. By building a functional connection between adjacent levels, the model ensures that a road identified at a high zoom level is correctly represented (and aligned) when the user zooms out.

The entire system is trained using a Generative Adversarial Network (GAN) setup, utilizing a multi-scale discriminator to ensure that details look sharp whether you are looking at a single city block or an entire neighborhood.

Datasets

The researchers introduced a new benchmark called the Multi-Level Map (MLM) dataset:

  • Scale: 18,700 paired samples of aerial images and ground-truth maps.
  • Diversity: Covers two major cities with distinct geographic features—Shanghai (Northern Hemisphere) and Rio de Janeiro (Southern Hemisphere).
  • Multi-Level: Includes four distinct zoom levels (15, 16, 17, and 18).
  • Comparison: Unlike previous datasets like the one used for Pix2Pix (which only had single-level data), MLM allows for the evaluation of cross-scale consistency.

Key Insights and Ideas

  • Semantic Knowledge is Key: Standard GANs often produce "blurry artifacts" because they don't "know" what a road is; they just try to match pixel distributions. By adding a semantic segmentation step, the model maintains the "topological relationship" of geographic elements.
  • High-to-Low Refinement: It is easier to maintain consistency by using high-resolution details to inform low-resolution representations than vice versa. The CAMG leverages the "richer information" of higher zoom levels to fix errors in the broader views.
  • Unified Training via Level ID: Instead of training four separate models for four zoom levels, the "Level Identification" channel allows the model to learn a universal mapping, which is more efficient and helps the model generalize across scales.

Key Results

The proposed method was tested against heavyweights like Pix2PixHD, CycleGAN, and SMAPGAN.

  • Quantitative: The model achieved significantly lower FID (Frechet Inception Distance) and KID (Kernel Inception Distance) scores, which are standard metrics for measuring how "real" generated images look compared to the ground truth. In Shanghai (MLM-SH), the model's average FID was 137.7, nearly 25% better than the next best competitor (TSIT at 181.2).
  • Qualitative: Visual results show that while other models often "lose" rivers or create broken road networks when zooming out, this framework maintains clean, continuous lines and consistent coloring across all four levels.

Limitations and Future Work

The authors note that the model still struggles to distinguish between different types of roads (e.g., distinguishing a white local road from a yellow highway) because they can look nearly identical in raw aerial imagery.

Future work aims to:

  • Expand the MLM dataset to include more cities and more diverse zoom levels (beyond level 18).
  • Investigate better loss functions specifically designed for cartographic accuracy rather than just visual "realism."
  • Improve the classification of road hierarchies by potentially incorporating auxiliary data.

Summary updated

Abstract

The multi-level tiled map service is widely used and serves as a kind of digital infrastructure. These map tiles are usually rendered from vector data, whose update needs to walk or drive with professional equipment to check every points of interest. This leads to inconvenience and expensive cost in timely updating maps. Compared with vector data, aerial images are much easier and cheaper to obtain. In this paper, we propose a novel multi-level map generation framework that can automatically generate accurate and consistent maps with multiple zoom levels from aerial images. It consists of a level-aware map generator and a consistency-aware map generator. The level-aware map generator is able to generate accurate initial maps with realistic details for each zoom level. The consistency-aware map generator regards initial maps at each zoom level as a sequence and builds the connection between them, so as to guarantee content consistency between maps at different zoom levels. Furthermore, we collect a large-scale high-quality dataset called Multi-Level Map (MLM) for map generation at multiple zoom levels. Experiments on our MLM dataset show that our method outperforms previous state-of-the-art map generation methods on both comprehensive quantitative metrics and perceptual quality.
BibTeX
@misc{chen2026consistency,
  author = {Linwei Chen and Zheng Fang and Ying Fu},
  title = {{Consistency-Aware Map Generation at Multiple Zoom Levels Using Aerial Image}},
  year = {2026},
  eprint = {2611.19576},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19576},
}
Figure