Mapping Historic Urban Footprints in France: Balancing Quality, Scalability and AI Techniques

This study presents a scalable dual-pass deep learning pipeline that successfully extracts the first open-access, nationwide urban footprint dataset for metropolitan France from historical maps (1925–1950), achieving 73% accuracy by effectively mitigating artifacts like text and contour lines to enable quantitative analysis of pre-1970s urban sprawl.

The Gist

Imagine trying to understand how a city grew over 100 years ago, but all you have are old, dusty, hand-drawn maps scattered across a library. Some are faded, some have coffee stains, and the ink is smudged. Trying to figure out exactly where the buildings were just by looking at these maps is like trying to find a needle in a haystack while wearing foggy glasses.

This paper is about a team of researchers who built a super-smart digital detective to solve this problem for all of France.

Here is the story of how they did it, broken down into simple steps:

1. The Problem: The "Foggy Glasses"

For a long time, we have great satellite photos of cities from the 1970s to today. But before that? We have a huge gap. We know France existed, and we know cities grew, but we didn't have a digital map showing exactly where the buildings were between 1925 and 1950.

The only clues were the Scan Histo maps: beautiful, high-resolution scans of old paper maps. But they are tricky. They are full of:

• Text: City names and labels that look like buildings.
• Roads: Lines that look like streets but aren't buildings.
• Shadows and Stains: Old paper gets dirty, and the ink fades differently in different places.

If you tried to teach a computer to read these maps using standard methods, it would get confused. It might think a city name written in big bold letters is a skyscraper, or that a contour line (a line showing a hill) is a wall.

2. The Solution: The "Two-Pass" Detective

The researchers didn't just throw a standard computer program at the problem. They created a two-step training process, kind of like training a new employee.

• Pass 1: The "Rough Draft" Detective
  First, they taught an AI (a type of computer brain called a U-Net) using a small set of examples. This AI made a first guess at where the cities were.

  • The Result: It was okay, but messy. It found the big black buildings but got confused by big text and roads.
  • The Trick: Instead of giving up, the researchers looked at where the AI made mistakes. They used those mistakes to teach the AI what not to look for.

• Pass 2: The "Refined" Detective
  They gave the AI a second chance, but this time with a better "textbook." They fed it the messy results from the first pass and showed it specifically: "Hey, this text isn't a building," and "This road isn't a house."

  • The Result: The AI learned to ignore the noise. It stopped thinking the word "Paris" was a building and started focusing only on the actual shapes of the cities.

3. The Scale: Cleaning the Whole Map

They didn't just do this for one town. They applied this method to 941 giant map tiles covering the entire country of France (except for a few islands).

• Imagine taking a puzzle of the whole country, cutting it into 941 pieces, and having a robot clean up every single piece to find the buildings.
• They used a supercomputer (a very powerful machine) to do this work, which would have taken a human hundreds of years to finish.

4. The Result: A Time-Travel Map

The final product is a digital map of France from the 1920s–1950s that shows exactly where the urban areas were.

• Accuracy: It got about 73% right. That's a huge success for such old, messy maps.
• Why it matters: Before this, we had to guess how cities grew before the 1970s. Now, we have a clear picture. We can see how Paris expanded, how small towns grew into cities, and how World War II destruction and rebuilding changed the landscape.

5. The "Open Source" Gift

The best part? The researchers didn't keep this secret. They released everything to the public:

• The code (the recipe for the detective).
• The data (the old maps and the new digital version).
• The results (the final map).

The Big Picture Analogy

Think of this like restoring an old, damaged photograph.

• Old Way: You try to guess what's in the photo by squinting. You might miss details or see things that aren't there.
• This Paper's Way: You use a smart AI tool that learns to "clean" the photo. It learns to ignore the scratches (text and roads) and highlight the people (the buildings). Once it's trained, it can clean up thousands of photos instantly, giving us a crystal-clear view of history that was previously blurry.

This work allows historians, urban planners, and anyone interested in the past to finally "see" the cities of 100 years ago with the same clarity we have for today's cities.

Technical Summary

Here is a detailed technical summary of the pre-print paper "Mapping Historic Urban Footprints in France: Balancing Quality, Scalability and AI Techniques."

1. Problem Statement

The study addresses a critical gap in historical urban analysis: the lack of nationwide, quantitative digital data for urban sprawl in France prior to the 1970s (the start of the satellite era).

• Data Scarcity: While historical maps exist (e.g., Cassini, Etat-Major, and the Scan Histo series from 1925–1950), they are unstructured scanned images. Converting these into exploitable vector formats (shapefiles) or rasters is difficult due to the inherent complexity of historical cartography.
• Challenges: Historical maps suffer from high radiometric variability, paper degradation, ink fading, and stylistic heterogeneity. Automated extraction is further complicated by the confusion between urban features and non-urban elements such as text labels, contour lines, roads, and hatching patterns, which often share similar visual characteristics (black ink on white paper).
• Limitations of Existing Methods: Traditional pixel-based classifiers, object-based image analysis (OBIA), and standard deep learning models (like single-pass U-Net) have proven inadequate at a national scale, often failing to distinguish between urban textures and map artifacts.

2. Methodology

The authors propose a scalable, dual-pass deep learning pipeline utilizing the U-Net architecture to extract urban footprints from the Scan Histo collection (941 high-resolution tiles covering metropolitan France).

A. Data Preparation

• Source: IGN's Scan Histo archive (1925–1950), 5m resolution RGB JP2 tiles.
• Ground Truth: A manually digitized dataset of 84 French cities (1–10 km buffers) converted into binary masks (Urban = 1, Non-urban = 0).
• Preprocessing: Images resized to 256×256 pixels, normalized, and paired with binary masks.
• Dataset Evolution:
  • Initial Set: 58 image/prediction pairs.
  • Refined Set: Expanded to 312 pairs by targeting problematic zones (roads, text, contour lines) identified in the first pass.

B. The Dual-Pass U-Net Approach

The core innovation is a two-stage inference process designed to handle radiometric noise and stylistic complexity:

1. First Pass (Native Map Input):
   • A U-Net model is trained on the original RGB scanned maps.
   • Goal: Generate a preliminary urban mask.
   • Outcome: Successfully identifies solid urban blocks but struggles with large text labels and hatched patterns, producing false positives.
2. Data Augmentation & Refinement:
   • The errors from the first pass (confusion zones) are analyzed to create a targeted, larger training dataset (312 pairs).
3. Second Pass (Binary Mask Input):
   • The model is re-trained and re-applied, but this time the input is the binary mask generated by the first pass, not the original RGB map.
   • Goal: By removing the original map's radiometric noise (colors, textures, text), the model focuses purely on spatial coherence to refine the shape and eliminate false positives (roads, text, contour lines).
   • Outcome: A cleaner, geometrically coherent urban footprint.

C. Implementation Details

• Framework: PyTorch.
• Architecture: Standard U-Net (Encoder-Decoder with skip connections).
• Loss Function: Binary Cross-Entropy (BCE).
• Hardware: Executed on the Austral HPC cluster (NVIDIA A100 GPU).
• Scale: Processed 941 tiles (53 GB total data) covering all of metropolitan France.

3. Key Contributions

1. First National-Scale Dataset: Creation of the first open-access, nationwide urban footprint raster for France covering the pivotal 1925–1950 period.
2. Novel Algorithmic Strategy: Introduction of a dual-pass U-Net approach. By using the output of the first model as the input for the second, the system effectively decouples radiometric noise from spatial features, significantly reducing false positives without requiring complex multi-modal inputs.
3. Open Science: Full open-access release of:
   • The nationwide urban raster dataset.
   • The training datasets (including ground truth shapefiles).
   • The Python code and pre-trained models.
4. Scalability: Demonstration that deep learning can be applied to historical cartography at a national scale, overcoming the limitations of manual digitization.

4. Results

• Overall Accuracy: The final mosaic achieved an overall accuracy of 73%.
• Performance Metrics:
  • Deep Learning vs. Traditional: The U-Net approach (73%) significantly outperformed semi-automatic radiometric thresholding and unsupervised K-means clustering (which showed lower accuracy and poor spatial coherence). DeepLabV3+ achieved 69% but was less effective in rural areas.
  • Regional Variance:
    • East-Central France: High accuracy (up to 85%) due to consistent map styles.
    • Western France: Lower accuracy (40–60%) due to heterogeneous map tones and higher residual noise from text/roads.
  • Urban vs. Rural: High precision in urban centers and suburbs. Rural areas with steep terrain or dense agricultural textures (contour lines) remain challenging, with accuracy dropping to 45–60% in these specific zones.
• Iterative Improvement: Analysis showed that 20 epochs were optimal for the first pass (high variability), while only 10 epochs were needed for the second pass (binary input). Exceeding 21 epochs led to overfitting.

5. Significance and Applications

• Bridging the Temporal Gap: This dataset allows researchers to analyze urbanization dynamics between the 1920s and the 1970s (the start of the Landsat era), a period previously lacking quantitative data.
• Historical Context: It enables the study of urban impacts from World War II (destruction and reconstruction) and the early automobile era.
• Comparative Analysis: The 5m resolution data can be resampled to match international datasets like GHSL (100m) or Landsat (80m), facilitating long-term diachronic studies of urban sprawl.
• Future Potential: The authors plan to extend this methodology to earlier map series (19th-century Etat-Major and 18th-century Cassini maps) to build a continuous, multi-century historical urban database.

In conclusion, the paper demonstrates that a tailored, two-stage deep learning pipeline can successfully extract high-quality urban footprints from complex historical maps, providing a foundational resource for understanding long-term anthropogenic changes in France.