Images satellites pour améliorer le repérage des logements dans les DROM

Raya Berova, Thomas Faria

3 juillet 2025

Outline

1️⃣ Introduction

2️⃣ Methodology

3️⃣ Data

4️⃣ Results

5️⃣ From experimentation to production?

6️⃣ Discussion

1️⃣ Introduction

Context

• Insee maintains a register of localized buildings (RIL)
  • Comprehensive housing data for cities with over 10k inhabitants 🏙️
  • Used to create sampling frames for census surveys 📋
• The RIL quality is good in metropolitan France ✅…
  • …but France also includes overseas territories (OTs)! 🌎
  • There, data quality is significantly lower ⚠️

Census in Overseas Territories

• A cartographic survey is conducted in OTs before each census

• Critical context, particularly in Mayotte and French Guiana:

  • Local authorities question official statistics
  • Rapid urban development
  • Difficult field conditions

• Requires significant work and is highly costly 💰

• Can we use satellite imagery to optimize census processes in OTs? 🛰️

• Satellite data as supporting evidence:
  • Independent information source 🔍
  • Continuous territorial monitoring 🔄
  • Transparent and reproducible methodology 📑

Our use cases

• Pre-cartographic survey:
  • Identify priority areas for fieldwork 🎯
  • Detect building changes since last survey 🏗️
  • Perform temporal comparisons using historical and recent imagery 📆
• Post-cartographic survey:
  • Automated change detection 🤖
  • Analysis of land-use evolution 🌳➡️🏘️
  • Assist population estimates in OTs 📊
• Extraordinary use case: Rapid response after devastating tropical cyclone “Chido”

Presentation Outline

1️⃣ Introduction

2️⃣ Methodology

3️⃣ Data

4️⃣ Results

5️⃣ From experimentation to production?

6️⃣ Discussion

2️⃣ Methodology

Semantic segmentation

Pleiades © CNES_2022, Distribution AIRBUS DS

Training a segmentation model

• Model trained for automatic segmentation from annotated examples
• Requirements:
  • Satellite image collection 🛰️
  • Production of annotations (building footprints or land cover, if available) 📍
• Model learns to reproduce annotations from images aiming to generalize on new images 🎯

From segmentation to change detection

Model used

• Model architecture 🧩:
  • Backbone: SegFormer (MiT)
  • Encoder: Transformer-based (efficient self-attention, no positional encoding) ⚙️
  • Decoder: Lightweight MLP head ✨
• Why SegFormer? 🚀:
  • No complex decoders → Efficient & scalable ⚡
  • Captures local & global context → High accuracy 🎯
  • No positional embeddings → Improved resolution generalization 📐
• Fine-tuned on our dataset 🗃️

Presentation Outline

1️⃣ Introduction

2️⃣ Methodology

3️⃣ Data

4️⃣ Results

5️⃣ From experimentation to production?

6️⃣ Discussion

3️⃣ Data

Pléiades (Very High Resolution) 🛰️

• 4 characteristics to look at:
  • Spatial resolution 📐
  • Acquisition frequency 🔄
  • Spectral resolution 🌈
  • Geographic coverage 🌍

• Characteristics:
  • 0.5m × 0.5m spatial resolution 🔍
  • 3 spectral bands (RGB) 🎨
  • Free archives, on-demand acquisition (6-8 months per department), Airbus © licensing 📅
  • Image size: 1 km² (2000 × 2000 pixels) 🖼️

Sentinel-2 (High Resolution) 🛰️

• Characteristics:
  • 10m × 10m spatial resolution 🔍
  • 13 spectral bands 🎨
  • 5-day revisit time, free access 🔄🆓
  • Image size: 6.25 km² (250 × 250 pixels) 🖼️

Reference Data (COSIA)

• Significant project by IGN colleagues 👏
• Land cover generated by AI as vector polygons for France and OTs 🗺️
• Based on IGN aerial photography at 20cm (!!) resolution
• Used as label for training data despite potential temporal misalignment

Presentation Outline

1️⃣ Introduction

2️⃣ Methodology

3️⃣ Data

4️⃣ Results

5️⃣ From experimentation to production?

6️⃣ Discussion

4️⃣ Results

• At this point we have built and trained our model, checked its performance
• Working on notebooks displaying maps… Key questions : how can we share our results ?
• Goals was not to publish an article but rather help colleagues from overseas territory => anticipate carto survey

Interactive Dashboard 📊

🌟 A picture is worth a thousand words! 🌟

👉 Access the interactive app: Click here 🚀

• Decided to build a tiny webapp … but we are datascientists be indulgent
• We computed building areas from our predictions this we get binary value for each pixel and aggregated by admin level

⚠️ Disclaimer : This is a an early stage of our application. 0 check or data quality process have been done.

• Standard leaflet map with open Street map layer show Pleiades 2023

97611 0332 : nuages in 2022

97610ﾠ0608 : different methodology of blurring for prison

97611 0237 : Our use case : detect building changes in precarious areas

• Administrative-level monitoring of building changes:
  • Selection by administrative units (municipalities, districts)
  • Quality control of detected changes
• Key features:
  • Side-by-side temporal comparison
  • Statistical indicators by area
  • Interactive visual validation

Presentation Outline

1️⃣ Introduction

2️⃣ Methodology

3️⃣ Data

4️⃣ Results

5️⃣ From experimentation to production?

6️⃣ Discussion

5️⃣ From experimentation to production?

Processing Pipeline 🛠️

1. Standardization of data acquisition—currently rudimentary. Need cloud-free, date-specific images with consistent methodology (challenge with Pléiades).
2. Training stage: cloud removal, dataset balancing, labeling, splitting.
3. PyTorch with pre-trained model from HuggingFace (requires GPU).
4. API building: prediction of images, clustering/bounding boxes, and statistics computation (building area).
5. Deployment via GeoServer and web application.
6. Results utilized by INSEE colleagues for statistical analyses.

Application Architecture 🧩

GeoServer dynamically serves satellite images.

Challenges & Perspectives

• High maintenance costs due to technical complexity 💸
• Need for specialized skills 🧑‍💻
• Complex technical environment due to:
  • Large data volumes 🗃️
  • High computational requirements ⚡
  • Reproducibility requirements ♻️
• Promising initial results supporting cartographic surveys ✅🗺️
• Potential improvements identified for each pipeline stage 🔧