Road segmentation with Sentinel data

[gregmaya]

Hi @giswqs and all!

I'm here to post a question that I would really appreciate any advice on.
My ultimate task: Create a change detection over roads in many regional areas of Colombia.

For that I would need:

• a free data source with a decent resolution. Most likely Sentinel 2.
  • given the nature of the roads, I probably need a way to increase resolution
  • I have found DeepSent model and its paper to be a good prospect but the training of it is not easy and I'm getting stuck there.
• a way to segment roads based on yearly composites (from 2019 to 2024).
  • ideally this would return vector geometries that I could then compare per year based on :
    • new built,
    • change of materiality

I'm looking for advice on how you'd address a task like this and potential references to pre-trained models for SuperResolution Sentinel 2 (if necessary)

For context, I do have some good experience with vector and graph models but the raster world I'd argue I'm still on the intermediate/beginner group. Comfortable with GEE and Python.

Any guidance would be much appreciated !

[giswqs]

10-m resolution is insufficient for detecting small roads. You need sub-meter level image for this.

Try out the AlphaEarth satellite embeddings, which use Sentinel-2.

• News release: https://deepmind.google/discover/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/
• Dataset: https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL#description
• Paper: https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/alphaearth-foundations.pdf
• Blog post: https://medium.com/google-earth/ai-powered-pixels-introducing-googles-satellite-embedding-dataset-31744c1f4650
• Tutorials: https://developers.google.com/earth-engine/tutorials/community/satellite-embedding-01-introduction
• Similarity search: https://earthengine-ai.projects.earthengine.app/view/embedding-similarity-search
• Clustering: https://code.earthengine.google.com/b0871454add885294f633f731b90f946

[giswqs]

If you don't have any images better than sentinel-2, then AlphaEarth is worth trying. It is probably OK for detecting for large roads. I don't think it will work for small roads.

[gregmaya]

Have you had any experience with SuperResolution models ?
Unfortunatetly my research question, points to more rural/smaller roads

[giswqs]

I would not trust super-resolution models for small roads detection.

[gregmaya]

interesting.
Do you mind elaborating a little on why?!

[giswqs]

Because super-resolution is not accurate. It may look nice visually, but the spectral information is not real. Also, the spatial detailed can be made up as well. If it were good, we would not need high-resolution satellite sensors anymore.