←Back to Blog
Geospatial AIUse Case

Zero-Shot Damage Assessment: Evaluating MBCTD on the 2025 LA Wildfires

I built a custom web app to test my MBCTD model on the 2025 LA wildfires. Here's a look at how it generalized to real-world disaster damage on the fly.

AM
Abdelrrahman Moubane
June 12, 2026
3 min read
Building Change DetectionDamage AssessmentMBCTDWildfiresZero-ShotGeneralization
Zero-Shot Damage Assessment: Evaluating MBCTD on the 2025 LA Wildfires

Zero-Shot Damage Assessment: Evaluating MBCTD on the 2025 LA Wildfires

Back in January 2025, the Palisades and Eaton wildfires caused immense devastation and tragic loss across Los Angeles. Because rapid damage assessment is critical for relief and recovery efforts, I wanted to see if the MBCTD (Multi-Label Building Change Type Detection) model I've been building could accurately map the extent of the destruction.

From a machine learning perspective, the question was whether a model trained on standard building changes (like urban redevelopment) could generalize to the severe, widespread structural loss of a major wildfire.

Could it do it zero-shot, without any extra training?

The Setup: On-the-Fly Inference

To test this out, I built a quick prototype app using Solara and Leafmap. The idea was to stream high-resolution aerial imagery directly from Pictometry's WMTS feeds (provided by LA County), run the model, and see the results instantly.

The workflow is pretty straightforward:

  1. Draw an area of interest over a neighborhood in the app.
  2. The app concurrently fetches the pre-fire (2024) and post-fire (Jan 2025) imagery tiles.
  3. The imagery is passed through the MBCTD model.
  4. It spits out a pixel-perfect mask of the damage directly onto the map.
MBCTD LA Wildfire App Demo
MBCTD LA Wildfire App Demo
The Solara app running MBCTD inference over a selected area of interest. Blue means the building is intact; red means it’s destroyed.

The Reality Check: How Did It Do?

Without ground truth data, I couldn't compute formal IoU or F1 scores and had to rely on a visual evaluation.

Despite the massive domain shift—from French cities to the scorched hills of California—the model successfully identified the vast majority of destroyed homes. While I noticed a few false negatives and false positives, the overall zero-shot performance was remarkably strong.

I attribute this robustness directly to the FOTBCD dataset. Training on over 220,000 diverse image pairs forced the model to learn the actual visual semantics of structural destruction rather than just overfitting to a narrow local environment.

Conclusion

This experiment proves that a well-trained model doesn't need to be explicitly fine-tuned for every new disaster type or geography. If the underlying training data is diverse enough, you can deploy the model to completely unseen, out-of-distribution scenarios like a California wildfire and get actionable results immediately.

For humanitarian organizations and disaster response teams, this capability is crucial. It means they don't have to wait days for manual mapping or custom model training—they can generate accurate damage assessments on day one, routing critical relief exactly where it is needed most.

Interested in using MBCTD for your own commercial applications, monitoring platforms, or disaster response workflows? Contact me to chat about licensing.