11 damage analysis products, one earthquake

3 references, 8 satellite damage products (6 evaluated) · A&A Cell - 11 August 2026

2026-08-11

What arrived, and where each product looked

Everything landed within about eight days. But analysed extents span three orders of magnitude, so the like-for-like arena is the red core region (61 km²) where all five extents and the expert reference coincide.

M7.5, coastal Venezuela, 24 June 2026; we evaluated the 6 of 8 products that published an analysed extent; numbers frozen to the 2026-07-15 snapshot. Timeline x-axis is our ingest date, an upper bound on availability; UH and LIST provider release dates still to confirm.

Three questions

Eight products arrived for the same earthquake within days of each other. A coordinator receiving several of them in the same week has no established way to decide which to trust, for what purpose, or at what spatial scale. So we asked:

Can a responder act on a single flagged building?
Do they rank neighbourhoods well enough to prioritise?
What is the sensible way to use them in the next event?

Performance metrics by model: building-level analysis

Precision & recall: single products vs combinations

Can they rank the worst-hit neighbourhoods?

Yes: products can support neighbourhood-level prioritisation. In the core region every product ranks the worst-hit cells far better than chance (ρ 0.48–0.74, n = 133 cells, all p < 10⁻⁸). As delivered the correlations are much lower (−0.09 to 0.58), and UH is statistically indistinguishable from zero.

Three questions, three answers

Can a responder act on a single flagged building? No. Precision runs 0.05–0.09 even in the damage-dense core region: most flags have nothing wrong underneath them.
Do they rank neighbourhoods well enough to prioritise? Yes, with a condition. Where the damage zone is known and confined (our core region), most products show real prioritisation value (ρ 0.48–0.74); as delivered the signal is much weaker (ρ −0.09 to 0.58).
What is the sensible way to use them in the next event? Still open. The promising lines:
- Combine products. Counting agreement between independent products roughly doubled the accuracy of the best single one here, with no reference data needed.
- Match the operating point to the moment. Different stages of a response may favour precision or coverage differently; field and first-responder feedback should establish where and when higher precision beats broader coverage (open).
- Improve the models. Evaluations like this one show providers where the failures are.