Insurance adjusters, FEMA inspectors and local emergency managers take two to four weeks to compile a damage picture after a disaster. Sentinel-2 photographs the whole county at 10 m resolution within five days of the event, for free. AlphaEarth, which I published on 7 September, is the pipeline between those two facts: before and after composites, index change detection, a damage tier for every property parcel, and a score against independent damage records.
Two scenarios so far: a flood in Perry County, Kentucky, and a wildfire in Los Angeles County. 247 km² and 24,608 parcels in total, 176 seconds of compute on one laptop core. Each configuration below adds one layer to the one above it, measured against complete ground truth.
| Configuration | Exact tier | Macro F1 | Wrongly flagged | Missed |
|---|---|---|---|---|
| Flood, Perry County, KY (15,635 parcels) | ||||
| Thresholds only | 81.8% | 0.325 | 1,816 | 446 |
| + mitigations | 93.3% | 0.437 | 17 | 603 |
| + supervised model | 99.3% | 0.906 | 22 | 10 |
| Wildfire, Los Angeles County, CA (8,973 parcels) | ||||
| Thresholds only | 93.1% | 0.445 | 554 | 0 |
| + mitigations | 95.4% | 0.489 | 345 | 0 |
| + supervised model | 98.8% | 0.671 | 42 | 3 |
The counts matter more than the percentages. For the flood, the mitigations cut wrongly flagged parcels from 1,816 to 17, at the cost of more misses, because every mitigation only subtracts. The supervised model, 43 features per parcel into gradient-boosted trees, brings the misses down to 10. The wildfire is harder: in leave-one-event-out validation the model beats the thresholds in only 2 of 4 held-out events.
Three details that decided whether it worked
- Flood detection needs both a level test and a change test. Without the change test, every permanent river is classified as flood on every run.
- Wildfire uses relative dNBR. Absolute dNBR measures how much signal was lost, so chaparral, which starts near NBR 0.38, is under-called against forest starting near 0.66, right across the wildland-urban interface.
- The model is cross-validated leaving one event out. Parcels inside one event share composites, sun angle and a hazard surface, so a random parcel split leaks all of it.
The report page renders every number from a data file. Nothing is typed into the page, so regenerating the data moves the page with it.