Istanbul · preparing, not predicting

Istanbul, before the next one

The model was tested on a city it had never seen. Istanbul has no post-earthquake images and no damage labels, so its accuracy there cannot be measured. What can be done now is to have everything else ready: every mapped building, the official scenario estimates per neighbourhood, and a pipeline that turns the first post-event images into a list of buildings to check first.

What a Mw 7.5 earthquake on the Marmara fault would do, building by building and street by street: the scenario →

964 neighbourhoods

Each shape is a mahalle. Scenario shows İBB's estimate of very heavily and heavily damaged buildings for a Mw 7.5 night-time earthquake. Old stock is the share of buildings built before 1980 (İBB, 2017). Footprints compares the footprints in the inventory with İBB's count: the image model can only score buildings that have a footprint.

The inventory starts from OpenStreetMap, whose outlines are drawn by people and carry storeys and building types. Where OSM has no building, Microsoft's machine-made footprints fill the gap: 456,023 of them, which brings the median mahalle from 67 % to 97 % of İBB's count. In the Antakya test the same footprints worked as an inventory made before the earthquake.

On the day

Before images arrive. The scenario layer and the building inventory say where heavy damage is most likely and how many buildings each mahalle has.
First post-event scene. Released scenes (for example through the Maxar Open Data Program, as in 2023) are cut into 600 m cells on the local UTM grid at 0.35 m and aligned to a pre-event reference image.
Every mapped building is scored with the model trained on the 2023 earthquakes. Probabilities are calibrated with the settings fitted on the validation city.
Output: a ranking, not a verdict. Buildings sorted by expected severity, a GeoJSON of the first few thousand for a field map, and a per-mahalle summary next to the scenario numbers. In Antakya, the top 10 % of such a ranking held about half of all destroyed buildings.

Tried on Antakya first

The same pipeline was run on Antakya as if it were the day after: footprints made in advance (Microsoft's machine-made building footprints, since OpenStreetMap had only 413 buildings there before the earthquake), the released Maxar acquisition of 11 February 2023, and the model and settings unchanged. On the 11,783 labelled buildings it could score, it did as well as the careful evaluation pipeline:

Same buildings	Macro-F1	AUROC	Destroyed in top 10 %	ECE
Evaluation pipeline	0.490	0.733	54%	0.083
Response pipeline	0.510	0.743	52%	0.053

23,397 buildings in 8 minutes on one GPU without alignment, 23 minutes with it. Most of the gain comes from scoring each crop in eight rotations and averaging. Details in the README.

python -m enkaz.respond --inventory data/istanbul/buildings.npz \
    --post scene.tif --pre reference.tif \
    --checkpoint post_only/best.pt --calibration post_only/eval.json \
    --out response/

What this is not

Not a forecast of the model's accuracy in Istanbul. Building stock, roof types and image conditions differ from the 2023 cities; a few hundred quickly graded buildings would be needed to check thresholds on the day.
The scenario is a planning estimate from İBB's loss model, not an observation, and it is shown here for one earthquake scenario only.
Footprints are not buildings one to one. In dense old districts (Beyoğlu, Bağcılar) both sources merge attached buildings into one block, so a row of houses is scored as one. In fast-growing districts (Arnavutköy, Çatalca) the inventory counts more than İBB's 2017 figure: new construction since then, and some sheds and small structures.
Names, not codes. İBB's tables are joined to OpenStreetMap boundaries by district and mahalle name; 937 of 959 rows match, the rest are left out rather than guessed.