Istanbul: The First Night

Question

I lived for two years as a student in Turgut Özal, a neighbourhood in Esenyurt. I was at home during the Mw 6.2 Silivri earthquake on 23 April 2025. No buildings collapsed, but the streets were more crowded than I had ever seen, and a trip that normally takes 15 minutes took an hour and a half.

The project started there: on the first night after a larger earthquake, where can people go on foot, and is there room for everyone there?

Data

• More than 25 open sources: TÜİK, the Istanbul Open Data Portal, OpenStreetMap, Overture Maps, USGS, GEM and the European Commission’s GHSL.
• I merged them into a single table of 959 neighbourhoods. Neighbourhood names are spelled differently across sources, so I matched them by official neighbourhood code.
• I documented every cleaning decision: 55 decisions and 56 automated quality checks. The whole table is rebuilt from the raw data with a single command.

Method

1. Shaking. I tested eight ground-motion models against real recordings of the 2025 Silivri (47 stations) and 1999 İzmit (25 stations) earthquakes and weighted them by fit.
2. People outside. I generated 1,000 possible scenarios for an Mw 7.5 rupture in the central Marmara and calculated, neighbourhood by neighbourhood, how many people would be outside on the first night.
3. Assembly assignment. I assigned about 49,000 population blocks to open spaces larger than 2,000 m² within 1 km walking distance, respecting their capacity. This is a linear program with about 170,000 variables; the HiGHS solver solves it in half a second. In rural neighbourhoods (under 300 people per km²) I also counted unbuilt land as open space, because people there can step out into the field next to their house.
4. Building heights. I built a LightGBM model for about 151,000 buildings with unknown floor counts; the map draws its 3D buildings from these estimates.

Findings

• On paper, Istanbul’s open spaces hold 25 million people. Once walking distance is taken into account, about 2.6 million of the roughly 9.1 million people outside on the first night of the median scenario find no space.
• The gap is largest in the dense districts on the European side: 64% of people outside in Bağcılar, 57% in Avcılar and 55% in Bahçelievler find no space.
• In Turgut Özal the share is 93%. The neighbourhood has 28,564 residents in 0.64 km², about 44,000 per km² and 3.7 times the Istanbul median.
• One of the ground-motion models (AS08) overestimated the 2025 recordings by about 2.3 times, which is why weighting matters.
• On districts it never saw in training, the building-height model has a mean error of 1.66 floors, against 2.26 for a rule-based estimate.

What I don’t know

• I published no death counts. I ran common damage and loss methods, including my own calibrations, with the real shaking of 6 February 2023; all of them consistently underestimated the actual losses. I chose not to publish a number that open data cannot test.
• The official list of assembly areas is not openly available in bulk, so I used potential areas derived from open data.
• How much of each area is usable, how much space each person needs and how many people go outside are assumptions. The result depends most on the last one.
• This is an independent study; it does not evaluate the plans of public institutions and does not replace them.

What I learned

I started the project to estimate deaths; when the test showed it did not hold, I changed the question. A less ambitious but testable question gave a more reliable result.