A disaster register for 165,411 buildings, from free data alone: we had an AI layer satellite, terrain, quake and land-price data onto Fujisawa's PLATEAU model

Japan has spent more than ¥16 billion of public money on PLATEAU, a nationwide set of 3D city models. In part one of this series, we found that PLATEAU still hasn't proven how it gets used.

So how do you get value out of it? Talking it over in the newsroom, what bothered DATA WORLD's editor-in-chief, Kaito Ogasawara, was that we kept looking at PLATEAU on its own.

PLATEAU's strength is that every building has a location and a shape. Most public data doesn't come in buildings. Satellites see a city as surfaces. Earthquake probabilities come in 250-meter grid cells. Land prices are known at a few dozen points across a city. On its own, each can only tell you about "this area".

Write those values onto each building, and you can say "this building: tsunami depth so many meters, so many percent odds of strong shaking, sitting on an old riverbed". That is exactly what insurers, property firms and anyone auditing their own sites want.

Until recently, that was hard. The data comes in different formats and resolutions, scattered across government and research websites. You'd have to parse 1.7 GB of 3D data, query a grid API more than a thousand times, fetch and composite a dozen-plus satellite scenes, and join it all to 160,000 buildings. That's a job for GIS software and someone who knows how to drive it.

Now you can hand it to an AI. PLATEAU has a trial MCP server, an interface that lets AI tools search and fetch its data. The AI can also write and run the code that fetches and joins everything else.

So Ogasawara asked our newsroom's AI:

We're leaning too much on PLATEAU alone. Combine it with satellite imagery and other open data, and show us things we couldn't see before. Go deep enough that engineers and researchers go "oh, nice".

The AI came back with a plan: use PLATEAU as a container for other data, pouring surfaces, grids and points into individual buildings. For the test city it picked Fujisawa, a seaside city in Kanagawa Prefecture with fresh PLATEAU data and both coastal and inland districts.

About an hour and a half after the request, the AI had gathered five public datasets, written them onto every one of Fujisawa's 165,411 buildings, and produced the tables and maps. In effect, a disaster register for an entire city.

With a register, you can put numbers on questions that staring at maps can't answer. Here's what it answered, up front.

Fujisawa was picked because its PLATEAU data is recent and carries five kinds of hazard projections. A seaside resort and inland residential and industrial districts in one city also make it a good test bed.

Asked for Fujisawa, the PLATEAU MCP Server returned the locations of 87 files, each covering roughly a square kilometer, about 1.7 GB in all. The AI wrote short programs to pull each building's footprint, use, height and hazard projections into one table.

Four more datasets were then written onto each building. The register's columns look like this.

The join itself is simple: check which polygon or grid cell each building's footprint centroid falls in, and copy the value onto the building. With a spatial index, 160,000 buildings against tens of thousands of polygons doesn't need a brute-force search.

We layered two very different maps.

The first is PLATEAU's flood and landslide projections: river flooding, urban flooding when drains overflow, tsunami, storm surge and landslides. It's the same information that colors in a city's hazard map. In Fujisawa, 48,828 buildings (29.5%) carry at least one of them.

The second is NIED's earthquake shaking map. For each location, it gives the odds of shaking of shindo 6-lower or stronger within the next 30 years, on Japan's seismic intensity scale, the level at which heavy furniture starts to topple. In Fujisawa, even the lowest value is 33.6% and the middle value is 59.0%. Like the rest of southern Kanto, the whole city rates high.

So do the two maps point to the same dangerous places? Split by district, clearly not.

Source: PLATEAU Fujisawa building model (FY2025) and J-SHIS Probabilistic Seismic Hazard Maps 2024 (average case, all earthquakes), DATA WORLD analysis.

In the north, nine in ten buildings are blank on the hazard map, yet their odds of strong shaking are higher than on the coast. Flood risk clusters along the coast and rivers; shaking risk leans toward the inland plateaus and valleys, just as the two maps at the top show. Among the tenth of buildings most likely to see strong shaking, only 26.9% would be colored on a flood hazard map.

The difference comes from how much the ground amplifies shaking. In J-SHIS's ground model, the middle amplification factor is 1.60 on the northern plateau and 1.44 on the coastal dunes. The highest, 1.77, is in valley-floor lowlands, where 8,267 buildings have a middle shaking probability of 69.1%.

If you're auditing your own sites or your suppliers', a flood hazard map alone isn't enough, and neither is a quake map alone. Only with both on each building can you rank which ones to check first.

Next, what the land is made of. The Geospatial Information Authority of Japan (GSI) classifies land from old maps and aerial photos: plateau, former river channel, man-made fill and so on. Matching it against PLATEAU's hazard projections turned into a cross-check between datasets.

Source: GSI landform classification (natural landforms) overlaid with PLATEAU Fujisawa building model (FY2025), DATA WORLD analysis. Each building is classified by the landform under its footprint centroid.

On former rivers and marshes, nearly every building has a flood or tsunami projection; on plateaus, 4%. Data made separately, by separate agencies with separate methods, lined up cleanly at the building level. It was also a check that our join was working.

It also shows what PLATEAU's projections leave out. GSI notes that former river channels and reclaimed land have "a very strong tendency to liquefy". According to a report by MLIT's Kanto Regional Development Bureau and the Japanese Geotechnical Society, in the 2011 Great East Japan Earthquake, liquefaction occurred in 20.2% of 250 m grid cells on reclaimed land and 10.4% on former rivers and lakes. Liquefaction is not among PLATEAU's hazard projections.

Fill is the same story. Fujisawa has 182 large-scale fill sites covering about 335 hectares, with 9,842 buildings on them. Only 14% of those carry any PLATEAU hazard projection. The national government and the city say this map "does not necessarily show places that are dangerous in an earthquake". In a register, it's best kept as its own column, flagging land worth checking.

PLATEAU's buildings reflect the city as it was when the aerial survey was flown. Anything that changes afterward waits for the next rebuild. So we looked for changes with Sentinel-2, part of the EU's Copernicus program. It images the same spot about every five days at 10 m resolution, and the images are free.

For each period, four to five low-cloud winter scenes were composited into one, then winter 2023–24 was compared with winter 2025–26 for large color changes. That found 58 clusters of change of 0.4 hectares (40 m × 100 m) or more, totaling 44.7 hectares.

Checked one by one, about a third were seasonal changes in fields or golf-course turf, and about a third were changes to buildings such as new roofs or solar panels. The rest included around ten land changes such as cleared woodland (a rough breakdown, sorted by eye). The machine finds candidates; a person still has to sort them.

The largest was a patch of woodland in Kuzuhara, in the north of the city.

Checked against the prefecture's urban planning documents, it is the "New Industry Forest, Zone 2" (about 8.4 hectares), which the city is developing to attract companies. On March 29, 2024, it was moved into a zone where building is allowed, and a land readjustment association was set up. In PLATEAU, the same spot still has no buildings. The satellite is seeing where the next buildings will go before the 3D city does.

The national government has tried this too. In FY2022, a pilot compared aerial and satellite imagery with AI to map where 3D city models most needed updating, with classification accuracy of about 84%. Changes to individual houses, it found, are hard to catch. The same holds at our 10 m resolution: you see big changes like factory lots and cleared land.

Last, the money data. Is disaster risk built into land prices?

Fujisawa offers a handy test. On March 22, 2021, Kanagawa Prefecture designated Fujisawa's coast a tsunami disaster alert zone, and property sales in the zone now require it to be explained as a material fact. So we overlaid the 52 residential points of Japan's official annual land price survey on the tsunami inundation projection.

From 2021 to 2026, the six points inside the zone rose 19.7–30.0%, a median of 26.2%. The 46 points outside rose a median of 15.4%. Even the three points projected to flood 3 meters or deeper (Katase, Katase-Kaigan and Kugenuma-Kaigan) rose 25.7–30.0%. The pull of the beach seems to outweigh the tsunami projection.

This is also where the register is weakest. Six points, valued by appraisal rather than actual sales, with no control for things like distance to a station. A 2022 Bank of Japan working paper using about 3,500 official land price points nationwide found residential land about 1.1% cheaper for every extra meter of projected flood depth. Risk is priced in, but only a little. We read Fujisawa's six points as an example of that "little".

The biggest lesson from building the register: the numbers on one building's row carry very different weight, column by column.

Then there's time. The tsunami projection is from FY2014, the use survey from FY2021, the satellite imagery from 2026. Putting data from different moments onto one building misaligns anything rebuilt in between. It's safest to store, alongside each value, which year and which resolution it comes from.

EDITOR'S TAKE ── Editor-in-Chief Ogasawara's reading

PLATEAU turned out to be a building ID that ties data together

Building a 160,000-building register, I came away thinking PLATEAU's biggest value isn't the 3D look. It's tying other data to individual buildings.

Satellites see a city as surfaces. J-SHIS uses 250 m grid cells; land prices are a few dozen points. Pour them into PLATEAU's buildings and they line up in one row, a register you can sort and filter. As part one showed, PLATEAU hasn't yet proven how it gets used. But used as a container for other data, it lets one company, working with an AI, audit 160,000 buildings in about an hour and a half. That's value the government hasn't fully counted yet.

This time, satellites caught a development site before it reached the 3D city. As part one counted, 73% of the 79 municipalities first modeled in FY2020–2022 have never received a new version.

Elsewhere, building registers are updated every year. The Netherlands' 3DBAG combines the national building register with laser survey data to generate 3D models of about 10 million buildings fully automatically. If PLATEAU used satellites and AI to find what changed each year and rebuilt only those parts, there would be plenty of room to bring update costs down.

Per-building risk can already be lined up in data. Meanwhile, the reference rates for flood cover in Japanese home fire insurance only moved to five tiers by municipality in the June 2023 revision. And in a pilot attaching "real estate IDs" to PLATEAU buildings, only 8% could be matched across 66 municipalities. The next job is building the plumbing that connects a register like this to contracts and prices.