Tokyo: land change, 2016–2025
Two Sentinel-2 satellite images of a 20 km square around Tokyo, from 2016-06-02 and 2025-06-17, were classified by a ResNet18 model trained on EuroSAT. Every 640 m cell is put in one of four land-cover groups, and the two dates are compared cell by cell.
- Before
- 2016-06-02
- After
- 2025-06-17
- Area analysed
- 394 km²
Change
-0.8 km²net change in built-up area
In cells the model is sure of on both dates (confidence ≥ 0.8): 0.8 km² became built-up, 2.9 km² changed in total.
Prior knowledge (Natural Earth): 6 cells in open water well away from the shore were counted as water; the model had seen another class there.
Land cover on both dates
| 2016 | 2025 | ± | |
|---|---|---|---|
| Agriculture no notable change | 0.3% | 0.2% | -0.1 |
| Natural vegetation no notable change | 2.6% | 3.1% | +0.5 |
| Built-up no notable change | 80.4% | 80.2% | -0.2 |
| Water no notable change | 16.7% | 16.4% | -0.3 |
Largest transitions (confident cells)
- Built-up → Water1.2 km²
- Built-up → Natural vegetation0.8 km²
- Natural vegetation → Built-up0.4 km²
- Water → Built-up0.4 km²
Limits
- Every value shown is a model estimate, not a measurement.
- Areas are not measurements but approximations counted from cells. The figures are always computed on 640 m cells; the scale only changes how the map is drawn.
- The model was trained on European images only; results for places tagged "experimental" are less reliable.
- Clouds, shadows and seasonal differences can show change that is not there.