Methods & limitations
See also: Evidence page and downloadable proof pack.
How it works
- Water area: Sentinel-2 NDWI/MNDWI with Otsu thresholding inside a JRC Global Surface Water mask; Sentinel-1 radar fills cloudy months.
- Level & storage: area inverted through a Copernicus DEM area–elevation curve; SWOT lake water-surface elevation where available.
- Alerts (default, "own baseline"): the monthly change in Sentinel-1 radar water area (same method every month since 2017) is compared with our own month-of-year normal from 2017–2022. A month is flagged when the drop is unusually large (threshold fixed in advance so 10% of 2017–2022 months would flag) and bigger than 1% of the reservoir's typical area. An ERA5 rainfall proxy adds a "may be low inflow" caveat. No third-party dam data is used to produce served alerts.
- Validation only: alerts are scored against storage-change events in the Stimson Center Mekong Dam Monitor (2023 onward, out of sample). That data is not served.
Alert accuracy
Alert accuracy was checked against a third-party reference dataset that we are not licensed to publish, so those scores are not shown here. Area-based alerts are weak at Manwan, Dachaoshan and Jinghong, small, steep reservoirs whose surface area barely changes with level.
Manwan, Dachaoshan and Jinghong are small, steep reservoirs whose surface area barely changes with level, so area-based alerts are weak there.
'});Honest accuracy
| Downstream forecasting | No useful skill — we do not forecast downstream river levels |
|---|---|
| SWOT bias | SWOT reads 2–4 m low at low reservoir levels; cause unknown, not a datum issue |
| Weekly alerts | Too noisy; monthly only |
| Rainfall caveat | Weak proxy (r 0.10–0.34), not validated against real inflow |
| Translations | Thai, Lao and Khmer texts are machine-drafted and need native-speaker review |
These results are indicative only. Not for operational or safety decisions.
Data sources & licensing
Served data comes only from Copernicus Sentinel-1/2, Copernicus DEM, JRC GSW, and NASA/CNES SWOT, processed by AI Cubed. Stimson Center data was used for research validation only and is not served (no commercial licence).
Experimental flood & drought scores (v2.4, logged-in only)
For each dam, we compare this month's radar water area with the same calendar month in 2017–2022 (a z-score, capped at ±4). We also compare catchment rainfall (ERA5) for the past 1 and 3 months with 2001–2022. Drought index = −(0.6 × area z + 0.4 × 3-month rain z). Flood index = 0.6 × area z + 0.4 × 1-month rain z. Score = 50 + 20 × index, clipped to 0–100; 60 or more is "elevated" and 80 or more is "high". The weights were fixed before testing.
Backtest (2023 on, out of sample): agreement with independent SWOT low water levels was weak. It was significant only at Nuozhadu (Spearman 0.42, p = 0.007). At the other four dams it ranged from −0.32 to 0.22, none significant. For predicting next month's area, the score did no better than simply assuming last month continues. These scores are descriptive, not forecasts.