River Monitor: evidence for MIT Solve judges
The problem
Large upstream dams on the Lancang–Mekong change when and how much water reaches downstream communities, but operating data is not published. River Monitor uses free satellite data and public river gauges to give an independent, cited view of what upstream reservoirs and the river are doing.
Who is affected
- Nearly 65 million people live in the Lower Mekong Basin; an estimated 80% depend on the river and its resources for their livelihoods. (Mekong River Commission, 'Mekong River Basin')
- About 40% of the Lower Mekong Basin's people live within 15 km of the Mekong mainstream. (Mekong River Commission, 'People')
- The Mekong's annual fish catch is about 2.3 million tonnes; MRC values Lower Basin capture fisheries at around USD 11 billion a year. (Mekong River Commission, 'Fisheries Monitoring')
- Fish provides 81% of protein intake in Cambodia and 48% in Lao PDR. (Mekong River Commission, 'Fisheries Monitoring')
All figures are quoted from the Mekong River Commission, with links.
What we cover now (counted live from our files)
| monitored sites | 18 |
| lancang dams with radar series | 5 |
| reservoir months of radar data | 590 |
| radar record | 2017-01 to 2026-10 |
| mrc gauge stations | 8 |
| gauge station days | 88055 |
| downstream towns with travel times | 10 |
| monthly dam reports available | 60 |
| stations very low now | 7 |
Our results, honestly
| Test | Result | Verdict |
|---|---|---|
| Monthly dam drop signals vs Chiang Saen rises | precision 0.57, recall 0.36; chance 0.55 | No skill above chance yet |
| Daily release-pulse travel to downstream towns (held-out 2018–2026) | 54 held-out pulses; validated towns: 0 of 6 | Experimental: no town beat persistence |
Known weaknesses
- Release-pulse detector (rise of at least 0.38 m in a day at Chiang Saen with little rain): on the held-out test (Nov 2018 – Oct 2026) no downstream town beat the persistence baseline. Labelled experimental.
- Monthly flow-alert detector: no skill above chance (see the table above).
- Radar reservoir outlines are raw and uncalibrated except at Xiaowan; at Nuozhadu, mixed shoreline pixels make radar and optical areas differ.
- Water area is a signal, not storage volume. We cannot see dam operating decisions directly.
Roadmap
- Native-speaker review of Thai, Lao, Khmer and Vietnamese alert text before any alert is sent.
- Field trip to meet river communities and partners, and collect ground-truth gauge readings with the offline field app.
- Improve the radar water-area method at dams where agreement with SWOT is weak (Manwan, Jinghong, Nuozhadu).
- Re-run the pre-registered detector tests on new data; keep labels 'experimental' unless a test beats its baseline.
- Seek written permission from the MRC before any commercial use of PMFM gauge data.
Field trip plan
Trip dates not yet set. Plan: visit downstream towns covered by our alerts, meet local partners, test the offline field app and gather ground-truth readings.