Lake Chad · 2026
Thirty million people lost their lake.
We watched it happen from orbit. AquaVeritas fine-tunes LFM2.5-VL-450M on Sentinel-2 imagery to classify the world’s freshwater bodies — Lake Chad, the Aral Sea, Tonle Sap, sixteen more — and produce structured, analyst-ready intelligence at a cadence the existing systems cannot match.
- Surface area lost
- 90%
- People dependent
- 30million
- Fine-tuned accuracy
- 85.4%
Since 1960. Irrigation, drought, and the basin’s collapse.
Across Chad, Niger, Nigeria, and Cameroon.
Across 10 structured fields vs 18.0% base model.
The gap
Satellites already see this. Nobody’s reading the imagery.
ESA Copernicus has flown Sentinel-2 over Lake Chad every five days since 2015. Eleven years of multispectral imagery at ten-metre resolution. The data exists. The interpretation does not.
Government water authorities can’t hire enough analysts to read every tile. Parametric insurers need a structured index, not raw rasters. Development finance institutions need to verify outcomes years after a project disburses. Today they all rely on expensive bespoke contractors, or they don’t verify at all.
AquaVeritas is the layer that turns the pixels into a verifiable water-stress index.
How it works
A 450-million-parameter vision-language model, fine-tuned on the crisis.
01
Capture
DPhi SimSat fetches Sentinel-2 tiles — RGB, SWIR, dual-zone — across twenty monitored sites on a daily cadence.
02
Classify
Fine-tuned LFM2.5-VL produces a structured eleven-field JSON for each observation. Water extent, flood risk, crop stress, shoreline encroachment, settlement, bare-soil expansion.
03
Compare
Outputs are cross-referenced against JRC Global Surface Water and GRACE-FO mass anomalies. Three independent sources agree before a signal is published.
04
Trigger
Parametric insurance customers receive webhook events the moment a verified stress condition crosses their contract threshold.