We built a synthetic weather monitoring network across 24 stations in Kenya, injected a cold front, and measured how quickly two independent temperature sources — compared against each other — could detect the event and tell you exactly which stations were affected. Here's what happened.
We generated 72 hours of synthetic temperature data across 24 weather stations. At hour 36, we injected a cold front that drops temperatures by 8 degrees C in 6 hours across a cluster of stations. The question: how quickly can you detect the event, and which stations are inside the anomaly zone?
24 weather stations with calibrated thermometers, recording hourly. High accuracy but sparse coverage — each station only sees its own location.
Satellite-derived temperature readings covering the whole region. Broad coverage but with systematic calibration biases and cloud-cover gaps. Lags behind ground truth during rapid changes.
The gap between what the ground stations measure and what the satellite estimates isn't random — it's information. When the two observers disagree sharply, something real is happening. Drag the timeline to watch.
Watching ground station readings alone, you notice temperatures falling around hour 37. But you can't tell whether it's a local sensor malfunction, normal overnight cooling, or the leading edge of a cold front. Every station looks the same — just numbers going down.
By comparing ground stations against satellite estimates, the convergent approach detected the cold front at hour 37, identified Kiambu-West as the first station to diverge from satellite predictions, distinguished real weather signal from satellite calibration noise (signal rose 6x during the event), and tracked exactly which stations were inside the anomaly zone as the front moved through the region.
| Capability | Single observer (ground only) | Convergent (two observers) |
|---|---|---|
| Detects temperature change | Yes (hour 37) | Yes (hour 37) |
| Tells you WHICH stations are affected | No — all stations just show numbers | Yes — stations where the gap spikes |
| Distinguishes real event from sensor error | No | Yes — signal/noise decomposition |
| Tracks the front's spatial spread | No | Yes — affected station count over time |
| Identifies where the front arrived first | No | Yes — first deviation station |
| Works when satellite has cloud gaps | Not applicable | Yes — cloud gaps become visible as structured disagreement |
This isn't a weather-specific tool. The same principle — measuring the gap between independent observers — applies to health systems (field data vs reported data), traffic networks (sensors vs ground truth), financial anomalies (price vs fundamentals), and any domain where you have two independent ways to look at the same thing.
CAAD is being built as a service. If you have a system with multiple data sources that should agree but sometimes don't — we should talk.
Get in touch