The sensor-vendor pitch
Every soil-moisture sensor vendor we have ever met has the same pitch deck: lab calibration to ±1.5%, NIST-traceable, ISO-something. Then you put it in the ground and it lies to you within four weeks. The lab number was honest. The deployment was not.
Why they lie
Three failure modes, in order of severity:
- Soil contact degrades. Roots grow into the gap, water etches the probe, the dielectric reading drifts. ~46% of the deployed-vs-lab error.
- Salinity confounds capacitance. Brackish-water installations — which is most of what Babylonian deals with — throw the dielectric reading off by a salinity term the sensor wasn't calibrated against. ~31%.
- Temperature compensation is wrong by the second season. Most temperature-comp curves are baked into firmware against a controlled lab. Real soil cycles colder and wetter than the lab. ~18%.
The trick that worked
We do not buy more expensive sensors. We buy three cheap ones per bench and run a Bayesian state-space model over them. The model has explicit terms for soil-contact decay, salinity confound, and a learned temperature-compensation residual.
The cheap-with-wrapper rig outperformed the precision sensor by 2.3× measured against a destructive ground-truth campaign across 24 benches over six months. It was also one-fifth the per-bench cost.
The protocol
The protocol we now ship is a checklist, not a model:
- Three sensors per bench, never one. Disagreement between sensors is signal, not noise.
- Recalibrate the Bayesian prior every 14 days against an in-bench reference (a soil-cube weighed dry/wet on a schedule).
- Replace probes on confidence, not calendar. The model tells you when a probe has decayed past the point of being usefully informative.
- Log everything. The dataset of sensor-vs-ground-truth disagreement is the most valuable single asset Babylonian owns.
Three things to take
- Lab calibration is real, but it's not what you measure in production.
- Three cheap sensors with a learned model beats one expensive sensor. Almost always.
- The data on disagreement is your moat. It is harder to reproduce than the model.