The models behind the field work
Three trained models run every week: two estimate crop stress, trained on our own pressure chamber readings matched to soil probes and weather, and one estimates the water each block applied from probes, weather and pump records. The Thursday draft water order comes from rules mined from our agronomist’s past orders. On Platform, your own Certified Crop Adviser or adviser makes the call. On Managed, our Certified Crop Adviser does.
On a block the model has never seen, a soil-and-weather model ranks stress far above the seasonal calendar, and one chamber reading improves it.
Held-out weeks, 2023 to 2026, on one citrus estate; readings 14 to 28 days old. Research result, not a customer promise.
We use trained models on our own data to help your agronomist decide what to check, with the farm’s own pressure chamber readings judged against each block’s stress line; between visits, alerts flag a missed set or a block that is dry and overdue.
Use the sensors and crew you already have with Platform, after we confirm the sensor connection. AgWaterAI adds crop-stress estimates, a draft water order and signed probe reading summaries. With Managed, we install and own the sensors and run the field program.
Your agronomist reviews the evidence and makes the irrigation call. The signed record shows what was recorded and whether it changed after signing; it does not prove a measurement was correct.
SENTRY
SENTRY is a trained classification model that estimates crop stress, trained on our own pressure chamber readings matched to soil probes and weather. It helps the agronomist choose where to check between visits. It needs local chamber readings and does not replace the chamber.
In the decision path
PULSE
PULSE is a trained regression model that estimates the water each block applied from probes, weather and pump records. An estimate stays an estimate; a model output is not a flow-meter reading.
In the decision path
WATCHDOG
WATCHDOG is a trained classification model that estimates tree water stress from soil, weather and rootstock data, with no pressure-chamber readings among its inputs. We use it on blocks without recent chamber readings. Its tests cover the farm it was trained on; a new farm needs its own assessment, and a block's own recent chamber readings come first. It is not a sensor-fault detector: separate rule-based sensor health checks flag probes that stop reporting or return error values.
In the decision path
PRESCRIBE
The Thursday draft water order comes from rules mined from our agronomist’s past orders. On Platform, your own Certified Crop Adviser or adviser makes the call. On Managed, our Certified Crop Adviser does. We test trained models against the current rules before using them to draft orders.
Rules from past orders are in use. Trained versions are being tested.
SALT FLUSH
We are testing a trained model for salt flushing. It does not make field recommendations, and we have not published an accuracy figure for it.
In testing. No field recommendations.
Testing forward in time asks how a model handles later readings. Holding out a site or soil texture asks a harder transfer question. Neither predicts results for a new customer. AUC describes classification; R² describes a regression fit.
| Model | Temporal test | Held-out test |
|---|---|---|
| PULSE | R² 0.7345 in-season / 0.7249 held-out site | Site median R² 0.7249 |
| SENTRY | Raw AUC 0.9226 walk-forward, leave-one-site-out 0.9287 median / 0.9153 pooled | Site median AUC 0.9287; pooled 0.9153 |
| WATCHDOG | AUC 0.8038 walk-forward, leave-one-texture-out 0.7873 mean / 0.6993 pooled | Leave-one-texture-out (LOTO) mean AUC 0.7873; pooled 0.6993 |
Research comparison, estate data
12,000+ pressure chamber readings, 4,000+ of them on blocks with their own soil probe logging that day (through Oct 6, 2026). Our models draw on years of pressure chamber readings matched to the same blocks’ soil probes.
On a new farm, our tests have not shown a head start from models trained on other farms. Your adviser weighs the model estimates against local readings. Chamber readings, when available, are judged against the block's stress line.
Results from one estate do not show how the models will perform on another farm.
The research and its qualifiers are available in the investor data room. We report each model's test method, evaluation data and limits separately, and we do not claim that a model trained on one farm carries over to another.
Probe reading summaries are signed and tamper-evident from signing forward. Each issued water order is kept with the readings behind it. A signature shows a summary has not changed since signing; it does not prove a measurement was correct.
Experiment, not a visit schedule
In the cadence experiment, a candidate raised more false alarms.
At the same false-alarm rate, this candidate caught fewer stressed readings than the last chamber reading. The candidate failed that gate. We withheld the cadence candidate from promotion.
This does not establish a reduction in field labor. Shadow results and agronomist review must come before a change to the chamber schedule.
Metrics come from the production model registry.
See how a field visit becomes a decision