Agriculture
Physics already knew the answer
A cooperative wanted a predictive model for grain spoilage across member silos. The measurement that detects spoilage earliest is a cheap sensor and a threshold. The model would have been an expensive way to be later.
A prediction model that a sensor and a threshold made unnecessary.
Problem
Spoilage is found during routine inspection, by which point a hotspot has usually spread. The board asked for a predictive model using weather data, intake moisture and historical loss records.
Context
Roughly two hundred member farms with on-farm storage of mixed quality, plus three central sites. Inspection is manual and infrequent at the member sites.
Current architecture
Temperature cables at the central sites, nothing at most member sites, and a spreadsheet of intake moisture readings.
Constraints
- Connectivity: many member sites have poor or no mobile coverage.
- Skills: no one at the cooperative maintains software.
- Capital: members fund equipment individually, so anything expensive will be adopted unevenly.
Evidence
Each statement placed on the ladder before it was used.
Carbon-dioxide monitoring detects hotspots from water ingress earlier than other available methods; CO₂ rises well before grain temperature does.
Temperature and humidity cables in the grain mass, with CO₂ sensors in plenum and headspace, are the established instrumentation; wireless sensors avoid the installation cost and structural load of roof-anchored cables.
Loss across member storage is material to the cooperative. Claimed in board discussion; the loss record is incomplete and self-reported.
A model over weather and intake data could predict spoilage earlier than instrumentation would detect it.
Questions that changed the answer
- What would the model use as a label, when losses are self-reported and inconsistently recorded?
- How much earlier than a CO₂ threshold would a prediction need to be to change what anyone does?
- What action follows an alert, and is anyone available to take it at a member site?
- Who replaces a flat battery in a sensor three years from now?
- How does an alert leave a site with no mobile coverage: store-and-forward, a LoRa gateway at the yard, or a poll when someone is next on site?
Options
Including the one nobody wanted to discuss.
Predictive spoilage model
Train on weather, intake moisture and loss history to forecast risk by bin.
What it costs: Depends on a loss record that is incomplete and self-reported, and competes with a physical signal that appears before any forecast would.
Instrument and alert
Wireless CO₂ and temperature sensors with a threshold alarm to the member’s phone.
What it costs: Capital per site, plus a battery-replacement routine that someone has to own.
More frequent inspection
Increase the manual inspection schedule.
What it costs: Labour the cooperative does not have, and still finds hotspots after they form.
Economics
Four horizons, not one estimate.
- Build
- Effectively nil for the sensor route. It is procurement and installation, not software.
- Run
- Batteries, replacements and a small alerting service. The model route would have added retraining and a data pipeline with no owner.
- Change
- A threshold is changed by editing a number. A model is changed by a project.
- Exit
- Low. Sensors are removable and the data is a time series in a standard format.
Decision
Buy instrumentation. Do not build a model. Revisit prediction only if alert data later shows a pattern a threshold cannot express.
Why
The biological process announces itself in CO₂ before it shows in temperature. A prediction would have to beat a direct measurement of the thing itself, and would be trained on a loss record too poor to support it.
Why not the alternatives
- Predictive spoilage model: No reliable label, and it competes against a physical signal that arrives earlier than any forecast.
- More frequent inspection: Costs labour that does not exist and still detects hotspots only after they form.
Trade-offs accepted
- Uneven adoption, because members fund their own equipment. Accepted, with the central sites instrumented first as the demonstration.
- A permanent maintenance obligation the cooperative had not previously carried.
Reversibility
Sensors can be removed and redeployed. Nothing about this decision forecloses building a model later. In fact it produces the first dataset that would justify one.
Complexity budget
A cloud platform, a data lake and a dashboard were all proposed. A threshold alarm and an SMS earn their place; the rest did not, on two hundred sites with no software staff.
Decision gate
GO on instrumentation at the central sites first, with a member rollout conditional on the alerting actually changing what people do.
Implementation
Wireless CO₂ and temperature sensors, a threshold alarm, and a documented response: who is called, and what they do when they arrive. Where there is no mobile coverage the gateway stores and forwards, and the member is told plainly that the alert is delayed and not live; sites with neither coverage nor a gateway stay out of scope until that is solved.
What the decision was expected to achieve
Hotspots detected before they spread, and a real dataset of alert-to-intervention outcomes. Time from first CO₂ rise to intervention is the measure.
No outcome is claimed. This is an illustrative example, so there is nothing measured to report, and a real engagement would state what happened and how it was verified.
Lessons
- When the physics gives you a direct early signal, a prediction has to beat it, and usually it cannot. ‘Predictive’ is not a synonym for ‘better’: earlier is better, and a sensor was earlier.
- Ask who changes the battery. Run cost in the field is people and logistics, not compute.