From Monitoring Crops to Anticipating Risk

"What happened in this field last week?" and "What is likely to happen in this field next week if nothing changes?" are two different questions, and most agricultural monitoring systems have historically only been built to answer the first one.

Monitoring is retrospective by design — it records a state. Prediction requires a model of how that state tends to evolve under a given set of conditions, built from enough historical cases that the pattern is more than a guess.

In practice, this shows up as a ranked list rather than a single alert: which fields, among many being monitored, are showing early combinations of stress indicators, weather exposure and crop-stage vulnerability that have historically preceded a measurable yield or quality impact.

Why anticipation deserves investment

FAO’s The Impact of Disasters on Agriculture and Food Security 2025 estimates agricultural losses of US$3.26 trillion during 1991–2023, averaging about US$99 billion annually. Asia accounted for 47% of those losses, or US$1.53 trillion. These historical estimates describe disaster exposure across agriculture; they do not represent losses that a single predictive tool could prevent. FAO’s report summary establishes the scale of the problem.

Prediction is useful when it changes an action while options remain available. A water-stress forecast can support an irrigation review; a disease-risk forecast can bring forward scouting; a harvest-risk forecast can help an enterprise review labour and logistics. The target must be specified. Predicting a canopy index, predicting a pest outbreak and predicting final yield are different tasks, with different evidence and validation requirements.

Define the event and the decision window

A credible model begins with a question precise enough to test. Which crop and growth stage are included? What event counts as a problem? How far ahead is the forecast issued? What information was available at that moment? These definitions prevent a retrospective pattern from being mistaken for an operational forecast. Data collected after the event must not enter the earlier prediction.

For a water-stress application, the forecast could combine current crop condition, expected atmospheric demand and estimated water available in the root zone. FAO’s established crop-water method links reference evapotranspiration with a crop coefficient. The FAO method provides a physical basis for demand estimation; it does not remove uncertainty about rainfall, soil properties or actual management. A model should state how those unknowns affect its result.

A probability also needs a time window. An illustrative ‘60% risk over the next seven days’ is more interpretable than an unexplained red score. Across many comparable cases assigned that probability, the event should occur roughly 60% of the time if the model is well calibrated. That example defines calibration; it is not a forecast for any current farm.

Prioritise by consequences and feasible response

The highest probability does not automatically deserve the first visit. A moderately likely problem in a sensitive crop stage may justify attention before a more likely but minor issue in a mature crop. Operational priority should consider expected severity, area exposed, time remaining to act and whether an effective response is available. The assumptions behind that ranking should be visible to the agronomist.

Consider an illustrative programme monitoring 1,000 fields with capacity to visit 50 in a week. A ranked queue can help allocate that capacity, but visiting only the highest-ranked fields creates a blind spot. Reserving some visits for lower-ranked or randomly sampled fields helps detect missed problems and assess whether the ranking works. The sampling plan should reflect local conditions and the cost of inspection, rather than an arbitrary universal percentage.

The practical benchmark is the existing workflow. Compare the model with ordinary scouting schedules, recent-condition rules or the agronomist’s current prioritisation. Measure false alerts, missed events, useful lead time and cost per actionable finding. A more complex model is justified only if it improves decisions enough to offset its operating and review costs.

Connect forecasts to people and institutions

Predictive services need an observation network as well as algorithms. FAO’s 2025 disaster report describes its Fall Armyworm Monitoring and Early Warning System operating across more than 60 countries. Its FAMEWS documentation explains how geolocated scouting and trap records are reviewed and shared. These examples demonstrate an organised evidence flow; they do not establish a universal prediction accuracy.

WMO’s 2025 climate update reported that 119 countries had multi-hazard early warning systems in 2024, compared with 56 in 2015, while 40% still lacked them. Its early-warning assessment identifies risk knowledge, forecasting, communication and preparedness as complementary requirements. Agriculture has the same practical dependency: a sound forecast can fail to protect production if nobody receives it, trusts it or has the means to respond.

When a Weather Forecast Is Not Enough explains how exposure becomes crop-specific risk. Where AI Can Actually Help the Agronomist sets out the human review that a prioritisation system needs. The enterprise objective is a repeatable process in which forecasts prompt appropriate checks, actions are recorded, and subsequent outcomes improve the next decision.

← Back to Intelligence Hub