Crop models and decision support

A simulated harvest

A model estimates yield for two possible irrigation schedules. That can help a grower examine a decision before either schedule has been used. Both results are simulations: their difference describes what the model predicts under its assumptions, not a harvest already observed.

A crop model expresses selected relationships in mathematical form. It may represent development, water use, biomass and harvestable output. The model’s purpose determines what it includes. An appropriate tool answers a defined question with adequate evidence; a more detailed model does not automatically make a better local prediction.

Inputs and represented mechanisms

FAO’s AquaCrop is a water-driven example. Its documented calculation links canopy development, transpiration, biomass and yield, using weather, crop, soil and management inputs. A harvestable output is not the same quantity as all crop biomass, just as transpired water is not the same as all water delivered to the field.

A model can represent some stresses while simplifying or omitting others. The selected AquaCrop overview excludes pests and diseases and does not simulate detailed nutrient cycles. Therefore a yield prediction should be read with its represented processes and limits, not as a complete diagnosis of everything that can affect the crop.

Calibration and validation

Calibration adjusts selected model parameters using observations. Validation evaluates predictions against observations not used for that fitting. Reusing the same observations for both tasks makes it harder to assess prediction beyond the fitted cases. Keep the model version, parameter choices and the two data roles identifiable.

A primary cassava study used two-thirds of its field data for calibration and the remainder for validation. Its authors also described limits involving missing observations, cultivars and the growth cycles represented. This illustrates a bounded evaluation, rather than proving that any crop model is accurate everywhere. New uses require evidence relevant to their conditions.

Look at more than an average error

Suppose three invented held-out observations are 5, 6 and 7 t/ha, and predictions are 5.5, 5.5 and 7 t/ha on the same stated yield basis. Define error as prediction minus observation. Errors are +0.5, −0.5 and 0 t/ha. Their mean is zero because opposite signs cancel.

Mean absolute error uses their magnitudes: (0.5 + 0.5 + 0)/3 ≈ 0.333 t/ha. The diagram compares these two summaries. Neither three cases nor either metric establishes reliable prediction over all future seasons. Inspect errors across the conditions and outputs that matter for the intended decision.

A scenario needs its boundary

In another invented case, predicted yield changes from 6.0 to 6.2 t/ha when a modelled management input changes. The difference is 0.2 t/ha, or 200 kg/ha. This is a modelled response under stipulated conditions. No field intervention, harvest gain or economic return has been measured.

An honest scenario states what changed and what was held fixed. Consider whether weather inputs are observations, forecasts or hypothetical sequences, and whether the soil and crop descriptions match the site. Testing several plausible inputs can reveal sensitivity, but it does not replace evidence about the actual values or justify an invented probability.

Use the output with evidence

Decision support combines model outputs with observations, practical constraints and judgement. A simulation that conflicts with local measurements may reveal a poor input, an unsuitable parameter or a missing process. Preserve the disagreement and investigate it rather than automatically changing the observations to fit the model.

Before relying on a prediction, ask whether its evaluation covered comparable crops, climates, soils, management and outputs. A model may be useful for comparing possibilities while uncertainty remains too large for a precise local recommendation. Field-trial design supplies complementary evidence about actual interventions and their resource and quality outcomes.

Worked examples

Check your understanding

1. A model reproduces the same observations used to adjust its parameters. Is that independent validation?

   Show answer and reasoning

   No.

   The fitting observations are calibration evidence; evaluation beyond fitting needs appropriately separate observations.

2. Why must a comparison state whether yield is fresh or dry mass?

   Show answer and reasoning

   They are different mass bases.

   A numerical agreement can be misleading if the model output and measurement do not describe the same quantity.

3. A model omits a process that caused a major field loss. What does its output establish about that loss?

   Show answer and reasoning

   It does not establish a prediction of the omitted process.

   The model’s represented mechanisms set limits; another measurement or model may be needed.

Key ideas

• Name the model, inputs, output basis and omitted processes.
• Keep fitting and prediction evaluation distinct.
• A simulated difference supports a conditional comparison, not an already measured outcome.