AI Irrigation Scheduling: When to Trust a Soil-Water Prediction

See how an AI model predicts root-zone depletion in maize and soybean, and check whether its error could change an irrigation decision.

A cutaway soil profile shows plant roots beside a moisture probe and a surface drip line.
A grower can check the AI model’s prediction against moisture measured at several depths.

An AI model estimates that a crop is close to needing water. Is it close enough to start the pump? In a field test on maize and soybean, the model predicted soil-water depletion with errors measured in tens of millimeters. Whether a grower can act on that prediction depends on how far the block is from its irrigation trigger and what the soil shows today.

How an irrigation model learns soil-water depletion

Precious Amori and colleagues used data from 2020 and 2023 to train several models on maize and soybean. The inputs brought together weather, soil properties, remote observations and management history. The target was soil-water depletion, the depth of water missing from the crop’s root zone relative to field capacity. Their best performer on an independent 2024 season was extreme gradient boosting, which builds a sequence of decision trees, each helping correct earlier prediction errors. Amori and colleagues’ field deployment and validation study.

An illustrated grower consults a tablet overlooking crop rows.
Ask the tool for the predicted quantity, its units and the observation date. An estimate of depletion can be checked against the profile in this block.

In 2024 the model’s root-mean-square error was 22 mm for maize and 24 mm for soybean. This statistic summarizes errors across observations and gives larger misses more weight. The model generalized less well to a separate 2021 dataset. Agreement between a model and another irrigation scheduler is a useful cross-check. To find out how much water a farm actually saves, a trial also needs metered applications and a harvest comparison.

Ground-check the model at the right depth

To test a prediction, match its date, depth and location with the observations. Begin with a unit the farm can water independently. Record the crop, growth stage, active rooting depth, soil layers and time of the last rain or irrigation. A seedling and an established crop can use different depths of the same soil. Compare measurements within each management zone before averaging across fields. The field-scouting example explains why a delivery fault deserves its own check.

Young plants and small points of light follow a row of dark soil.
A probe samples its surroundings. Record its depth and position relative to the row and emitter so the next reading is comparable.

Field capacity describes how much water the soil retains after excess water has drained. The permanent wilting point is the lower reference used in estimating plant-available water. Their difference, multiplied by the thickness of a soil layer, gives that layer’s available water capacity. The difference between field capacity and the current water content gives its depletion. Add the layers inside the active root zone. Local soil values, crop stage and the irrigation system determine the depletion that should trigger action. Minnesota Extension on sensors, soil water and placement.

Can an independent water budget reproduce the model’s call?

Return to the fictional block where the model reported 35 mm. Two days earlier the grower took soil readings. The block has an effective root depth of 600 mm at the current crop stage. It has two 300 mm layers. The moisture readings are volumetric fractions, so 0.25 means water occupies one quarter of the measured soil volume. The example uses millimeters of water depth throughout. Its soil values and trigger are teaching assumptions.

  1. Upper layer. Field capacity is 0.30 and current water content is 0.25. The depletion is (0.30 − 0.25) × 300 mm = 15 mm.
  2. Lower layer. Field capacity is 0.29 and current water content is 0.25. The depletion is (0.29 − 0.25) × 300 mm = 12 mm.
  3. Starting balance. The measured root zone is 27 mm below field capacity. Suppose the available fraction between field capacity and wilting point is 0.14 in each layer. Total available water is 0.14 × 600 = 84 mm. A locally chosen trigger of 40% depletion would be 33.6 mm.
Young plants grow in furrows beneath an open sky.
As roots deepen during the season, update the soil layers used in the balance. An old rooting depth changes the calculated storage.

Carry the balance forward from the measurement time. Add estimated crop evapotranspiration (ET), subtract rain retained in the root zone, and subtract net irrigation retained there. Reference ET from a weather station describes a reference surface. When using a crop-coefficient method, use the coefficient appropriate to the crop and growth stage to estimate crop ET. We assume negligible groundwater contribution in this example. Runoff and drainage beyond the roots do not refill the active zone. Start a new balance when a field measurement establishes a better starting point. Extension’s irrigation checkbook method.

Account for the time the system needs to reach the block, any change in rooting depth and the lead time of the rain forecast. If the readings already include yesterday’s rain, enter only events since those readings. The satellite ET guide explains why a monthly map alone cannot settle a daily schedule.

After accepting a recommendation, measure what the system delivers

Suppose the dry scenario occurs and a field check supports the 35 mm deficit. For this example the grower wants to leave 15 mm of storage below field capacity. That means adding 20 mm net. If the estimated application efficiency for this irrigation is 0.80, the gross depth is 20 ÷ 0.80 = 25 mm. Across one hectare, 25 mm is 250 cubic meters at the point where gross delivery is measured. A different area, wetted pattern or efficiency estimate changes the result.

A cutaway landscape shows soil layers beneath an irrigated surface.
Follow the applied water into the root zone. Water running off or draining below active roots cannot supply the planned refill.

Check the block area, meter volume and operating flow before converting a depth to runtime. After the run, compare timed emitter output across the block. Recheck moisture at the relevant depths after redistribution. If the inlet meter shows the planned volume but one end stays dry, turn back to delivery uniformity. Extending the whole run can push water below the roots in the wetter rows.

Download the field water record for checking a prediction (Markdown) to keep soil layers, observed rain, forecast scenarios, meter totals and the next field check together.

Illustrated field notes

These conceptual illustrations show the other evidence a scheduler may use. Open a subject for the large images.

Weather, remote observations and drought
Irrigation pipes stand over a wet surface in the evening light.
Record forecast rain separately from measured effective rain. Update the decision after the event.
A small aircraft passes over a multicolored map of agricultural fields.
Write down the acquisition time and what the colors estimate before using a map to place a moisture sensor.
An irrigated furrow reflects a setting sun while a dry landscape extends beyond.
A seasonal supply limit changes how a farm allocates water among blocks. Each block still needs its own balance.
Automation, connectivity and thresholds
An illustrated valve or controller sits beside orderly fields.
Evaluate an adaptive controller against a named schedule, using the same crop stage, water source and operating constraints.
Glowing paths follow a curving channel in a field at dusk.
Keep the last successful reading time beside a connected sensor value. Treat a missing transmission as missing data.
Small plants and a traced curve appear beside the edge of a cultivated row.
Record why a depletion trigger changed, such as a new crop stage, a shallower rooting depth or a longer irrigation interval.
Costs, previous seasons and shared records
A tractor works at night amid graphical curves on a field.
Metered water, pumping energy, labor and marketable yield belong in the same cost comparison.
A monitoring display overlays a wooded agricultural landscape.
Test a model on a different season. A strong fit to the years used for training can conceal errors in an unusually wet or dry year.
Two silhouetted people stand beside a translucent landscape and instrument graphics.
Keep units, sensor depth, soil type and calibration history with shared data so a second grower can interpret it.