Spraying only the weed: how to audit agricultural AI field by field

On August 22, 2024, Iowa State University published the results of a field-scale demonstration: a camera-equipped sprayer reduced the area receiving post-emergence herbicide by an average of 76 percent across five soybean fields. The figure is promising, but it does not mean the machine will always cut product by three quarters or that artificial intelligence has solved weed control. It means something narrower: given the plant pressure found across 415 Iowa acres, a commercial system opened only the nozzles associated with areas where it detected target vegetation.

That distinction matters to any farm considering the technology. Savings are not a fixed property of the algorithm; they emerge from weed density, the preceding agronomic program, detection, travel speed, boom height, nozzles and tank mix. The transferable skill in this case is turning a marketing percentage into an on-farm test protocol: measure avoided coverage, control efficacy, errors, total cost and crop effects before extrapolating.

From one field decision to thousands per second

In a broadcast application, the command is essentially binary across the working width: the boom moves and sprays. A selective system divides that action. Cameras observe the surface; onboard processing classifies each zone; the controller associates a detection with a nozzle or nozzle group; and a valve opens for the required interval. The result is not a tractor that “understands” agronomy. It is a perception-and-action chain built for a bounded objective.

Iowa State’s technical account separates two problems. “Green on brown” systems detect vegetation against soil or residue and can be used, for example, before the crop emerges. “Green on green” systems must distinguish a weed from an already visible crop. The latter needs richer visual information and finer classification. A reported accuracy rate that does not identify the evaluated setting hides a decisive part of the difficulty.

The camera does not work alone. Iowa State lists stable boom-height control, typical speed limits of 10 to 15 miles per hour, individual nozzle control, compatible nozzles and recirculation to maintain coverage while the number of active outlets changes continuously. Every element creates a possible gap between seeing and spraying. A correct detection can land in the wrong place if the vehicle travels farther than expected or the boom geometry changes.

A 2024 study registered by the US Agricultural Research Service helps isolate another layer. A YOLOv4-based prototype calculated a 79 to 80 percent reduction, but that result excluded algorithm and synchronization failures. This is a methodological warning, not a direct comparison with John Deere. “Savings when the system succeeds” and “operational savings including every failure” are different measures.

What the five fields actually measured

The Iowa State demonstration used a John Deere See & Spray Ultimate on five soybean fields in Boone and Story counties, all receiving similar pre-emergence residual programs. Their areas were 110.7, 71.9, 104.7, 37.1 and 90.1 acres. The share on which product was avoided was 87.2, 43.9, 71.2, 90.6 and 87.6 percent, respectively. The reported weighted average was 76 percent.

Across the project, 4,700 gallons of tank mix and nearly $6,500 in product were avoided, equivalent to $15.70 per acre. Drone images on application day and subsequent scouting confirmed differences in weed pressure, treatment of emerged weeds and clean fields through canopy closure. Those are valuable field observations. They are not, however, a randomized, multi-season trial comparing yield, drift, environmental effects and full machine payback.

The field-to-field contrast explains the system better than the average. Three fields where the residual herbicide had worked well entered the second pass with low weed pressure and averaged 88.4 percent savings. Field two, tilled on April 13 and planted on April 15, received substantial rainfall, showed poorer residual control and avoided only 43.9 percent. The machine did not necessarily become worse when it crossed a boundary; it encountered more areas requiring treatment.

The percentage of product saved should therefore not be treated solely as an AI score. It also measures the field’s prior state. A farm with effective residual management may record high selective savings; another with uniform infestation will approach broadcast coverage even with perfect detection. Ranking the two machines by that percentage without controlling initial weed pressure would credit the algorithm for a difference created by the field.

Correcting the manufacturer figure

John Deere reported performance at a much larger scale, but the figure needs a date and label. Its 2023 business impact report says See & Spray was used on more than one million acres that year, achieved a solution reduction of nearly two thirds and avoided more than eight million gallons. Those are aggregated manufacturer figures, not findings from the 2024 university demonstration.

The cited documents do not support turning them into “one million acres in 2024 at 59 percent.” Changing the year or blending a metric from another source prevents readers from reconstructing the comparison. The word “solution” matters too: avoided tank-mix volume is not automatically the same as active-ingredient mass, toxicity reduction or a proportional decline in risk. A tank mix contains water and products applied at different rates.

The manufacturer supplies useful details about its machine — 36 cameras, computer vision and selective actuation — but has a commercial interest. Iowa State provides an independent, localized demonstration. USDA provides a different architecture whose experiment exposes the role of synchronization failures. These sources should not be averaged into a universal number. They should be triangulated to ask whether they observe the same physical chain under different conditions.

The five metrics missing from a headline

The first is avoided coverage: what share of the field did not receive the post-emergence application? The second is efficacy: how many target weeds died and how many survived? The third is actuation error: false negatives that leave a weed untreated, false positives that spray crop or soil, and offsets between detection and nozzle. The fourth is total economics: product, license, equipment, fuel, time, refills, maintenance and corrective passes. The fifth is agronomic consequence: yield, crop injury, the future seed bank and resistance pressure over subsequent seasons.

An as-applied map can audit part of that chain, but it is not ground truth. It records where the controller commanded a spray and needs independent observation. A useful minimum design divides comparable areas, documents initial density, keeps an appropriate broadcast reference where legally and agronomically justified, scouts after treatment and records failures by species, plant size, light, speed and field zone.

Lower volume also does not replace resistance management. The US Environmental Protection Agency recommends integrating modes of action and other tactics, rotating crops, and scouting before and after treatment to determine whether control worked. Selective spraying can reduce unnecessary exposure, but if it systematically misses a species or repeatedly applies the same chemistry, it can preserve poorly understood selection pressure. The labeled rate and use requirements still govern every time a nozzle opens.

How to decide without buying a percentage

Before signing a contract, a farm can ask the vendor for definitions rather than slogans: whether “savings” means area, tank mix or active ingredient; whether failures are included; which crops, growth stages, weeds, speeds and lighting produced the result; and what per-acre cost includes the license. It should then set a success threshold and an intervention plan before scouting reveals escapes.

The Iowa demonstration shows that the concept can work in real fields and can avoid substantial product when weeds are distributed unevenly. Its 43.9-to-90.6-percent range also shows why there is no single promise. The mature conclusion is not “AI saves 76 percent,” but “selectivity turns actual variability into a measurable opportunity.” A reader who keeps that distinction can test the tool under local conditions and know whether the farm is buying precision or merely someone else’s average.