Crop Disease and Yield Forecasting AI: Promising Trials, Fragile Transfer, Thin Field Evidence

The use case

A 2026 benchmark trained two machine-learning yield models on United States county data for 2017 to 2021 and tested them on 2022. In some regions they predicted well; in others the models did worse than simply guessing the average yield, a result statisticians report as a negative R2². The finding frames the whole use case: models that look accurate where they were trained can fail elsewhere, and most public documentation does not say how often.

Two tasks are covered. The first is crop disease diagnosis, where a phone app classifies leaf photos. The second is yield forecasting, where satellite and weather data are used to estimate harvests before they happen.

Scope note. Here AI means machine learning, including deep learning. Older methods do similar work and are not credited as AI: manual crop-cutting experiments (sample harvests of small plots), trend lines of past yields, and expert diagnosis by extension officers. The European Commission's Joint Research Centre (JRC) calls its ridge regression and Gaussian process models machine learning, but ridge regression is a regularised linear model that many would call statistical, and the JRC set-up can select a trend-only model with no weather input³. The Indian system described below is not described as machine learning in the sources read⁴. Out of scope: drones, robotic spraying, input-price forecasting and precision fertiliser application.

Where it is running

The Nuru trial is older than three years and no newer evaluation was found. No yield-forecasting deployment outside India and the JRC programme was documented in the sources read.

What the evidence shows

Yield models lose accuracy in unseen regions. The strongest evidence is a peer-reviewed benchmark covering corn and soybean in more than 1,200 US counties, trained on 2017 to 2021 and tested on 2022 (measured on held-out data, no field deployment)². For one model on soybeans, the error rose from 6.15 to 9.75 bushels per acre when the model predicted a different region from the one it was tuned for, a gap of 58.53%. Across models, crops and scenarios, the gaps ranged from 1.23% to 58.53%². The authors report negative R2 values frequently in unseen regions, and name the Prairie Gateway region as consistently hard². The mechanism is that crop yield depends on local soil, irrigation and weather patterns that the model has not seen; the authors suggest unobserved irrigation and heat-dry extremes but label these as hypotheses².

Reproducing published results was hard. For one model, the reproduced corn R2 was 0.78 against 0.89 reported originally; the authors say not all aspects of the original set-up may have been fully documented².

Phone diagnosis beat people, but was far from perfect. In a December 2020 comparison, Nuru diagnosed cassava disease symptoms with 65% accuracy against 40 to 58% for extension agents and 18 to 31% for farmers (measured in a comparison study, abstract)¹. Accuracy rose to 74 to 88% when six leaves per plant were assessed (abstract)¹. The result depends on the scanning procedure, not only on the model. Two weeks of use gave extension workers only a slight increase in their own accuracy (abstract)¹.

The large public systems state no measured accuracy. The JRC methodology, dated 20 June 2023, selects among 12 model configurations per crop and country using leave-one-year-out testing and accepts a model if its cross-validated R2 exceeds 0.3, with about 20 yearly samples per country; it gives no accuracy values (official methodology, no measured results)³. The Indian programme says block-level yields arrive three weeks before crop-cutting results (programme-reported) and gives only the qualitative statement that forecasts are aligned with manual results⁴. No accuracy figure for either system appears in the documents reviewed.

What it takes: cost, integration, governance

Data. The JRC series start in 2002 because of satellite availability, and country grouping was partly done by hand using analyst knowledge³. India's AgriStack farm registry had mapped 253 million plots and issued 60 million Farmer IDs by 2024-25 but remains at pilot stage, and crop-insurance yield data reach central systems several weeks after harvest (preprint, not peer reviewed)⁸.

Compute. In the US benchmark, training one model took 31.5 hours against 14 minutes for the other on one consumer-grade GPU (authors' figures, 2026)². No price per hectare or per farmer was found for any system.

People. The Nuru result depends on training users and on the number of leaves checked¹. Smallholder AI research rarely studies users: a report on a systematic review of 182 studies says only 9 examined adoption, perceptions or practical use (secondary report, 15 July 2026, not specific to disease or yield tools)⁷.

Rules. No regulator guidance specific to crop disease or yield AI was found. One legal analysis reads the EU AI Act as covering mainly safety components in agricultural machinery and, on a literal reading of Annex III(2), not the food system as such⁹. Where yield estimates set insurance payouts, governance is a matter of programme design: Madhya Pradesh gave the system full weight from kharif 2023⁵.

Where it falls short

The challenge question is what makes the use case look better than it is. Four answers emerge.

Narrow test design. The US benchmark uses a single test year (2022), a five-year training window and county-level totals, and omits process-based crop models as baselines (authors' stated limitations)². One bad or good weather year can dominate the result.

Old or absent field evidence. The only controlled disease comparison is from December 2020 and covers one crop in East Africa¹. No multi-year evidence that gains persist was found.

Disputed field results. In Madhya Pradesh, large soybean farmers were reported to be less satisfied with YES-TECH settlements than with crop-cutting-based ones, and a Maharashtra speaker described large variations between the two methods in some districts, with no figures given⁵. The state agency called YES-TECH the lesser evil because its errors are smaller than manual ones, a claim without published numbers⁵. A Centre for Science and Environment report of 18 May 2026 describes critical gaps and calls for rollback (description on the publisher's landing page; the report itself was not read)⁶. The sources attribute low claims partly to historical inaccuracies in farmers' loss reporting, which the article does not verify⁵.

Dependence on people and process. Hand-made groupings in the JRC set-up, scanning several leaves in the Nuru trial and parallel crop-cutting experiments all mean results rest on process as well as on the model¹³⁵. In the searches run, no programme withdrawals, lawsuits or regulator enforcement actions were found.

Phase: Pilot with results — why

The hypothesis was In early use. The evidence supports a lower phase. Confirmed machine-learning deployments at known scale are limited to the JRC set-up, whose models are partly linear and whose accuracy is not published³. The Indian rollout is large but is not described as machine learning in the sources read⁴. The one measured disease result is from 2020 and lacks a deployment scale¹. Measured results exist¹², so the phase is Pilot with results rather than Proven elsewhere.

What to watch next

• Whether Madhya Pradesh keeps full weight on YES-TECH after the 18 May 2026 report calling for rollback⁵⁶.
• Whether the JRC publishes hindcast accuracy for its national forecasts, which its 20 June 2023 beta methodology did not provide³.
• Whether the independent third-party comparison of crop-cutting and YES-TECH yields requested in Maharashtra is published⁵.
• Whether a multi-year, multi-region test of the kind the US benchmark authors say is missing appears².

Questions to ask before adopting

• Has the model been tested on regions and seasons it was not trained on, and what was the error there?
• What was the baseline: manual sampling, a trend line or an older statistical model, and how much better is the AI?
• Who checks the output when it sets money, such as an insurance payout, and what is the appeal route?
• How many leaves, images or days of data does a reliable result need, and who collects them?
• What does it cost per farmer or hectare, including data, training and parallel manual checks?