From Pixels to Decisions: Practical AI for Crop Health and Farm Operations – InCrop White Paper

Executive summary

A widely cited deep-learning study reached 99.35% accuracy identifying crop diseases on held-out images from its own dataset, but only 31.4% on images collected from other sources.¹ That gap — between lab benchmark and real field — is the central risk in agricultural AI.
The cure is method, not a bigger model: start from a farm decision, collect data on the farm, validate on held-out seasons and sites, deploy with confidence thresholds and human review, and monitor for drift.
AI use cases differ widely in maturity. We rate them so clients invest first where evidence is strongest.
Farmers own their data. Systems should work offline, in local languages and on modest phones.
Value is proven with comparison plots or zones and farm KPIs (yield, losses, inputs, labour hours), not with model accuracy alone.

1. The lab-to-field gap

In 2016, Mohanty, Hughes and Salathé trained deep convolutional networks on 54,306 images of healthy and diseased leaves covering 14 crop species and 26 diseases. On held-out images from the same collection, the best model reached 99.35% accuracy.¹

The same paper reports a sobering second test: on images gathered from trusted online sources, taken under different conditions, accuracy fell to 31.4% — still far better than the 2.6% expected by chance, but nowhere near the headline figure. The authors note that more diverse training data is needed.¹

Peer-reviewed

Lab images (controlled background, single leaf, consistent lighting): 99.35%. Images from other sources: 31.4%; 47.93% when the crop type was supplied.¹

The lesson generalises beyond disease detection. Models learn the conditions of their training data — camera, light, background, variety, growth stage, region. When any of those change, performance can fall sharply and silently. This is known as domain shift, and farms are full of it: every season, field and variety is a little different.

2. Use cases and maturity

Use case	Typical data	Maturity	Main caution
Sensor-based irrigation scheduling	Soil moisture, weather, evapotranspiration	Established	Sensor placement and calibration
Climate control in greenhouses / indoor farms	Temperature, humidity, CO₂, light	Established (rule-based) → Emerging (predictive)	Needs hard safety limits
Yield and harvest-date forecasting	Sensor history, images, weather, past yields	Emerging	Needs several seasons of local data
Disease and pest detection from images	Phone, fixed-camera or drone images	Emerging	Domain shift (section 1)
Weed detection for targeted spraying	Machine-mounted cameras	Emerging	Crop/weed similarity, speed
Demand forecasting & harvest planning	Orders, sales history, calendar	Established in retail; emerging on farms	Short history on small farms
Generative-AI farmer advisory	Questions, farm records, vetted guidance	Emerging	Confident but wrong answers (section 7)

Maturity ratings are LIDAVI’s assessment for planning purposes. They describe how proven a category of technique is, not how well any particular product performs.

3. A seven-step deployment method

Start from the decision. Define the farm decision the AI supports (“spray block 4 this week?”), who makes it, and what an error costs in each direction.
Collect data on the farm. Capture images and sensor data with the cameras, varieties, growth stages and lighting the model will meet in use.
Label with agronomists. Use local experts and a written labelling guide; measure agreement between labellers.
Validate honestly. Hold out whole fields, zones or seasons — not random images from the same day — so the test reflects the next season, not the last one.
Deploy with thresholds. Act automatically only above a confidence threshold; route uncertain cases to a person; always show the reason.
Monitor for drift. Track input statistics and spot-check predictions; retrain when performance drops.
Keep a safe fallback. If a model or connection fails, revert to agreed rule-based practice and alert the team.

Figure 1. The deployment loop. A model is never “finished”; it is maintained like any other farm equipment.

4. Data, ownership and governance

Ownership. Farm data belongs to the farm. InCrop uses it to train the farm’s models under written agreement and does not sell it.
Consent for pooling. Pooling anonymised data across farms can improve models; it is opt-in and documented.
Personal data. Worker and farmer information is minimised, protected and handled under applicable privacy law.
Portability. Data can be exported in open formats; clients are not locked in.
Security. Encrypted transport and storage, role-based access, audit logs — drawing on LIDAVI’s cyber security practice.

5. Designing for farmers

A model that is accurate but unused delivers nothing. InCrop applies these design requirements to every farmer-facing tool:

Offline-first. Edge inference on gateways or phones; sync when connectivity returns.
Local language and voice. Interfaces and advice in the language the farmer thinks in, with voice input where literacy or typing is a barrier.
Modest devices. Apps that run on mid-range Android phones and tolerate poor cameras.
Actionable output. “Inspect rows 12–15 today” beats a probability score.
Explain and invite feedback. Show why, and let farmers correct the AI — corrections become new training data.

6. Measuring real value

Model accuracy is a means, not an outcome. Each InCrop pilot defines farm-level KPIs up front and measures them against a comparison:

Comparison plots, zones or time periods managed without the AI recommendation.
Outcome KPIs: marketable yield, crop losses, water and fertiliser per kg, pesticide applications, labour hours, energy per kg.
Model KPIs: precision and recall on local data, share of cases sent for human review, time to detection.

Design target

No InCrop AI feature moves from pilot to production without evidence on the client’s own farm that it improves at least one outcome KPI without harming the others.

7. Generative AI advisory: handle with care

Large language models make it possible to answer agronomy questions conversationally, in many languages, at any hour. They can also produce fluent, confident answers that are wrong — a well-documented limitation. In agriculture, a wrong dose or a missed disease costs money and can harm people and the environment.

InCrop advisory assistants are therefore built with:

Grounding in vetted local guidance and the farm’s own records, with sources shown.
Guardrails on chemicals, dosages and safety-critical topics, which route to qualified agronomists or official labels.
“I don’t know” as an acceptable answer when evidence is missing.
Review loops, where agronomists sample and grade answers regularly.

8. Roadmap

Discovery (weeks, not months). Map decisions, existing data and connectivity; choose one or two use cases with clear KPIs.
Data foundation. Sensors, cameras and the edge-to-cloud platform; labelling with local agronomists.
Pilot season. Deploy with thresholds and comparison plots; measure outcome KPIs.
Scale what worked. Extend proven features to more fields, crops or sites; retire what did not.

9. Conclusion

The promise of AI in agriculture is real, and so is the gap between benchmark and field. Farms that succeed with AI treat it like any other critical equipment: specified for a purpose, tested on site, maintained, and judged by results. InCrop brings that discipline, together with LIDAVI’s IoT, cloud and app engineering, so that pixels and sensor readings turn into decisions growers trust.

References

Mohanty S.P., Hughes D.P., Salathé M. (2016). Using Deep Learning for Image-Based Plant Disease Detection. Frontiers in Plant Science 7:1419. https://doi.org/10.3389/fpls.2016.01419; preprint: https://arxiv.org/abs/1604.03169

This white paper is provided for information. Figures labelled “Established” or “Peer-reviewed” are taken from the cited sources; figures labelled “Industry claim” have not been independently verified by LIDAVI; anything labelled “Design target” or “Hypothesis” is LIDAVI’s own and must be confirmed in a site-specific feasibility study. Nothing here is a performance guarantee.