fruitpunch

Guide · AI & data

AI for Food and Agriculture: From Smallholder Farms to Supply Chains

By the FruitPunch research deskUpdated November 15, 20269 min read

In short

Five of every six farms in the world are smaller than two hectares and produce roughly 35% of the world's food, according to FAO research, so useful farm AI has to run on cheap phones. PlantVillage Nuru diagnosed cassava disease with 65% accuracy against 40 to 58% for extension agents, and up to 74 to 88% when six leaves were checked. After harvest, about 13% of food is lost in the supply chain and another 1.05 billion tonnes, or 19% of food available to consumers, was wasted in 2022.

This address once hosted FruitPunch AI, an Eindhoven AI for good community. FruitPunch is a new, independent publication and is not that organization. This guide was written from public sources; the history page explains the change.

AI helps food and agriculture most where farmers lack expert advice and records: it can diagnose crop disease from a phone photo, estimate yields from satellite images, guide tractors and turn paper ledgers into searchable data. The design constraint is farm size. According to FAO, five of every six farms in the world are smaller than two hectares, operate only around 12 percent of agricultural land and produce roughly 35 percent of the world's food. Tools built for large mechanized farms rarely fit those farmers, so the most useful projects work offline, on cheap Android phones, in local languages. This guide covers what works, with figures, named projects and open datasets.

Key facts

  • 5 of 6 farms are under two hectares and produce roughly 35% of the world's food (FAO, 2021).
  • Up to 40% of global crop production is lost to pests each year, and plant diseases cost the world economy over $220 billion (FAO).
  • 65% accuracy for PlantVillage Nuru on cassava disease, against 40 to 58% for extension agents and 18 to 31% for farmers (Mrisho et al., Frontiers in Plant Science, 2020).
  • Well over 50% of US acreage planted to six major crops uses automated guidance (USDA ERS, 2023).
  • 1.05 billion tonnes of food were wasted in 2022, 19% of food available to consumers, on top of about 13% lost in the supply chain (UN News on the UNEP Food Waste Index Report 2024).

Who grows the world's food, and why does farm size matter for AI?

The FAO study published in World Development breaks the numbers down further. Around 70 percent of all farms are smaller than one hectare and operate just 7 percent of agricultural land. At the other end, the largest one percent of farms, those above 50 hectares, operate more than 70 percent of the world's farmland. FAO also counts more than 608 million family farms, producing around 80 percent of the world's food in value terms.

Farm size also tracks income. FAO reports that 99 percent of farms in high-income countries are larger than five hectares, compared with only 28 percent in low-income countries. That split explains why AI in agriculture has two quite different faces. On large farms in North America and Europe it lives inside machinery: auto-steer, yield monitors and variable rate application. On small farms in Africa and South Asia it lives on a phone, as a diagnosis app, a chatbot or an SMS alert, and it has to cope with poor connectivity, small irregular fields and little historical data.

Area What AI does Real figure or project Source
Crop disease Classifies leaf photos 99.35% accuracy on a held-out set of 54,306 lab images Mohanty et al., 2016
Crop disease in the field Offline phone diagnosis Nuru 65% vs extension agents 40 to 58% (cassava, Tanzania) Mrisho et al., 2020
Yield estimation Satellite imagery plus field data R² up to 0.4 against survey yields, western Kenya Burke and Lobell, PNAS, 2017
Precision agriculture Automated guidance Well over 50% of planted acreage for six US crops USDA ERS, 2023
Farm advice LLM chatbot with retrieval Farmer.Chat: over 15,000 farmers, over 300,000 queries Singh et al., arXiv 2409.08916
Records and credit OCR of Nepali paper ledgers Heifer and FruitPunch AI challenge, 2024 Heifer International press release

Can a phone app really diagnose crop disease?

Yes, within limits. The stakes are large: FAO estimates that "annually up to 40 percent of global crop production is lost to pests," and that plant diseases cost the global economy over $220 billion a year.

The starting point for most of this work was a 2016 paper from EPFL and Penn State. Using a public dataset of 54,306 images of diseased and healthy leaves, the authors trained a deep convolutional neural network "to identify 14 crop species and 26 diseases (or absence thereof)" and reported "an accuracy of 99.35% on a held-out test set" (Mohanty, Hughes and Salathé, Frontiers in Plant Science). Those images were taken under controlled conditions, so the number says little about a muddy field at noon.

Field results are lower and more honest. PlantVillage Nuru, built by PlantVillage at Penn State, runs a cassava model on Android phones. Its model was trained on 2,756 annotated images of cassava leaves collected in coastal Tanzania. When researchers tested it against people, Nuru "could diagnose symptoms of cassava diseases at a higher accuracy (65% in 2020)" than agricultural extension agents (40 to 58%) and farmers (18 to 31%), and its accuracy rose to between 74 and 88% when six leaves per plant were assessed (Mrisho et al., Frontiers in Plant Science, 2020). The practical lesson: an app like this beats an untrained eye, but users need to photograph several leaves, and the app works best as a support for extension staff rather than a replacement.

PlantVillage now describes a wider system. Its project page says its extension officers "now work across 12 countries," and that once a farmer enters "3 critical details (crop type, location, planting date)" its engine can send advice by smartphone, SMS, TV or local social networks. It also built eLocust3m for FAO to collect data on locust sightings.

How does AI forecast crop yields?

Yield forecasts tell governments, insurers and traders how much food to expect before harvest. On large farms, combines with yield monitors record output metre by metre. On small farms there is usually no such record, and survey data is expensive and often wrong.

Satellite imagery plus machine learning is the main workaround. In a widely cited PNAS study, Marshall Burke and David Lobell of Stanford tracked smallholder maize yields in western Kenya "using a combination of 1-m Terra Bella imagery and intensive field sampling on thousands of fields over 2 y." Agreement with field surveys was highest, "R 2 up to 0.4," when plot areas were measured precisely and fields were larger. The paper also shows why small farms are hard: GPS measures from household surveys in four African countries indicate that 25% of fields are smaller than 0.5 acre and more than half are smaller than 1 acre, and farmers over-reported the area of tiny fields by as much as a factor of 5.

Two takeaways follow. First, satellite yield estimates for smallholders are useful for comparing regions and evaluating programs, but they are not yet precise enough to judge a single field. Second, the quality of ground data limits the model, so better field measurement matters as much as better algorithms.

What is precision agriculture, and how widely is it used?

Precision agriculture uses GPS, sensors and data to apply seed, fertilizer, water and chemicals where they are needed instead of uniformly. The best adoption data comes from the US Department of Agriculture's Agricultural Resource Management Survey. In its 2023 report Precision Agriculture in the Digital Era, the Economic Research Service found that automated guidance is used on "well over 50 percent of the acreage planted to corn, cotton, rice, sorghum, soybeans, and winter wheat." Yield maps, soil maps and variable rate technology spread more slowly and cover only between 5 and 25 percent of planted acreage for winter wheat, cotton, sorghum and rice.

These tools assume capital, connectivity and large fields, which is why most smallholders are reached through phones instead.

How can OCR and LLMs help farmers get credit?

Many farmer cooperatives keep sales, loan and production records on paper. Without digital records, members struggle to show a lender their history, join traceability schemes or analyze their own business.

This was the problem behind a 2024 project in Nepal. In January 2024, Heifer International and FruitPunch AI, the Eindhoven AI for Good community that previously used this domain, announced the AI for Women Farmers Challenge. The release said the project would "use AI Optical Character Recognition (OCR) models to extract, digitize and translate Nepali handwritten and typeset text from photos and scans of existing paper records." A planned second phase would build data structures from the results, "such as financial data to stimulate formal lending to farmers." The release also stated that Heifer Nepal had supported over 400,000 families organized into 277 agricultural cooperatives. The current FruitPunch site is a separate rankings and guides publisher and is not connected to that challenge.

The general pattern applies anywhere: OCR turns ledger photos into text, a language model or rules engine extracts fields such as dates, amounts and borrower names, and the output feeds a spreadsheet or a lender's credit model. Handwriting in low-resource scripts is the hard part, so projects need locally labeled samples and human review of extracted figures before they reach a loan decision.

Language models also answer farmers' questions directly. Digital Green's Farmer.Chat uses retrieval over research papers, crop tables and videos; deployed in four countries, it "has engaged over 15,000 farmers and answered over 300,000 queries" (arXiv 2409.08916). A 2026 follow-up from the same group notes that vanilla models "exhibit unsupported recommendations" and generic advice, and found that fine-tuning on curated data from Bihar, India improved fact recall (arXiv 2603.03294).

How much food is lost or wasted, and where can AI help?

After harvest, losses continue. Reporting on the UNEP Food Waste Index Report 2024, UN News states that 1.05 billion tonnes of food went to waste in 2022, about 19 per cent of food available to consumers, "in addition to around 13 per cent of food lost in the supply chain" from post-harvest to the point of sale, as estimated by FAO. Households accounted for 631 million tonnes, food service 290 million and retail 131 million. Food loss and waste generates up to 10 per cent of global greenhouse gas emissions.

The data points to where AI can help and where it cannot. Forecasting and inventory tools sit with retailers and food service, which together account for 421 million tonnes in the UNEP figures. Households, the largest source, are harder to reach with business software, so tools for shops and kitchens address only part of the problem.

Which projects and open datasets can people use?

  • PlantVillage Dataset: the leaf image dataset from the 2016 paper, published by EPFL's Digital Epidemiology Lab, with color, grayscale and segmented versions.
  • iCassava 2019: labeled and unlabeled cassava leaf images released as a Kaggle challenge at the CVPR 2019 fine-grained visual categorization workshop. Its authors note that at least 80% of smallholder households in Sub-Saharan Africa grow cassava.
  • PlantVillage Nuru: the Penn State app and advisory platform described above.
  • Farmer.Chat: Digital Green's advisory chatbot, with design details published on arXiv.
  • USDA ARMS precision agriculture data: the best public time series on farm technology adoption in the US.
  • FAO farm size research: the baseline figures on who farms and how much land they hold.

Frequently asked questions

What share of the world's food do small farms produce?

According to FAO research published in 2021, farms under two hectares make up five of every six farms, operate about 12 percent of agricultural land and produce roughly 35 percent of the world's food. Family farms of all sizes produce around 80 percent of food by value.

How accurate are crop disease apps?

On lab images, deep learning reached 99.35% accuracy in a 2016 study. In real cassava fields in Tanzania, PlantVillage Nuru reached 65%, better than extension agents and farmers, and 74 to 88% when six leaves per plant were checked.

Can satellites measure yields on small farms?

Partly. A PNAS study in western Kenya found satellite estimates agreed with field surveys at R² up to 0.4 when plot sizes were measured precisely. That is useful for regional comparisons but not for judging a single small field.

How are OCR and language models used in agriculture?

They digitize paper records and answer farmer questions. A 2024 Heifer International and FruitPunch AI challenge used OCR on Nepali handwritten cooperative records to support lending, and Digital Green's Farmer.Chat had answered over 300,000 farmer queries when its paper was published.

How much food is wasted globally?

UNEP's Food Waste Index Report 2024 puts food waste at 1.05 billion tonnes in 2022, or 19% of food available to consumers. FAO estimates a further 13% is lost between harvest and retail.

Data behind this guide

Spotted an error in this guide? Email editorial@fruitpunch.ai

Keep reading