The Challenge
You can’t fix what you can’t see. Across low- and middle-income countries, some of the crops farmers grow never reach a plate. Food is lost to heat, handling, and the absence of a cold chain between the field, the market stall, and the consumer. And even when not totally lost, handling issues can cause food to lose value, negatively affecting incomes of farmers, traders, processors, or retailers who own them. In other words, just knowing loss occurs is not the same as knowing where it happens within the supply chain, and what these losses actually cost value chain actors. Without that level of clarity, interventions aimed at reducing food loss are just guesswork.
Measuring food loss with precision has traditionally required rigorous surveying and physical sampling — these methods are thorough but expensive, slow, and difficult to repeat. This makes them a less-than-ideal tool in a market that changes with every season and price swing. Take a crop like tomato, which moves fast, bruises easily, and passes through many hands. A recent estimate suggests that between farm and retail, about 13% of tomatoes are lost. To accurately determine where quality degrades within the supply chain, decision-makers need real-time insights. Nigeria’s tomato value chain concentrates all of this.
The Action
IFPRI is developing an AI-based system that assesses tomato quality from images — turning an ordinary photograph into a quantified measure of loss.
Researchers trained an image recognition model to sort tomatoes by maturity using visual reference charts documenting five stages: unripe (breaker and turning), ripe, fully ripe, overripe, and rotten or discarded. Applied at successive points along the chain, the model makes it possible to see not only the produce already lost, but the produce on a trajectory toward loss — the difference between counting waste and preventing it.
The approach was initially tested in a January 2026 pilot in Jos, Plateau State, and testing continued for differences during the wet season in Lagos in August. The team captured 4K video and photographs in wholesale and retail markets, and paired the visual record with systematic weight measurements of fresh and rotten tomatoes. Five varieties were studied — two local (UTC and Roma) and three hybrid (Belfast, Bellfort, and Dingyanfen No. 2) — alongside greenhouse farm visits and consultations with input suppliers on which varieties dominate. Throughout, the team worked directly with the people who move the crop: farmers, traders, market leaders, transporters, and input suppliers.
The work is a collaboration between IFPRI, DAJRHAS Health and Agric Development in Nigeria, the World Vegetable Center, and the Japanese agricultural technology company Inaho, Inc., with support from the Government of Japan. An operational pilot application is targeted for the end of 2026.
Crucially, the method is designed to complement primary data collection rather than replace it. Robust sampling and up-to-date market data — particularly on prices — remain essential, both to validate what the model sees and to give value chain actors a reason to act on the causes of loss once they are identified. By helping traders, processors, or retailers better identify inherent qualities of tomatoes that are priced in the market, they can adopt practices or technologies that retain that value and reduce both physical and quality losses. “We believe image-based assessments have the potential to reduce the cost of identifying losses among perishable products, which can help reduce transaction costs and eventually improve diets,” said Alan de Brauw, lead for IFPRI’s research on Markets for Better Diets.
The Outcome
The goal is to create an AI driven app that can identify differences in quality and ripeness. At present, the app can identify edible versus nonedible tomatoes at 95% accuracy relative to human grading, but is not quite as good at measuring ripeness. The new training data set from Lagos will help improve accuracy, since it will work with wet season tomatoes. A further program is being developed to measure the volume of discarded tomatoes in markets. What is documented so far is the method's reach rather than its results: after optimizing the app for tomatoes, the approach can be adapted for other crops, including peppers, onions, mangoes, and oranges, with potential applications across Nigeria and beyond. However, further training is necessary in any of these cases.
Learnings
One of the key learnings is that it is important to train the AI app using all sorts of variation—including differences in lighting (e.g., morning, noon, evening), weather conditions, as well as the way that tomatoes are arranged. The variation helps the AI “learn” how to judge quality even when tomatoes do not look the same. Early on, we learned that video was more effective than photos, because videos taken at 30 frames per second include 1800 pictures per minute, all with slightly different information. We run a program to make sure that the computer learns not to double-count tomatoes, and then assesses the quality of each one. In doing so, we learned to break up the problem into steps—first the AI must identify a tomato, and then it needs to judge the quality. All of the learning would not be possible without strong collaboration within markets with market agents or managers, who then explain to traders or others that we are conducting research so they agree to participate (or have their tomatoes participate) in the videos.
Champions 12.3 is a coalition of leaders from governments, businesses, international organizations, research institutions, farmer groups and civil society dedicated to accelerating progress toward UN Sustainable Development Goal Target 12.3 — halving global food loss and waste by 2030.