A Banana’s Passport in One Click: Using AI and Citizen Science to Trace our food

Food fraud is a growing global problem, especially with highly valued and expensive products such as saffron, vanilla, olive oil and organic goods. For instance, “organic” bananas from Ecuador sometimes turn out not to be organic at all, but simply mislabeled. Often the motive is financial—cheap products sold at higher prices—but sometimes the risk is more serious, such as alcohol being replaced with dangerous methanol. That is why importers and inspection agencies worldwide are searching for smarter ways to detect fraudulent products more quickly. “In Europe alone, the damage caused by food fraud is estimated at 8 to 12 billion euros per year,” says researcher Yamine Bouzembrak.

For Bouzembrak, the question of how artificial intelligence (AI) can contribute to safe and fair food is part of his daily work. The idea of coupling AI technologies and citizen science to detect food fraud arose when a colleague pointed him to iNaturalist, an online social network where users worldwide share photos and biodiversity information. 

Yamine uses a system called convolutional neural network (CNN) which learns to recognize patterns in the color, shape and texture of bananas “similar to how people recognize things by eye,” he explains. “That’s when I realized such a system could also work for food products.” His team is now training a prototype with thousands of banana photos to detect subtle regional differences. That’s an important development, because traditional laboratory tests are expensive and slow.

The research started only a few months ago and on a small scale. Using iNaturalist, Bouzembrak and his colleagues asked citizens on the platform to upload photos of bananas from 20 countries, together with basic details such as the country of origin. After collecting and processing the images, the team used the data to train and validate the first version of their model. They then tested it by letting the model predict the origin of a banana based solely on a photo. “The accuracy was surprisingly high—over eighty percent,” Bouzembrak says. “That’s very promising, but for a usable prototype we aim for at least ninety-five percent.”

Bouzembrak recently presented his idea at the International Food Fraud Conference. Reactions ranged from enthusiasm to slight hesitation. “One of the questions was whether AI will replace the lab analysis,” he says. It’s an understandable concern, but according to Bouzembrak not relevant in this case. “With AI system like this, you can more quickly identify which batches of bananas deserve extra attention. Inspectors and laboratories remain essential to conclusively prove fraud.”

The AI model is mainly a smart pre-selection tool. “Instead of sending every sample to the lab, inspectors can quickly scan which batches look suspicious. That way, expensive tests are reserved for the highest-risk products.” And if a case goes to court, laboratory research will always be required as formal evidence.

What’s the logical next step for this promising research? The prototype is not yet fully reliant, and Bouzembrak wants to look beyond bananas. “Because the AI model works with photos, it’s especially suitable for expensive products with recognizable visual features,” he explains. Products like saffron, vanilla or nuts are well-suited for image recognition because they show clear visual variation. “Saffron, for instance, is an expensive spice – from 3,000  to 10,000 euro per kilogram - that often turns out not to be of authentic origin.”

For liquid products, the situation is different. “With olive oil or honey, determining origin through photos is much more difficult because they simply show too little visual variation.” In the coming period, researchers will explore which products are best suited for inclusion in the app. Accuracy also remains a key focus. “Eighty percent certainty is nice, but our goal is at least ninety-five percent.”

In the coming months, Bouzembrak hopes to gauge interest in further developing the tool among importers, inspectors and other players who deal with food fraud daily. “My focus is on the scientific research,” he says, “but I hope to secure funding through a public-private partnership to move it forward.” His ultimate goal is for a start-up or other private party to eventually bring the tool to market.

And the potential impact goes well beyond detecting fraud. “In the long run, similar AI techniques could also help assess freshness or quality,” Bouzembrak adds. “In that way, the technology could evolve into a widely accessible tool that helps consumers make better, more reliable and fairer choices.”

University of Twente, The Netherlands, MISTEA, INRAE, France