The past few years have demonstrated just how vulnerable our food systems are to large-scale disruptions, from pandemic lockdowns and social distancing to labour issues and extreme weather events. For the consumer, these can result in product shortages and empty supermarket shelves. Such uncertainty looks set to continue, as food producers start to feel the effects of climate change, and global political systems remain fragile. Consequently, a key part of the STFC Food Network (SFN)’s work is to research ways we can build resilience into food supply chains, so that they can keep functioning even when situations (and people’s shopping behaviour) change quickly.
 
Potentially, this could include harnessing the power of information contained in social media posts. One of the winning projects in the 2020 SFN Sandpit competition is investigating whether it could be possible to apply deep learning techniques to build a computer model that can forecast specific food-system disruptions (such as flour running out of stock) on the basis of the text and images that individuals post online (e.g. comments about flour).

Project lead Dr Laura Wilkinson (Senior Lecturer in Psychology at Swansea University) explains the rationale: “The COVID-19 pandemic saw an unprecedented level of conversation around food on social media, with many people experiencing shortages of key staples such as flour for the first time. We suggest that understanding how people reacted during the pandemic may help us to understand what might happen to the food system if other events occur that lead to uncertain times.”

The project’s ultimate aim is to train a machine learning programme to automatically detect patterns in people’s social media activity that indicate changes to food systems, and how consumers react to these. Training such a model, however, requires a detailed, annotated dataset, so that it can ‘learn’ which information is relevant to spot trends. To compile this, Laura and her colleagues created a database of just over a million anonymised Twitter posts, using searches with food-related keywords.
 
The trouble is that some of these posts may not actually be about food at all, as Laura explains: “Many food words can also refer to other things. For example, the word ‘coconut’ could have been used to describe someone wanting to be able to eat a coconut or could have been used to describe their favourite shampoo scent. On the other hand, a tweet might be referring to food but not actually mention a food word like 'banana'. For instance, it might be referring to home deliveries and mentioning supermarkets. Unfortunately, computers are not very good at interpreting these situations.”
 
This means that, in order to advance the project to the next stage, each tweet in the database first needs to be manually checked and classified as either being about food or not. As a small project team, this would have taken them an unfeasibly long amount of time… so they decided to call in reinforcements, by launching the project on Zooniverse!
 
Founded in 2009, Zooniverse is perhaps the best-known citizen science platform, and gives any member of the public the opportunity to contribute valuable data for research projects. These typically involve categorising data that computers would struggle to do automatically, such as picking out animals in camera trap footage, transcribing historical documents, or identifying cell features from microscope images. For this project, participants were asked to read through the Twitter posts and decide whether they were about food or not.
 
Since launching on Zooniverse in July this year, the response has been phenomenal. By the beginning of August, over 209,000 Tweets had been classified by 1,200 volunteers, reaching a high of over 36,500 classifications in a single day. But besides the raw data, Laura has been impressed by how the platform encourages dialogue between contributors and the research teams. “What makes citizen science projects distinct from many other research projects is that there is a high degree of knowledge sharing: the Zooniverse contributors are not passive participants” she says. “On the discussion boards for the project, participants have often reflected on what particular tweets mean to them and how they relate to their own food experiences during the pandemic. This really adds value to the study, giving us a much greater breadth of lived experiences.”
 
The discussion boards also revealed how social and cultural differences can impact the understanding of food-related content, which could help inform the design of future studies. As Laura explains: “Our database is restricted to UK-located Tweets, but Zooniverse contributors are based all around the world. This has caused some questions about certain terms, for instance about the names of some UK supermarkets that aren’t widely known abroad, and whether Easter eggs in the UK are edible. The two-way discussion boards are a great feature, as they enable us to respond rapidly to these issues, and share information with the whole community of contributors.’
 
The team hope to start training the first machine learning model in September 2022. Once the workflow is established, Laura hopes it can be adapted to process information on an international scale and from different social media platforms.
 
“I’m talking about Zooniverse to anyone who will listen – I think it’s an absolutely brilliant platform, which I definitely hope to use more of in the future” she says. “For instance, it would be very interesting to train an algorithm to classify photographs of meals that people post on social media for specific properties, such as the number of food components and portion size.”
 
“As recent events have shown, we live in uncertain times and we can’t take our food systems for granted, presuming that we can always get exactly what we want from the supermarket, whenever we want” she added. “It is imperative that when global disruptive events occur, we have the information and the means to ensure that the most vulnerable people in our societies who may be food insecure aren’t hit the hardest.”

The COVID-19 pandemic brought to light just how fragile food supply chains can be: with traditional trade disrupted, many producers struggled to find alternative routes to sell their products. Consequently, across the world shoppers were faced with empty shelves at supermarkets, even though farmers were having to throw away mountains of unsold produce.

Some of the worst hit were the many thousands of smallholder farmers and urban microenterprises trading fresh, perishable produce in low- and middle-income countries. Typically, these rely on in-person interactions at marketplaces, trade hubs and bazars, meaning that the COVID-19 lockdowns effectively shut down their trade.
Sarang Vaidya, co-founder of fresh produce agritech venture Go4fresh, describes the situation in his native India: “Disruptions to traditional marketplaces during the COVID-19 pandemic resulted in extensive on-farm food wastages, unfulfilled demand, and major setbacks to the livelihoods of millions of small grocery shops, street vendors, and hawkers. If we want to help communities to become more resilient to future shocks, we need to provide access to alternative markets.”

A marketplace on a mobile phone
With mobile phones being almost ubiquitous across India, Sarang believes that one solution could be to create ‘virtual’ marketplaces that directly connect smallholder farmers to buyers (such as retailers, wholesalers, restaurants, and canteens). In 2019, this vision inspired him and his team at Go4fresh to develop a collaborative, digital platform to connect producers and businesses in food supply chains.

However, currently very little is known about how such a platform should be designed to best meet the needs of smallholder farmers and microenterprises. To address this, Sarang is leading a SFN-backed project which is engaging end-users to identify the key data sources and features required to build a scalable, user-friendly virtual marketplace for affordable smartphone devices. The UK-India partnership involves experts in data science from STFC, besides a variety of Indian-based food-sector partners, including farmer groups and fresh produce buyers.

The central objective was to build a prototype virtual marketplace for four vegetables (tomato, cabbage, okra, and green chilli) along an established trade route in India. This ran from the village of Otur in the Pune district of Maharashtra, to a market in the Kandivali neighbourhood of the state capital Mumbai (a distance of 200 km, or roughly five hours of driving).

Through the Indian team’s links, the project consulted over 60 smallholder famers, microenterprise owners, and transporters. Due to ongoing COVID-19 lockdowns and social distancing restrictions, many of the interviews were conducted using phone/video calls or online chat functions. This qualitative research was supplemented with quantitative data, including live market prices, delivery dates, crop production data, and consumer prices on ecommerce platforms.

Mapping the challenges against widespread adoption of digital technologies
“We found that both the supply and market ends of the chain faced distinct challenges hindering the adoption of digital technologies” says Sarang. For those supplying fresh produce, a key issue was low current use of digital technologies, with less than a third using digital sources of information, such as market data and weather alerts. Another concern was that current data sources were often updated only infrequently, limiting their accuracy over time. In combination, this meant that smallholder farmers were often unaware of the current market prices for their products, resulting in them selling their goods for below-average price.

Meanwhile, for the microenterprises purchasing fresh produce, the most prominent issue was the large variation seen among different lots of the same food product (including quality, weight, and packaging) and the lack of standardisation. This inconsistency meant that most traders visited the market every day to ensure they could access the best-quality products available.

From their findings, the team mapped out the different variables currently influencing transactions between buyers and sellers, and used these to build a prototype digital marketplace. This also sought to address the low access smallholders currently have to pricing information by linking to benchmark markets with live data feeds. STFC co-PIs in the project Dr Jens Jenson and Dr Tom Kirkham, contributed their expertise in data architecture, selecting a platform and user interface, and effective use of data science & sensor technology.

A promising prototype
The result was an ecommerce platform designed to be resilient to network failures, provide strong data security, and be compatible with mobile devices. “The buyer interface has an end-to-end order management, with an integrated payment gateway. Meanwhile, the seller interface is a simpler version of a multivendor marketplace with features to add products, quantities, and target prices” says Sarang. “Based on over 300 transactions, we found that when farmers used the app to check the forecasted price for their product, this could increase their income by 15-22%, by helping them to find the best place to sell their products” he says. “This also enabled them to decide whether it was worth transporting their produce to the market in Mumbai to get a better price, or if they would receive the same amount at a local market.”

In the future, Go4fresh envisages that the apps could also allow smallholders to capitalise on growing food trends, including the rising demand in India for healthy and nutritious food, besides food that is produced more sustainably. “Digital marketplaces can allow smallholder farmers to directly connect with the buyers who are most interested in their products, resulting in the best price for them and higher-quality products for consumers” Sarang says.

For the next stage, the team will refine the template, based on input from a wider range of stakeholders covering different food products and supply routes. A particular aim is to simplify the interface and introduce an Indian language version for those with limited English. Ultimately, however, Sarang hopes this approach will have an impact far beyond India.

“In many countries across South-East Asia and Africa, we see a similar scenario. If rolled-out at scale, digital marketplaces could help provide a more sustainable and resilient livelihood for millions of marginal communities” he concludes.