Machine Learning to Illuminate Smallholder Farmers’ Problems

As any social scientist who has worked in rural areas of developing countries, including our founder, can tell you, data that inform policy decisions affecting smallholder farmers are anything but precise. Could cutting-edge machine learning techniques help?

A method, if not a technology

The term “machine learning” doesn’t necessarily point to any specific technologies or products. It’s more of a discipline.

Scientists working with machine learning are exploring ways to build computational tools that will allow computers to gather data, assess it, and then make optimal use of this data in specific contexts, sort of “learning” as new input is received. Machine learning can be employed in a variety of applications, most famously speech recognition where a computer can “learn” the cadence or pronunciation peculiarities of a person’s voice to produce text.

Two recent studies argue that machine learning can be relied upon to better understand the needs of smallholder farmers, as well. Researchers out of India claim machine learning is helping them organize a system for farmers to share or lease farm machinery or equipment when needed. Meanwhile, a separate study from South Africa is exploring how machine learning tools can be put to use to monitor smallholder maize growing.

“Learning” the right way to loan equipment

In India, smallholder farmers’ problems are not only severe, but can also be deadly. “The main issue of the farmers is they are not sensitized with modern equipment and tools, and on the flip side, they are falling under the burden of debt that forces them to commit suicide,” Rakhra et al. explain in their article in the Journal of Food Quality. To help farmers cope with their challenges, the team says it built an “Uberized model” whereby farmers in a region can share or lend farm implements to one another, the idea being that this will help them grow more food inexpensively and boost incomes to pay off debts. They surveyed 562 farmers in Punjab to gather as much data as they could on what implements were available and how farmers might best afford them. Then they put the machines to work, to determine which machine learning tool might best illuminate what an ideal farm equipment rental and sharing arrangement could look like.

The Indian researchers practiced analyzing their data using three machine learning tools: nearest neighbors, logistic regression, and decision trees. For their purposes, they found the decision tree machine learning model yielded the best results “given the large number of input factors, such as the kind of crop, the time/month, and the type of equipment necessary for the crops.” They also had to contend with farmers’ educational attainment, the volume of land they owned, their familiarity with technology like mobile phones, how much debt they owed, who they owed it to, and the interest rates they were paying. For technical reasons, they determined that decision tree analysis produced “excellent performance on our data set”.

An eagle’s eye view

Machine learning is also being relied upon to predict how well a cyclical harvest might go and where interventions might be needed if signs of trouble emerge.

“Machine learning proved to have a high capacity to estimate smallholder maize planted areas.”

Writing in Sustainability, a team of South African agricultural scientists said they experimented with using machine learning to better assess satellite imagery of smallholder maize plots. They said they narrowed in on this subject because, although maize is critical to sub-Saharan Africa’s food security, many countries in the region don’t actually know how many hectares are devoted to growing this staple, let alone how much is produced each year. “The disparity between declining maize supply and increasing demand for maize makes it necessary to develop a methodology to map smallholder maize farms and their sizes,” they argued. “Information about the real extent of smallholder farms will guide the government when dispersing aid to them, inform land-use policies, and provide an indication of the current food security status, especially in vulnerable rural communities”.

Do machine learning tools help achieve those above goals? Emphatically yes, said the South African researchers. Their algorithms were designed to improve upon satellite imagery taken by the Sentinel 1 and Sentinel 2 remote sensing platforms, and they reported great success. “Machine learning proved to have a high capacity to estimate smallholder maize planted areas,” they reported. The next step is finer-grained machine learning-empowered satellite image analysis to help predict cyclical yields and perhaps even detect drought or disease outbreaks.

Machines cannot “learn” like humans per se, but they can be programmed to process huge volumes of data and then discern meanings and patterns from this information swarm far better than humans could acting alone.

Machine learning is an evolving category of innovation that may help smallholders grow more food and earn more money. Perhaps most exciting for Grow Further, supporting machine learning projects is a way to reach smallholder farmers without the long and expensive sales cycle associated with many technologies that require retail-level adoption on individual farms.

Grow Further

Photo credit: Maize harvested from smallholder farms in Mozambique. Stevie Mann, International Livestock Research Institute.

Subscribe to Our Newsletter