Three ways that machine learning can bring precision agriculture to small-scale farms
CGIAR Initiative on Digital Innovation
- Impact Area
The success of every harvest depends on three things: the crop grown, the environment and the actions of the farmer. Yet the factors behind this simple formula are far more numerous and interconnected: an early rain, fertilizer applied, the genetics of the plant itself. For this reason, machine learning algorithms – specialized at finding patterns and interactions in large, unstructured datasets – may offer a new opportunity to optimize farming practices, especially in the Global South.
Ongoing research from CGIAR and partners, including AgriLAC Resiliente and the CGIAR Initiative on Digital Innovation, has identified case studies showing three ways in which machine learning can optimize farming.
Why machine learning?
Farmers have always experimented and relied on past experience to get the most out of their farms. In the modern era, experimental trials have been used to reduce complexity, controlling factors associated with environment and management to isolate the effect of changing each one and unpick the key interactions. While this has been successful on a global scale, it is also limiting. Industrialized agriculture has flourished by growing a handful of modern varieties in similar and adapted environments around the world. When it comes to the small-scale farms that provide around a third of the world’s food, adapting to a more unpredictable climate, reducing the environmental impact of farming or exploring lesser-known crops, new approaches may have greater potential.
Every harvest is an experiment, repeatedly conducted by an estimated 600 million small-scale farmers around the world. With digital and remote sensing tools now available to collect data on crop performance, farm management practices and climatic or environmental conditions, it is possible to generate the large unstructured datasets that machine learning thrives on.
“Machine learning methods allow us to accelerate research in important areas,” said Daniel Jimenez, data scientist with the Alliance of Bioversity International and CIAT, and one of the study co-authors. “For example, to study high-value perennial crops such as fruit trees would normally take decades of field observation, whereas now we can conduct crop performance experiments using these models.”
1. Growing the right crop in the right place
Lulo is a crop originating in Colombia, Ecuador and Peru that produces a nutritious and tasty fruit, long seen as a potential cash crop for farmers in tropical regions. To identify where else lulo could be grown, one study collected data from current growing areas and machine learning to identify groups of farms that shared similar environments. A combination of machine learning models combined with other modelling techniques was able to explain 80% of the yield variation across farms, describing the ideal conditions for growing lulo. The model also identified well-managed farms even though no data was available on farm practices. It showed how machine learning can overcome the challenge of distinguishing between environmental and management effects even with relatively crude data.
Coffee prices are often tied to the specific region in which it is grown, which is understood to have an impact on the perceived quality and taste of the final product. Taste is a complex trait, measured by an industry-standard subjective test. One study in Southwest Colombia found that a specific machine learning model could identify combinations of environmental factors that would lead to great-tasting coffee, which would have confounded traditional factor-by-factor experimental trials.
2. Better crop management
More precise fertilizer recommendations are needed for farmers in diverse environments: otherwise, farmers will invest in inputs with little return and cause unnecessary damage to the environment. Using data from 6,000 field trials, a machine learning model has been developed to predict how much fertilizer is needed for optimal yields in different environments of Ethiopia, providing more precise recommendations than those used by the national extension system.
To improve crop management recommendations for maize farmers in Northern Colombia, data on farming practices and yields was fed to a machine learning algorithm, which identified five practices used by high-performing farms. Field trials confirmed that yield gains of up to 2.5 tons per hectare were possible, and some of the recommendations contradicted those given by extension agents. The guidelines also substantially reduced fertilizer costs and provided advice on how to reduce risks related to variation in the weather patterns, with an emphasis on reducing the negative impacts of heavy rainfall. In another study, this time of farmers in Chiapas, Mexico, it was found that recommendations derived from machine learning could increase yields by 1.8 tons per hectare.
3. Weather risk management
Farmers are often given recommendations based on climate data or average weather conditions, but the actual weather experienced on-farm is often different. One study found that variation in the El Niño climate phenomena up to 24 months before harvest could explain 75% of the difference in yields on cacao farms in Indonesia, with the machine learning model also offering the potential for fertilizer recommendations. As climate change causes climate to become more predictable, machine learning can help farmers respond to these weather events.
Challenges and prospects
The case studies show that machine learning can take advantage of new methods to collect farm management and climate data, and even in some cases where farm management data was missing. This has the potential to bring precision agriculture to small-scale farmers, facing varied growing environments and uncertainty due to climate change, but also many unexplored opportunities to improve their livelihoods.
The case studies also highlighted some limitations, for example when farmers collectively adopt different practices in different environments it becomes difficult to separate these factors in the data. “Machine learning models only work well within the range of training data and cannot be generalized to situations that weren’t captured in the dataset,” said Jiménez who, along with the other authors, emphasized in the study that machine learning should be seen as complementary to other knowledge rather than a universal solution.
Read the full study: Cock, J., Jiménez, D., Dorado, H., and Oberthür, T. (2023) Operations research and machine learning to manage risk and optimize production practices in agriculture: good and bad experience. Current Opinion in Environmental Sustainability 62. https://doi.org/10.1016/j.cosust.2023.101278
Top photo credit: Axel Fassio/CIFOR-ICRAF