A food secure future is vital for all of us, but food production is facing many challenges, both in the global north and the global south. Researchers from the University of Leeds are tackling these major issues and working together to drive innovation in climate-smart food systems. The work, using complex machine learning is enabling a greater understanding of crop yields and knowing what is grown where using remote sensing data, and is underpinned by a combination of existing work on crop models, AI technology, and deep learning (a subset of machine learning). Some of this work was part of the University of Leeds’ ground-breaking Institute for Climate and Atmospheric Science (ICAS) work which has recently been awarded the Queens Anniversary Prize - the country’s most prestigious higher education honour, recognising excellence, innovation and the impact of research onsociety.
Andy Challinor, Professor of Climate Impacts at University of Leeds, explains that a combination of machine learning with the large amounts of data available has really provided the step forward in applying crop science and climate science into an understanding of how crops respond to climate, and into predictive modelling. “It really is a better way of harnessing the power of computers and data to work far faster than even large groups of people could,” he says.
Prof. Challinor leads the Climate Impact Group (CIG) which generates understanding of the impacts of climate variability and change on food security. His work focusses on developing and running crop models, which are used with climate models to generate process-based simulations of how crops respond to weather and climate. His previous work (add link https://climate.leeds.ac.uk/transforming-food-systems-needs-strong-partnerships-and-cycles-of-innovation/) has shown that the rate of climate change is already beginning to outpace the ability of crop breeding to keep up with that change. It showed the process of breeding, delivery and adoption of new maize varieties in Africa is lagging behind the continent’s changing climates. As Prof. Challinor explains, “This means seed is being produced that is, in a sense, mismatched by the time it reaches farmers’ fields.” This work used more standard techniques and fundamental theory, but after seeing a Van Gogh painting which was created by AI, he became more interested in how AI could improve the existing methods he was using. This led him to work with Samuel Bancroft, one of the first wave of PhD students funded by the Satellite Data in Environmental Science – Centre for Doctoral Training (SENSE CDT), who is researching deep learning (a subset of machine learning), and remote sensing.
The Earth Observation CDT is an exciting new centre funded by the Natural Environment Research Council and UK Space Agency. It will train fifty PhD students to tackle cross-disciplinary environmental problems, applying pioneering data science methods to the vast amounts of satellite data collected daily.
“My work is a little different but along the same path as the rest of my colleagues in the Climate Impacts Group,” Sam explains. “The CIG is integrating machine learning, explicitly calculating crop yields. I’m coming in from a remote sensing angle, which can also be used for crop yield predictions. I’m looking at crop type classification. What that means is, you take satellite imagery, feed it into a model, it outputs a map telling you what fields correspond to which crop.”
Machine learning requires knowledge about what exists on the ground in order to make a successful model. When mapping crop type from satellite imagery, a popular approach is to use machine learning because you don’t need to explicitly programme how it makes predictions. “You can just give it the ‘ground truth’ and it will work out what it needs to detect for the best result,” says Sam. Traditional approaches to crop-type classification using satellite imagery rely too much on this ground truth. In developing regions of the world, where information is limited, models are hard to develop successfully. There are disadvantages, because less ground truth means the models are less capable.
“My research is exploring a new way of approaching crop type classification,” says Sam. “It includes the best from traditional supervised approaches but also allows the model to interpret satellite imagery without ground truth, to work out the important signatures of crops growing through time, generating additional (synthetic) satellite data. This dramatically reduces dependence on labels, still gets good performance, and has additional benefits such as being able to score geographical regions by how confident the model is that the predictions made can be relied upon as true.”
Sam also helped develop the open-source and modular Leeds Farm app (https://spen-farm.herokuapp.com/), which allows the health of crops to be analysed over time. It provides a ‘digital twin’ to the CIEL Spen Farm project, the 9,500m² extension to the existing farm owned by the University of Leeds. Sam is building an infrastructure to have satellite imagery rolling automatically. Sam uses data from the EU Copernicus programme Sentinel 2, which is used for other things including the EU Common Agriculture Policy.
Professor Challinor is using this complex machine learning described by Sam, in the CIG to link crop climate science with biology and even genetics, to be able to inform breeders about what traits to look for when breeding new varieties of crops. He explains why this collaborative work is so important: “If you know the yields of a crop and you know what is grown where, those are the two components of food production – you know how much food there is and what the food types are.”
This work is part of the University of Leeds’ Climate Change, Agriculture and Food Systems (CCAFS) research, driving innovation in climate-smart food systems with CGIAR (formerly the Consultative Group for International Agricultural Research), a global research partnership for a food-secure future dedicated to reducing poverty, enhancing food and nutrition security, and improving natural resources.
For more information:
Prof. Andrew Challinor - A.J.Challinor@leeds.ac.uk
Samuel Bancroft - email@example.com