Using machine learning to unlock the secrets of plant productivity

Crop yields are affected by a huge range of different factors, which can make it immensely difficult to work out which set of conditions result in the best growth. But a project supported by the STFC Food Network (SFN) demonstrates that artificial intelligence and machine learning are promising approaches to reveal ‘ideal plant growth recipes’ hidden in large datasets.  

Farming already accounts for around 38% of the global land surface (Food and Agriculture Organization), but this will need to expand even further if current methods have to scale to feed the rapidly growing population. And with cultivated soils degrading at a frightening rate, agriculture is increasingly encroaching into pristine, biodiverse habitats. With ‘business as usual’ clearly being an unsustainable option, we urgently need to rethink the way we farm.

This has sparked huge interest in vertical farming, where plants are grown stacked in layers within a protected environment (typically indoors). In these soil-free systems, plant roots are either immersed in a nutrient solution (hydroponics) or exposed to the air and irrigated with a nutrient-rich mist (aeroponics). Vertical farming offers many benefits, including high space-efficiency, reduced water usage, shorter growing times, reduced need for pesticides/herbicides, and shelter from extreme weather. In addition, since vertical farms can be set up practically anywhere (even underground), they could enable hyper-localised production, thus shortening supply chains and providing fresh, nutritious food all year round. 

But there are currently many challenges with vertical farming, including high capital costs, considerable energy consumption and inefficient resource use. According to Chungui Lu, Professor of Sustainable Agriculture at Nottingham Trent University, a crucial first step to address these issues would be facilitating greater precision control of plant growing conditions. “Young crop plants are incredibly sensitive to a large range of different environmental factors, including air and water temperature, light, relative humidity, and CO2 levels” he says. “Even small differences can have a great effect on growth. However we currently have very little data on the optimum conditions for indoor farming systems.”

With so many different factors affecting plant growth, it would be virtually impossible to manually work out the best combination in real-time to maximise yields. So as part of an SFN scoping project, Chungui joined forces with Dr Gadelhag Mohmed and Dr Steven Grundy from Nottingham Trent University, and Professor Wantao Yu (University of Roehampton) to explore how big data approaches could be used to derive ‘optimum growing recipes’.
​
“Essentially we had two main questions” says Chungui. “Can we monitor in real-time all the factors that affect plant growth? And can we then use artificial intelligence, big data and internet of things methods to predict the conditions that give the best yield and quality for different crops?”

The team’s approach centred on using neural networks, so-called because they mimic the way the human brain processes information. Structurally, neural networks are made up of a number of layers of nodes (neurones), each connected to nodes in the next layer. The connections between the nodes are weighted, and these weights are adjusted during a training process using a labelled dataset. The resulting network of connected neurones can then automatically (and accurately) process novel data to achieve the correct output.

Traditionally, developing a neural network relies on prescribing important features within the training dataset, which requires expertise and can be time-consuming. But what sets deep learning apart is that the algorithm doesn’t need important features to be specified in advance: given enough labelled samples, it will discover the most relevant attributes on its own. “It would take forever to experimentally test each individual combination of the factors that affect plant growth, so being able to automatically infer this would be a tremendous advantage” says Chungui.

But this intuition comes at the cost of considerable processing power: a single neural network can be made up of tens to hundreds of layers of artificial neurones connected by millions of weights, and in a single experiment go through over a billion computational operations. Since this is far beyond the capabilities of most computers, the team collaborated with STFC’s Tom Kirkham to access the Hartree Center at STFC Daresbury Laboratory in Cheshire. This enabled them to construct a cloud computing and data processing platform to develop the model. “A real strength of being part of the SFN network is being able to access experts and facilities to support really cutting-edge techniques. It makes a really good platform for research” says Chungui.

To train the model, the team collected data from sensors which measured the temperature, day length, relative humidity, CO2 levels, and light levels. The output data included measurements of shoot fresh/dry weight, root dry weight, and leaf area, and photosynthetic efficiency. “The algorithms in the resulting trained model had a strong predictive power, demonstrating that we can effectively extract hidden patterns in these huge datasets to identify the best growth conditions” says Chungui.

For their next steps, the team are looking to combine this approach with high-quality, real-time data captured using a state-of-the-art plant phenotyping scanner: the PlantEye multispectral 3D scanner. “These impressive machines can measure a whole range of factors related to plant crop and productivity, including biomass, plant height, 3D leaf area, and greenness– all without having to destroy the plant” says Chungui.

Chungui is particularly keen to extend this work to explore tailored ‘light recipes’ for different crops. Plants mainly use the red and blue parts of the visible light spectrum to power photosynthesis, and differences in the relative proportions of these can have a big impact on growth. Since blue light uses considerably more energy than red, working out the optimum recipe can result in great cost savings. Besides productivity, the team intend to investigate how different light combinations also affect flavour and nutritional value.

But ultimately, Chungui’s research group wants to see this technology in the hands of farmers, empowering them to make evidence-based decisions. “We are now exploring ‘Smart Green Grow Vertical Farming’ models, where the growing systems have built-in data transfer nodes and communicate with remote data bots to monitor and control environmental and plant performance factors. Ultimately, this would link farmers and their crops together with instant food production enhancement data.”