The eyes don’t have it: Machine vision to improve cotton breeding
When it comes to cotton breeding, it’s important to know how ‘hairy’ cotton leaves are. But just how do you measure ‘hairiness’?
It takes years to become proficient in the art of estimating leaf hairiness for cotton breeding. But it is ultimately just an estimate – as fallible as any manual task.
Unlike DIY projects where you can eyeball a measurement, the stakes in cotton breeding end up having global ramifications for the world’s fibre needs.
Why do the hairs on the back of a cotton leaf matter?
CSIRO cotton breeder Dr Warwick Stiller said breeders cross parental lines with characteristics that are potentially valuable. Breeders then select offspring that have characteristics from both parents to further develop improved crop varieties.
“Cotton breeders frequently use their experience of what is optimal by visual assessments of plant attributes when they make selections for progression towards commercial varieties,” Dr Stiller said.
“It just happens that cotton leaf hairiness is a characteristic that impacts insect resistance, fibre yield and fibre value,” he said.
“We’re accurate, but after years of breeding we’re chasing diminishing returns. We need to improve our accuracy to find further gains.”
At stake is a $2 billion per annum industry in Australia. Together CSIRO and its commercial partner Cotton Seed Distributors have released more than 100 new cotton varieties in the past 30 years. CSIRO developed all cotton grown in Australia, and Australian cotton yields are the highest in the world.
Faster and more precise plant breeding is critical for producing future crops. These crops will need to meet human needs, be resilient to climate change and produce higher quality food and fibre products.
It’s no surprise that in plant breeding, just as in many traditional work areas, there is great interest in developing digital and machine-based visual decision systems that can improve tasks typically undertaken by humans.
So, are the days of traditional plant breeding techniques numbered?
Not accurate enough – cotton breeding requires a new way
Dr Vivien Rolland is a CSIRO plant biologist with an established skillset in plant function and microimaging.
Essentially, he combines biotechnology and high-powered microscopes to study plants with the goal of boosting crop productivity and sustainability.
Aware of the problem faced by our Australian Cotton breeders, Dr Rolland wanted to apply his skills to this problem.
“The problem is that human eyes ultimately aren’t accurate enough. One person’s assessment may differ from another’s,” Dr Rolland said.
“That’s a problem in a high-throughput setting like a commercial breeding program. We needed to find a way to improve accuracy at scale,” he said.
In 2019, he was awarded a CSIRO Julius Career Award. It enables early to mid-career scientists to learn new skills and technologies to grow a new science area.
His idea was simple, yet potentially impactful. There are lots of microscopic plant characteristics which are hard to visualise, but that are very important for crop breeding.
Cotton leaf hairs is one but there are others, like stomata. Stomata are the tiny pores on leaves which are important to regulate water use efficiency.
“To capture microscopic plant traits, we needed to combine a simple microimaging system that can be used in the field, with an automatic way to score these traits based on artificial intelligence.”
Finding the cotton breeding sweet spot
The first target was the assessment of leaf hairiness. Cotton breeders know that there is a sweet spot for optimal leaf hairiness.
Selected plants need leaves that are not too hairy. This is because it can increase plant ‘trash’ (impurities) in the harvested cotton fibres, which reduces economical value of the fibre. Hairy leaves also expose plants to certain insect infestations. On the other hand, very low leaf hairiness is linked to a lower fibre yield. So that’s not good either.
In the first year of his Julius Award Dr Rolland experimented with a range of simple, cheap and practical imaging set-ups to collect a baseline cotton leaf image dataset. He also expanded his skills in data science. He participated in a deep learning workshop at European Molecular Biology Laboratory (EMBL) in Germany to build proof-of-principle deep learning models to score leaf hairiness.
“I built deep-learning models which showed very promising results. But to turn this into a tool the breeders could use, it became obvious that we needed to step it up a notch,” Dr Rolland admits.
Enter the HairNet model
It was just at this time that CSIRO established the Machine Learning and Artificial Intelligence Future Science Platform (MLAI FSP).
Through the MLAI FSP, Dr Rolland connected with fellow researcher Dr Moshiur Farazi, who specialises in computer vision and machine learning. Dr Farazi brought his cutting-edge machine learning skills to bear on this computer vision challenge.
“In computer vision, we often work on fundamental problems using very large, curated image datasets of trains, cats and cars. I was really excited to get to work on a real industry problem,” Dr Farazi said.
Together, the team built ‘HairNet’: a sophisticated deep learning model that can analyse leaf images and automatically score leaf hairiness.
“It’s a robust model that can now be integrated into a cotton breeding program to give the breeders increased confidence in their scoring,” Dr Farazi said.
Imagery can also be saved for later analyses. This is not possible with a purely visual inspection.
Proven accuracy with wider applications
New research published in Plant Methods shows the model has proven to be very accurate in reproducing breeder scoring of leaf hairiness.
“We found that we could score leaf hairiness in the field or in the glasshouse. This is useful because breeders use both environments at different stages of the commercial crop development pipeline,” Dr Rolland said.
“Overall, HairNet achieves 89 per cent accuracy when we take one image per leaf. We can push that up to 95 per cent when we take several images per leaf,” Dr Rolland said.
“We can even get 100 per cent accuracy when we test plants grown in the glasshouse.”
There are only a few visual recognition systems globally that are currently used in agriculture, with fruit harvesting and the identification and control of weeds being the most frequently reported.
With improved performance and reduced price of hardware and software, many new visual recognition systems will be used in the future to increase agricultural productivity in cotton and beyond.
HairNet shows that combining existing successful biological teams with new data science and machine learning skills will be important to breed future crops.