Key points
- CSIRO researchers are using specialised drone-mounted cameras to detect wheat traits invisible to the human eye.
- This technology could track complex processes, like photosynthesis, respiration and grain-filling rate, with the potential to unlock yield with less water.
- The project has proved these instruments can work reliably at scale, gaining interest from breeders.
Aswin Singaram Natarajan rolls a stem of wheat in his hands. It looks perfectly healthy, but the senior experimental scientist at CSIRO, Australia’s national science agency, knows his eyes can deceive him.
A plant could look green and vital and still be losing - losing too much energy or losing too much water or filling its grain to ripeness too slowly. And while these weaknesses are invisible to the human eye, they’re plain as day to the right instrument.
“We use X-ray to see through our body to see our bones, similarly there are things we cannot see that hyperspectral vision technology can see,” said Mr Singaram Natarajan.
Where our eyes see light as just red, green and blue, a hyperspectral camera splits it into hundreds of narrow bands. It sees wavelengths from 350 to 2500 nanometres – a range more than seven times wider than the human eye can see, stretching from ultraviolet deep into the infrared. Each leaf reflects those bands differently depending on what’s going on inside it, leaving a fingerprint the camera can read.
Mr Singaram Natarajan is one of a team of researchers on AcceleTraitPlus, a CSIRO project aiming to put that technology at the service of Australian breeders by working to deliver more robust and abundant varieties of wheat for farmers.
What the camera can see that we can't
Among the traits the team is chasing are optimised photosynthesis and respiration. The former is how a plant uses sunlight and atmospheric carbon dioxide to produce sugars that fuel growth and, eventually, produce grains. Respiration is the other side of the ledger: it is how much of those sugars a plant uses to keep its cells functioning and growing, particularly overnight when photosynthesis has stopped.
A third trait, the rate of grain filling, is the process by which final grain size is achieved to sustain yield and it might be the project’s strongest result so far. Measured by hand, it means visiting a trial site 10 to 12 times over a season, plucking thousands of wheat heads, then drying and weighing each one.
Breeders are normally focused on yield, but looking at characteristics like this could reveal why one variety of wheat copes with difficult conditions like extreme heat or drought better than another, long before yield ever comes into question.
This is a question the Grains Research and Development Corporation (GRDC) has been engaged with for decades.
GRDC Senior Manager for Genetic Technologies (cereals) Jessica Hyles said GRDC was investing in this research to help breeders develop wheat varieties that could deliver higher yield potential, better water use and more reliable performance in tough seasons for Australian grain growers.
The team is exploring whether phenomics – the large-scale measurement of plant traits using sensors and imaging – could do this work faster. By testing whether these traits can be read from hyperspectral signatures, the team hopes to give breeders a way to identify promising plants more quickly in the field.
“Testing these tools with a commercial breeding partner is important to ensure they are practical to use. Ultimately, we want growers to have varieties with higher yield potential, better water use and more reliable paddock performance.”
Dr Anton Wasson, a project leader in CSIRO’s Agriculture and Food research unit, said there had been significant investment in plant physiology to revolutionise breeding – but very little of that work has actually been adopted by commercial breeding programs.
He added that the traits with the most promise were also the hardest to adopt: measuring them at the scale of a breeding trial, across hundreds of plots, was simply too slow and expensive.
Then the team acquired a new tool: a high-specification hyperspectral camera mounted on a drone – it was a gamechanger.
From leaf clip to flyover
Up until now, these traits often had to be measured painstakingly by hand.
Essentially, a researcher would walk along the trial rows carrying a heavy spectrometer, clipping one leaf at a time and then waiting to take a point reading. But now they had a drone that could image an entire trial site in a single flight, gathering in minutes what used to take hours, and capturing the whole crop, not just individual plants.
It isn’t simply a matter of strapping a camera to a drone – a hyperspectral sensor is a sensitive scientific instrument that needs careful calibration, said CSIRO’s Ryan Lagerstrom. The project’s data analysis and computer vision expert, Mr Largerstrom explained that a satellite can cover the same ground, but it passes on its own schedule, not that of the breeders. In addition, its images are far grainier than those a drone can capture close up. A handheld sensor is accurate, but having people walk trial rows clipping leaves was never going to be adopted at scale.
“The drone is the middle ground between what you see from a satellite and what you see right up close,” said Mr Lagerstrom.
GRDC backed the new project, partnering with breeding company AGT. They agreed on a narrow goal: take the three most promising traits and test whether the new hyperspectral tools could measure them faster than manual methods, and fast enough to be useful to breeders.
Interest in the drone-mounted camera held, and became a second, expanded project, AcceleTraitPlus, running from 2024 to 2025 across breeding trials in Western Australia and New South Wales.
Getting there has meant confronting how difficult it is to collect quality field data.
“The planning, resourcing and expertise needed to acquire and analyse high-quality data, especially hyperspectral, are critical,” said Dr Gonzalo Estavillo, a Team Leader with CSIRO’s Agriculture and Food research unit. “Generating robust data at scale is one of the biggest challenges in crop research, but it's the foundation for every discovery we make.”
The team presented at the Australian Crop Breeders Week conference showing the models could predict grain-filling rate and photosynthesis.
Dr Wasson credited the progress so far to the team’s long-running physiological and phenotyping research – and said it was already changing minds.
“One of the breeders was already giving up on hyperspectral tools, and I think we turned that thinking around.”
A problem the numbers can't dodge
What's different this time, Dr Wasson said, is that the project has proven the hard operational part: getting hyperspectral instruments to work reliably inside a real breeding trial, at scale, plot after plot.
Part of the reliability comes down to the work most of those outside the lab will never see: the terabytes of hyperspectral imagery that have to be processed and run through statistical models until you get the information you’re looking for.
“There’s a continually evolving way of attacking the problems,” Mr Lagerstrom said, describing the shift from simple linear models to, more recently, deep learning and AI. At CSIRO, cloud computing has made it possible to process the volume of the data and share it across the team in real time – a real advantage.
“We’re still scratching the surface of what’s possible,” Dr Estavillo said. “We’ve shown hyperspectral imaging can work at scale. But there’s still a lot to learn from the data, and the technology itself has a long way to go.”
What matters just as much as more research, he said, is listening to breeders – building tools they actually want, not just ones that are clever.
“If we get that right,” Dr Estavillo said, “breeders will have better tools for selection in the field and growers will have better varieties in the paddock.”