A field with a future: Artificial intelligence for better seed

The TerraSentia robot of the agricultural start-up EarthSense drives through the parcels of a wheat trial field.

Artificial intelligence is gaining in importance in seed breeding. KWS is currently developing and conducting tests with a field robot in the U.S. to find out how plant traits can be identified automatically and precisely to support variety selection decisions and enable conclusions that will help improve yields and resistances in agricultural crops.

Just a thin mast with a black pipe rocks to and fro between the ears of wheat – that’s all that can be seen of TerraSentia from the edge of the field. The knee-high robot, which is electrically powered and guided by GPS, moves through a precisely laid-out trial field in Champaign near Chicago in the U.S. state of Illinois. With its four jagged-edged wheels, the robot moves almost effortlessly over dirt clods and furrows and wends its way incessantly over the field.

The robot is helping breeders produce plants that are less prone to damage from diseases, are better adapted to climate change, and secure and increase yields for farmers.

Breeders face many challenges in their work. One of them is that they need as much information as possible on the hundreds of varieties and more than 20 crops KWS offers and breeds. They include sugarbeet, wheat, corn, rapeseed and sorghum. As is company tradition, KWS’ experts test millions of plants a year on tens of thousands of trial plots.

That costs a lot of time, work and effort. “And yet breeders can only watch over a small part of a plot,” says Jia Yan, the project manager responsible for digital innovations at KWS. “It would be better to keep a constant eye on the plants and see how they fare in the field.” But where do you get the time to keep tabs on even more ears, leaves or shoots? And all that at short time intervals and in the many countries where KWS conducts breeding?

Images from autonomous robots

Breeders are helped in the field in Illinois by the robot TerraSentia. It has been developed and built by the start-up EarthSense from the University of Illinois in Urbana-Champaign.

The pipe on the mast contains two cameras that take detailed pictures of the wheat field non-stop as the robot moves around the field. The robot also stores the exact location where the pictures are taken. The breeders then know what stage of development the ears have reached (such as ear emergence) on which of the many plots. Up to now, that inspection work has been done by people out in the field, come rain or shine.

Training for artificial intelligence

Yet the heart of the system is not the four-wheeled robot, but an artificial intelligence software solution on KWS’ computers and EarthSense. The term artificial intelligence is used when a machine delivers results that are otherwise ascribed only to intelligent beings, such as humans. It analyzes the robot’s pictures and identifies things of interest to the breeders – for example if an ear on a stalk is already fully developed or if it is still partly enclosed by protective leaves. That information is relevant for breeders.

The software must first be trained to do this – after all, it can’t identify plants on its own to begin with. Wheat breeders like Jana Murche and Mark Christopher screen the images taken by the robot. Important features on them include ear emergence, plant height and awn type. Also of great interest to the breeders are visible leaf diseases or those affecting the ears.

And then data specialists “feed” the software with information from the breeders. If the images show completely or incompletely emerged ears, for instance, the software is taught that it is looking at complete or incomplete ear emergence for a plot at a given time.

Good results with artificial intelligence

The software generates knowledge from experience by means of repetition. Its neural network thereby creates a new mathematical model, an algorithm. It does what people call “learning.” Once the artificial intelligence has obtained enough knowledge from humans, it uses it to compare new images and take action. In the case of the robot, it evaluates the pictures of plants without the need for human assistance.

However, if the crop to be analyzed is not wheat, but sugarbeet, say, humans have to teach the machine the key differences all over again. The neural network has to be retrained and the algorithm adapted for every new application.

The results of the system’s first version dating from 2018 show that the algorithm already works accurately. The artificial intelligence detects emerged ears with 96-percent reliability. It can identify whether an ear is completely awned – the long beardlike extensions at the end of florets – with an accuracy of 92 percent.

The system is now being constantly improved. The objective is for TerraSentia to travel effortlessly and independently over the field every day. And twice if necessary. Or three times. “That will allow us to make more soundly based decisions as part of selection,” says wheat breeder Mark Christopher. “Especially in our breeding nursery with its hundreds of thousands of individual rows, where it hasn’t been easy for us to collect that data up to now.”

People are still vital

Yet the example also shows that “technology won’t replace the experience of breeders,” says project manager Jia Yan. Artificial intelligence and robots can assist breeders by providing broader information on which to base their decisions. “But only our breeders can train the system to deliver the right information precisely.” Combining human and artificial intelligence will make the breeding process faster and more reliable. Work with artificial intelligence and autonomous robots is therefore an important part of KWS’ research strategy.


The number of robots can be increased as desired

The more robots are used in the fields, the more data the breeders acquire. TerraSentia can maneuver on many fields and is also relatively easy to build. The predecessor model even came out of a 3D printer. The navigation technology and digital cameras are now widespread and standardized. That means the number of robots can be increased quickly. The computing power needed to train and operate a neural network can be found in the cloud at the click of a mouse, as and when required.

At the same time, many researchers worldwide are working to expand the capabilities of artificial intelligence – for example to reduce the number of training images required. “We’re still training our system and don’t yet use it commercially,” says Jia Yan. “But it’s just a matter of time until robots and artificial intelligence will help us in breeding.”

Phenotyping – a look at the plant

To ensure plants are doing well and growing, breeders have to inspect them repeatedly – out in the field, i.e. where they grow with their genetic makeup (genotype) under environmental influences.

That requires a lot of time, but also a lot of breeding experience to assess existing or desired characteristics of the plant – their phenotype – and respond with appropriate breeding measures.

Modern technology can help in all of that and deliver additional information. KWS is therefore putting a great deal of work into developing new methods to automatically identify specific traits of plants. Images of fields or plots are taken from the ground or air, for example. They can be used on the computer to deduce information on traits. That requires close cooperation between IT specialists and experienced breeders.

What is artificial intelligence?

Artificial intelligence (AI) has long been a field of research. Decades ago, it delved into a distant future where machines assume important human tasks, a vision that has taken on more concrete form in the past years – as evidenced by voice assistants, voice translators, autonomous vehicles or detection of cancer and other diseases. For artificial intelligence to work, it needs huge volumes of data, commonly called big data.

AI can roughly be divided into three areas:

  • Perception – such as voice, text, picture or facial recognition.
  • Learning – such as deep learning and machine learning
  • Application – such as the use of robots like TerraSentia.

Back to the “World of Farming” main page

Your contact person

Thilo Resenhoeft
Thilo Resenhoeft
Corporate Communications
Send e-mail
CONTACT