AI Breakthrough in Crop Breeding: New Self-Learning Model Revolutionizes Agricultural Research

New approach makes AI adaptable for computer vision in crop breeding

Scientists developed a machine-learning tool that can teach itself, with minimal external guidance, to differentiate between aerial images of flowering and nonflowering grasses -- an advance that will greatly increase the pace of agricultural field research, they say. The work was conducted using images of thousands of varieties of Miscanthus grasses, each of which has its own flowering traits and timing. Accurately differentiating crop traits under varied conditions at different points in the growing cycle is a formidable task, said Andrew Leakey, a professor of plant biology and of crop sciences at the University of Illinois Urbana-Champaign, who led the new work with Sebastian Varela, a scientist at the Center for Advanced Bioenergy and Bioproducts Innovation, which Leakey directs. The new approach should be applicable to numerous other crops and computer-vision problems, Leakey said. The findings are reported in the journal Plant Physiology. "Flowering time is a key trait influencing productivity and the adaptation of many crops, including Miscanthus, to different growing regions," Leakey said. "But repetitive visual inspections of thousands of individual plants grown in extensive field trials is very labor intensive." Automating that process by collecting images via aerial drones and using artificial intelligence to extract the relevant data from those images can streamline the process and make it more manageable. But building AI models that can distinguish subtle features in complex images usually requires vast amounts of human-annotated data, Leakey said. "Generating that data is very time-consuming. And deep-learning methods tend to be very context-dependent." This means that when the context changes -- for example, when the model must distinguish the features of a different crop or the same crop at different locations or times of year -- it likely will need to be retrained using new annotated images that reflect those new conditions, he said. "There are tons of examples where people have provided proof-of-concept for using AI to accelerate the use of sensor technologies -- ranging from leaf sensors to satellites -- across applications in breeding, soil and crop sciences, but it's not being very widely adopted right now, or not as widely adopted as you might hope. We think one of the big reasons for that is this huge amount of effort needed to train the AI tool," Leakey said. To cut down on the need for human-annotated training data, Varela turned to a well-known method for prompting two AI models to compete with one another in what is known as a "generative adversarial network," or GAN. A common application of GANs is for one model to generate fake images of a desired scene and for a second model to review the images to determine which are fake and which are real. Over time, the models improve one another, Varela said. Model one generates more realistic fakes, and model two gets better at distinguishing the fake images from the real ones. In the process, the models gain visual expertise in the specific subject matter, allowing them to better parse the details of any new images they encounter. Varela hypothesized that he could put this self-generated expertise to work to reduce the number of annotated images required to train the models to distinguish among many different crops. In the process, he created an "efficiently supervised generative and adversarial network," or ESGAN. In a series of experiments, the researchers tested the accuracy of their ESGAN against existing AI training protocols. They found that ESGAN "reduced the requirement for human-annotated data by one-to-two orders of magnitude" over "traditional, fully supervised learning approaches." The new findings represent a major reduction in the effort needed to develop and use custom-trained machine-learning models to determine flowering time "involving other locations, breeding populations or species," the researchers report. "And the approach paves the way to overcome similar challenges in other areas of biology and digital agriculture." Leakey and Varela will continue to work with Miscanthus breeder Erik Sacks to apply the new method to data from a multistate Miscanthus breeding trial. The trial aims to develop regionally adapted lines of Miscanthus that can be used as a feedstock to produce biofuels and high value bioproducts on land that is not currently profitable to farm. "We hope our new approach can be used by others to ease the adoption of AI tools for crop improvement involving a wider variety of traits and species, thereby helping to broadly bolster the bioeconomy," Leakey said. Leakey is a professor in the Carl R. Woese Institute for Genomic Biology, the Institute for Sustainability, Energy and Environment and the Center for Digital Agriculture at the U. of I. The U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research; the U.S. Department of Agriculture, Agriculture and Food Research Initiative; and Tito's Handmade Vodka supported this research.

New approach makes AI adaptable for computer vision in crop breeding

Scientists developed a machine-learning tool that can teach itself, with minimal external guidance, to differentiate between aerial images of flowering and nonflowering grasses -- an advance that will greatly increase the pace of agricultural field research, they say. The work was conducted using images of thousands of varieties of Miscanthus grasses, each of which has its own flowering traits and timing. Accurately differentiating crop traits under varied conditions at different points in the growing cycle is a formidable task, said Andrew Leakey, a professor of plant biology and of crop sciences at the University of Illinois Urbana-Champaign, who led the new work with Sebastian Varela, a scientist at the Center for Advanced Bioenergy and Bioproducts Innovation, which Leakey directs. The new approach should be applicable to numerous other crops and computer-vision problems, Leakey said. The findings are reported in the journal Plant Physiology. "Flowering time is a key trait influencing productivity and the adaptation of many crops, including Miscanthus, to different growing regions," Leakey said. "But repetitive visual inspections of thousands of individual plants grown in extensive field trials is very labor intensive." Automating that process by collecting images via aerial drones and using artificial intelligence to extract the relevant data from those images can streamline the process and make it more manageable. But building AI models that can distinguish subtle features in complex images usually requires vast amounts of human-annotated data, Leakey said. "Generating that data is very time-consuming. And deep-learning methods tend to be very context-dependent." This means that when the context changes -- for example, when the model must distinguish the features of a different crop or the same crop at different locations or times of year -- it likely will need to be retrained using new annotated images that reflect those new conditions, he said. "There are tons of examples where people have provided proof-of-concept for using AI to accelerate the use of sensor technologies -- ranging from leaf sensors to satellites -- across applications in breeding, soil and crop sciences, but it's not being very widely adopted right now, or not as widely adopted as you might hope. We think one of the big reasons for that is this huge amount of effort needed to train the AI tool," Leakey said. To cut down on the need for human-annotated training data, Varela turned to a well-known method for prompting two AI models to compete with one another in what is known as a "generative adversarial network," or GAN. A common application of GANs is for one model to generate fake images of a desired scene and for a second model to review the images to determine which are fake and which are real. Over time, the models improve one another, Varela said. Model one generates more realistic fakes, and model two gets better at distinguishing the fake images from the real ones. In the process, the models gain visual expertise in the specific subject matter, allowing them to better parse the details of any new images they encounter. Varela hypothesized that he could put this self-generated expertise to work to reduce the number of annotated images required to train the models to distinguish among many different crops. In the process, he created an "efficiently supervised generative and adversarial network," or ESGAN. In a series of experiments, the researchers tested the accuracy of their ESGAN against existing AI training protocols. They found that ESGAN "reduced the requirement for human-annotated data by one-to-two orders of magnitude" over "traditional, fully supervised learning approaches." The new findings represent a major reduction in the effort needed to develop and use custom-trained machine-learning models to determine flowering time "involving other locations, breeding populations or species," the researchers report. "And the approach paves the way to overcome similar challenges in other areas of biology and digital agriculture." Leakey and Varela will continue to work with Miscanthus breeder Erik Sacks to apply the new method to data from a multistate Miscanthus breeding trial. The trial aims to develop regionally adapted lines of Miscanthus that can be used as a feedstock to produce biofuels and high value bioproducts on land that is not currently profitable to farm. "We hope our new approach can be used by others to ease the adoption of AI tools for crop improvement involving a wider variety of traits and species, thereby helping to broadly bolster the bioeconomy," Leakey said. Leakey is a professor in the Carl R. Woese Institute for Genomic Biology, the Institute for Sustainability, Energy and Environment and the Center for Digital Agriculture at the U. of I.