Photo credit: www.sciencedaily.com
Researchers have created a machine-learning tool capable of autonomously learning to differentiate between aerial images of flowering and nonflowering grasses. This advancement promises to significantly accelerate agricultural field research. The study utilized images from numerous varieties of Miscanthus grasses, each characterized by unique flowering traits and timing.
Identifying crop traits in diverse conditions throughout different stages of growth is quite challenging, as noted by Andrew Leakey, a plant biology and crop sciences professor at the University of Illinois Urbana-Champaign. He collaborated with Sebastian Varela, a scientist at the Center for Advanced Bioenergy and Bioproducts Innovation, on this project.
According to Leakey, the new methodology has potential applications across a range of crops and computer-vision challenges.
The research findings are documented in the journal Plant Physiology.
“Timing of flowering plays a crucial role in crop productivity and adaptation to varying growing conditions,” Leakey explained. “Traditional visual inspections of numerous individual plants in extensive field trials require significant labor.” Automating this process through aerial drone imagery combined with artificial intelligence can facilitate the extraction of essential data, thus making the task more efficient. However, creating AI models capable of detecting subtle distinctions in complex images typically relies on extensive human-annotated datasets, a process that is notably time-consuming. “Deep-learning techniques are often context-sensitive,” he added.
This means that any changes in context, such as the need to distinguish between different crops or the same crop in varying locations or seasons, may necessitate retraining with fresh annotated datasets.
“There are numerous instances where AI has been proposed to enhance the use of sensor technologies – from leaf sensors to satellites – in applications related to breeding and crop sciences. Yet, the adoption rate remains lower than expected. We believe one significant hurdle is the substantial effort required to train AI tools,” stated Leakey.
To minimize the dependence on human-annotated data, Varela employed a popular technique involving two competing AI models, known as a “generative adversarial network” (GAN). In this setup, one model generates synthetic images of a particular scene, while a second model evaluates these images to identify which are authentic and which are fabricated. Over time, both models enhance each other’s capabilities, with the first generating more realistic images and the second improving its discerning abilities.
Varela theorized that he could leverage this self-improvement mechanism to reduce the quantity of annotated images necessary for training models that differentiate various crops, leading to the creation of an “efficiently supervised generative and adversarial network,” or ESGAN.
In a series of trials, the researchers assessed the accuracy of the ESGAN against existing AI training methods and discovered that ESGAN significantly decreased the need for human-annotated data, achieving reductions by one to two orders of magnitude compared to “traditional, fully supervised learning approaches.”
This breakthrough signifies a substantial decrease in the effort required to develop custom-trained machine-learning models for determining flowering times, even when involving varying locations, breeding populations, or species. The approach represents a pathway to addressing similar challenges across other fields in biology and digital agriculture.
Leakey and Varela plan to collaborate further with Miscanthus breeder Erik Sacks to implement their new method in a multistate breeding trial aimed at developing regionally suited Miscanthus lines. These could serve as feedstock for biofuels and valuable bioproducts on currently unprofitable agricultural land.
“We aspire for our innovative approach to facilitate the integration of AI technologies in crop improvement efforts across diverse traits and species, ultimately contributing to the enhancement of the bioeconomy,” Leakey remarked.
Leakey holds professorships in the Carl R. Woese Institute for Genomic Biology, the Institute for Sustainability, Energy and Environment, and the Center for Digital Agriculture at the University of Illinois.
This research received support from 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.
Source
www.sciencedaily.com