Â鶹ÒùÔº

Hyperspectral sensor pushes weed science a wave further

By combining artificial intelligence and sensors that can see beyond visible light, Arkansas researchers have developed a system that exceeds human discernment when it comes to measuring herbicide-induced stress in plants.

Scientists with the Arkansas Agricultural Experiment Station, the research arm of the University of Arkansas System Division of Agriculture, a study in Smart Agricultural Technology providing proof-of-concept that hyperspectral sensors like a spectroradiometer can help in quantifying herbicide effectiveness, a critical element of weed management that helps curb herbicide resistance.

While normal cameras use three visible light bands—red, green and blue—to create images in the spectral range of 380 to 750 nanometers, hyperspectral sensing captures bands ranging from 250 nanometers to 2,500 nanometers and thermal infrared.

The researchers used this technology to evaluate how common lambsquarters responded to glyphosate. They also turned up empirical evidence that photosynthesis in the plant actually increased when exposed to a sub-lethal dose of the herbicide. Common lambsquarters—Chenopodium album L.—is a weed in agricultural and garden settings.

"Plant response to herbicide application is measured using visual ratings, but accuracy varies with the quality of training and years of practice of the rater," said principal investigator of the study Aurelie Poncet, assistant professor of precision agriculture in the crop, soil and environmental sciences department for the Division of Agriculture and the Dale Bumpers College of Agricultural, Food and Life Sciences.

"We thought, if we could have a sensor that automates some of this decision, we might be able to implement it into applications down the road."

Weed scientists are trained to rate herbicide efficacy within a 10% margin of error, plus or minus 5%. The researchers were able to use machine learning models on data collected with a spectroradiometer to reach a margin of error of 12.1%. Their goal is to get below 10%.

The researchers used a random forest machine learning algorithm to analyze thousands of vegetation index data points collected in the experiment. The algorithm combines the output of multiple decision trees to reach a single result.

"Our success using random forest to describe common lambsquarters response to glyphosate application opens the possibility of moving beyond the development of vegetation indices, another approach gaining traction in the published literature," said Mario Soto, lead author of the study and a crop, soil and environmental sciences master's student in Bumpers College.

Next steps

Once refined, hyperspectral sensing could be used to measure specific weed response to herbicide application and overcome the limitations of a human's visual assessment. Further development of the method and validation may also be used to create a platform for high-throughput categorization of weed response to herbicides and screening for herbicide resistance, the study's authors noted.

While training can overcome lack of experience for evaluators, mental and physical fatigue from long workdays evaluating treatments in harsh environmental conditions can affect judgment for even the most experienced evaluator, said Nilda Roma-Burgos, professor of weed physiology and molecular biology for the experiment station and Bumpers College.

"This method, in principle, could remove the human factor in herbicide efficacy evaluations and will be an invaluable research tool for weed science," said Burgos, a co-author of the study. "Meanwhile, much work still awaits to validate the method across key weed species, herbicide modes of action, time after herbicide application and environmental conditions."

Co-authors of the study included Kristofor Brye, University Professor of applied soil physics and pedology; Wesley France, program associate, and Juan C. Velasquez, weed science graduate research assistant, of the crop, soil and environmental sciences department.