Growing high-demand ornamentals and specialty crops is a time and labor-intensive profession requiring constant monitoring of inventory and health, while also handling routine maintenance. A Mississippi State scientist is looking to the sky for solutions to these ground-level tasks.
Patricia Knight, director of coastal horticulture research at MSU’s Coastal Research and Extension Center and scientist in the Mississippi Agricultural and Forestry Experiment Station, is exploring how drone technology and artificial intelligence, or AI, might help growers save time and money on labor-intensive tasks. She has partnered to test cost-effective, industry-ready solutions with Siva Kumpatla, research leader for the USDA Agricultural Research Service’s Thad Cochran Southern Horticultural Laboratory, and Prabha Sundaravadivel, associate professor of electrical engineering at the University of Texas-Tyler.
The three-part study’s first phase is training a computer vision model to conduct plant inventories.
“It’s important to have an accurate inventory, but during the active shipping season, that gets pushed to the side,” said Knight, also a research professor in MSU’s Department of Plant and Soil Sciences. “Sending out a drone to count while your staff is busy with more immediate tasks could help you avoid selling plants you don’t have.”
The team has visited three South Mississippi nurseries, and the UT group has shot footage of different shrubs and trees including magnolias, azaleas, roses and more. The images are being used to train an AI model to identify every possible variation of the plants, beginning with magnolias.
“In this method, called manual labeling, we label the images for features such as ‘plant’ and ‘soil.’ After being fed enough images, the system learns to quickly distinguish between plant and soil and then between individual plants and counts them,” Kumpatla said. “We have repeated this process over and over, always with a manual check until we got to a high level of accuracy.”
To date, the model has achieved over 96% accuracy in identifying magnolias, and the team is working on transferring and tweaking the models to achieve similar accuracy in other species.
Source : msstate.edu