The biologist, Lance Cadle-Davidson, Ph.D. ’03, an adjunct professor in the School of Integrative Plant Science (SIPS), is working to develop grape varieties that are more resistant to powdery mildew, but his lab’s research was bottlenecked by the need to manually assess thousands of grape leaf samples for evidence of infection.
Powdery mildew, a fungus that attacks many plants including wine and table grapes, leaves sickly white spores across leaves and fruit and costs grape growers worldwide billions of dollars annually in lost fruit and fungicide costs.
Cadle-Davidson is also a research plant pathologist with the U.S. Department of Agriculture’s Agricultural Research Service (USDA-ARS). He works in the Grape Genetics Research Unit in Geneva, New York, and his team developed prototypes of imaging robots that could scan grape leaf samples automatically – a process called high-throughput phenotyping – through the USDA-ARS funded VitisGen2 grape breeding project and in partnership with the Light and Health Research Center. This partnership led to the creation of a robotic camera they named “BlackBird.”
But extracting relevant biological information from these images was still a critical need.
Enter the engineer and computer scientist: Yu Jiang, an assistant research professor in SIPS’ Horticulture Section at Cornell AgriTech. Jiang’s research focuses on systems engineering, data analytics and artificial intelligence. The BlackBird robot can gather information at a scale of 1.2 micrometers per pixel – equivalent to a regular optical microscope. For each 1-centimeter leaf sample being examined, the robot provides 8,000 by 5,000 pixels of information.
Extracting useful information from such a large, high-resolution image was Jiang’s challenge, and his team used AI to solve it. Using breakthroughs in deep neural networks developed for computer vision tasks like face recognition, Jiang applied this knowledge to the analysis of microscopic images of grape leaves. In addition, Jiang and his team implemented the visualization of the network inferential processes, which help biologists better understand the analysis process and build confidence with AI models.
