VitisGen2 scientists train neural networks to identify Powdery Mildew

By Anna Underhill

Leaf disk and powdery mildew.

A grape leaf disk, showing powdery mildew infection in lower left.

‘Practice makes perfect’ is undeniably cliché, but its truth is evident to everyone who has ever tried to learn a new skill.  Whether it’s hitting baseballs, pushing piano keys, or pipetting off supernatant, this inevitable cycle of trial and error eventually yields to increasing expertise.

What’s true for humans is also true for computer systems.  Machine learning –teaching a computer to generate an algorithm to recognize and classify items – begins with a training set made up of the information the system is supposed to classify.  For example, training a neural network to recognize photos of people could involve providing it with images of faces along with other non-facial images and telling it which is which.

phenotyping robot with a tray of grape leaf disks.

“Blackbird”, the phenotyping robot, imaging grape leaf disks.

For machines, just like for us, in order for practice to really make perfect (or, at least, make better), you need a good teacher.  Learning things incorrectly the first time often means a difficult adjustment later, unlearning the bad habits developed early. By telling the network what to look for in a picture, new, unknown images can be properly sorted.  But if this first training set is mislabeled or doesn’t adequately represent images the system will be given, the resulting classifications won’t be very useful.  As the classic computer science saying goes—garbage in, garbage out.

VitisGen2 team members and the Rensselaer Polytechnic Institute have been collaborating to implement machine learning technology to automatically evaluate photos of leaves for the presence of the powdery mildew (Erysiphe necator) pathogen.  By training a neural network to determine whether or not images show infections – and, if they do, their level of severity – a large amount of time can be saved in the lab.

Leaf disk showing difference between hyphae and leaf trichomes.

Leaf disk sample imaged 3 days after inoculation with Powdery Mildew. Inset at right shows closeup of fungal hyphae and leaf hairs. From Bierman et al. (2019) Plant Phenomics, in press.

Instead of days hunched over a microscope counting hyphal transects, lab members are using robots and image analysis to work more efficiently and accurately.  Experiments have shown that this method is just as good at identifying powdery mildew as a human is.  To be exact, in our studies they agreed up to 91.7% of the time (learn more by watching the VitisGen2 webinar: Advanced computer vision techniques).

This novel method has helped identify new areas of the genome involved in powdery mildew resistance, and will aid in efforts to breed grapes that are robustly resistant to the disease.  Ultimately, plants that are better natural pathogen fighters need fewer fungicide sprays, lessening the impact on the environment and on a grower’s pocketbook.

Future work will include training the neural network to differentiate between fungal structures: for example, how can we teach it the difference between hyphae and spores?  Can we teach it to quantify other leaf diseases, like downy mildew?  Our hopes are that practice – or, in this case, training – really does make perfect, even for a machine.