Using AI to Spot Wine Grape Disease from the Air

NASA’s JPL and Cornell University team up to apply spectroscopy and machine learning to detect leafroll.

In August, researchers from NASA’s Jet Propulsion Laboratory and Cornell University shared that NASA’s Airborne Visible/Infrared Imaging Spectrometer – Next Generation (AVIRIS-NG) can detect invisible symptoms of a harmful virus that affects wine grape vines. When machine learning models analyze images of the vines, they are able to determine the extent of infection before further loss occurs to the crop. Having this data will empower vineyard managers to remove infected vines early. This could put a dent in the $3 billion in losses that the U.S. wine industry suffers every year from plant disease.

AVIRIS-NG accomplishes this task by using its optical sensor to find deficiencies in how grape vines are performing. AVIRIS-NG can measure the wavelength range from 380 nm to 2510 nm with 5 nm sampling. The virus, grapevine leafroll-associated virus complex 3 (GLRaV-3), may cause no visible symptoms in infected vines for up to a year. Yet the virus will change how the sun’s rays interact with a vine’s leaves and canopy. To make the task harder, the spectroscopic signals for red grape varieties like Cabernet Sauvignon differ from those for white grape varieties like Chardonnay.

The area where AVIRIS-NG collected data when flying over vineyards (left) and  a zoomed-in view of the flight lines that collected spectroscopic imagery over vineyards (right).  Image source: JGR Biosciences.

The area where AVIRIS-NG collected data when flying over vineyards (left) and a zoomed-in view of the flight lines that collected spectroscopic imagery over vineyards (right).
 Image source: JGR Biosciences.

After AVIRIS-NG captures images, it processes them on an edge computer on the farm and finishes on the public Azure cloud. Then researchers use radio frequency models to isolate the vines from other objects in the images. Next, they use a series of executable Python scripts to train and apply models to find the vine’s deficiencies.

Early detection through imaging spectroscopy is much cheaper than molecular testing, the current method of asymptomatic testing, Molecular testing costs between $40 and $300 for each vine. A small vineyard usually has 1,000 vines, while a large vineyard can have over 30,000 vines. Vines can be asymptomatic for roughly a year. During that time, they can infect nearby vines. In addition, when vines start to show visible signs of infection, only red grape varieties like Cabernet Sauvignon may display one of the tell-tale markers: red blotches on leaves. White grape varieties like Chardonnay tend not to display these blotches.

The researchers collected spectrometer images for this study by flying AVIRIS-NG, contained in the belly of a Twin Otter research plane, over approximately 11,000 acres of Cabernet Sauvignon vineyards in Lodi, Calif. in 2020. That year and the next, vineyard managers scouted the imaged vines for visible symptoms of GLRaV-3. They also collected a subset of leaves for molecular testing. The best-performing machine learning models could distinguish between noninfected and asymptomatic vines with 87 percent accuracy.

Using AVIRIS-NG to detect GLRaV-3 does not guarantee a perfect dataset. Co-occurring biotic stresses, such as other pathogens like trunk disease, and abiotic stresses, like drought, can confuse the detection of GLRaV-3. It is not uncommon for a vine to experience multiple stresses at once. Research efforts need to further examine how abiotic stressors lead to detection errors.

This project on GLRaV-3 is a case study that indicates how NASA’s new technologies can support ground-based pathogen surveillance. The second phase of NASA’s Surface Biology and Geology (SBG) study will involve a number of missions to collect data with JPL’s hyperspectral satellites. Researchers could combine this information with machine learning to identify other crop diseases in different regions.