William & Mary

At the GRS Symposium: Drones, machine learning & better oyster reefs

  • Drone oyster data:
    Drone oyster data:  Sofya Zaytseva is part of a team trying to determine if there is an ideal shape for an oyster reef. They’re using drone-gathered images of successful reefs, combined with machine-learning techniques to analyze the images.  Courtesy photo
Photo - of -

Sofya Zaytseva is working to engineer a better oyster reef.

Oysters are important organisms to coastal economy as well as coastal environment. But Zaytseva notes that the population of this important species has dropped to less than one percent of its historic level. She is part of a team working to restore oyster habitat through a better understanding of how reefs develop.

Zaytseva is a fifth-year Ph.D. student in the Department of Applied Science at William & Mary. She is the recipient of the W&M Graduate Studies Advisory Board Award for Excellence in the Natural and Computational Sciences.

Zaytseva will speak on her work using drone-gathered data to analyze oyster reefs during the 18th Annual Graduate Research Symposium, held at the Sadler Center on the campus of William & Mary on March 15 & 16, 2019. The symposium is hosted by William & Mary's Graduate Student Association of Arts & Sciences and the Arts & Sciences Office of Graduate Studies and Research.

“Oyster restoration efforts focus on constructing artificial reefs to provide more oyster habitat and promote successful settlement and persistence of the oyster population,” she explains.

Zaytseva works with a team that includes Leah Shaw, associate professor of mathematics, and two faculty members from William & Mary’s Virginia Institute of Marine Science, Donglai Gong and Rom Lipcius. Their idea is simple: Study the characteristics of successful natural oyster reefs to understand how to create the best artificial oyster habitat.

“Besides considering water temperature, salinity, bottom hardness and food availability when constructing and placing artificial oyster reefs, the importance of initial reef shape and size should not be underestimated,” Zaytseva wrote in her entry for the GRS award, “Analysis of oyster reef patterns in remotely sensed data.”

She points out that it’s not clear if it’s better for any one reef to start off rounded or more elongated, but it seems that the best initial shape is the one that best fits the unique water conditions.

“For example, it is known that reef height matters,” Zaytseva explained. “Reefs should not be too high, where the oysters do not get enough time to feed during high tide — or too low, where they get buried by sediment.”

Zaytseva and her team used a 230-acre portion of a successful natural reef in the intertidal area near Wachapreague. They surveyed the area at low tide using a DJI Phantom 4 Pro drone at an altitude of about 400 feet, followed by lower flights for groundtruthing.

Data from the drone camera is stitched together, then Zaytseva begins processing the JPEG images, using ArcMap 10.4. Her goal is to classify the component of the reef that each pixel in the file represents.

“Let's say you have an oyster reef and a body of water in an image,” she explained. “The oyster reef is made up of many pixels of various gray tones while the water is made up of mostly blue pixels.”

There is more to an oyster reef than oysters and water — algae, rocks and sediment are part of the mix and so Zaytseva’s images are more complex than shades of gray.

It would be just about as tedious to classify an oyster reef pixel by pixel as it would be to drift over it in a boat, taking soundings and samplings. So Zaytseva uses machine-learning techniques to train a computer to parse the pixels.

“This is where image classification comes in,” she said. “There are two approaches to image classification that come from machine learning — unsupervised classification and supervised classification.”

Supervised classification begins with Zaytseva or another researcher preparing training data, manually assigning pixels from a known section of reef, carefully separating the gray (oyster) from the blue (water) and dealing with the blue-gray and the gray-blue and the countless other shades by truthing the pixels to the known features they represent in the reef.

“The computer then uses the training data to classify the rest of the pixels,” she said. “This way clearly results in better classification as you are helping to define what the classes should be. However, the drawback to this is the collection of training data, as one has to be careful and make sure to obtain good quality training samples.”

Unsupervised classification is left to the computer to sort the pixels according to algorithm-governed classes based on color. No training data is necessary, but can yield less accurate results.

Zaytseva was able to improve the accuracy of both her unsupervised and supervised data by harnessing the texture in her images.

“An image of one mostly solid color would have very low texture,” she said. “A highly speckled image would have higher texture due to the variation of the color pixels.”

Oyster reefs are usually composed of live oysters and the shells of dead oysters, along with sediment and some algae. Zaytseva explains that a portion of an image showing sediment is low-texture, as it is mostly beige. The mix of living oysters and shell makes a more “speckled” set of pixels — a higher texture.

She added texture bands to her RGB images by assigning values to each pixel in a small window based on the pixel’s deviation from the colors of all the pixels that surround it.

Zaytseva tested her results by groundtruthing her images against higher-resolution images, and found that the addition of the texture bands improved the accuracy of both supervised and unsupervised data.

She also noted that supervised classification was superior overall, with and without texture, the unsupervised-with-texture approach was best at identifying the oyster pixels. Supervised classification is dependent on the training data, and Zaytseva says it’s only the better choice when the researcher has extensive prior knowledge of the sample area.

“Unsupervised classification with texture is more efficient, does not require training samples, and may be better considering that the intertidal area is a complex landscape where identifying features on the ground visually may be difficult, therefore compromising the quality of the training data,” she reports.

Zaytseva says preliminary results indicate that the addition of texture bands to her images greatly increase her ability to distinguish some of the knottier features, such as the bleached surface of dead shells.

“Given the importance of oyster reef shape and size in successful reef development, using aerial imagery provides an efficient and cheap way to study the structure of individual reefs,” she said.