1Running Tide Technologies, Portland, ME, United
States
ORCiD: 0000-0003-0176-1859
Keywords: Computer vision, 3D reconstruction, Structure from
motion, 3D point cloud, Object detection, Plant phenotyping, Automated
harvesting, Carbon sequestering
Computer-vision based phenotyping approaches have shown to be successful
in terrestrial based agricultural crops, however little attention has
been paid towards computer-vision based phenotyping in marine based
aquaculture. Here, we investigate 2D and 3D computer vision based
methods for plant phenotyping in Saccharina latissima (sugar
kelp). For our 2D approach, we developed a farm-level (20ft row),
out-of-water photobooth camera system that captures images of harvested
kelp. We then trained a Faster-RCNN model to automatically detect kelp
blades. This model output fed into our image processing algorithm to
segment each blade and compute phenotyping traits such as: blade area,
width, height, perimeter, blade count. Additionally, this algorithm
detects blade overlay to identify both blade area and count for blades
overlapping in the image. For our 3D approach, we leveraged a revolving
camera system and open-source structure-from-motion software to
construct 3D representations of individual kelp blades. Using point
cloud processing techniques, we developed an algorithm to compute volume
(and 2D measurements as listed above). Correlations of this preliminary
dataset showed an R^2=0.985, p=2e-16 when comparing computer vision
to hand-measured blade length measurements as well as an R^2=0.87,
p=2.32e-15 when comparing total area calculated through computer vision
against total weight measured by hand. These preliminary results convey
the effectiveness of computer vision based phenotyping and the need to
increase research efforts in this space.