by Abhisesh Silwal
The precision vineyard sensing team works through the year during different growth stages of the vines. Starting early season for shoot count, leaf area measurement and fruit cluster count all the way to berry sizing and fruit quality assessment from images. In addition to this, the team has started to analyze images captured during the dormant season to automate the process of estimating vine vigor. From two different field trips to vineyards in New York (Cornell Lake Erie Research and Extension Laboratory), a total of three varieties of dormant vines (Riesling, Concord, and Vignoles) were imaged. These imaged vines were manually pruned and all canes were weighed and taken back to the lab for ground truthing. For ground truthing, nearly 28,000 buds and 80K inches of cane length were manually measured by two people. This large set of ground data would provide a great way to evaluate the performance of computer vision algorithms that detects, counts, and measures buds and cane length from images.
Counting buds in dormant vines from images is a challenging task because of the small size of buds that occupy few pixels even in high resolution images. The task of detecting buds was accomplished by detecting nodes in the canes assumed to have a single bud. A state-of-the-art Deep Neural Net was trained with images acquired during the dormant season for detecting individual buds. This Deep learning-based approach has shown promising results as the correlation coefficient between ground truth and computer count is as high as 0.87. We plan to test the system at large scale during the next dormant season.