Early-season Vineyard Shoot and Leaf Estimation Using Computer Vision Techniques

This paper describes computer vision techniques for early-season measurement of vine canopy parameters; leaf count, leaf area and shoot count. Accurate and highresolution estimation of these key vineyard performance components are important for effective precision management. We use a high-resolution stereo camera with strobe lighting mounted on a ground-vehicle that captures high-quality proximal images of the vines. For shoot image segmentation, we apply the Frangi vessel filter (originally developed for medical imaging processing) in conjunction with custom filtering to extract shoot counts. We also present an incremental leaf count estimation algorithm, that proposes leaf candidates for incremental leaf sizes and then removes the repeating candidates to accurately assess leaf count. The specified algorithms are robust to partial occlusion and varying lighting conditions. For shoot count measurement we observe an F1 score of 0.85 for image shoot count and R correlation of 0.88 for ground-truth shoot counts. The R correlation for leaf count estimation between ground truth sample images and algorithm output is 0.798. Whereas the R correlation between the data collected by a PAR sensor and leaf area estimation algorithm is 0.69.