Low-Cost, Computer Vision-Based, Prebloom Cluster Count Prediction in Vineyards

Abstract

Traditional methods for estimating the number of grape clusters in a vineyard generally involve manually counting the number of clusters per vine in a subset of the vineyard and scaling by the total number of vines; a technique that can be laborious, costly, and with an accuracy that depends on the size of the sample. We demonstrate that traditional cluster counting has a high variance in yield estimate accuracy and is highly sensitive to the particular counter and choice of the subset of counted vines. We propose a simple computer vision-based method for improving the reliability of these yield estimates using cheap and easily accessible hardware for growers. This method detects, tracks, and counts clusters and shoots in videos collected using a smartphone camera that is driven or walked through the vineyard at night. With a random selection of calibration data, this method achieved an average cluster count error of 4.9% across two growing seasons and two cultivars by detecting and counting clusters. Traditional methods yielded an average cluster count error of 7.9% across the same dataset. Moreover, the proposed method yielded a maximum error of 12.6% while the traditional method yielded a maximum error of 23.5%. The proposed method can be deployed before flowering, while the canopy is sparse, which improves maximum visibility of clusters and shoots, generalizability across different cultivars and growing seasons, and earlier yield estimates compared to prior work in the area.