Automated yield estimation in viticulture by computer vision

Abstract

Currently, industry standard yield predictions in viticulture are generated by manual sampling of the weight and number of components such as bunches and berries. The data collection takes significant time and despite requiring trained labour still results in inaccurate yield estimates which in turn lead ton substantial inefficiencies in scheduling and logistics through the entire wine industry supply chain.

Therefore, this thesis introduces an automated visual yield estimation system using computer vision and data mining with the aim of improving crop estimation. Firstly, a biological baseline for yield estimation is proposed and implemented together with novel algorithms for shoot and bunch detection. A method for transforming the object counts extracted from videos of individual rows to a prediction of crop yield without requiring GPS is introduced, with the additional benefit of generating a production variation map during the season. In addition, a novel fully automated 3D bunch reconstruction algorithm has been developed which is based on a single input image to assist berry counting in vivo.

Shoot detection using only a low-cost camera (GoPro) was achieved using a combination of a novel unsupervised dynamic feature selection procedure and unsupervised clustering and shown to be robust to varying light conditions, multiple varieties and shoot densities. Bunch detection was achieved using an SVM classifier including the introduction of two novel features and shown to outperform existing approaches. The bunch reconstruction algorithm achieved an accuracy of 13-15% across multiple cultivars at multiple stages of maturity. The shoot detection algorithm was used to generate yield forecasts which were within 15% of the final fruit tonnage delivered to the winery, despite being calculated five months prior to harvest.

The overall automated visual yield estimation system has thus been demonstrated to be feasible and to reduce the error and cost associated with existing manual sampling procedures. Together, the contributions of this thesis bring not only economic benefits for grape growers and winemakers, but also provide valuable inputs for future work on vine-level management that will improve quality, yields and optimise resource usage, leading to better wine.