Automated Infield Grapevine Inflorescence Segmentation based on Deep Learning Models

The world wine sector is a multi-billion dollar industry with a wide range of economic activities. To be able to compete in domestic and global markets, wine producers must reduce operational and production costs, and optimise production cycles, aiming to improve crop yields and quality. Yield forecasting is of immeasurable value in modern viticulture to optimise harvest scheduling and quality management. Traditionally, this task is carried out through manual and destructive sampling of vineyard yield components and their accurate assessment is expensive, time-consuming, and error-prone. Therefore, yield estimation is based on a low number of samples, resulting in erroneous projections as the yield variation is very high and unevenly distributed across the vineyard. The number of inflorescences and flowers per vine is one of the main yield components and serves as an early predictor of vineyard productivity. The adoption of new non-invasive technologies can automate this task and drive viticulture yield forecasting to higher levels of accuracy. In this study, a digital phenotyping approach, based on Deep Learning (DL) models, is proposed for grapevine inflorescence detection and segmentation, with the aim of validating and subsequently implementing the solution for counting the number of inflorescences and flowers. Different Single Stage Instance Segmentation models from the state-of-the-art You Only Look Once (YOLO) family, such as YOLOv5, YOLOv7, YOLOv8 and YOLOACT, will be benchmarked on a dataset of RGB images, collected under field conditions by a mobile device in a vineyard.