Comparative analysis of single-view and multiple-view data collection strategies for detecting partially-occluded grape bunches: Field trials

Summary

Extracting phenotypic traits of grape bunch is crucial for accurately monitoring grape quality, health, and yield estimation. This is important for optimising resources, enhancing marketing strategies, and boosting overall agricultural productivity. While most research concentrates on data processing algorithms, this study focused on the preceding step: collecting reliable data. Object detection and tracking enable precise monitoring and quantification of fruit, facilitating agricultural management. This study compares two data acquisition methodologies for grape bunch detection and tracking in a commercial vineyard where leaf removal was not performed: a traditional single-view approach and a multiple-viewing method designed to mitigate fruit occlusion issues. The PointTrack algorithm, trained and validated using MOTS annotations, was employed to evaluate detection and tracking performance through metrics of three trials. The multiple-view method achieved i) higher ratio between tracked and GT detections of 74 % compared to 23 % for the single-view approach and ii) enhanced tracking metrics, with the multiple viewing trials metrics ranging from −1.35 to 3.84 for MOTSA (Multiple Object Tracking and Segmentation Accuracy) and sMOTSA (soft MOTSA), and iii) higher correlation and lower RMSE of grape bunch phenotypic traits (OIV codes 202 and 203) compared to ground truth measurements (R2 = 0.53, RMSE = 19.13). Nonetheless, the multi-view technique was compromised by motion blur due to UAV movements, complicating the tracking process. This study underscores the importance of strategic data acquisition in improving performance for fruit detection and tracking. Future work should extend this methodology to other fruit varieties and environments to validate its broader applicability, enhancing the reliability of yield estimation in precision agriculture