Deep learning modelling for non-invasive grape bunch detection under diverse occlusion conditions

Abstract

• Development of four models using YOLOv4 for accurate detection of grape bunches under varied occlusion. • Training and testing were conducted on over 1700 images from 11 distinct vineyards. • The ComprehensiveModel achieved high accuracy with R 2 values of 0.83 in validation and 0.81 in the testing phases. • The most effective models necessitated images depicting various occlusion conditions. Accurately and automatically estimating vineyard yield is a significant challenge. This study focuses on grape bunch counting in commercial vineyards using advanced deep learning techniques and object detection algorithms. The aim is to overcome the limitations of conventional yield estimation techniques, which are labour intensive, costly, and often inaccurate due to the spatial and temporal variability of the vineyard. This research proposes a non-invasive methodology for identifying grape bunches under different occlusion conditions using RGB cameras and deep learning models. The methodology is based on the collection of RGB images captured under field conditions, coupled with the implementation of the YOLOv4 architecture for data processing and analysis. Statistical indicators were used to evaluate the performance of the developed models. The comprehensive model produced a favourable outcome during validation, with an error rate of 1.12 bunches (R 2 = 0.83). In the test dataset, the model achieved an error rate of 1.12 (R 2 = 0.81). The results highlight the potential of emerging technologies to significantly improve vineyard yield estimation. This approach has the potential to assist vineyard management practices, enabling more informed and efficient decisions that could increase both the quantity and quality of grape production intended for winemaking.