Accurate grape yield estimation is essential for effective vineyard management, crop planning, and resource allocation. Traditional methods often involve time-consuming and labour-intensive processes, which may introduce errors due to the large size and inherent spatial variability of the vineyard blocks. In this sense, artificial intelligence (AI) can contribute to increasing the number of samples by using images or videos which can be proposed by trained AI models reducing the uncertainties of the yield estimation, especially those associated with the number of bunches per vine, being one of the most import yield components. This study explores the use AI as a novel approach to improve yield estimation for both wine and table grapes in the context of the standard viticulture practices of South Africa. Four field experiments were conducted to capture RGB images of selected vines of Cabernet Sauvignon and Chenin Blanc (wine grape cultivars) and Crimson Seedless and Sugar Crisp™ (table grape cultivars). Different protocols of image acquisition considering the particularities of the trellising systems and multiple phenological periods were analysed. RGB images captured under field conditions were used to train the AI models based on YOLO architecture for bunch detection. Statistical indicators were used to evaluate the performance of the calibrated models by comparing the results with images labelled by human experts. The preliminary results obtained with this methodology show errors of the AI models varied from the different conditions analysed, presenting error rates ranging from 0.8 to 1.8 bunches per vine. These results demonstrate the potential of AI to enhance vineyard yield estimation. However, further studies and technical adaptations are needed to scale the methodology to a vineyard block level and advance its practical application.
[0000-0001-8025-5879] et al. (Wed,) studied this question.