Model comparison and parametrization strategies for accurate RGB‑Based grapevine phenology classification

Abstract

Accurate monitoring of grapevine phenology is essential for vineyard management and for understanding seasonal responses to climate variability (Reis et al., 2020). However, phenological assessment in commercial vineyards still relies largely on manual field observations, which are time-consuming, subjective, and difficult to scale across large and heterogeneous blocks. In this sense, automatic classification of grapevine phenology from RGB imagery has shown strong potential (Íñiguez et al., 2025); however, reported accuracies depend strongly on model architecture and parametrization choices that are often underexplored. This study presents a systematic comparison of deep learning models and training configurations for grapevine phenological stage classification under realistic vineyard conditions in Spain and South Africa. Vineyard canopy images were acquired in two sites located in different hemispheres (Spain and South Africa), covering contrasting cultivars, training systems, and growing conditions. A dataset of 4,381 images was organised into four major phenological classes: shoot and inflorescence development (E-L 12–18), flowering (E-L 19–26), berry formation (E-L 27–33), and berry ripening (E-L 35–38). Multiple convolutional neural network architectures were evaluated using identical datasets, with particular attention to input resolution, network complexity, and transfer-learning strategies. Deep learning models were trained to classify canopy images into these broad stages, enabling objective comparison of model performance across configurations. Results demonstrate that classification accuracy is highly sensitive to parametrization decisions, with clear trade-offs between model complexity, robustness, and cross-site generalisation. Optimized configurations achieved substantial accuracy improvements relative to baseline models, highlighting the importance of methodological tuning over architectural novelty. Among the tested architectures, YOLO-Classification provided the best balance between classification performance and computational efficiency. These findings provide practical methodological guidance for the design of accurate and reliable image-based phenology classifiers and support more informed use of deep learning models under operational vineyard conditions, in line with broader advances in image-based monitoring of grapevine traits under field conditions (Íñiguez et al., 2024). The practical value of this approach lies in providing rapid, repeatable, and scalable information to support viticultural decision-making, including irrigation scheduling, canopy management, phytosanitary timing, and planning of harvest-related operations. Beyond management applications, this methodology also provides a foundation for analyzing phenological shifts associated with climate variability and climate change through higher-frequency observations over large vineyard areas (Reis et al., 2020).

References

Íñiguez, R., Gutiérrez, S., Poblete-Echeverría, C., Hernández, I., Barrio, I., & Tardáguila, J. (2024). Deep learning modelling for non-invasive grape bunch detection under diverse occlusion conditions. Computers and Electronics in Agriculture, 226, 109421. https://doi.org/10.1016/j.compag.2024.109421

Íñiguez, R.; Wolela, F.; Gonzalez Pavez, M.I.; Barrio, I.; Tardáguila, J.; Venter, T.; Poblete-Echeverría, C. Artificial intelligence-driven classification method of grapevine major phenological stages using conventional RGB imaging. OENO One 2025, 59(2), 9306. https://doi.org/10.20870/oeno-one.2025.59.2.9306

Reis, S.; Fraga, H.; Carlos, C.; Silvestre, J.; Eiras-Dias, J.; Rodrigues, P.; Santos, J.A. Grapevine phenology in four Portuguese wine regions: modeling and predictions. Applied Sciences 2020, 10(11), 3708. https://doi.org/10.3390/app10113708