Automatic vineyard health control: A review of existing systems and datasets for disease identification

In recent years, computer vision and machine learning technologies have enabled new approaches to automatic vineyard health monitoring. This review synthesizes findings from 30 highly relevant studies covering automatic vineyard disease identification systems, examining system architectures ranging from ground-based robots and unmanned aerial vehicles (UAVs) to IoT-integrated platforms, mobile applications, and embedded devices. We analyse the computer vision and machine learning methods employed, including convolutional neural networks for classification, semantic segmentation, object detection, and multimodal sensor fusion techniques. We identify and describe publicly available and custom datasets used for model training and validation, highlighting important differences between controlled laboratory datasets and field-acquired imagery that affect model generalizability. The restricted generalizability is documented by training an EfficientNet-B2 classifier on controlled-background PlantVillage images and testing it on an independent dataset of field-acquired images. Although the model achieved 99.61 ± 0.17% accuracy on the internal test set, its accuracy decreased to 66.80 ± 8.38% and its mean F1-score to 58.33 ± 5.90% on the external field dataset, demonstrating a substantial domain shift between laboratory and vineyard conditions. Our analysis reveals that, while many systems achieve classification accuracy above 90%, significant challenges remain in the early detection of pre-symptomatic infections, robust multi-disease differentiation, and practical deployment across diverse real-world vineyard conditions. We discuss current limitations and outline priority directions for future research, including the need for standardized benchmarks, that require large annotated and balanced datasets, integration of emerging deep learning architectures such as Vision Transformers and foundation models, and development of low-cost, accessible solutions for widespread adoption.