Automated detection of downy mildew in vineyards using explainable deep learning

Abstract

Traditional methods for identifying downy mildew in commercial vineyards are often labour-intensive, subjective, and time-consuming. Artificial intelligence offers an opportunity to streamline and standardize this process, increasing sampling efficiency and consistency. This study presents an interpretable, automated approach to detect downy mildew symptoms in vineyard conditions. RGB images of grapevine canopies were collected across multiple commercial vineyards using a ground-based mobile phenotyping platform. Image analysis was conducted using a sliding window technique to classify sub-images into zones with and without symptoms. Techniques such as transfer learning, fine-tuning, and data augmentation were used to automate classification, comparing the performance of convolutional neural networks (CNNs) and vision transformers (ViTs). The trained model, integrated into the sliding window system, accurately localized symptomatic areas, with predictions interpreted through explainable artificial intelligence (XAI). The EfficientNetV2S model achieved a classification accuracy of 91% and an F1-score of 0.92 in localizing symptomatic zones. This approach enables reliable and interpretable downy mildew detection under diverse field conditions, representing a significant advancement in precision vineyard protection.