Selective Detection of Black Rot in Grapevine Leaves Using Convolutional Neural Networks

Abstract

Grapevine diseases represent a major threat to vineyard productivity, with Black Rot being among the most destructive due to its rapid spread and visual similarity to other diseases. These diseases are associated with a diversity of pathogenic agents, namely fungi, oomycetes, bacteria and pests. While prior work frequently reports high accuracy in controlled multi-class classification, practical deployments commonly require selective detection of a target disease against a contaminated negative class. In this work, Black Rot detection is formulated as a binary classification task, where the negative class includes healthy leaves and other visually similar diseases. This study employed ImageNet pretrained Convolutional Neural Network (CNN) backbones, MobileNetV2, DenseNet121, ResNet50 and VGG16, using a two-stage transfer learning protocol. The ability of the CNN models to accurately identify Black Rot cases was evaluated using standard classification metrics, namely accuracy, precision, recall and F1-score. The results show clear differences in detection behaviour across architectures. ResNet50 achieves the highest overall performance, obtaining 100.0% precision, with no false positives while maintaining a high recall 96.3% and a F1-score of 98.1%, with an accuracy of 98.9%. Overall, the achieved performance is competitive and exceeds values reported in the literature, while addressing a more realistic contaminated-negative scenario.