Grading for grapevine downy mildew and feature extraction methods for predicting abaxial lesions from adaxial leaf images.

Abstract

Grading grapevine downy mildew severity is essential for the precise application of pesticides. Since typical symptoms appear on the abaxial (underside) surface of grape leaves, and lesion area proportion determines severity, it is necessary to analyze lesion characteristics and develop adaxial-to-abaxial lesion inversion methods to build lightweight yet accurate grading models. This study proposes a comprehensive disease grading framework for grape downy mildew. First, a convolutional neural network (CNN)-based classification model is developed with specialized modules and coordinate attention to enhance feature extraction and semantic richness for improved lesion identification. Second, a novel K-Means++-CNN-Vote Consolidation lesion extraction method is introduced. In this framework, K-Means++ segments leaf sub-images, CNNs classify lesion types, and a voting mechanism consolidates results-addressing challenges posed by irregular lesion shapes and blurred boundaries. Finally, an abaxial lesion inversion framework is established by constructing a morphological feature mapping between the adaxial and abaxial surfaces, utilizing mapping functions and lesion generation techniques to infer the abaxial lesion distribution from the adaxial images. Experimental results showed disease grading accuracies of 82.16% (combined adaxial and abaxial), 79.74% (adaxial only), and 84.59% (abaxial only), with a model size of 5.08 MB. Lesion segmentation accuracies reached 89.29% (adaxial and abaxial), 76.92% (adaxial), and 64.47% (abaxial), while the adaxial-to-abaxial lesion inversion achieved an 80% similarity. This study provides methodological support for the online grading of grapevine downy mildew and offers a scientific basis for precise disease control.

Plant phenotyping relevance

ブドウ葉の病斑面積・分布という植物病害状態を画像から抽出・推定する手法を開発し、病害重症度の評価精度も検証しており、フェノタイピング手法が中心である。

Code and data availability

The supplied blocks describe the authors' GDCData/GDSData datasets and GDCNet/K-CNN-VC methods, but contain no public deposit, repository, or availability URL for the images, datasets, or code. The only URL present is the article DOI itself, which is not a paper-specific asset.

No evidence-backed public reproduction asset is currently recorded.