A Generative Data Augmentation Enhanced Plant Disease Classification Combined Diffusion Model

The early detection of plant diseases is of paramount importance for the optimal functioning of agricultural production. However, the performance of models is limited by the amount of data as well as impacted by the imbalance distribution between categories. This study proposes a novel generative data augmentation strategy to enhance classifier performance in plant disease classification tasks. First, elucidated diffusion models (EDM) were proposed to synthesize high-quality synthetic images of grape leaf diseases. This approach effectively mitigates the overfitting and unbalanced problems caused by insufficient data. Secondly, the three pre-train CNN-based models, ResNet50, VGG16, and AlexNet, were employed to classify grape leaf diseases based on the enhanced dataset by combining the generative image. These models effectively capture global features using larger convolutional kernels with overlapping pooling methods and progressively refine these features to extract detailed information from the images. At last, the framework could achieve high performance in classifying healthy grape leaves and three disease types, where the average precision (PR), recall (RE), F1-score (F1), and accuracy (ACC) across four classes were black rot, black measles, healthy and isariopsis leaf spot.This study demonstrated that the generative strategy could enhance the data with a limited amount and sported effective learning in grape leaf disease classification. This novel framework has a high potential for application in the field of plant disease protection and can assist in the development of intelligent agriculture.