Improvement of plant disease classification accuracy with generative model-synthesized training datasets

Abstract

Digitalization in agriculture requires critical research into applications of artificial intelligence to various specialization domains. This work aimed at investigating the application of image synthesis technology to the mitigation of the data volume constraint to digital plant disease phenotyping accuracy. We designed an experiment involving the use of a deep convolutional generative adversarial network (DC-GAN) to synthesize photorealistic data for healthy and bacterial spot disease-infected tomato leaves. The training dataset contained 1,272 instances per class. We further employed a 3-block visual geometry group (VGG) convolutional neural network (CNN) model with dropout regularization and 1 epoch to compare classification accuracies of the original dataset and various synthetic datasets. Our results showed that the third DC-GAN synthesized training dataset containing 3,816 synthetic examples of both healthy and bacterial spot infected tomato leaf classes outperformed the original training dataset containing 1,272 real examples of both tomato leaf classes (77.088% accuracy with the former dataset on a 3-block VGG CNN model with dropout regularization and 1 epoch, as compared to 76.447% accuracy with the latter dataset on the same classifier).

Plant phenotyping relevance

トマト葉の健全・細菌斑点病状態を画像から分類するため、DC-GANによる合成データ生成とCNN分類精度を比較検証しており、植物病害フェノタイピング手法が中心である。

Code and data availability

The authors explicitly deposited their analysis scripts and research dataset link in a public GitHub repository, and the paper's phenotyping input images (tomato healthy/bacterial spot classes) come from the public PlantVillage dataset. Both are paper-specific, public, and actionable.