Application of Hyperspectral Imaging and Generative Adversarial Network for Powdery Mildew Severity Detection on...

Powdery mildew in melon reduces fruit quality and yield, and may cause plant death. Accurate severity diagnosis supports disease control, resistance evaluation, and precision management. This study applied hyperspectral imaging to assess disease severity in two cultivars (Zhetian 105 and Zhetian 501) under five stress levels. To explore the feasibility of generative models for spectra generation for disease severity classification, three generative models (Conditional Generative Adversarial Network (CGAN), Deep Convolutional Generative Adversarial Network (DCGAN), and Conditional Deep Convolutional Generative Adversarial Network (CDCGAN)) were used to generate different numbers of samples for each stress level. To evaluate the generated data quality, classification models (logistic regression (LR), support vector classification (SVC), eXtreme Gradient Boosting (XGBoost), and Convolutional neural network (CNN)) models were built under three conditions: using real training samples to predict real test samples; using generated samples to predict the real test samples; and using real training samples with generated samples to predict real test samples. The results showed that DCGAN and CDCGAN can generate the samples close to the real training samples. Classification models using generated samples, as well as the combination of real training samples and generated samples, showed increase in the classification performance on the real test samples. The different numbers of generated samples did not show specific patterns on the prediction performance. Some models using generated samples and real training samples with generated samples could improve the prediction accuracy up to 10% compared with the models using real training samples. Class-wise analysis indicated that not all prediction performance of different classes increased. The overall results indicated that generative adversarial networks have the potential to generate reflectance spectra of different melon cultivars under varying disease severities. More real samples and generation strategies are needed to better improve the generated data quality.