Application of visible/near-infrared hyperspectral imaging with convolutional neural networks to phenotype aboveground parts to detect cabbage Plasmodiophora brassicae (clubroot)

Cabbage often suffers from various diseases as one of the most favorable vegetables in China. Among all the diseases, clubroot is one of the most severe diseases. Due to the ability of hyperspectral technology to reveal hidden changes in plant physiology, it can be used to assess the often nonspecific damage caused by root pathogens. This study explored the feasibility of detecting clubroot by using hyperspectral imaging (HSI) combined with convolutional neural network (CNN). A highly susceptible cabbage variety of clubroot disease, ECD-05 was used in this study. Infected samples were inoculated by the injection vaccination method. To explore the influence of soil types on clubroot detection, three types of soil, including nutrient soil, mixed soil and red clay were studied. Hyperspectral images were acquired six weeks after inoculation of plant with the disease. Qualitative analysis by score images formed by principal component analysis (PCA) indicated differences existed between healthy and infected samples. Convolutional neural networks (CNN) and support vector classification (SVC) models using the object-wise spectra of different treatment of cabbage were built. Analysis of variance (ANOVA) was conducted on each wavelength of healthy and infected samples to explore the spectral regions related to the samples' healthy status. Vegetation index-anthocyanin reflectance index (ARI) and physiological reflex index (PhRI) successfully distinguished infected samples from samples. CNN model obtained good performance on predicting object-wise spectra with the classification accuracy of the training set and the test set over 85%. This study demonstrates the feasibility of cabbage clubroot detection by phenotyping of aboveground parts using near-infrared HSI combined with CNN.