Classifying early-stage soybean fungal diseases on hyperspectral images using convolutional neural networks

Using convolutional neural networks (CNNs) to detect plant diseases has proven to reach high accuracy in the classification of infected and non-infected plant images. However, most of the existing researches are based on RGB images due to the availability and the comparably low cost of image collection. The limited spectral information restricts the detectability of plant diseases, especially in the early stage where often symptoms of pathogen infection have not yet become visible. To this end, in this study, hyperspectral imaging (HSI) data are combined with deep learning models to test the classification ability of two soybean fungal diseases: Asian soybean rust (Phakopsora pachyhizi) and soybean stem rust (Sclerotinia scleroriorum). Different CNNs employing 2D, 3D convolution, and hybrid approaches are compared. The influences of the depth of the convolutional layer and the regularization techniques are also discussed. Besides, image augmentation methods are investigated to overcome the problem of data scarcity. The results indicate the 6-convolutional-layer depth hybrid model to have the best capacity in classifying Asian soybean rust in the early-mid to mid-late stage when there are over 2 % visible symptoms but a limited detectability in the early stages when there are below 2 % visible symptoms on leaves. On the other hand, the optimized CNN model shows a limited capability to detect both diseases when there are no visible symptoms observable. Overall, this study suggests a hybrid 2D-3D convolutional model with augmentation and regularization methods has a high potential in the early detection of fungal diseases. This research is expected to contribute to a new cropping system that vastly reduces the chemical-synthesis plant protection products, where a continuous pathogen disease monitoring plays a key to manage the crop stands.