Early detection of Wheat Stripe Mosaic Virus using multispectral imaging with deep-learning

Abstract

Wheat Stripe Mosaic Virus (WhSMV) is a soilborne virus that threatens wheat yields in South Africa. Traditional WhSMV diagnosis methods rely on visual inspection, which is labour-intensive, time-consuming and prone to errors. This study explores the application of deep-learning algorithms employing various spectral filters for the early identification of WhSMV. The models were tested for classifying healthy, early, and diseased stages, where early-stage specifically included images taken before any visible disease symptoms appeared. DenseNet121 demonstrated the highest accuracy of 91.23% with the K590 filter, which can capture 590-1000 nm, including parts of the visible and near-infrared spectrum. Further, the K590 filter showed the most significant precision values with most of the tested Convolutional Neural Networks, Vision Transformers, and hybrid and Swin Transformer models. This result suggests filters that capture visible and near-infrared spectrum ranges perform better in identifying WhSMV. These findings show that multispectral images combined with deep-learning models are viable for WhSMV detection in wheat fields, especially for identifying early-stage infections. • A new multispectral dataset, obtained in natural environmental settings using six spectral filters, was evaluated using multiple deep-learning models, including CNN, ViT, hybrid ViT, and Swin Transformer architectures. • Deep-learning models were evaluated to determine their effectiveness in identifying various stages of WhSMV infection. • Multiple filters were assessed across different disease stages to identify the most suitable spectral ranges for accurate disease detection.