Assessing the severity of cotton Verticillium wilt disease from in situ canopy images and spectra using convolutional neural networks

Abstract

Verticillium wilt (VW) is a common soilborne disease of cotton. It occurs mainly in the seedling and boll-opening stages and severely impairs the yield and quality of the fiber. Rapid and accurate identification and evaluation of VW severity (VWS) forms the basis of field cotton VW control, which has great significance to cotton production. Cotton VWS values are conventionally measured using in-field observations and laboratory test diagnoses, which require abundant time and professional expertise. Remote and proximal sensing using imagery and spectrometry have great potential for this purpose. In this study, we performed in situ investigations at three experimental sites in 2019 and 2021 and collected VWS values, in situ images, and spectra of 361 cotton canopies. To estimate cotton VWS values at the canopy scale, we developed two deep learning approaches that use in situ images and spectra, respectively. For the imagery-based method, given the high complexity of the in situ environment, we first transformed the task of healthy and diseased leaf recognition to the task of cotton field scene classification and then built a cotton field scenes (CFS) dataset with over 1000 images for each scene-unit type. We performed pretrained convolutional neural networks (CNNs) training and validation using the CFS dataset and then used the networks after training to classify scene units for each canopy. The results showed that the DarkNet-19 model achieved satisfactory performance in CFS classification and VWS values estimation (R2 = 0.91, root-mean-square error (RMSE) = 6.35%). For the spectroscopy-based method, we first designed a one-dimensional regression network (1D CNN) with four convolutional layers. After dimensionality reduction by sensitive-band selection and principal component analysis, we fitted the 1D CNN with varying numbers of principal components (PCs). The 1D CNN model with the top 20 PCs performed best (R2 = 0.93, RMSE = 5.77%). These deep learning-driven approaches offer the potential of assessing crop disease severity from spatial and spectral perspectives.