Application of thermal and visible imaging to estimate stripe rust disease severity in wheat using supervised image classification methods

Abstract

Reliable and accurate estimation of plant disease severity at the field scale is a key factor for predicting yield losses, disease management and food security. A field experiment was designed and conducted during 2017–18 and 2018–19 with 24 wheat cultivars to estimate the stripe rust severity by supervised classification of thermal and visible images using parallelepiped, minimum distance, mahalanobis distance, maximum likelihood, support vector machine and neural network methods of image classification. Results demonstrated the potential of thermal and visible imaging techniques to estimate wheat stripe rust severity with good accuracy. For both visible and thermal images used in this study, support vector machine gave the best estimates of the rust severity, while the parallelepiped method was the worst-performing method. Support vector machine and neural network methods showed d-index, Nash-Sutcliffe efficiency and coefficient of determination values above 85%, with accuracies above 98% and kappa coefficient above 0.97 for both thermal and visible images. Comparison of thermal and visible image classification performance revealed that for all the methods except support vector machine, the estimated rust severity, overall accuracy and kappa coefficient of thermal images were better than visible images. The present study clearly showed that both thermal and visible image analysis can be applied as a rapid non-destructive technique to estimate the wheat rust severity under field conditions. The study also provided a comparative insight into thermal and visible image classification methods that have great potential for sustainable plant disease management in modern agriculture.