RGB-based phenotyping of foliar disease severity under controlled conditions

Abstract

Plant diseases induce visible modifications on leaves with the advance of infection and colonization, thus altering their spectral reflectance pattern. In this study, we evaluated the visible spectral region of symptomatic leaves of five plant diseases: soybean rust (SBR), Calonectria leaf blight (CLB), wheat leaf blast (WLB), Nicotiana tabacum-Xylella fastidiosa (NtXf), and potato late blight (PLB). Ten spectral indices were calculated from the RGB channels (red, green, and blue) of images of leaves varying in percent severity, which were obtained under controlled lighting and homogeneous background. Image processing was automated for background removal and pixel-level index calculation. Each index was averaged across pixels at the leaf level. We found high levels of correlation between leaf severity and the majority of the spectral indices. The most highly associated spectral indices were overall hue index, visible atmospherically resistant index, normalized green red difference index, primary colors hue index, and soil color index. The leaf-level mean value of each of the ten indices and digital numbers on the RGB channels were gathered and used to train boosted regression tree models for predicting the leaf severity of each disease. Models for SBR, CLB, and WLB achieved high prediction accuracies (>97%) on the testing dataset (20% of the original dataset). Models for NtXf and PLB had prediction accuracies below 90%. The performance of each model may be directly related to the symptomatology of each disease. The method can be automated if the images are obtained under controlled light and homogeneous background, but improvements should be made in the method for using field- or greenhouse-acquired images which would require similar conditions.