Spectral signatures of wheat rust from hyperspectral data: the potential of machine learning methods

Abstract

The study presents an approach for detecting rust and differentiating lesions using hyperspectral imaging and machine learning methods. Analysis of the differences in the spectral characteristics of lesions on the wheat ear, leaf, and stem revealed patterns between light reflectance and the structure of plants and pathogens. Healthy areas have a high reflectance coefficient due to their normal cellular structure and chlorophyll content. Lower intensity was detected in desiccated areas, which is associated with moisture loss and disrupted plant structure. Areas affected by rust exhibit low reflectance coefficients, which are related to tissue degradation and the accumulation of dark pigments by the pathogens, contributing to their resistance to external stresses. This forms a distinct spectral profile, allowing for clear visual identification of the disease. As a result, with the aim of identifying the presence of lesions caused by Puccinia graminis and Puccinia triticina, a classification model based on the Random Forest algorithm was developed to recognise rust-affected zones, achieving an overall classification accuracy of 94%. The obtained values indicate the model's high potential for detecting rust lesions, confirming the promise of synergistic analysis of hyperspectral data combined with ensemble machine learning algorithms for the non-invasive detection of wheat rust diseases.