Study on the application of hyperspectral imaging for identifying barley loose smut at different stages of development using machine learning methods

Abstract

This paper presents a study of the potential of hyperspectral imaging for identifying barley infection by Ustilago nuda at various stages of disease development using machine learning methods. The study focused on barley samples (Hordeum vulgare L.) collected from agrocenoses in the northeastern part of Pavlodar region. Analysis of spectral characteristics revealed significant differences between healthy, infected, and desiccated plant areas, including reduced reflectance and the absence of a distinct red edge in infected tissues. The Maximum Entropy algorithm was employed for classification. The training set comprised 243 samples, while the testing set included 352 samples with pronounced class imbalance. The model demonstrated high classification accuracy of up to 95%. The results confirm the effectiveness of hyperspectral imaging combined with machine learning for disease detection and monitoring, including at early stages, which can be applied in precision agriculture systems to enhance phytosanitary control efficiency and promote sustainable grain production.