VNIR Hyperspectral Signatures for Early Detection and Machine-Learning Classification of Wheat Diseases

This article presents the results of a comprehensive study aimed at developing automated diagnostic methods for identifying spring wheat phytopathologies using hyperspectral imaging (HSI). The research aimed to create an effective plant disease detection system, including at the early stages, which is critically important for ensuring food security in regions where wheat plays a key role in the agro-industrial sector. The study analyses the spectral characteristics of major wheat diseases, including powdery mildew, fusarium head blight, septoria glume blotch, root rots, various types of leaf spots, brown rust, and loose smut. Healthy plants differ from diseased ones in that they show a mostly uniform tone without distinct spots or patches on hyperspectral images, and their spectra have a consistent shape without sharp fluctuations. In contrast, disease spectra, differ sharply from those of healthy areas and can take diverse forms. Wheat diseases with a light coating (powdery mildew, fusarium head blight) exhibit high reflectance; chlorosis in the early stages of diseases (rust, leaf spot, septoria leaf blotch) exhibits curves with medium reflectance, and diseases with dark colouration (loose smut, root rot) have low reflectance values. These differences in reflectance among fungal diseases are caused by pigments produced by the pathogens, which either strongly absorb light or reflect most of it. The presence or absence of pigment production is determined by adaptive mechanisms. Based on these patterns in the spectral characteristics and optical properties of the diseases, a classification model was developed with 94% overall accuracy. Random Forest proved to be the most effective method for the automated detection of wheat phytopathogens using hyperspectral data. The practical significance of this research lies in the potential integration of the developed phytopathology detection approach into precision agriculture systems and the use of UAV platforms, enabling rapid large-scale crop monitoring for the timely detection. The study's results confirm the promising potential of combining hyperspectral technologies and machine learning methods for monitoring the phytosanitary condition of crops. Our findings contribute to the advancement of digital agriculture and are particularly valuable for the agro-industrial sector of Central Asia, where adopting precision farming technologies is a strategic priority given the climatic risks and export-oriented nature of grain production.