Оценка применимости конвейера обработки гиперспектральных данных, разработанного для ранней диагностики ржавчинных заболеваний зерновых, для фенотипирования пшеницы, ржи и ячменя

Abstract

Hyperspectral sensing data processing pipeline, originally developed for the early diagnosis of rust diseases in grain crops, was assessed for its applicability for the task of phenotyping of healthy plants of wheat Triticum aestivum, barley Hordeum vulgare, and rye Secale cereale. Hyperspectral images of healthy plants, obtained under laboratory conditions using a Cubert Ultris 20 camera (450–874 nm range, 106 channels), were utilized. The effectiveness of various preprocessing schemes was compared: full (including normalization, smoothing, calculation of derivatives, and identification of extreme features), reduced, and minimal. Machine learning models were exploited for classification: logistic regression, support vector machine, and gradient boosting, trained on averaged spectra. It is shown that the use of a full pipeline optimized for phytopathological diagnostics leads to reduced classification accuracy in phenotyping tasks. The best results (F1 = 0.97 ± 0.025) were achieved using the original averaged spectral curves without additional transformations. It is concluded that for healthy wheat, barley, and rye phenotyping, absolute reflectance levels are informative, whereas for disease diagnostics, changes in the shape of the spectral curve are more important. The obtained results clarify the applicability limits of pipelines developed for phytosanitary purposes and can inform the development of remote monitoring and phenotyping systems for cereal crops.