Hyperspectral imaging of symptoms induced by Rhizoctonia solani in sugar beet: comparison of input data and different machine learning algorithms

The fungal pathogen Rhizoctonia solani is one of the most important soil-borne diseases in sugar beet production worldwide. Root and crown rot caused by this fungus are traditionally recognized later in the cropping season by rating the above-ground symptoms like wilting and chlorosis on foliage, and dark brown lesions at the base of petioles. The present study was designed to evaluate noninvasive sensors and machine learning for measuring disease incidence and early detection. Eight-weeks-old plants were inoculated with the pathogen in two different concentrations and under controlled conditions. Hyperspectral images in the visible and near-infrared range from leaf were obtained in time-series. One hundred thirty and fifteen spectral features were selected in two levels by using the recursive feature elimination method (RFE) and a clustering approach. Subsequently, five types of machine-learning methods were employed to train four types of spectral data containing reflectance values, vegetation indices, selected variables of the RFE process and selected variables of an RFE-clustering process. The best classifier was obtained from a partial least squares modeling process and required a number of 15 spectral features, which include first and second derivatives of the wavelength spectrum as well as the Ctr3, EVI and PSSRa vegetation index. This investigation proves that under controlled conditions early detection of indirect symptoms caused by Rhizoctonia root rot in sugar-beet plants is possible. Scoring of disease incidence of Rhizoctonia root rot at 10 dai was 3 to 5 times higher with a machine-learning classifier in comparison with the human visual rating.