Robust fitting of fluorescence spectra for pre-symptomatic wheat leaf rust detection with Support Vector Machines

Early recognition of pathogen infection is of great relevance in precision plant protection. Pre-symptomatic disease detection is of particular interest. By use of a laserfluoroscope, UV-light induced fluorescence data were collected from healthy and with leaf rust inoculated wheat leaves of the susceptible cultivar Ritmo 2–4 days after inoculation under controlled conditions. In order to evaluate pathogen impact on fluorescence spectra 215 wavelengths in the range of 370–800 nm were recorded. The medians of fluorescence signatures suggest that inoculated leaves may be separated from healthy ones, but high-frequency oscillations and individual reactions of leaves indicate that separability is difficult to achieve. The misbalance between the high number of measured wavelengths and the low number of training examples induces a high overfitting risk. For a pre-symptomatic pathogen identification a small number of robust features was desired which comprise most of the information relevant for the given classification task. Instead of choosing only the most relevant wavelengths, the coefficients of polynomials fitting the spectra were used for classification. They specify the global curve characteristics. Piecewise fitting by polynomials of fourth order led to high classification accuracy. Support Vector Machines were used for classification. Cross validation demonstrated that the achieved classification accuracy reached 93%. This result could be attained on the second day after inoculation, before any visible symptoms appeared. The described method is of general interest for pre-symptomatic pathogen detection based on fluorescence spectra.

► Presymptomatic disease detection is of great relevance in precision plant protection.
► Fluorescence data were collected from healthy and leaf rust infected wheat leaves.
► Coefficients of fourth order polynomials specified global curve characteristics.
► Support Vector Machines identified infected leaves before symptoms were visible.
► Classification accuracy was 93% only 2 days after inoculation.