Non-invasive and early detection of tomato spotted wilt virus infection in tomato plants using a hand-held Raman spectrometer and machine learning modelling

Abstract

Tomato spotted wilt virus (TSWV) is a polyphagous thrips-transmitted pathogen inducing significant economic losses in agriculture, particularly on tomato plants. The leading management and containment strategies to fight TSWV infection rely on growing resistant cultivars and spraying insecticides for thrips control. Therefore, its early detection is fundamental in sustainable crop management. Aim of the present work is to reveal TSWV infection using a hand-held Raman instrument and Machine Learning (ML) approaches. Artificially inoculated tomato plants were scored for symptom development for one month, while Raman spectra were collected 3 and 7 days after virus inoculation. After preliminary spectral pre-processing, a filter method based on Partial Least Squares Discriminant Analysis (PLS-DA) coefficients was applied to remove redundant and irrelevant variables. The resulting condensed dataset was checked with multivariate exploratory methods and exploited to build multiple PLS-DA models, using different random splitting of the samples between training and test sets. By interpreting the classification metrics, Raman spectroscopy coupled with ML techniques allowed us to discriminate infected from healthy tomato plants within the first 3–7 days after inoculation, with average accuracy of 90–95 % in validation. The model was also validated on two different sets of susceptible and resistant plants, achieving average accuracy higher than 85 %. Early detection of TSWV infection well before visual symptom occurrence represents an important advantage in a sustainable agricultural system. Notably, the use of a portable Raman spectrometer, much less expensive and cumbersome than benchtop instruments, allows the direct in-field execution of these diagnostic measurements.