Detection and differentiation between potato (Solanum tuberosum) diseases using calibration models trained with non-imaging spectrometry data

Abstract The proportion of light at wavelengths across the electromagnetic spectrum that is either absorbed, transmitted or reflected from a plant leaf is dependent on leaf structure, physiology and biochemistry. Since these elements are influenced by pests, pathogens and their associated induced diseases, the detection, differentiation and diagnosis of plant diseases is theoretically possible by non-destructive analysis of the light reflected from plant leaves. In this study the utility of analysis of light over the visible and near-infrared (400–1000 nm) portion of the spectrum to detect and distinguish between several economically important potato diseases, using either Partial Least Squares and BackPropagation Neural Network spectral calibration models was explored. Models could detect and distinguish between diseases with obvious foliar symptoms (blackleg and late blight), even pre-symptomatically, correctly classifying spectra from greenhouse experiments with an accuracy of 84.6%. When these diseases were analysed separately, models could distinguish between spectra from healthy and pre-symptomatic leaves, plus three classes of late blight lesion advancement with 92% accuracy. For blackleg, models distinguished between spectra from healthy, pre-symptomatic foliage and plants expressing blackleg symptoms with a 74.6% classification accuracy. However, models trained on spectra from whole-plant readings from field trials did not have this level of accuracy, with an r2 between target and model values of 0.66 for late blight, 0.31 for blackleg symptoms and 0.41 for healthy foliage. Regardless of greenhouse or field environment, models failed to detect or distinguish between diseases with subtle foliar impacts (black dot, powdery scab and Rhizoctonia diseases). While deployment of hand-held spectrometers for disease detection on a broad-acre scale is impractical, these findings could underpin methods to analyse hyperspectral imaging data with sub-plant resolution for incorporation into precision agriculture and Integrated Pest Management programmes for potato blackleg and late blight management.