Detection of biotic stress (Venturia inaequalis) in apple trees using hyperspectral data: Non-parametric statistical approaches and physiological implications

The use of hyperspectral approaches for early detection of plant stress caused by Venturia inaequalis (apple scab) was investigated to move towards more efficient and reduced application of pesticides, fertilizers or other crop management treatments in apple orchards. Apple leaves of the resistant cultivar, Rewena and the susceptible cultivar, Braeburn, were artificially inoculated with conidia of V. inaequalis in a controlled greenhouse environment. The research focused on (i) determining if leaves infected with V. inaequalis could be differentiated from non-infected leaves, (ii) investigating at which developmental stage V. inaequalis infection could be detected, and (iii) selecting wavelengths that best differentiated between infected and non-infected leaves. Hyperspectral data were used because of their contiguous nature and the abundance of narrow wavelength bands in the electromagnetic reflectance spectrum, thereby providing the spectral sensitivity needed to detect subtle variations in reflectance. Processing the data, however, presented challenges, given the need to avoid data redundancy, identification of data extraction techniques, and maintaining modeling accuracy. Statistical techniques therefore had to be robust. Logistic regression, partial least squares logistic discriminant analysis, and tree-based modeling were used to select the hyperspectral bands that best defined differences among infected and non-infected leaves. Results suggested that good predictability (c-values > 0.8) could be achieved when classifying infected plants based on these supervised classification techniques. It was concluded that the spectral domains between 1350–1750 nm and 2200–2500 nm were the most important regions for separating stressed from healthy leaves immediately after infection. The visible wavelengths, especially around 650–700 nm, increased in importance three weeks after infection at a well-developed infection stage. Identification of such critical spectral regions constitutes the logical first step towards development of robust stress indicators based on hyperspectral imagery.