Abstract
Apple proliferation is an important disease of apple of which, one of the only management strategies is the eradication of infected trees, which are usually identified by the expression of symptoms. Reliable identification and eradication of infected trees early in the vegetation period strongly reduces the risk of further disease spread. In-field, spectrally-based diagnosis offers an alternative to cumbersome and error-prone human-based approaches and could potentially be more sensitive, cost effective and easily automated. Here we investigate the use of a field spectroradiometer in combination with machine-learning assisted data modelling for the detection of apple trees infected with ‘Ca. Phytoplasma mali’, which is the bacterial agent, associated with Apple proliferation. Furthermore, we identified factors that may affect successful detection. We found that the method could be used to detect infection with high sensitivity at accuracy levels (80%) comparable, if not slightly better than human symptom-based diagnosis (77%). The methods used here were able to correctly diagnose all infected trees in the test set expressing symptoms as well as many of the infected but asymptomatic trees that would have been missed by human diagnosis. In general, models often erred in favor of sensitivity at the expense of precision and specificity. Training models using green leaves improved specificity and precision without sacrificing overall accuracy. Model transferability between orchards was assessed with mixed results, which may be improved by accounting for systematic differences between orchards.