Potential of laboratory hyperspectral data for in-field detection of Phytophthora infestans on potato

Researchers have shown increasing interest in hyperspectral imaging for detecting potato late blight disease ( Phytophthora infestans ). Because it is difficult to get accurate spectral signatures of disease development in field conditions, especially at early disease stages, previous works focused on laboratory measurements under controlled conditions. However, the extrapolation of results from a laboratory to a field setting has proven difficult. The current work evaluates the use of laboratory hyperspectral data to train an in-field detection model for potato late blight. A hyperspectral training library was constructed from six detached leaf trays, containing 8585 spectra labelled into a healthy class and five progressive stages of disease development. After smoothing and normalisation, a logistic regression model was trained on 70.0% of this data, with 30.0% reserved for validation. Twelve hyperspectral images taken in field conditions were then classified, for two potato cultivars (susceptible and resistant to late blight), at high and low disease pressure. The classification accuracy of laboratory data was 94.1%, which was not sufficient to detect field symptoms, using infield collected dataset. When spectra pre-processing was changed by including first derivation and adopting a new normalisation strategy, a new model resulted in a lower classification accuracy of 80.8%, validated on labelled laboratory spectra, but was able to detect symptoms in field conditions. The correlation between visual disease scoring and the classification result of the field disease model yielded an R 2 value of 0.985. It could be concluded that it was possible to train a model on laboratory data for in-field disease detection.