Early detection of Phytophthora infestans in potato plants using hyperspectral imaging, local comparison and a convolutional neural network
Janne Kool, A. Evenhuis
Abstract
An important primary inoculum source of Phytophthora infestans are latent infected seed potatoes. To remove those inoculum sources, it is necessary to detect these plants before symptom expression. Often, when a human can detect it by eye, sporulation has occurred and secondary spread might have commenced. Moreover, it is not feasible to check all potato plants by manual labour. It is possible to detect a P. infestans infection early, using hyperspectral imaging, however, models trained on the spectral information only, do not generalize, because the spectrum is very much dependent on the local conditions like soil humidity. A method is developed to detect P. infestans early by comparing spectra from infected plants with the spectra of healthy plants that have grown under similar conditions, i.e., in the same plant row, and creating an image that quantifies deviations. On those images in turn a convolutional neural network is trained, which is able to detect infections on plants grown in another field or a year later in a similar manner but under different conditions. The method has been tested in the field in a dry year with little potato late blight infection. In this experiment it has not been possible to detect the infection using hyperspectral imaging.