Assessing machine learning methods to detect bois noir in vineyards - Towards a systematic monitoring approach with spectroradiometry and RPAS data to reduce farmers’ manual detection time

Remotely piloted aircraft systems (RPAS), also known as unmanned aerial vehicles (UAVs) are used more and more in every part of life, from movie cameras, to package delivery systems, and to monitor plant health. Their use allows for more consistent and reproducible results in all these sectors. Bois noir (BN) and flavescence dorée (FD) are two grapevine yellowing diseases which reduce yields for vine growers and increase management time for their plots by requiring manual monitoring of each plant and removal of infected plants. Using RPAS to monitor BN and FD would reduce labour and allow for more targeted intervention when a disease is detected. The aim of this thesis is to examine the relationship between phenotypic, imageable symptoms and the infection of BN to inform future monitoring systems of BN and FD. To identify the most relevant wavelengths in detection of BN, field spectroscopy measurements were taken which span from the visible (VIS) to the near- and shortwave-infrared wavelengths (NIR/SWIR). Additionally, RPAS images covering the VIS and NIR regions were taken from the air to assess feasibility of transferring results of the spectroscopy to the RPAS. From the spectroscopy measurements, the most relevant wavelengths were calculated using a recursive feature elimination (RFE) technique based on a support vector machine (SVM) model. Other models were used in tandem to evaluate the consistency of the results. To evaluate the effectiveness of the feature selection and to see how accurately infected plants can be identified, machine learning (ML) models were employed. This thesis is part of the project SmartGrape and acts as a pilot study to inform later stages of the project. There was only limited data available from the first year of collection, so comparisons to literature were made to verify the feasibility of the results. Wavelengths relating to pigments (520 and 650 nm), photosynthetic activity like fluorescence (710 nm), and non-pigment dry matter and water content of the leaf (1500-1600 and 2000+ nm) were identified to be the main differentiators when trying to classify infected and non-infected leaves. Affordable, commercially available RPAS-based sensors can be used to monitor the VIS wavelengths (520, 650, 710 nm), but not those in the SWIR. This means that classification is possible, but with less accuracy than if the higher wavelengths were included. In the future, a sensor covering higher wavelengths could help improve the in-situ detection of BN and FD.