Abstract:
Phytophthora root diseases are of great concern to commercial forests. A South African Phytophthora root disease, caused by P. alticola is increasing in incidence across an expanded range. One of the most effective ways to manage the spread of this and other plant diseases is early detection, before the disease has caused significant damage. Traditional detection methods such as visual detection have many disadvantages and are not effective on larger scales. Therefore, there is an increasing focus on developing non-destructive, rapid and easily scalable methods for early disease detection. One such method is hyperspectral sensing. The first chapter of this dissertation provides a literature review which aims to outline all the current knowledge of hyperspectral sensing as a technology that can be used for early detection of plant diseases. The review further highlights the challenges in processing hyperspectral datasets and outlines numerous ways in which these challenges are overcome through machine learning. Lastly, this chapter explores the potential future uses of hyperspectral sensing as an alternative way to detect plant metabolites.
The second chapter focuses on the development of a hyperspectral sensing method for the detection of Phytophthora root diseases using the Eucalyptus benthamii - Phytophthora alticola pathosystem as a model system. This chapter also aimed to investigate algorithms for identifying wavelengths which could play an important role in the infection process. Lastly the chapter assess the potential of using self-attention networks as a classification algorithm and leveraging the attention mechanism to do wavelength selection. There was a gradient of susceptibility among the 19 different families included in this study. The results show that detection of P. alticola root rot is possible using any of the three machine learning algorithms included in this study. This research also shows that self-attention networks perform on par or better than other algorithms such as the random forest and support vector machine models. Investigation of selected vegetation indices showed that diseased trees accumulated secondary metabolites, such as anthocyanins and carotenoids. Indices also revealed that diseased trees experienced moisture stress, which is consistent with the symptoms caused by P. alticola root disease.
This research highlights the potential of using hyperspectral sensing technologies for plant disease detection and understanding certain aspects of plant-pathogen interactions. In future this research can be used as a basis for the use of machine learning algorithms, not only for disease detection but also for wavelength selection. This research also demonstrated that it is feasible to detect a root infection using leaf measurements. The ability to detect wavelengths and associate them with certain plant metabolites and other physiological aspects of plants, such as leaf wax composition, will serve as a powerful tool to learn more about the molecular interactions between pathogens and their host plants.
