Detection of grapevine viral diseases in Australian vineyards using remote sensing and hyperspectral technology

Abstract

Grapevine viral diseases cause substantial productivity and economic losses in the Australian viticulture industry. Two economically significant grapevine viral diseases - Grapevine Leafroll Disease (GLD) and Shiraz Disease (SD) - affect numerous vineyards across major wine regions in Australia. Accurate and quick diagnosis of the virus infection would greatly assist disease management for growers. Current detection methods include visual assessment and laboratory-based tests that are expensive and labour-intensive. Low-cost and rapid alternative methods are desirable in the industry. Recent advances in low-altitude remote sensing platforms such as unmanned aerial vehicles (UAVs or “drones”) in conjunction with high-resolution multiand hyper-spectral cameras now enable large spatial-scale surveillance of plant stresses. My thesis therefore focuses on developing fast and reliable methods for GLD and SD detection on a vineyard scale using optical sensors including RGB and hyperspectral and low-altitude remote sensing technology. The thesis is constituted by a review article and three result parts, it begins with a general introduction for the background and is followed by the research goals and significance of the project that is described in Chapter 1. In order to be familiar with all possible technologies that can be potentially used for GLD and SD detection, Chapter 2 includes a comprehensive overview of methodologies for the detection of any plant viruses reviewed from laboratory-based, destructive molecular and serological assays, to state-of-the-art non-destructive methods using optical sensors and machine vision, including use of hyperspectral cameras. A key contribution of the review is that, for the first time, a detailed economic analysis or cost comparison of the various detection methodologies for plant viruses is provided. In my research, various detection methods with different degrees of complexity were attempted for GLD and SD detection. Firstly, a simple and novel detection method using the projected leaf area (PLA) calculated from UAV RGB images is proposed in Chapter 3 for the disease symptom that alters the growth of the vine such as SD in Shiraz. The PLA is closely related to the canopy size. There are significant differences in PLA between healthy and SD-infected vines in spring due to retarded growth caused by SD, which offers a simple, rapid and practical method to detect SD in Shiraz vineyards. However, for diseases that cannot be easily detected by RGB images such as GLD in the white grape cultivars, different approaches are needed. Hyperspectral technology provides a wide spectrum of light with hundreds of narrow bands compared to RGB sensors. The advanced technology can detect imperceptible spectral changes from the disease and is particularly valuable for asymptomatic disease detection. A new approach using proximal hyperspectral sensing is described in Chapter 4. Using a handheld passive (sunlight is the radiation source) hyperspectral sensor to detect GLD in the vineyard presents a simple and rapid measurement method to detect the diseases using the spectral information from the canopy. An assessment was done for the disease's spectral reflectance throughout the grape growing season for both red and white cultivars. The partial least squares-discriminant analysis (PLS-DA) was used to build a classification model to predict the disease. Prediction accuracies of 96% and 76% were achieved for Pinot Noir and Chardonnay, respectively. The proximal hyperspectral sensing technique is readily applicable to a low-altitude remote sensing method to capture high-resolution hyperspectral images for large-scale viral disease surveillance in vineyards. The subsequent study in Chapter 5 presents an advanced method to quickly detect disease using an UAV carried hyperspectral sensor. The study evaluated the feasibility of UAV-based hyperspectral sensing in the visible and near-infrared (VNIR) spectral bands to detect GLD and SD in four popular wine grapevine cultivars in Australian vineyards. The method combined the spectral and spatial analysis to classify disease for individual pixels from the hyperspectral image. The model predictions for red- and white-berried grapevine cultivars achieved accuracies of 98% and 75%, respectively. For each viral disease, unique spectral regions and optimal detection times during the growing season were identified. The spectral difference between virus-infected and healthy vines closely matched the spectral signal from the proximal sensing method in Chapter 4, which demonstrated the reliability of the low-altitude hyperspectral sensing for grapevine disease detection. Lastly, a summary of the outcomes and remaining challenges and limitations of the existing technology is discussed in Chapter 6, followed by suggestions for further research for further improvement.