Evaluation of Sensor-based Early Detection Methods for Grapevine Diseases like Palatinate grapevine yellows, Bois noir, Grapevine leafroll disease and Esca

Abstract

Grapevine (Vitis vinifera ssp. vinifera) as a perennial crop is typically grown over several decades. During their lifespan, vines may be infected by a number of different pathogens, some of which remain inside the vine and thus accumulate over time. At the moment, only prophylactic measures are available to reduce the spread of many endogenic diseases, which include visual ratings and subsequent uprooting of infected grapevines in the field or mandatory pathogen tests in nurseries to provide healthy planting material. Sensor-based approaches could significantly contribute to early disease diagnosis. Hyperspectral sensors detect plants’ reflection objectively and non-invasively in the visible range of light (400 – 700 nm) but also in the near infrared (700 – 1000 nm) and short-wave infrared region (1000 – 2500 nm). Biochemical and biophysical changes induced by pathogen infestation cause deviations in the reflectance spectra, which can be analyzed by different machine and deep learning models enabling disease detection at early infection stages. In this study, the suitability of ground-based hyperspectral analyses for the detection of the grapevine phytoplasma diseases Palatinate grapevine yellows (PGY) and Bois noir (BN), the virus infection grapevine leafroll disease (GLD), and the fungal Esca disease has been evaluated.

Disease detection models for both BN and PGY could be developed under controlled conditions using greenhouse plants. These models were able to classify plants correctly as either healthy or infected with an accuracy of up to 96%. However, identification of infected but symptomless plants needs further improvements. Since symptoms of both diseases may vary strongly depending on environmental factors, shoots collected in the field from different cultivars were also analyzed. Again, high classification accuracies of up to 100% could be achieved leading to the assumption that both diseases might also be detectable directly in the field.

GLD detection was at first tested using different greenhouse plants. Thereby, 83 – 100% of symptomatic and 85 – 100% of infected but symptomless vines could be correctly identified. Moreover, approximately 500 grapevines were analyzed directly in the field during the years 2016 – 2018 leading to similar classification accuracies. Furthermore, the potential of hyperspectral analyses for the in-field detection of infected but symptomless vines could be shown. However, results strongly differed between experimental years, therefore, further analyses are necessary for a final evaluation of this aspect. Moreover, symptoms caused by the two main GLD agents Grapevine leafroll-associated virus-1 and Grapevine leafroll-associated virus-3 could successfully be discriminated.

Esca disease detection was also performed directly in the field during three consecutive years. Thereby, hyperspectral detection models could successfully be established for original field data as well as for manually annotated data. In addition, first results clearly showed the potential of pre-symptomatic disease detection. Moreover, model transferability to unknown data was tested but remains challenging and will require to include further experimental years.

Based on hyperspectral data, most important wavelengths were determined for every disease and every analysis approach in order to simplify this complex system. Multispectral sensors could eventually be developed using these wavelengths being faster, cheaper and more flexible than a hyperspectral application. In the case of Esca, additional airborne multispectral data were acquired during this study and compared to a multispectral simulation based on hyperspectral approaches. Although, the simulated multispectral data achieved good results, thus, showing the potential of this method, airborne disease detection needs to be improved.