Evaluating the Potential of High-Resolution Visible Remote Sensing to Detect Shiraz Disease in Grapevines

,


Introduction
Shiraz disease (SD) is a devastating viral disease of grapevines that was frst reported on Merlot from South Africa [1]. SD disrupts the physiological development of grapevines and causes signifcant yield loss in specifc cultivars, including Shiraz, Merlot, Malbec, and Sumoll [2]. Te symptoms of SD infection in Shiraz include delayed budburst with restricted spring growth, lack of lignifcation on some canes, and delayed leaf senescence well into the dormant season [3,4]. SD symptoms are latent (no symptoms) in tolerant cultivars such as Chardonnay and Cabernet Sauvignon; however, the viruses can be transmitted to susceptible cultivars (Shiraz and Melot) by mealybugs and soft scales [5,6]. Grapevine virus A (GVA) group II variants were associated with SD [7,8]. GVA also causes a rugose wood disease known as "Kober stem grooving" [9]. GVA often coexists with grapevine leafrollassociated viruses (GLRaVs) [6,[10][11][12], which is a group of viruses that causes Grapevine leaf disease (GLD) [13]. In Australia, GLRaV-1, GLRaV-3, and GLRaV-4 strain 9 (GLRaV-9) are commonly associated with GVA in SDinfected vines [14]. Tere are only a few efective methods to control grapevine viral diseases including roguing infected vines, replanting with certifed, virus-free material, and controlling the vectors to stop the virus from spreading [15,16]. It is therefore critical to accurately detect the patterns and extent of viral infections in vineyards to stop the virus from spreading further.
Standard detection methods for SD include serological methods, nucleic acid-based methods, and visual assessment [17,18]. Lab-based methods are costly, thus limiting the number of grapevines tested and, consequently, an underestimation of the true incidence of virus infection in vineyards [17,19]. Currently, the recommended minimum test rate by commercial diagnostic labs is fve vines per thousand (0.5%) across the block [20]. Conducting onground visual assessments is labour-intensive, subjective, and sometimes unreliable. Low-altitude airborne remote sensing enables the capture of high detail with greater potential to rapidly survey the vineyards. Various optical sensors including red-green-blue (RGB), multispectral, hyperspectral, and thermal sensors have been used on the ground or platforms like unmanned aerial vehicles (UAVs) and manned fxed-wing aircraft for grapevine disease detection [21][22][23][24][25][26]. RGB imagery acquired through UAV-based remote sensing was used for the current study due to its relative simplicity compared to multi and hyperspectral images. A vertical projection of the canopy from the aerial image, the projected leaf area (PLA), for each vine was calculated from the image to compare the canopy size between healthy and SD-infected vines. PLA acquired from remote sensing imagery has a positive correlation to the canopy area. For example, Raj et al. [27] achieved an R 2 of 0.84 and RMSE of 0.36 by using PLA calculated from UAV RGB image and compared to leaf area index of maize.
In this study, we used high-resolution RGB remote sensing imagery to systematically assess PLA of individual healthy and diseased vines to predict SD infection in Shiraz grapevines in the feld. Te specifc objectives of this study were: (1) to develop a simple remote sensing methodology that can consistently assess grapevine canopy size (using PLA as a surrogate) as a visual indicator of SD infection; (2) to confrm PLA-based disease status classifcation with labbased tissue analysis; (3) to evaluate the time series of remote sensing imagery in order to conduct a spatial-within-season temporal analysis of canopy size diferences between healthy and infected vines; and (4) to evaluate the temporal consistency of seasonal patterns of canopy development across multiple growing seasons. Our overarching goal was to develop a rapid and low-cost surveillance platform for SD detection at the vineyard scale.

Study Sites and Visual Estimation of Virus Infection.
Two virus-infected Shiraz blocks (some vines previously tested positive with GVA and GLRaVs) were selected in diferent climatic wine regions in South Australia (SA) for this study. Te frst vineyard was in Monash, located in the warm inland region of Riverland (34°13′28″S, 140°33′01″E). A block of 1.5 ha of Shiraz was selected for the study. Te soil type of vineyard was sand over limestone. Te block was drip-irrigated with 7.5 ml·ha −1 of water per year. Approximate 50 kg ha −1 N and 50 kg·ha −1 P fertiliser were applied through fertigation annually. Te vines were consistently machine spur pruned with a same size box shape each winter. Vineyard management was consistent between seasons. Integrated pest management was as per convention in this region, which generally has low disease pressure due to its warm-to-hot climate. Te second vineyard was in the Barossa region, located in Lyndoch, SA (34°35′28″S, 138°53′01″E). A 1.5 ha block was chosen for the study. Te soil type of the block was Calcic on red Sodosol. It was drip-irrigated with approximate 1 ml·ha −1 water per year. Both solid fertiliser and fertigation were applied at the rate of 130 kg·ha −1 N, 55 kg·ha −1 P, and 9 kg·ha −1 K annually. Shiraz was consistently two-bud spur pruned to 20 buds per m each winter. Details of the study sites (vineyards) are provided in Table 1.

Virus
Testing. Laboratory-based tissue testing was used for ground-truthing (Figure 1(b)). Tissue samples were collected based on visual symptoms for virus testing, of which half the vines were symptomatic and half were asymptomatic. Leaf petioles were sampled near harvest time [28]. Te leaves were carefully selected from the base of the shoots to avoid errors associated with sampling from a potential long shoot coming through from a neighbour vine. Four petioles near the base of the shoots (two from each side of the canopy) were sampled and transported with chilled ice packs.
All samples were virus-tested in the lab using an enzymelinked immunosorbent assay (ELISA) [29]. Te ELISA test kits produced by Bioreba (Reinach, Switzerland) were used to test GVA, GLRaV-1, GLRaV-3, and GLRaV-4 strains. 20% of these leaves samples were tested with reversetranscription polymerase chain reaction (RT-PCR) [17,30] for confrmation of the ELISA results. Te RT-PCR test was conducted by a commercial diagnostics lab that routinely tests for grapevine viruses. Six commonly occurring grapevine viruses in Australia [31] were tested: GVA, GLRaV-1, GLRaV-3, GLRaV-4, GLRaV-4 strain 6, and GLRaV-4 strain 9. Te result showed a 100% match between PCR and ELISA, confrming the reliability of the ELISA test. Te number of vines in each class is shown in Table 2. Because GLRaV-1, -3, and -4 complexes cause similar GLD symptoms in grapevines, vines infected with either a single or combination of any GLRaVs were treated as a GLRaV infection. In total, there were four classes: (i) healthy, (ii) GVA only, (iii) GLRaVs only, and (iv) GVA + GLRaVs.

High-Resolution Remote Sensing: Data Collection and
Processing. DJI Mavic 2 Pro (SZ DJI Technology Co., Ltd, Shenzhen, China) was used for image collection in this study (Figure 1(a)). Te UAV uses a Hasselblad RGB camera with a 28 mm focal length and f/2.8-f/11 aperture. Te feld of view is approximate 77°and the image size is 5472 × 3648. Flight planning was automated by the Pix4D app (Pix4D S.A., Prilly, Switzerland) with the setting of nadir view, side and forward overlapping at approximate 80%, altitude at 45 m above ground level, and forward fight direction. Te calculated spatial resolution of the images was approximate 1 cm pixel −1 .
Aerial image data were collected between October to April in S1 and September to April in S2. Data were captured at approximate monthly intervals (one fight per month) based on weather conditions (low wind and sunny) which resulted in six fights in Riverland and ten fights in Barossa (Table 1). camera, the mosaicked images produced a 1 cm pixel −1 ground sampling distance. Te image geo-processing was conducted with ArcGIS Pro V2.8 (Esri, Redlands, California, US). Individual vines were geolocated using the image at dormancy when the shadow of vine trunks was clearly visible. Grapevine locations were manually digitised, and square bufers were created along with the orientation of row lines (Figure 1(c)). Te size of the bufer was adjusted to about 90% of vine spacing to avoid the overlapping area  between vines. Te vine spacing was larger in the Riverland Shiraz block (3.5 × 3.5 m vine and row spacing) compared to Barossa and Adelaide Hills vineyards (spacing 1.5 × 3.0 m), which results in a larger canopy therefore a larger bufer area per vine. Orthomosaics from each date were georeferenced to the dormancy image in order to accurately coalign vines. Te grapevine canopy was mapped using a supervised random forest classifer [32] (also called the random tree method in ArcGIS Pro). We used the ArcGIS Pro V2.8 random trees method with a maximum number of trees of 30 and a maximum tree depth of 15. Pixels in the image were classifed as "Grapevine," "Soil," "Shadow," and Weeds." We manually labelled 5-7 training polygons in each training class and found the training data was sufcient to train Random Tree for classifying all pixels in the images. Te "Soil," " Shadow," and "Weeds" classes were combined into a "nongrapevine" class to obtain a binary image for canopy area calculation (Figure 1(d)). To improve classifcation accuracy, diferent training data sets were created for early, middle, and later seasons as changing colour in the canopy over time. As undervine weeds were well controlled in all blocks, the grapevine was visually clearly distinguishable from the nongrapevine. Te classifcation results were visually assessed by comparing the RGB and classifed images, and results were consistent in all images, thus quantitative accuracy assessment of classifcation results was not required.
Te projected leaf area (PLA) per individual vine was calculated as the sum of pixels that classifed to "Grapevine" within square reference areas that were adapted to the vine and row spacing of the diferent vineyards. We used a square area of 3 × 3 m in Riverland, and 1.4 × 1.4 m in Barossa.

Statistical
Analysis. Two-way ANOVA was used for statistical analysis using GraphPad Prism v9.0.0 (San Diego, CA, US). Te PLA value of all virus-tested vines was used for analysis. Mean PLA values between each class (healthy, GVA only, GLRaVs only, and GVA + GLRaVs) at each time point were compared. Tukey's multiple comparisons test was used as a post hoc test (p < 0.05).

Symptoms of Shiraz Disease.
Te ground visual observations showed that SD-infected Shiraz vines had delayed budburst by approximate 15-20 days and smaller canopies in spring as indicated visually (Figure 2(a)). However, by midsummer (approximate fruitset stage), healthy and infected vines had indistinguishable canopies (Figure 2(b)). However, the canes of infected vines showed a lack of lignifcation, as shown in Figure 2(c). SD-infected vines were clearly identifed in winter due to delayed leaf fall (delay approximate 15-20 days), which shows red leaves attached to the vine, while healthy vines had no leaves (Figure 2(d)). Te SD symptoms consistently showed in two seasons and locations, this matched with observations in other studies [2,4,33].

PLA Diference between SD Symptomatic and Asymptomatic Canopy.
Te average PLA was calculated for each class (healthy, GVA only, GLRaVs only, and GVA + GLRaVs) in both blocks and seasons at each time point (Figure 3). In Riverland, the average size of coinfected (GVA + GLRaVs) vines was consistently approximate 1 m 2 smaller than healthy vines at 25 days after budburst (2.24 m 2 for healthy and 1.32 m 2 for coinfected vines) and fowering stage (4.42 m 2 for healthy and 3.62 m 2 for coinfected vines) in S1 (Figure 3(a)). Te statistical analysis showed the GVA + GLRaVs classes were signifcantly (p < 0.0001) diferent from healthy in the early season. However, the diference in PLA between the two classes decreased after fowering. Figure 3(b) shows PLA of coinfected vines was approximate 1.3 m 2 smaller than healthy vines at 24 days after budburst (1.92 m 2 for healthy and 0.72 m 2 for coinfected vines) and fowering stage (4.26 m 2 for healthy and 2.82 m 2 for coinfected vines) in S2. Similar to S1, the diference in S2 between healthy and coinfected vines decreased after fowering; however, it still has a signifcant diference before veraison (with p < 0.0001).
In the Barossa vineyard, the PLA of coinfected Shiraz was also signifcantly smaller than that of healthy in the early season, especially at the fowering stage. In S1, the average PLA of the healthy and coinfected vines at the fowering stage was approximate 1.5 m 2 and 1.0 m 2 (p < 0.0001), respectively, thus coinfection resulted in 33% smaller PLA (Figure 3(c)). However, the diference between the two classes started to decrease at veraison and no signifcant diferences were observed in PLA in the latter part of the growing season. Te PLA diference between diseased and healthy vines was reduced by veraison although still signifcant (p � 0.0307). Te p-values for the diference between healthy and coinfected vines were more signifcant around the fowering stage than at other times in both seasons.
Te results indicated the symptomatic SD infection in Shiraz could be predicted using PLA calculated from RGB remote sensing images. Te PLA of healthy and SD-infected vines had the highest diference between 20 and 70 days after bud burst, which unveils the optimum time window for SD detection as symptoms could be easily identifed due to the signifcantly smaller PLA of the diseased vines. Te PLA of SD-infected vines were 30%-70% smaller than the average healthy vines. We suggest setting a PLA threshold of 70% in healthy vines to classify as an SD infection in Shiraz. Terefore, PLA values at or less than 70% are classifed as being SD infected. Tis threshold works between 15-45 days Table 2: Te ELISA test results. Samples classifed as "healthy" tested negative for GVA, GLRaV-1, -3, and -4; GVA only is grapevine virus A positive (single infection) but GLRaVs negative; GLRaVs only is single or any combination of grapevine leafroll-associated virus-1, -3, or -4 positive but GVA negative; GVA + GLRaVs is coinfection of both GVA and one or more GLRaV-1, -3 or -4.

Healthy
GVA only Australian Journal of Grape and Wine Research after the budburst in the Riverland region and between 30-60 days after the budburst in the Barossa region. From veraison onwards, this method appeared to be less efective as canopy size diferences between infected and healthy vines become smaller. However, our early PLA results could not distinguish SD from Grapevine trunk disease (GTD), a debilitating fungal disease that afects grapevines worldwide, causing devigourated shoots, and sometimes dead cordons [34]. Te PLA of GTD-infected vines would likely remain low throughout the season since dieback results in very few growing shoots and death of the cordons [35,36]. In contrast, SD-infected vines appear similar in growth to GTD-infected vines, but in contrast to GTD vines, have fully-developed canopies by the veraison stage; this key diference can be used to diferentiate SD infection from GTD or dead vines. Te PLA of SD-infected vines were 5%-15% smaller than the average healthy vines at this stage. Tus, we suggest that an 85% PLA threshold be used at the veraison stage to distinguish between SD-and GTD-infected or dead vines. Terefore, if the PLA is at or below 85% of the PLA of healthy vines between 90-120 days after budburst, the vine could possibly have GTD or be dead. Terefore, a minimum of two data collection timepoints are suggested per season, one in the early season and one in the mid-to-late season for determining SD using remote sensing. However, as the technique is an indirect detection method, which measures the canopy response to the virus, we cannot exclude the possibility that various other factors could be altering the phenotype. For example, other biotic stresses (fungal diseases), abiotic stresses (drought, salinity, heat stress, and mechanical damage), and virus strains and coinfections could infuence vegetative growth and alter PLA [37]. Terefore, this remote sensing technique is indicative but not a conclusive method for SD infection. Te current results were based on two study sites and years, the further assessments and virus testing validations are needed for the diferent regions, years, the age of vines, and cultivars. As additional information is acquired, diferent recommendations of the PLA threshold can be used for vineyards that have similar conditions. If validated, this method can potentially be scaled to larger regions using RGB imaging from manned aircraft, or even satellite imagery in the future as their camera resolutions continue to increase.

Diference between Coinfection and Single Infections.
Canopy development of coinfected vines (GVA + GL-RaVs) lagged behind healthy vines due to delayed budburst in spring. Tis pattern was consistent in both vineyards and seasons (Figure 3). In comparison, the development of GVA and GLRaV (single infection) infected vines had no signifcant diference from healthy vines in both blocks or seasons. Despite previous studies showing that GVA and its variants are associated with SD [8], there is little systematic information between coinfection and SD symptoms. As the coinfection of GLRaVs and GVA is commonly found in vines, it is important to consider both coinfection as well as environmental factors when studying disease symptoms. Te study found that SD symptoms in both Shiraz blocks only occur in vines that are coinfected with GVA and one or more GLRaVs, which in the vineyards we surveyed, were found to be mostly GLRaV-1 or GLRaV-4 strain 9 (GLRaV-9). We did not observe any typical SD symptoms when Shiraz vines were infected with GVA only (i.e., without GLRaVs). Similarly, Goszczynski and Habili [8] reported that SD symptoms in Shiraz were always associated with GVA group II and GLRaV-3 in South Africa. Consistent with the results of the present study, the same authors also observed that some vines did not exhibit any SD symptoms when infected with GVA group II alone; however, only visual evidence, but no quantitative evidence, was provided.
GVA variants of group II have been closely associated with SD, but not groups I and III [3]. As the ELISA serological method is unable to discriminate between virus variants, the asymptomatic GVA-infected vines in our study could belong to group I and/or III. A previous study also reported that the variant GTR1-2 in GVA group II did not produce SD symptoms in Shiraz; however, other group II variants (BMO32-1, KWVMo4-1, and P163M5) produced SD symptoms in both Shiraz and Merlot [33]. Te GVA variants in our study were unknown because the GVA primers used for the RT-PCR test in our study were not variant-specifc. However, if the GVA variants in the present study did not belong to either group I or III, or the GTR1-2 variant (in group II), we could then infer that coinfection of GVA and GLRaVs is a requisite for SD symptoms in Shiraz. We are  Figure 3: Average PLA for each lab-tested class at diferent times for both seasons in two vineyards. Te p-value of healthy vs GVA + GLRaV shows in the graph, with * p ≤ 0.05, * * p ≤ 0.01, * * * p ≤ 0.001, * * * * p ≤ 0.0001, and nonsignifcant with a blank. (a, b) Riverland Shiraz in S1 and S2; (c, d) Barossa Shiraz in S1 and S2.
unaware of any systematic studies that have been done to understand the relationship between SD symptoms and the combination of various viruses and their variants. Tis hypothesis requires a comprehensive investigation, potentially by using next-generation sequencing techniques to screen all GVA and GLRaVs strains in the samples.

Conclusion
Reliable detection of grapevine viruses in the feld remains challenging due to varying symptomology. Tis study systematically compared the canopy growth response of SDinfected vines to healthy vines and proposed a rapid method to predict the SD infection in the Shiraz blocks using visible remote sensing technology. Tis technique has the potential to rapidly detect SD in the feld, thereby providing prompt guidance for sampling locations for tissue testing of viruses as well as vineyard management. Further validation studies including various sites, seasons, cultivars, and virus strains are needed for this emerging technology. An additional, but important fnding was that coinfection of GVA and GLRaVs results in signifcant vine devigoration in Shiraz, which does not occur with GVA or GLRaV alone. Tis observation was consistent across diferent soils and seasons under diferent weather conditions.

Data Availability
Te raw and processed data used to support the study are available from the corresponding author upon request.

Conflicts of Interest
Te authors declare that there are no conficts of interest regarding publication of this work.