*2.4. Post-Processing: Clustering and Reference Data*

Single-band orthomosaics were extracted from the multispectral orthomosaic and tested for band correlation. Green and red bands were de-correlated from the NIR band using the formula *Radj* = *Rini* − *RNIR* ∗ 0.8, where *Radj* is the corrected reflectance value, *Rini* is the initial value of the red or green bands and *RNIR* is the reflectance value in the NIR spectrum. This correction reduced the R<sup>2</sup> correlation from 0.89 to 0.64 between the red and NIR band, and from 0.92 to 0.81 between the green and NIR band. Temperature and RGB values were not modified after the stitching.

**Figure 1.** Overall processing pipeline. In (**a**) processes are represented with a darker background, while data have a light background. The green steps represent data and processes, used for the unsupervised classification of multispectral and thermal data; the blue steps represent factors used for reference data estimation based on RGB clips; the yellow steps are common between the two processing pipelines. In (**b**) details of the segmentation process are depicted: starting with mask positioning along planting lines, proceeding with rectangular shape clipping and concluding with circular mask extraction. In (**c**) the classes derived by the expert assessment are shown for each vigour and plant health shift class. On the left, the four vigour classes are represented from highly vigorous and asymptomatic (V4) to dead (V1). On the right, the classification of the health shifts that occurred between 2017 and 2018 in plants that were initially highly vigorous and asymptomatic (V4) is reported: from those that remain highly vigorous and asymptomatic in the second year (S1), to the ones that in 2018 were weakened (S2), diseased (S3) or dead (S4). Abbreviations: ROI, region of interest; GeoRef, georeferenced.

Identification of the stump positions and determination of the canopy boundaries is very difficult to achieve when performing in-field UAV surveys over kiwi vines because, firstly, T-bar or pergola training systems create a flat and dense canopy preventing stump identification, and secondly, the canopies of two neighbouring plants usually overlap due to the disordered growth of shoots. For these reasons, a rough segmentation algorithm was applied to the RGB, thermal, and each single-band multispectral orthomosaic. Firstly, the orthomosaics were clipped along the planting lines with rectangular masks. These digital masks were 5 m wide and 1.5 m long and were placed with the longer side perpendicular to the direction of planting (Figure 1b). The rectangles were placed 1.5 m apart from each other to avoid overlapping. Secondly, a circular shape with a radius of 1 m was applied to the centre of each rectangle clip, retaining only the area within the circle. The shape resulting from the rectangle's and circle's intersection was selected to reduce the border effect, while the mask dimensions and the distances between the mask centres were set to minimise the interference from background pixels of the inter-row grass, and to reduce the canopy overlap within the same clip. Segmentation of RGB data

was stopped at the rectangular shape clipping, while multispectral and thermal data were subjected to the complete segmentation workflow. The RGB clips were used to estimate reference values for plant vigour and plant health shifts via expert assessment, while multispectral and thermal data were used for unsupervised clustering. Maintaining the entire rectangular clips for RGB data improved expert awareness of the canopy dimensions, and consequently the precision of plant vigour classification. On the other hand, multispectral and thermal data was reduced to the assumed canopy extent to reduce background bias, since the central portion of the canopy is usually flat and orthogonal to the acquisition of the images (Figure 1b).

Two types of reference data were evaluated from the rectangular RGB clips via expert assessment: plant vigour and plant health shifts (Figure 1c). Plant vigour is a good indicator to distinguish asymptomatic and symptomatic plants, but it is not directly associated with the status of the root system [2]. Therefore, the a posteriori evaluations of vigour shifts and the death rates between years were the only available means to perform realistic evaluations of plant health status in 2017 without uprooting plants.

Reference data for canopy vigour was estimated based on RGB clips derived from the segmentation process of the 2017 orthomosaics. Each clip was evaluated individually and independently by two experts and ranked in four classes based on canopy vigour, considering: (i) the presence/absence of symptoms on the leaves (yellowing, scorching or wilting), (ii) disease severity (percentage of canopy with wilting symptoms), and (iii) fractional vegetation coverage (Table 2 and Figure 1c).


**Table 2.** Criteria used by experts to evaluate plant vigour and health shift classes. Plant vigour was estimated from RGB clips for 2017 and 2018 surveys. Plant health shifts between the two years were then derived only for plants that were highly vigorous (V4) in 2017.

Reference data for plant health shifts were derived by comparing plant vigour classes in 2017 with those observed in 2018. Vigour classes for 2018 were evaluated via the same procedure described above. Shifts that occurred in highly vigorous plants from 2017 (V4) were the only ones considered due to their relevance for aetiological studies (Table 2 and Figure 1c). Indeed, plants with canopies resembling the diseased classes (V2 and V1) usually possessed a heavily deteriorated root system that may have already been colonised by secondary invaders, while the vigour reduction of plants in class V3 could have been caused by several aspects unrelated to the disease being studied.

Data reduction was applied to multispectral and thermal data by averaging the pixel values within each masked area and storing the results in a matrix associated with the coordinates of the masking centres. K-means [48] and Ward's hierarchical [49] clustering were then performed separately for multispectral (NIR and red bands) and thermal data, using two, three and four clusters. The highest number of clusters was selected based on the capability of the experts to correctly classify the reference

data, while lower numbers were tested to better understand the clustering prediction capability (see below for details).
