**3. Results**

### *3.1. Multispectral and Thermal Data as an Indicator of Plant Vigour and Disease Outbreaks*

For both methods tested, multispectral data were mostly correlated with plant vigour (Figure 2a–f), but it was found to not be a good predictor for disease outbreaks (Figure 2g–l). Indeed, the spatial distribution of multispectral clusters resembles that of the reference data for vigour classes (Figures 3c–h and 3a, respectively). With four clusters, it was possible to estimate all the vigour classes needed for the evaluation of the disease. With three, the weakened plants (V3 and V2) were separated from the asymptomatic and the dead ones, while with two clusters, plants with leaves were distinguished from those already dead. Overall, K-means performed better than the hierarchical method, since the di fferences between the middle vigour classes V3 (weakened) and V2 (diseased) could be better identified. However, hierarchical clustering showed higher correlations towards the highly vigorous plants (V4).

**Figure 2.** Correlation between clusters and reference data. (**<sup>a</sup>**–**f**) and (**g**–**l**) show the correlations of multispectral clusters with plant vigour and health shift classes, respectively. Results obtained from the K-means method are reported in (**<sup>a</sup>**–**<sup>c</sup>**) and (**g**–**i**), while those from hierarchical clustering are in (**d–f**) and (**j–l**). (**m–r**) and (**s–x**) picture the correlation of thermal clusters with plant vigour and health shift classes, respectively. (**m–o**) and (**s–u**) report the results from K-mean clustering, while (**p–r**) and (**v–x**) report those from the hierarchical method. White represents non-significant correlation. Positive and negative correlations are highlighted in blue and red, respectively, and shaded in relation to their confidence level (Pearson test): 95% confidence in brighter colours, 99% in darker tones. The bases of the rectangles are proportional to the total number of images within each reference class (numbers at the bottom of the graph), whereas heights are the proportion of reference data within the specific cluster.

**Figure 3.** Spatial distribution of the results obtained after clustering and the references used for the interpretation. (**a**) In the blue scale, expert assessment ratings for the vigour classes (V4 high, V3 medium, V2 diseased, V1 dead). (**b**) In the brown scale, plant health shifts between 2017 and 2018 (S1 asymptomatic plants that preserved high vigour, S2 asymptomatic but weakened plants, S3 diseased plants but still alive, S4 plants that died in 2018). In the green scale, K-means (**<sup>c</sup>**–**<sup>e</sup>**) and hierarchical (**f**–**g**) clustering derived from multispectral bands. In the red scale, K-means (**i**–**k**) and hierarchical (**l**–**<sup>n</sup>**) clusters resulting from the analysis of thermal images: triangles indicate the temperature of the clusters'

centres. In (**b**) and (**i**–**<sup>n</sup>**) white represents plants that were not highly vigorous in 2017 and were therefore removed from the plant health shift analysis. Yellow rectangles in (**b**) and (**f**-**h**) identify the hottest area within the orchard that happened to have the highest disease incidence in 2018.

When K-means was applied to four clusters to classify multispectral orthomosaics (Figure 2a), each cluster was positively correlated with only one vigour class, and negatively correlated with all the others. Almost all the correlations were significant or highly significant, suggesting that cluster KM4 can be used for the detection of apparently asymptomatic plants with high vigour (V4), cluster KM3 for asymptomatic plants with low vigour (V3), cluster KM2 for plants that are heavily compromised but still alive (V2), and KM1 for the dead plants (V1). Misclassification errors occurred mostly between the narrower cluster classes KM3 and KM4. Hierarchical clustering showed a higher correlation between the V4 class and the HM4 cluster, and performed similarly regarding the HM1–V1 and the HM2–V2 associations (Figure 2d). However, the cluster HM3 was not clearly correlated with the V3 class, since it also contained some diseased plants belonging to the V2 class.

Using K-means with three clusters on the multispectral data (Figure 2b), the extreme vigour classes V4 and V1 were still well discriminated by KM3 and KM1, respectively, while cluster KM2 was mostly correlated with classes V2 and V3, corresponding to plants with a weakened status. Considering this association, errors can occur when assuming cluster KM3 to be solely associated with the highly vigorous plants in V4. By using hierarchical clustering, the association HM3–V4 and HM2 (V2 and V3) was maintained (Figure 2e), but most of the misclassifications occurred within the cluster HM1, which was not highly correlated solely with dead plants (V1) and was also correlated with the V2 class.

Finally, with two clusters only, both methods performed equally (Figure 2c,f). The plants with an asymptomatic canopy (V4 and V3) were clearly separated from the dead plants (V1), and highly correlated with clusters KM2 or HM2, and KM1 or HM1, respectively. Major errors occurred in the classification of plants that were diseased but still alive (V3), which seemed to be better correlated with clusters KM1 and HM1.

For both methods tested, thermal data were mostly correlated with health shifts (Figure 2s–v), but could not be associated with any vigour classes (Figure 2m–r). The hottest area in 2017 within the orchard happened to be the one with the highest disease incidence in 2018 (Figure 3b,j–n). Canopy temperature ranged between 22 ◦C and 30 ◦C, with the clusters averaging between 23 ◦C and 26 ◦C, depending on the number of clusters used. K-means centres were on average slightly higher (+0.3 ◦C) than the means evaluated for each hierarchical cluster. The mosaic plots suggested that the temperature data were able to predict disease outbreaks by clustering the plants into two groups: one where disease symptoms appeared (S3 and S4) and another where the plants remained asymptomatic (S1 and S2). Indeed, among the tested combinations, the two-cluster model showed the best discrimination between the plant health shift classes. Both algorithms performed similarly, with K-means manifesting slightly better correlations between the plants that remained highly vigorous in 2018 (S1).

Nevertheless, the plants that died in 2018 (S4) were always positively associated with the hottest clusters, regardless of the number of clusters used (Figure 2s–v). Furthermore, plants that were heavily compromised, but still alive, in 2018 (S3) were usually abundant in the hottest cluster, although not always with a significant correlation for all numbers of clusters tested (Figure 2s–v). Lastly, the coldest clusters, KT1 and HT1, were consistently positively associated with plants that also maintained high vigour in 2018 (S1), and negatively correlated with plants that became diseased (S3 and S4) (Figure 2s–v).
