**4. Discussion**

Our results sugges<sup>t</sup> that UAVs combined with multispectral and thermal imaging can be useful tools for scouting activities at a field scale in kiwifruit orchards, and can greatly improve monitoring activities. For instance, the sole observation of an RGB map is highly valuable, because it provides experts with an overview of the entire field, improving their awareness of the disease spread (Figure A1 in Appendix A).

Multispectral data can be used to speed up the assessment of the symptomatic plants due to their strong correlation with plant vigour. By identifying dead or highly compromised plants, it was certainly possible to precisely delimit and monitor the spread of the disease, while the detection of very vigorous plants made it possible to identify homogeneous areas, where other analyses should be focused for the prediction of disease outbreaks. Our findings are in line with those of other studies that used NIR (850–1700 nm), green (495−570 nm) and red (620–750 nm) wavelengths to develop vegetation indices for the estimation of plant biophysical traits [14,53–55]. Indeed, healthy plants are usually characterised by a higher reflectance in the NIR bands [56]. Studies of multispectral data regarding soil-borne diseases identified NIR bands as an important factor for the detection of plant vigour and the estimation of disease severity indices [57–61]. However, as observed in our study, multispectral data seem to have a poor predictive capability for this kind of diseases, and they are usually useful once symptoms are visible [34,36,57–61].

The lack of predictive capability within the multispectral data can be o ffset by thermal data. The canopy temperature was found to be a reliable predictor of plant health status, being able to indicate the spread of the disease one year before the actual outbreak. It is interesting to note that the best prediction metrics were obtained precisely for plants that were completely asymptomatic in 2017, far exceeding the predictive capability of any expert. Indeed, these plants were still apparently healthy in both October 2017 and in the first half of the vegetative season in 2018. Until June 2018, they produced a wide canopy, but suddenly after the first heat waves of July, they showed scorching, leaf drop and, in the worst cases, complete defoliation. Results obtained from the clustering of thermal data confirmed that plant responses to KD might be similar to those induced by drought stress, as also observed for other soil-borne diseases [23,34,36]. Indeed, just like abiotic factors, pathogens may a ffect the stomatal response by influencing the temperature gradient between the plant tissue and the air [62]. Soil-borne diseases cause the reduction of water absorption translocation and transpiration functions as a net result of their infection [63]. These alterations induce a closure of the stomata, and consequently the increase of leaf temperature as a consequence of the reduced evaporative cooling e ffect [24,63,64]. Indeed, for soil-borne diseases, canopy temperature has been shown to be particularly useful in detecting compromised plants, even during the early stages of infection when visual symptoms were invisible [23,34,36,63,64]. It can be speculated that the multispectral data probably failed to predict disease outbreaks because KD does not cause internal structure modifications until a critical point is reached, at which time the root system can no longer support the plant's transpiration rate [65]. It seems that the kiwifruit has a very high root/canopy ratio [25,66], and thus can cope with a substantial (80%) loss of the root system before the growth of shoots and leaves is a ffected [25]. Conversely, a justification for the predictive capability of canopy temperature data may reside in the physiological response of kiwi vines to drought stress or root loss, which induce a rapid reduction in gas exchange fluxes and consequently an increase in leaf temperature [24,25,67]. Finally, it should also be noted that even flooding conditions can induce a reduction in the transpiration rates of kiwi vines [27], and thus the water content of the soil cannot be overlooked during temperature data acquisition for KD detection.

Based on our results, the use of unsupervised clustering can be a reliable method for quickly exploring and identifying sampling areas that are suitable for aetiological studies. By using multispectral data, symptomatic plants can be easily identified, simplifying the disease assessment process. The clustering of thermal data is even more useful, since the early detection of diseased plants allows us not only to identify the best areas where samples for laboratory analysis should be collected, but also to observe the evolution of symptoms over time, and improve the understanding of root degradation dynamics. However, we believe that the unsupervised clustering of remote sensing data might also be implemented for other diseases. To the best of our knowledge, unsupervised clustering has not been used previously for forecasting purposes, although it has been used for the assessment of cotton root rot [68], for structuring the aggressiveness of fungal infections on peas [69], and for the segmentation of plant backgrounds for weed detection [70]. With our experimental set up, K-means gave better results than hierarchical clustering in both assessing and predicting the disease spread. However, it cannot be ignored that under other conditions the performances of these two methods might di ffer. The biggest drawback of the unsupervised classification methods is the need for an expert to assign meaning to each cluster. Nevertheless, for KD, cluster interpretation can be easily achieved by overlapping cluster results with an RGB orthomosaic. The use of supervised machine learning algorithms might be a solution to the problem of directly classifying plant health status, but when the aetiological background is not clear or the labelling is impractical (such as in the KD syndrome), the development of a reliable training dataset is di fficult because rapid tests to discriminate between diseased and healthy plants are currently unavailable. Therefore, unsupervised clustering is an appropriate method of seeing or finding groups within unknown data when labels are not available, especially when we do not know what kind of sensor data is needed to find patterns and groups [71].

The study was limited by low cost technologies, and it was performed by only analysing shifts that occurred between two seasons, so several future improvements can be envisaged with this approach. A better segmentation algorithm, based on sensor fusion processes and improved geometric accuracy of the orthomosaics, is needed to remove background noise and to improve the quality of the analysis. In this regard, a similar approach was successfully adopted to detect olive plants infected by *Verticillium* wilt, using airborne data captured simultaneously with thermal, fluorescence and hyperspectral detectors [34]. Moreover, the advantages derived from sensor fusion are already evident in fruit safety and quality control studies [72–74]. Segmentation could also be improved by using automatic tree segmentation with object-based image analysis [75]. In our study, the low-cost multispectral sensors available precluded the possibility of correctly segmenting the inter-row grass from the kiwifruit canopy. Nevertheless, testing flexible and low-cost methods is important to ensure the real application of remote sensing technologies to the field scale. A better understanding of the relationship between temperature and plant health status is also needed. Models that are able to predict leaf temperature using data from weather stations (e.g., air temperature, air relative humidity, radiation, soil water content) should be developed to set more objective thresholds for discriminating healthy from diseased plants, and thus use temperature as a more reliable predictor. Temporal analysis of simple vegetation indices might also improve our understanding of the disease dynamics and increase the detection accuracy of KD. Studies on a wider range of wavelengths are also needed to identify the best bands for disease detection and prediction. In particular, it would be interesting to study the possible correlations between KD appearance and short-wave infrared wavelengths (1100 to 2500 nm), as these bands are influenced by leaf chemical composition and water content [76,77].
