**1. Introduction**

Italy is, beside New Zealand, one of the two major kiwifruit producers and exporters worldwide, accounting for 15% of the total world demand on its own [1]. Since 2012, a major new disease known as Kiwifruit Decline (KD) has been threatening the future of kiwifruit production in Italy [2]. Up to now, KD has already destroyed more than 25% (~6600 ha) of the Italian kiwifruit growing area, and it has become a serious threat to the future of the crop in this country (Tacconi, personal communication updating [3]). To our knowledge, KD outbreaks have only been reported in Italy, but similar diseases

have also been described in other countries, including New Zealand [4], Japan [5], and more recently in Turkey [6–8].

The main damage caused by KD is the almost complete degradation of the white feeding roots. Coarse roots usually show red discolouration under the cortex, and the external cylinder easily detaches from the centre, creating the so-called "rat tail" symptom [2]. Symptoms on the root might not immediately a ffect the canopy, which may remain completely asymptomatic for an unknown amount of time, even if the root system is considerably deteriorated [3]. Usually, diseased plants go through an irreversible and fast wilting process, which is suddenly visible late in the vegetative season (July, August, northern hemisphere) when heat waves occur [9]. Once the wilting appears, plants usually lose all their leaves and die within a few weeks. In the following years, plants that manage to survive have considerably reduced canopy vigour, may show di fferent degrees of nutrition deficiencies, and never recover from their weakened condition [3]. KD spreads following an oil-spot pattern, with an astonishing speed that can compromise the whole orchard in only one or two growing seasons [2]. Currently, studies on KD are mostly focused on understanding the aetiology of the disease, which has not ye<sup>t</sup> been fully clarified, but it seems that an interaction between waterlogging and several soil-borne pathogens is necessary to induce the disease [2,10–13].

The timing and positioning of sampling are key factors for all aetiological studies. The lack of knowledge about the dynamics of root degradation severely compromises the quality of sampling activities, which cannot be properly scheduled. In particular, the core samples (feeding roots) that are collected for laboratory analysis of KD are usually irreversibly deteriorated, and almost useless for aetiological investigations once canopy symptoms appear [10]. Currently, there are no e fficient methods to survey orchards. Field surveying presents several limitations, mostly associated with the obstruction of inter-row spaces by shoots that reduce the field of view (One to three metres only, depending on plant vigour) and can increase the time required even just for observation of the canopies. In addition, there are no methods to predict disease outbreaks, because wilting appears abruptly, even within healthy plants that are apparently devoid of any warning signs [2,3,10]. The only way to confirm the presence of the disease before plant dieback is to observe the symptoms in the roots. This strategy, though, is highly unfeasible; firstly, because there are no reliable methods to select which plant must be uprooted, and secondly because it is time-consuming and requires hard work.

Remote sensing can be a real aid in monitoring activities, especially for field surveys. The high impact of the disease on plant vigour, if approached appropriately, can be exploited non-invasively to assess the spread of KD. Indeed, multispectral imagery has been widely exploited for the evaluation of canopy vigour indices [14], drought stress indices [15], plant pathogen detection [16] and, on kiwifruit, to predict the dry matter content of fruits from satellite data [17].

RGB cameras have also been largely adopted to estimate geometrical parameters of field crops [18–21], but their performance on plant vigour estimation is generally secondary to multispectral sensors, due to the higher correlation of near-infrared (NIR) wavelengths with plant biophysical and biochemical traits (e.g., chlorophyll content) [22,23]. However, for KD studies, RGB cameras have the advantage of representing the scene exactly as observed by the human eye, allowing experts to recognise the wilting symptoms, evaluate the disease spread and develop reference data for machine learning algorithms.

Finally, since KD causes fine root decay, the water absorption capabilities of kiwi vines are highly compromised [10]. Consequently, it is highly probable that the physiological response of the plant is similar to that induced by a drought stress: closure of the stomata and rising of the leaf temperature. It has been reported that kiwifruit vines are sensible to water-related stresses, and quickly close their stomata in response to drought conditions [24], severe root pruning [25] and even waterlogging [26,27]. The increase in leaf temperature can be significant in drought-stressed kiwi vines, and reported to be between +1 ◦C and +6 ◦C, depending on the exposure of the leaf to the sun, the sampling time and the intensity of the stress [28,29]. Leaf temperature might be the key to detect a compromised plant, which would be otherwise undetectable by the naked eye. Since leaves emit radiation according to

their temperature in the thermal infrared band (TIR; 3 to 15 μm) [23,30], this wavelength range has been widely used to estimate the water status of several plants [31–33], and to detect a number of plant diseases that influence leaf stomatal conductance and transpiration rates directly or indirectly [34–36]. Regarding kiwifruit, thermal imaging has been previously adopted for the early detection of another vascular disease caused by the bacterium *Pseudomonas syringae* pv. *actinidiae* [37].

Among the remote sensing technologies, drone-based surveys provide several advantages in KD studies: (i) they allow us to generate high resolution images that can be registered to image the whole field and quickly evaluate the health status of the plants; (ii) they can be easily equipped with di fferent sensors, collecting data in non-visible wavelengths such as the near infrared region (NIR) and thermal infrared; (iii) they can fly at lower altitudes, reducing the influence of the atmosphere; (iv) they are cost e fficient; and (v) they have a high scheduling flexibility, which is crucial for surveying plant diseases that are strictly related to climatic condition [38,39]. Considering KD, the latter aspect is probably the most important, since the survey must be performed not only with favourable weather conditions (radiation, clouds, wind), but also with optimal soil water content, since both waterlogging and drought stress can a ffect the transpiration rates [26,27,29].

Images acquired with these systems can be classified using either supervised or unsupervised machine learning algorithms. However, the application of supervised algorithms is limited, since it is di fficult to produce su fficient high-quality labelled data to create a robust supervised classifier that is able to account for all the variability associated with KD and kiwifruit orchards. The best method for labelling the data is the observation of the disease spread between two consecutive growing seasons, since protocols to rapidly and objectively detect diseased plant are still missing [2]. However, the swiftness of the disease spread make the use of supervised algorithms impractical. Indeed, by the time the classifier has been trained, the field could have already been irreversibly compromised. Unsupervised learning may prove to be an e ffective alternative, since it can reveal the underlying patterns and structure of the data without the need for labelling [40].

Thus, this study aims to understand the feasibility and reliability of thermal and multispectral imaging in predicting and assessing disease outbreaks, via an unsupervised classification approach. To the best of our knowledge, the present study is the first one that introduces remote sensing as a reliable tool for addressing aetiological and epidemiological studies of Kiwifruit Decline.
