*2.2. Image Acquisition*

The UAV surveys were carried out on 8 August 2017 and 18 July 2018, respectively at 11.00 a.m. and 10.00 a.m. on hot sunny days. The timing of the flight was based on a previous report suggesting this time frame as one that would maximise the di fferences between well-fed and drought-stressed

plants [28]. Weather condition were favourable for the acquisition: no clouds, wind speed <2 km/h, relative air humidity ~44%, radiation ~2900 MJ/m<sup>2</sup> and temperature ~29 ◦C. The acquisitions were performed with a single flight lasting less than 20 min, including take-o ff and landing. The flights were undertaken, respectively, two days after an irrigation event in 2017 and three days after a heavy rainfall (20 mm) in 2018, to exclude water deficit from the variables influencing the canopy temperature and to make sure that the soil water content was not a limiting factor. The gravimetric soil moisture was measured the day before the flight at nine locations in the surveyed area. Three sampling areas were selected inside an apparently healthy area, three in a heavily compromised area and three in a transition area between these two. The surveys were performed if the soil humidity was between 70% and 80% of the field capacity, representing an optimal condition for kiwifruit growth [41]. The field capacity value was derived from the soil texture using soil-plant-atmosphere-water field and pond hydrology (SPAW) models [42]. SPAW models are pedo-transfer functions, which use soil texture (percentage of sand, silt and cay) and percentages of soil organic matter to estimate the soil hydrodynamics properties relevant to agronomical practices, such as volumetric field capacity and wilting point [43].

The UAV flights were performed with a hexa-copter using Real Time Kinematic–Global Navigation Satellite System (RTK-GNSS) and equipped with a gimbal system on two axes (Adorn-technologies, Italy). Thermal, Multispectral and RGB images were acquired simultaneously at a flying height of 35 m, at a flying speed of 5 m/s. Frame rate was adjusted for each sensor in order to achieve a forward image overlap of 80%, while sidelap overlap was 80%, 88% and 90% between thermal, RGB and multispectral images, respectively. In 2017, the acquisition combined the three sensors: GoPro Hero 4 (GoPro, San Mateo, CA, USA) as the RGB camera, Tetracam ADC snap (Tetracam, Chatsworth, LA, USA) as multispectral sensor and FLIR TAU2 640 for thermal imaging (FLIR System, Wilsonville, OR, USA). In 2018 the RGB and multispectral sensors were exchanged for Sequoia+ sensors (Parrot, Paris, France). For detailed information about the sensors used see Table 1. Tetracam ADC snap is a low-cost multispectral sensor, modified from an RGB sensor equipped with a blue absorbing glass filter to eliminate the blue sensitivity, and to use the blue pixels in the sensor to measure NIR bands. FLIR TAU2 640 is a thermal camera capable of detecting temperature di fferences of ±0.05 ◦C with a sensibility of 2 ◦C. It has the advantage of performing automatic flat-field corrections (FFC) while recording, to account for microbolometer variation and remove artefacts from the 2D images. For this study, the FFC was performed every 100 frames. FLIR TAU2 640 also measures the internal temperature of the camera to frequently update the non-uniformity coe fficients used to convert raw data into radiometric values [44]. Air temperature and relative humidity were measured before the flight and inserted in the camera settings, together with the plant emissivity value, which was set to 0.98 as suggested in Maes et al. (2014) [37]. Sequoia+ is a camera equipped with a sunshine sensor for image calibration, an RGB camera and four single-band cameras measuring reflectance in the green, red, red-edge and near-infrared (NIR) bands.

### *2.3. Pre-Processing: Creation of Orthomosaics and Alignment Checking*

An overall visualisation of the workflow is presented in Figure 1. The analysis was performed using all data (RGB, multispectral and thermal) acquired in 2017, while in 2018, only RGB data were used to evaluate the disease spread, as detailed below. After the acquisition, raw RGB and multispectral data were converted to reflectance and thus normalised from 0 to 1 using Pix4D software (Pix4D., Lausanne, Switzerland). In 2017, calibration was performed using a white Teflon plate included in the Tetracam ADC snap packaging as white reference, and as dark reference a picture acquired with the shutter closed. White and dark references were acquired at the beginning of the flight. Raw measurements were then calibrated as reported in Elmasry et al. (2012) [45]. In 2018, RGB images were instead normalised following the manufacture-recommended methods for the Sequoia+ sensor, using the one-point calibration plus sunshine sensor method as reported in Poncet et al. (2019) [46]. Thermal data was not calibrated; however, rather than absolute leaf temperature, it was of higher interest in the relative di fference occurring between plants with di fferent health statuses. Temperature

values (◦C) were derived from the TIR wavelengths using FLIR-studio (FLIR System Inc., USA). Stitching, georeferencing and orthorectification were performed independently for each set of data (RGB, multispectral and thermal data from 2017 and RGB from 2018) with Pix4D. Image processing is based on structure from motion (SfM) algorithms to reconstruct the three-dimensional scene based on shared features detected across the images. The four orthomosaics (multispectral and thermal from 2017; RGB from 2017 and 2018) were combined in Qgis [47], and planting lines and control points were drawn separately over each orthomosaic to check alignment of the datasets. A total of 11 control points were selected as follows: four points in both the northern and southern part of the field, and three in the orchard. Errors between planting line slope were minimal (on average 0.01◦), while the mean error between control points was 4.97 cm horizontally and 4.51 cm vertically, suggesting that no major deformation occurred during the stitching. Alignment was deemed good enough for the analysis purposes, therefore co-registration of the orthomosaics was not performed. Errors linked to temperature drifts were checked at two pairs of points: two points diagonally collected on a grass-free compacted portion of the earth roads delimiting the field on the northeast and southwest sides; and two points were collected on the water contained in an irrigation canal which flanks the north side of the field. Temperature values for each point were extracted from the thermal orthomosaic and respectively compared to check for temperature drifts, which were 0.2 ◦C on the bare soil and 0.04 ◦C on the water.


