Using Aerial Thermal Imagery to Evaluate Water Status in Vitis vinifera cv. Loureiro

The crop water stress index (CWSI) is a widely used analytical tool based on portable thermography. This method can be useful in replacing the traditional stem water potential method obtained with a Scholander chamber (PMS Model 600) because the latter is not feasible for large-scale studies due to the time involved and the fact that it is invasive and can cause damage to the plant. The present work had three objectives: (i) to understand if CWSI estimated using an aerial sensor can estimate the water status of the plant; (ii) to compare CWSI from aerial-thermographic and portable thermal cameras with stem water potential; (iii) to estimate the capacity of an unmanned aerial vehicle (UAV) to calculate and spatialize CWSI. Monitoring of CWSI (CWSIP) using a portable device was performed directly in the canopy, by measuring reference temperatures (Tdry, Twet, and canopy temperature (Tc)). Aerial CWSI calculation was performed using two models: (i) a simplified CWSI model (CWSIS), where the Tdry and Twet were estimated as the average of 1% of the extreme temperature, and (ii) an air temperature model (CWSITair) where air temperatures (Tair + 7 °C) were recorded as Tdry and in the Twet, considering the average of the lowest 33% of histogram values. In these two models, the Tc value corresponded to the temperature value in each pixel of the aerial thermal image. The results show that it was possible to estimate CWSI by calculating canopy temperatures and spatializing CWSI using aerial thermography. Of the two models, it was found that for CWSITair, CWSIS (R2 = 0.55) evaluated crop water stress better than stem water potential. The CWSIS had good correlation compared with the portable sensor (R2 = 0.58), and its application in field measurements is possible.


Introduction
In recent years, increased frequency of extremely high temperatures and low rainfall [1], together with the intensification of agricultural practices, has required that more attention be paid to the use of natural resources [2]. According to the most unfavorable scenario forecast by the Intergovernmental Panel Climate Change (IPCC) related to the representative concentration pathway (RCP) 8.5, water availability will decrease throughout the north of Portugal, and approximately 20% of the Vinho Verde Region, accompanied by a gradual increase in temperature [3]. In this scenario, grape quality and wine production [4,5] and the overall cycle of vines [6,7] will be affected by low soil water availability [3,5,6]. Strategies to achieve the full potential of a vineyard should include frequent monitoring of vine water status [7] and estimates of cultural evapotranspiration [8][9][10][11] to adopt the best irrigation practices.
The vineyard was planted in 2001 in a north-south orientation, 1103 p rootstock, with 3.0 m between rows, and 2.0 m between vines (1666 plant/ha) trained to a single upward cordon. It has a drip irrigation system with one lateral per row (self-compensating drippers with 4 L h −1 flow rate). The experimental design consisted of five blocks (B1, B2, B3, B4 and B5) of four rows of seven vines each measuring 1500 m 2 located 50 m from the beginning of a row ( Figure  1b). Only the vines of the two central lines were monitored, and this was conducted on three days of the year (DOY), i.e., day 182 (1 July), day 190 (9 July) and day 194 (13 July) for the following parameters: (i) soil water content (mm) using a capacitive probe (diviner); (ii) midday stem water potential (MPa) in four vines per block using a Scholander pressure chamber; (iii) the temperature of the canopy, using a portable thermal camera (FLIR e75, USA) in each of the 20 vines, on the three data recording dates for a total sample of 60 vines; and (iv) the temperature of the canopy using an aerial thermal camera (Zenmuse XT2, USA) supported by a UAV, (DJI Matrice 210, Frankfurt, Germany).

UAV Platform, Thermal Camera and Data Acquisition
The aerial images were collected using a DJI UAV, model Matrice 210, coupled with an aerial thermal camera [60,61] (Zenmuse XT2), with an 8 mm lens, resolution 640 × 512 pixels, frequency of 30 Hz, spectral range of 7.5-13.5 µm and a temperature range from −40 to 550 °C. The portable thermal camera (FLIR, e75) had a 17 mm lens, a resolution of 320 × 240 pixels, frequency of 30 Hz, spectral range of 7.5-14 µm and a temperature range from −20 to 120 °C. Sensors were used simultaneously. The images were taken at noon at the sampling site, and each pixel represented temperature in degrees Celsius. In the The vineyard was planted in 2001 in a north-south orientation, 1103 p rootstock, with 3.0 m between rows, and 2.0 m between vines (1666 plant/ha) trained to a single upward cordon. It has a drip irrigation system with one lateral per row (self-compensating drippers with 4 L h −1 flow rate).
The experimental design consisted of five blocks (B1, B2, B3, B4 and B5) of four rows of seven vines each measuring 1500 m 2 located 50 m from the beginning of a row (Figure 1b). Only the vines of the two central lines were monitored, and this was conducted on three days of the year (DOY), i.e., day 182 (1 July), day 190 (9 July) and day 194 (13 July) for the following parameters: (i) soil water content (mm) using a capacitive probe (diviner); (ii) midday stem water potential (MPa) in four vines per block using a Scholander pressure chamber; (iii) the temperature of the canopy, using a portable thermal camera (FLIR e75, USA) in each of the 20 vines, on the three data recording dates for a total sample of 60 vines; and (iv) the temperature of the canopy using an aerial thermal camera (Zenmuse XT2, USA) supported by a UAV, (DJI Matrice 210, Frankfurt, Germany).

UAV Platform, Thermal Camera and Data Acquisition
The aerial images were collected using a DJI UAV, model Matrice 210, coupled with an aerial thermal camera [60,61] (Zenmuse XT2), with an 8 mm lens, resolution 640 × 512 pixels, frequency of 30 Hz, spectral range of 7.5-13.5 µm and a temperature range from −40 to 550 • C. The portable thermal camera (FLIR, e75) had a 17 mm lens, a resolution of 320 × 240 pixels, frequency of 30 Hz, spectral range of 7.5-14 µm and a temperature range from −20 to 120 • C. Sensors were used simultaneously. The images were taken at noon at the sampling site, and each pixel represented temperature in degrees Celsius. In the portable thermal camera, images were taken two meters from the canopy and perpendicular to the direction of the line. Aerial images were taken at a height of 60 m from groundwater potential, with a longitudinal and lateral overlap of 90% to obtain a photogrammetric mosaic ( Figure 2). All images were acquired on clear days with minimal wind, and were evaluated on site. In the same location, a PMS Model 600 pressure camera was used to monitor stem water potential, because this is a sensitive indicator for vine water status and is frequently monitored to drive irrigation management. portable thermal camera, images were taken two meters from the canopy and perpendicular to the direction of the line. Aerial images were taken at a height of 60 m from groundwater potential, with a longitudinal and lateral overlap of 90% to obtain a photogrammetric mosaic ( Figure 2). All images were acquired on clear days with minimal wind, and were evaluated on site. In the same location, a PMS Model 600 pressure camera was used to monitor stem water potential, because this is a sensitive indicator for vine water status and is frequently monitored to drive irrigation management.

Image Processing
Aerial thermal images analysed by Agisoft Metashape professional, Version 1.6.3 software to produce final orthophotos with a spatial resolution of 7 cm. To compare temperatures, the portable thermal camera was used to take images at the same time.
Temperature calibration parameters such as emissivity, distance to a target, reflectivity temperature and relative humidity were entered into the Flir Tools, Version 5.13.18031.2002 software [62]. Production of the thermal and RGB orthophotography of the work area was carried out in eight main phases according to the flow diagram shown in Figure 2. To support georeferencing and geometric correction of the aerial image, a set of 16 points was assigned as control with a Root Mean Square Error (RMSE) less than 0.5. It was necessary to build a dense cloud of points by which, through photogrammetric restitution, it was possible to derive an extensive set of points that had information regarding an image's latitude, longitude and altitude.
With this base, the texture of the image and the construction of a Digital Elevation Model (DEM) provided visual improvement and orthophotography production, by orthorectification, of the photogrammetric mosaic previously produced ( Figure 2).
In the final phase, segmentation of the RGB aerial image was carried out using IDRISI software. The objective was to differentiate and isolate the areas between the rows, which were occupied by herbaceous vegetation, from areas occupied by vines so that they could be processed separately. Soil area extraction was necessary for the CWSI calculation to be carried out only within the vineyard area. For this procedure, and to delimit the segmentation polygons, a similarity of about 40% between the pixels was considered for analysis ( Figure 2).

Calculation of Portable and Aerial CWSI
The calculation of the CWSI was based on the equation proposed by [13] and modified by [35] (Equation (1): where T c is the canopy temperature obtained from the thermal image, and T dry and T wet are the reference temperatures ( • C). For the portable CWSI (CWSI P ) the average T c and average T wet and T dry were calculated using Flir Tools software. The reference temperatures (T wet and T dry ; • C) were obtained by selecting two healthy leaves close to each other in the canopy. Vaseline was applied to both sides of the T dry leaf for 30 min before taking the readings to force the stomata to close, thereby preventing evapotranspiration, leading to a consequent increase in leaf temperature. For T wet , the leaf was sprayed with water, two minutes before taking readings to simulate maximum evapotranspiration rate [13,35].
CWSI obtained from the aerial thermal sensor was processed by two methods. The first was CWSI S based on the equation. The reference temperatures were obtained from the image histogram (Figure 3b), in which 99% of the values were assumed to correspond to normal temperatures and the remaining 1% to represent extremes, as discussed by [49], assuming that the T dry represented the maximum temperature and T wet the minimum temperature [49].
The air temperature method (CWSI Tair ) was also based on Equation (1) but differed in the method for obtaining temperatures. T c was the value of each pixel of the canopy temperature, T wet was calculated from the average of 33% of the minimum temperatures within the histogram (Figure 3b), and T dry was obtained from the air temperature plus 7 • C [55, [63][64][65] (Figure 2).

Soil Water Content
The water content in the soil (ϕ) was measured at different depths (from 10 to 80 cm) by a capacitive probe to determine if soil water availability influenced the water status of the crop.

CWSI Validation with Stem Water Potential
To analyze the ability of the aerial CWSI to estimate the water status of the vine, coefficients of determination (R 2 ) were calculated for each DOY and the data from the different CWSI (portable, simplified and air temperature), stem water potential (MPa) and CWSI P with CWSI (simplified and air temperature). To validate the CWSI from the aerial sensor, georeferencing of the vines in each block was carried out. At these sampling sites, a 50 cm buffer was created to obtain the average value of the interior CWSI for later correlation with the field sample.   Table 1 shows the temperatures recorded in the five blocks on the three sampling dates by aerial thermal imaging. The wide canopy temperature range (26-30 • C) on all data collection days, highlighted the variability within each block and the ability of infrared thermography to support analysis and temperature variation from the vine canopy. All data followed a normal distribution and had a homogeneous dispersion on the different dates. In practically all cases, the median and average temperatures overlapped. On different DOY, the temperatures captured by the aerial thermal camera relative to the canopy and the ground varied between 19.  From the temperature map (Figure 3a), on DOY 182 and 190 the temperatures inside each block were very close. Block 4 (B4) had the highest and block 5 (B5) had the lowest temperature. On DOY 194, block 1 (B1) had the highest and B5 the lowest temperature.

Calculation and Spatial Representation of Crop Water Stress Index (CWSI)
For this process, image segmentation was used, so that only the temperature affected by the vine canopy was used. It was based on this new image that the CWSI was processed (Figure 4). After obtaining the reference temperatures necessary to calculate CWSI, a spatial representation of the index was elaborated along the five blocks, where each pixel of the new image had its associated CWSI value, to determine if there was variability within and between blocks ( Figure 5). The values for the two methods showed a range between 0 and 1 over the three DOY: the higher the value, the greater the water stress. CWSIs visually and numerically smoothed the data, whereas the CWSITair, by normalizing Tdry, shifted the values close to the extremes. In general, it appears that the values of the two CWSIS increased over the DOY, with a greater emphasis on the first block. It was also possible to verify a variation in CWSI within the sample block by changing the values represented, which may indicate plants with greater water stress in some areas.    To calculate the CWSI (CWSI S and CWSI Tair ), different reference values were needed, depending on the method, to identify the difference in T wet and T dry values ( Table 2). The T wet for the CWSI S had values of 25.5, 26.1 and 25.5 • C for DOY 182,190,194, respectively, all of which were lower than the air temperature. The CWSI Tair , presented values of 28.9, 29.0 and 27.8 • C, all of which were higher than the air temperature. The CWSI S method resulted in lower T wet values compared to CWSI Tair , which may have been due to the lower percentage of the histogram used in the first method. The T dry for the CWSI S had values of 38.2, 38.4 and 36.0 • C, and for the CWSI Tair these were 35, 35 and 33 • C. The CWSI S method resulted in higher T dry values compared to the CWSI Tair method, probably because the air temperature method, by normalizing this indicator, made it more stable than in the simplified method, which depended on temperature variability existing within and between the blocks. After obtaining the reference temperatures necessary to calculate CWSI, a spatial representation of the index was elaborated along the five blocks, where each pixel of the new image had its associated CWSI value, to determine if there was variability within and between blocks ( Figure 5). The values for the two methods showed a range between 0 and 1 over the three DOY: the higher the value, the greater the water stress. CWSIs visually and numerically smoothed the data, whereas the CWSI Tair , by normalizing T dry , shifted the values close to the extremes. In general, it appears that the values of the two CWSI S increased over the DOY, with a greater emphasis on the first block. It was also possible to verify a variation in CWSI within the sample block by changing the values represented, which may indicate plants with greater water stress in some areas.
The CWSI P values of the sampling point, calculated for validation, varied in amplitude over the three dates from 0.36 to 0.70. The averages increased over the reading days, decreasing the CWSI P amplitude. The same happened with the aerial sensors: the CWSI S had an overall amplitude between 0.35 and 0.78, and the average values of the CWSI increased while the amplitude decreased. For the CWSI Tair , the amplitude increased from 0.45 to 0.88, while the average values of the CWSI increased from 0.57 to 0.72, decreasing its amplitude ( Table 3).
The portable and aerial CWSI values remained close among the three methods. There was a slight increase from DOY 182 to DOY 194, which was explained by the lower soil water content in the latter, confirmed by a slight increase in stem water potential (Table 3). It should be noted that the CWSI Tair had a very large coefficient of variation (0.42 to 0.44), whereas that of the CWSI s , on average, was less than half.   The soil water content varied between 110 and 199 mm for DOY 194 and DOY 182, respectively. The averages were 165, 159 and 144 mm for DOY 182, 190 and 194, respectively. All had low coefficients of variation (0.01-0.10). Thus, the water available in the soil decreased by about 20 mm overall (Table 4).

Correlation between Portable CWSI, Aerial CWSI and Stem Water Potential
To evaluate the relation between the proximal and remotely sensed water stress measurements with the vine water status, we determined the correlation between all the CWSI values (CWSI Tair , CWSI S and CWSI P ) and the stem water potential (Ψ st ).
For each studied DOY (182, 190 and 194) the CWSI P had the highest correlation, and the CWSI Tair had the lowest correlation, with the Ψ st (Table 5). These results are consistent with the global correlations, R 2 = 0.59 for the CWSI P and R 2 = 0.49 for the CWSI Tair . When we compared the CWSI, the highest correlation was obtained between CWSI S and CWSI P (R 2 = 0.58) ( Table 5). The fact that there were no changes in the determination coefficients when the UAV-supported thermal CWSI was incorporated (CWSI Tair and CWSI S ) suggested that the spatial structure represented the pattern of the stem water potential estimated from the simplified CWSI model.

Discussion
This study focused on the evaluation of thermal information to estimate the water status of a vineyard based on aerial and portable thermal images of the vine canopy. The method was fast and non-invasive compared to traditional methods such as using a pressure chamber to determine stem water potential. Using aerial thermal images captured with a UAV it was possible to produce an orthophotograph that comprised the temperature values of each pixel of the canopy. Segmentation of aerial RGB images isolated the canopy, discarded the area between the lines, facilitated calculation, and reduced thermal amplitudes that can influence the CWSI calculation model [16,17,49,55].
From analyzing the thermal images of three flights, it was possible to predict the water status of the plant (CWSI), taking into account the coefficient of determination when correlated with the stem water potential. Likewise, a moderate coefficient of determination was observed with the portable CWSI; however, this model did not allow a continuous surface of values. The results were in line with previous studies [17,49,55], but adding the ability to spatialize data, with the advantage of obtaining information from a vertical rather than a horizontal view.
Analyses of three days of capturing and processing data from portable thermal cameras and calculating respective CWSI P concerning the stem water potential (R 2 = 0.59) showed a moderate correlation, which agreed with the work developed by Garcia-Tejero et al. [35]. The results for each day, DOY 182 (R 2 = 0.69), DOY 190 (R 2 = 0.60) and DOY 194 (R 2 = 0.58) had moderate correlations, which enhanced estimates of plant water status from the CWSI P . According to [41], the variability observed between the stem water potential and the CWSI for different DOY can be affected by parameters related to the phenological phase of the crop. The CWSI P method, like use of the Scholander pressure chamber, is very expensive for determining an overall distribution because each reading only provides values for each plant in isolation.
Our results showed was possible to obtain crop water stress indices, calculated using date from two methods of aerial thermography, that made it possible to estimate and interpret the water status of the plant. Therefore, the method is as accurate as the Scholander pressure chamber method.
Given that it only needs to monitor air temperature during UAV flight, the calculation of the CWSI Tair is a quick method for obtaining thermal data and, consequently, calculation of the CWSI. With this method, a moderate coefficient of determination was obtained with the stem water potential (R 2 = 0.49). It was observed that DOY 182 (R 2 = 0.49), DOY 190 (R 2 = 0.53) and DOY (R 2 = 0.52), which had moderate correlations, were in agreement with the results obtained by [55].
The CWSI S calculation method required the least amount of time because it did not require field data. The correlation between the CWSI S and stem water potential (R 2 = 0.55) showed a moderate correlation, which was in line with the findings of [16,49]. When the analysis was aggregated by days, DOY 182 (R 2 = 0.60), DOY 190 (R 2 = 0.59) and DOY (R 2 = 0.54) had moderate correlations and with approximations superior to that of CWSI Tair .

Conclusions
In this work, the development of two CWSI spatialization methods was explored, and the results were compared with those obtained by measuring stem water potential. It was found that CWSI S could potentially estimate the water status of the vine using the reference temperatures (T wet and T dry ) from a histogram of image temperatures after segmentation.
The ability to obtain and spatialize thermal data has become extremely important, given that it is possible to image the plot as a whole, not just isolated plants. Furthermore, it is a non-invasive method, which allows taking a large number of measurements without weakening the plant in its phenological cycle. It also permits the collection and analysis of data on a large scale, and in a shorter time, whereas the CWSI S method more closely estimates water stress of the crop. The CWSI P method produced viable results, which were very close to those of the CWSI S , but it did not allow spatialization of the data.
Vegetation cover has a great influence on soil water preservation and thermal regulation. Where there is no turfgrass on the plot, the probability of obtaining a bimodal histogram can influence the CWSI results when applying the CWSI S or CWSI Tair methods.
Future research should consider information focused on leaf area, phenological phase, variety and age.