The study took place during the 2017–2018 growing season (November–January) at a field located within the King Abdulaziz University Agricultural Research Station in Hada Al-Sham, Saudi Arabia (21° 47ʹ48ʺN, 39° 43ʹ35ʺE, Figure 1). The field was divided into four separate plots, each approximately 40m x 40m, with 15 rows of 20 tomato plants. Two plots were established as controls, with freshwater irrigation (approx. 900–1,000ppm NaCl). The other two plots were irrigated twice daily (except Fridays) with saline water of increasing concentrations (Figure 1). In developing the diversity panel, 200 accessions (199 wild Solanum pimpinellifolium and one commercial S. lycopersicum) were screened for salt tolerance via randomized planting of three replications of each accession for each treatment (i.e., three salt-treated and three control plants per accession, producing a total of 1,200 plants). At the beginning of November, 1,200 seedlings were transplanted into the field (after 1month of greenhouse growth), with harvesting taking place between 16 and 26 January (Figure 1). Additional details of the site and trial design information can be found in Aragon et al. (2020)) and Johansen et al. (2019). The focus of this study was to understand whether TIR data can identify differences in plant stress between control and salt-treated Solanum pimpinellifolium plants.

A weather station was installed toward the middle of the field (Figure 1) to collect meteorological data throughout the growing season. T_a and relative humidity (RH) were recorded every minute at 2.3m above ground level (AGL) using an HMP155 humidity and temperature probe (Vaisala, Helsinki, Finland), from which the VPD was calculated (May et al., 2008). Wind speed and direction were also recorded every minute at 2.2m AGL with a WindSonic anemometer (Gill, Hampshire, United Kingdom). Meteorological data were augmented by four distributed stations in each of the plots that measured point-scale thermal infrared temperature via an Apogee radiometer (SI-111, Apogee, Logan, United States), which facilitates interpretation of the UAV-collected TIR data (see locations in Figure 1). The Apogee sensors were installed in each plot approximately 1m above a plant, representing a footprint of around 0.40m². As our study occurred in an arid desert environment, sandstorms impacted the site on December 8 and 16, 2017, and January 4 and 8–10, 2018. To combat the impact of the sandstorms on results, field staff washed the plants with non-saline water after each event.

An object-based image analysis (OBIA) approach was applied to the RGB orthomosaics to identify each plant’s extent in the TIR orthomosaic (Figure 2, Step 1). A full description of the workflow used to create the OBIA RGB delineations can be found in Johansen et al. (2019). In order to omit pixels within the delineated plants that were associated with white identification tags (attached to individual plants), pixels with blue reflectance above the 99.5th percentile were removed. Next, green vegetation was discriminated within the delineated objects by applying a k-means clustering to the green–red vegetation index (GRVI) (Motohka et al., 2010). The GRVI was calculated as per Eq. 1 using the collected RGB data, as this index produced good results in Johansen et al. (2019; Figure 2, Step 5a). We applied a k-mean unsupervised approach run with two clusters, k-means++ initialization, ten different centroid seeds, and a maximum iteration of 300 in the scikit-learn package of the Python 3.5 software (Pedregosa et al., 2011). We set two clusters since the plants had already been delineated with the OBIA approach, and we were merely interested in discriminating vegetation from the sandy background, which had distinct spectral characteristics. For the classification of vegetation, a threshold value of GRVI > 0 was also used (Motohka et al., 2010). The distribution of temperature for vegetation classified with both the k-means approach and the GRVI threshold was subsequently compared to determine the most suitable approach (Figure 2, Step 6).

Even after the GRVI mask was applied, there were a number of pixels with T_p that was considerably higher than that expected for vegetation, indicating mixed pixel or classification issues. Therefore, the approach of Rud et al. (2014) was adopted to determine a realistic estimate for the maximum deviation of T_p from T_a. In this case, a threshold of T_a+9°C was used after analyzing both the field-installed Apogee radiometer and UAV data for the growing season. Subsequently, any pixels that had positive GRVI values but were warmer than T_a+9°C were removed to allow the formation of the final vegetation mask, from which T_p was ultimately retrieved (see Figures 2, 3).

Following Poblete et al. (2018), a k-mean clustering using a five-cluster a priori and k-means++ initialization was also applied on the blue band in order to differentiate sunlit and shaded areas of the tomato plants. The selection of a five-cluster a priori was also verified by applying the elbow method to identify the optimum number of clusters (Thorndike, 1953). The maximum blue reflectance value of the first cluster was used as the threshold above which vegetation was identified as sunlit. From the final vegetation mask (Figure 2), we retrieved descriptive statistics of T_p (minimum, maximum, mean, median, standard deviation, and pixel count). If the vegetation mask had a pixel count of <10% of the original number of pixels in the OBIA delineation, we removed the plant from further analysis, assuming the plant was dead or that the canopy had senesced and was thus too sparse for accurate T_p retrieval.

To consistently compare plant temperature across the five flights, we calculated the deviation of T_p from ambient temperature (dT_p=T_p - T_a), a measure often used in field phenotyping studies of heat tolerance (Balota et al., 2007). To further normalize for meteorological conditions, we calculated the CWSI using Eq (2) (Idso et al., 1981; Jackson et al., 1981), where dT_p is the actual difference between T_p and T_a, dT_pLL is the lower limit that represents transpiration at the maximum rate (theoretically a non-stressed plant cooled via latent heat exchange), and dT_pUL is the upper limit that represents a halt in transpiration (theoretically a stressed plant, where sensible heat exchange determines T_p).

Traditionally, there have been two ways to derive these transpiration baselines: empirically (CWSI_E) and theoretically (CWSI_T), with many researchers reviewing and debating the various limitations of each (Gardner et al., 1992; Maes and Steppe, 2012; Gerhards et al., 2019). The main limitation of the CWSI_T is the complex meteorological data required to solve the energy balance equation. CWSI_E has seen broad application, as it only needs three variables (T_a, T_p, and RH) to be calculated. However, the CWSI_E approach requires dT_p and VPD measurements to be collected across an entire growing season to calculate robust baselines (Gardner et al., 1992). More recently, UAV studies have proposed a simplified statistical method (CWSI_S) using the temperature distribution in the image scene to set the baselines (Gonzalez-Dugo et al., 2013; Rud et al., 2014; Bian et al., 2019). This simplified approach is appealing, as it only requires measurements of T_a, which facilitates applications in precision agriculture (Cohen et al., 2017). However, both stressed and non-stressed plants need to be present in the imagery using the simplified approach.

As our study occurred in Saudi Arabia, where there is a paucity of studies applying the CWSI, we tested all three approaches. For CWSI_E, we calculated the baselines using the intercept and slope values for tomato plants in Idso (1982). For CWSI_T, we calculated dT_pLL as presented in O’Shaughnessy et al. (2011). As the calculation of dT_pUL in CWSI_T is error-prone due to the estimation requirements of aerodynamic resistance and roughness length (Idso et al., 1981), we did not calculate it. Instead, we adopted T_a+9°C as an estimate for dT_pUL (see Retrieving Plant Temperature). For the simplified statistical approach (CWSI_S), we examined the T_p histogram distribution and set dT_{p LL} as the mean of the lowest 5% of plant temperatures in the control plots, while dT_pUL was set as T_a+9°C. (Meron et al., 2013; Rud et al., 2014; Bian et al., 2019).

We applied a standard independent two-sample T-test (α=0.01) in the SciPy package of the Python 3.5 software language (Virtanen et al., 2020) to assess whether there was a difference in thermal indices between salt-treated and control plots. To understand the change in thermal indices across the season, we calculated the percentage difference between the treatments and plotted the thermal indices as a box plot for each treatment to determine the optimum time to detect stress.

A field-based visual assessment of plants in poor condition was performed on January 4, which identified 30 dead plants. To assess whether T_p could be used to identify the dead plants earlier in the season and prior to senescence, 30 healthy plants were also selected from a visual assessment of the January 7 RGB data, with those plants distributed across the two control and two salt plots. That allowed comparison of the plants from the two groups, i.e., healthy and dead in the beginning of January, to determine whether T_p could be used for early detection of plant stress, while all plants were still green in December.

To determine the best approach to discriminate vegetation in the TIR orthomosaics to retrieve T_p, pixel-based temperature distributions within all tomato plants in the field trial were plotted. The presence of pixel-based temperatures >50°C (i.e., approximately T_a+20°C) within the OBIA delineations (Figure 4) indicated that some pixels represented soil or non-photosynthetic vegetation. When pixel-based temperature was retrieved using k-means clustering of the GRVI with a two-cluster a priori to separate background and vegetation, the frequency of pixels with temperatures >40°C reduced significantly (Figure 4). Therefore, it was assumed that this method was predominantly retrieving temperature from pixels representing vegetation rather than a mixed pixel response. A limitation of the k-means classification was attributed to vegetation being discriminated with a dynamic threshold of the GRVI value for the different campaigns to separate the two classes (Table 2), making a multi-temporal comparison of T_p challenging. Using a fixed threshold of GRVI > 0 to discriminate vegetation produced a similar temperature distribution across the five campaigns to that of the k-mean approach (Figure 4). However, the frequency of pixels with positive GRVI values decreased as the percentage of senesced vegetation increased. For example, the k-mean threshold for GRVI that separates vegetation and background was 0.02 on December 6. However, as non-photosynthetic vegetation increased, the threshold became −0.04 by January 14, which was the date exhibiting the largest difference between the two approaches in the number of retrieved vegetation pixels (Table 2). As a consistent comparison across the five flight dates was of most interest, a fixed threshold of GRVI > 0 was adopted for the final mask to retrieve T_p. However, a flexible clustering approach may produce better discrimination for single campaigns, which can be seen in the reduced number of pixels >40°C in the k-mean approach on December 6 (Figure 4).

As shown in Figure 4, the number of plant pixels increased through the growing season, peaking on January 7 with a subsequent reduction due to increasing plant senescence prior to harvest. Counter to this trend was the reduction in the number of vegetation pixels on December 20. The fact that this occurred in both the OBIA and GRVI retrievals suggests that the decline may be attributed to the plant damage and decrease in plant area caused by a sandstorm before the UAV capture (Johansen et al., 2019).

There is a tendency toward a negative relationship between GRVI and T_p, as increased GRVI values (greenness) result in T_p decreases due to latent heat exchange during transpiration. In our study, this trend held within the OBIA delineations, which included background soil and non-photosynthetic vegetation (Figure 5). However, there was no clear relationship between T_p and GRVI for GRVI > 0. The large range in T_p values for pixels with GRVI > 0 and the fact that there were pixels with positive GRVI values that have unrealistically high temperatures for vegetation demonstrated that the GRVI co-registration method did not fully resolve mixed pixel issues. Therefore, we set a more realistic threshold of T_a+9°C for the maximum deviation of T_p from T_a to mask pixels further. The need for the T_a+9°C threshold is shown with the reduction in the maximum standard deviation (σ) of T_p before and after the threshold was applied (Table 2). The mean of the maximum σ of T_p was 7°C for the five dates with GRVI > 0 but decreased to 2.9°C with the GRVI > 0 and T_p<T_a+9°C (Table 2). The drop in the σ of T_p indicates that GRVI > 0 and T_p<T_a+9°C effectively classified vegetation pixels and omitted background and mixed pixels, which is essential to ensure confidence that changes in T_p are an indication of a response to salt stress.

The number of plants from which T_p was able to be retrieved with the final vegetation mask (GRVI > 0 and T_p<T_a+9°C) compared to the number of plants as identified with the initial OBIA RGB delineation is shown in Table 3. As the growing season progressed, the sample size of the salt and control plots started to differ due to increased deterioration of plant condition in the salt plots based on the GRVI < 0 and T_p>T_a+9°C thresholds. Note also that T_p was extracted from more plants on December 6 than November 16, due to the small plant size of the initial vegetative growth stage, as well as and soil background effects (i.e., the T_a+9°C threshold).

To determine differences in plant response to either fresh or saline water irrigation, we assessed the deviation of T_p from the ambient temperature in both the salt and control plots. As shown in Figure 7, the mean temperature of tomato plants in the salt-treated plots consistently deviated from the ambient temperature more than the control plots across all five collection dates. The mean dT_p was above 5°C in both the salt-treated and control plots during the first collection on November 16, indicating that the plants may have been too small or sparse for accurate T_p retrieval. For instance, the mean plant area based on the OBIA RGB delineation was 0.06m² on November 16, but increased to 0.42m² by December 6. From December 6 to January 14, mean dT_p increased from 2.2 to 4.1°C in the control plots and from 3.6 to 4.7°C in the salt plots, demonstrating that the salt treatment led plants to have a higher T_p above the ambient temperature (Figure 7). The biggest difference in dT_p between salt and control plots occurred on December 20, with a difference of 1.3°C. Interestingly, on this day, plants also had the smallest deviation from T_a, with only one outlier in the control plot exceeding 6°C. The UAV flight on December 20 had a higher VPD (atmospheric demand for water) than on December 6 and January 7 and 14. Often, increasing VPD can lead to an initial increase in stomatal conductance, which decreases as the plant regulates its water exchange (Damour et al., 2010). The influence of VPD on tomato stomatal conductance may have caused the smaller dT_p values for this date and may also be contributing to the larger dT_p difference between salt-treated and control plants (Patanè, 2011). The difference in dT_p between treatments was less apparent on January 14 (4days before harvest), which may have been the result of plant aging and senescence being a larger factor in determining T_p than salt stress, as will be discussed in UAV-Derived Plant Temperature.

In order to compare results across the data collections, T_p had to be normalized for the variable weather conditions. To do this, the CWSI was calculated in three ways, as presented in Identifying Plant Stress and Calculating Thermal Indices (also see Supplementary Figure S1). Here, we only discuss CWSI_S, as it only required measurements of T_a and showed similar characteristics to CWSI_E and CWSI_T (Also, a full season of accurate daily dT_p was not available to calculate robust local transpiration baselines.) A smaller difference in CWSI_S between the control and salt-treated plots occurred on January 14 compared to the preceding dates. The smaller difference in CWSI_S between treatments closer to harvest suggests that T_p was better at discriminating stress between the fruit formation and ripening/mature stages (Figure 1), when plants in both plots had more developed canopies. From December 6 to January 7, mean CWSI_S in the control plots ranged between 0.23 and 0.27, whereas the salt plots ranged from 0.36 to 0.44, indicating that CWSI_S>~0.36 may be an indicator of stress. It is worth noting that CWSIs values <0 represent plants that are cooler than the mean of the lowest 5% of plant temperatures in the control plots, which was used to set the lower limit in the CWSI that represents transpiration at the maximum rate. As CWSIs was overestimated if T_p was retrieved from non-vegetation surfaces (Irmak et al., 2000), we omitted CWSI_S values for November 16 due to large dT_p values on that date, which represented T_p retrievals integrated the soil background.

It is apparent from Figures 7, 8 that there is a large range in T_p (and consequently dT_p) and CWSI_S values within both the control and salt plots, which may be due to different stomatal responses to stress in each of the 200 accessions, as well as spatial variations within the trial. The spatial variations are plotted in Figure 9, with individual CWSI_S shown for both the control and salt treatment for December 20, 2017, and January 7, 2018, which represented the time from fruit formation to mature, ripe fruit. As can be seen, there is a clear tendency for higher CWSI_S values in the two salt treatments, relative to the control, with a larger number of plants with CWSIs values >0.35 in the salt-treated plots. For instance, on December 20, only 19% of control plants had a CWSI_S>0.35, compared to 57% for the salt-treated plants. On January 7, the proportion of plants with CWSI_S>0.35 for the control and salt plots increased to 24 and 68%, respectively (Figure 9). It is, of course, important to recognize that spatial variability in real-world trials is more than just a function of plant stress, with other soil and environmental factors playing a role. However, while not all aspects of the spatial variation (e.g., the December sandstorms with northeasterly winds) in CWSIs observed in Figure 8 can be attributed to salt-induced stress alone, Figure 9 provides some additional insights to help interpret the influence of irrigation treatments.

A field-based assessment of plant condition was undertaken on January 4, with 30 plants identified as dead. An equivalent number of healthy plants were separately identified from the RGB imagery collected on January 7. The CWSI_S values for plants in the healthy and dead categories are shown on December 6 in Figure 10 to understand whether CWSIs values measured earlier in the season were indicative of the plant condition in early January.

Plants in the salt plots that were dead by January 4, but in good condition on December 6, generally had higher CWSIs values than those control plants that were still healthy at the beginning of January (Figure 10). Of the plants that were classified as healthy, the ones in the control plots exhibited lower CWSI_S values than in the salt plots (median=0.46 and 0.23, respectively). Interestingly to note is that for salt-irrigated plants on December 6, the difference in median CWSI_S values between plants that were dead and healthy by the beginning of January (0.57 and 0.46, respectively) is much smaller than for the control plants (0.47 and 0.23, respectively). This is most likely because the salt irrigation caused some level of plant stress early in the growing season, i.e., December 6, irrespective of plant appearance. These differences in CWSI_S values on December 6 indicate that at least some plants that appeared green and visibly healthy with GRVI > 0 and T_p<T_a+9°C showed early stress warning signs with high CWSIs values almost a month prior to plant death.

Here, we explored the retrieval of T_p from a UAV TIR camera in a tomato field trial, demonstrating that it is possible to detect differences between salt-treated and control plants, which may help identify salt-tolerant tomato germplasm in future research. In our study, T_p was retrieved where GRVI > 0 and setting a maximum pixel threshold of T_a+9°C. The latter condition was required because the presence of pixels with T_p>T_a+20°C in the OBIA plant delineation showed that object-based methods alone are insufficient to retrieve accurate T_p, at least from the tomato plants explored herein. This finding aligns with Cohen et al. (2017), who also suggest that while object-based approaches work well for tree crops, they fail to retrieve T_p from field crops due to their less defined canopy structure. We observed that even when T_p is extracted from pixels with GRVI > 0, temperatures that are unrealistically high for vegetation still occurred, demonstrating that the use of GRVI alone does not fully resolve mixed pixel issues. Our findings align with recent UAV TIR studies that could not eliminate all mixed pixels. For example, Zhang et al. (2019) used red and green reflectance together with TIR data to retrieve T_p for a maize crop and concluded that better methods for eliminating mixed pixels are required to facilitate accurate extraction.

In our study, the mixed pixel issues were alleviated by combining RGB data with this empirical method (i.e., T_a+9°C), which estimates the maximum temperature possible for non-transpiring vegetation. Researchers commonly report this empirical upper baseline in studies of drought stress for inclusion in CWSI calculations, e.g., T_a+5°C in cotton (Cohen et al., 2005), T_a+7°C in potato (Rud et al., 2014), and T_a+5°C in wheat (Jackson, 1982) have all been used. The fact that our upper baseline was larger than those published could be attributed to the higher solar radiation and T_a of the arid field site or potentially an extreme isohydric behavior (Han et al., 2020), with closed stomata required to maintain turgor. As the field installed Apogee TIR radiometers used for setting the T_a+9°C threshold make an integrated measurement of T_p from their field of view, vegetation movement driven by wind may have occasionally led to the integration of soil temperature, but it was not possible to fully resolve or remove the impact of soil background (Aubrecht et al., 2016).

The successful retrieval of T_p using a co-registration approach between the RGB and TIR imagery was dependent on good pixel alignment of the whole study area (Meron et al., 2013). While the datasets in the study were collected with two different sensors (Zenmuse X3 and TeAx 640) having differing resolutions and viewing geometries, they showed good alignment at the GCPs. Future research could identify whether the processing of RGB and TIR data together, as in Javadnejad et al. (2020), leads to better T_p retrieval than processing datasets separately with co-registration to GCPs. While new strategies for processing TIR data and identifying vegetation within the orthomosaic would likely improve results, research advances are inevitably constrained by available UAV TIR camera resolutions (640×480 pixels) and precision (Aragon et al., 2020). Although lower flying heights can increase pixel resolution, the downwash from a multirotor UAV may influence measured T_p (Tang et al., 2020). Lower flying height also increases flying time to cover the site, increasing the chance of temperature changes occurring during a flight, which could further influence results. The precision of uncooled microbolometers, together with the potential impact of adjacency effects from background scattering (Aragon et al., 2020), adds further uncertainly to derived T_p measurements. While the adjacency effect on high-resolution satellite data has recently been explored (Zheng et al., 2019; Duan et al., 2020), the influence on UAV-based data remains under-explored and should be the focus of future work, especially in regard to phenotyping studies, where sub-degree accuracies may be required.

The detection of plant stress via UAV TIR data can be sensitive to the level of solar radiation due to its influence on stomatal conductance, with many studies showing the need to consider variation between sunlit and shaded plant components (Jones et al., 2002; Meron et al., 2013; Poblete et al., 2018; Zhang et al., 2019). However, these studies predominately occur in tree or vineyard crops with developed canopies where intra- and inter-plant shading can be significant compared to low profile well-spaced tomato plants. Nonetheless, we examined the temperature difference between high (sunlit) and low (shaded) blue reflectance areas of the plants and found, as opposed to Poblete et al. (2018), that shadowing did not increase the range in T_p. Therefore, the separation of sunlit and shaded plant components did not improve results in our study. It also meant that methods incorporating the standard deviation of T_p as a proxy for transpiration differences between sunlit and shaded areas to detect stress such as in Han et al. (2016), could not be applied to our study.

While there are many unresolved questions and inherent sensor limitations for T_p retrievals from UAV TIR data, our research demonstrates a detectable difference in T_p between the salt-treated and control plots. Differences are apparent across all data collections following the initial salt application on November 14, 2017. Results suggest that T_p best discerns plant stress between the stages of fruit formation and ripening (i.e., between December 20 and January 7), an outcome most likely related to canopy cover, which was shown to peak approximately a month before harvest (Johansen et al., 2019). Increased canopy closure reduces soil background influence and increases the plant area over which transpiration is occurring. Once senescence begins, and photosynthesis reduces, and so too does transpiration and canopy cover. This result aligns with Perich et al. (2020), which, although based on a wheat crop, also showed that the optimal time to make TIR measurements is before the onset of senescence. The smaller difference in TIR indices (dT_p and CWSIs) between salt and control plots on January 14, together with the broad range in plant condition in both treatments, demonstrates that the morphometric methods of Johansen et al. (2019) present a better approach for identifying stress-tolerant germplasm close to harvest.

Our results suggest that a threshold of CWSIs >0.36 may indicate stress, based on mean differences between salt-treated and control plants and the fact that this threshold applied to 57 and 68% of plants in the salt plot, but only 19 and 24% in the control plots on December 20 and 7 January, respectively. While studies applying CWSI to tomato plants are limited, our results are similar to Anconelli et al. (1993), where CWSI >0.35 led to yield reduction in processing tomatoes (i.e., tomatoes that are canned and machine harvested). Many studies have suggested that CWSI values around 0.3 represent an optimum threshold for commencing irrigation in response to water stress (Reginato, 1983; da Silva and Rao, 2005; González-Dugo et al., 2006). While there are observable differences between the salt and control plots, there is a broad range of dT_p and consequently CWSIs values in both treatments. This range may be inherent to the data collection method due to thermal drift or the creation of the orthomosaic. However, compared to previous research we applied a novel orthomosaic generation method by Malbéteau et al. (2021), wherein the temperature of overlapping pixels was averaged along each swath and normalized between-swath temperatures to reduce the impact of standard orthomosaic generation approaches (which integrate overlapping flight lines collected minutes apart and exposed to different wind directions).

Presuming the ranges in CWSIs are reflective of real temperature differences between plants, we suggest that these differences are due to the 200 accessions exhibiting a range of stomatal conductance responses to salt stress. While T_p has been used to detect plant stress since the 1960s (Fuchs and Tanner, 1966), it is based on the assumption that plants show an isohydric reaction to stress, reducing stomatal conductance to limit transpiration. A growing body of evidence suggests that plants within the same species exhibit both isohydric and anisohydric responses to stress (Sade et al., 2012). The mechanism employed by tomato varieties with different salt tolerance levels to regulate water use is also unclear (Han et al., 2020). For example, the commercial variety “Moneymaker” (Lycopersicon esculentum Mill., cv) is anisohydric and maintains stomatal conductance in response to stress (Sade et al., 2012). The domesticated variety “Brigade” (Lycopersicon esculentum Mill.) reduces stomatal conductance under drought stress. However, it also opens stomata within a day of irrigation (Patanè, 2011). In comparison, wild types of tomato plants can keep stomata closed for up to 6days after irrigation to maintain turgor (Torrecillas et al., 1995). The variation in stomatal conductance response among the 200 wild genotypes in our trial is still to be determined. Therefore, even with very accurate T_p retrievals, cooler plants may not necessarily be the least stressed in terms of agronomically desirable traits such as yield. Plants that had a higher temperature soon after salt application may maintain turgor and produce comparatively higher yields. Resolving this complexity and determining whether T_p can be used to differentiate the performance of accessions in our trial are the focus of ongoing research. Identification of inter-accession differences was not the intent of the research presented herein, as the combination of accuracy limitations in current TIR cameras (Kelly et al., 2019; Aragon et al., 2020), the complex role of environmental interactions with plant response, and the uncertainty and complexity in the mechanism employed by Solanum pimpinellifolium plants in response to salt stress are all aspects that impact the discrimination of accession-based behavior. Ongoing work will seek to explore some of the genotype–phenotype interactions, and the thermal infrared data may provide some insights into this effort. As UAV-based T_p results are confounded by a plant’s morphology (canopy density, leaf inclination), there also needs to be focused research into how to account for morphological variation to increase confidence in the association between observed T_p and stomatal conductance (Perich et al., 2020).

The fusion of TIR information with broadband spectral (Johansen et al., 2019) or hyperspectral (Angel, in review) data will likely provide more in-depth insight than TIR data alone to elucidate the challenges in observed T_p associated with plant physiological response (Hernández-Clemente et al., 2019). Building upon the results herein and integrating TIR data into the development of turnkey UAV phenotyping solutions could provide a method to enable the early detection of salt impacts by detecting changes in T_p in the initial ion-independent response to stress. While our study would have been improved by ground-based visual scoring of plant health during November and December (after the initial salt application), our results showed that CWSI_S values were higher in salt-treated than control plants from December 6. Early detection of stress before observed changes in plant form would enable breeders to select germplasm for future breeding studies rapidly and farmers to balance irrigation with brackish water while maintaining yields.