Introduction

Water is often the main limiting resource for agricultural production in arid and semi-arid regions, with increasing climate variability and demand for crop products further exacerbating pressure on hydrological resources. Vineyards occupy a particular cultural and economic importance in the Mediterranean region (Limier et al. 2018) and its relevance has also increased globally in other semi-arid regions such as USA, South America, South Africa and Australia (OIV 2022). Irrigation practices are becoming increasingly widespread in vineyards, particularly to control deficit irrigation practices and, especially, in light of the increase of extreme heat and drought events due to climate change (Rienth and Scholasch 2019; Romero et al. 2022).

As such, the precise spatio-temporal detection of vine water status can provide valuable guidelines for agronomic management practices to optimize water use and grape production. This is notably relevant for regulated deficit irrigation practices, a common strategy in viticulture to control wine quality and production, which needs an accurate assessment of crop water needs at different phenological periods (Coombe and McCarthy 2000; Girona et al. 2006). Indeed, water availability is a major factor for regulating both quality and productivity in grapevines with photosynthesis being primarily affected by stomatal closure (Escalona et al. 2000). Remote sensing, especially based on thermal infrared (TIR) information, has shown great promise to monitor vine water stress in both space and time (Bellvert et al., 2014; Kustas et al. 2022). It is well established that the canopy temperature is a sensitive indicator of vegetation water status (Jackson et al. 1981), through the relationship between canopy temperature, stomatal conductance and transpiration. Indeed, transpiration, consuming part of the energy from incoming solar irradiance, results in the cooling of canopy leaves. Water stress causes stomatal closure, inducing a reduction of transpiration and an increase in canopy temperature. As such, canopy temperature based on TIR has been widely applied in remote sensing to monitor water stress, most commonly through the crop water stress index (CWSI) as proposed by Jackson et al. (1981). The CWSI is related to the difference between actual evapotranspiration (ET) and potential ET without water limitation (i.e. CWSI = 1−ET/ETpot) and is most commonly estimated empirically by applying threshold limits normalizing between canopy temperature conditions for a well-watered crop and that with maximum stress (Maes and Steppe 2012). However, this empirical approach is often limited to depict the spatial variability of a given area with the empirically derived thresholds being dependent of the structural characteristics of the vegetation, phenological stage and surrounding environmental conditions (Bellvert et al. 2015; Gonzalez-Dugo and Zarco-Tejada 2022; Maes and Steppe 2012).

Other remote sensing-based methods use TIR observations to estimate ET through physically-based approaches such as surface energy balance models. In recent studies, notably through the GRAPEX project (Kustas et al. 2018, 2019, 2022), great advances have been made in estimating ET in vineyards at different spatial scales using various adaptations of the Two-Source Energy Balance (TSEB, Norman et al. 1995) with input data ranging from tower-based proximal sensors (e.g. Kustas, et al. 2022) to spaceborne imagery (e.g. Knipper et al. 2019). TSEB is particularly useful to detect crop water status as it inherently partitions the energy balance between the vegetation canopy and soil, thus providing an estimate of vegetation transpiration. Indeed, Bellvert et al. (2020) demonstrated that TSEB-derived canopy transpiration had a more robust relationship with stem water potential compared to bulk (i.e., canopy + soil) fluxes, which may be influenced by the background soil/substrate conditions. In light of this, Nieto et al. (2022) proposed a set of alternative crop water stress indicators that took advantage of TSEB’s ability to estimate canopy fluxes as well as describing a method to estimate leaf stomatal conductance (gst) by inverting flux estimates from TSEB. Through this, the Crop Stomatal Stress Index (CSSI), relating actual gst based on TSEB estimates and maximum gst, was proposed using tower-based inputs and obtained stronger correlations with in-situ water stress proxies compared to the more traditional CWSI using bulk fluxes (Nieto et al. 2022).

While bulk flux estimates from TSEB are generally well modelled when compared to EC measurements, the modelled ET partitioning has shown less consistency with mixed results observed in the literature (Gao et al. 2023; Kool et al. 2021; Kustas et al. 2019). Indeed, Kool et al. (2021) showed that ET partitioning was sensitive to the estimation of soil heat flux, the simulated radiation partitioning between canopy and soil and the transpiration initialization method. The original TSEB algorithm (Norman et al. 1995) implemented a Priestley and Taylor (1972) framework for an initial transpiration estimate (TSEB-PT) but other TSEB refinements have implemented a Penman–Monteith (PM) initialization (Colaizzi et al. 2012, 2014) or dual-source Shuttleworth and Wallace (1985) (SW) initialization (Kustas, et al. 2022a, b). In certain cases, TSEB-PT was shown to effectively estimate canopy fluxes in vineyards (Kool et al. 2021; Nieto et al. 2022) but other studies reported larger uncertainties (Kustas et al. 2019). Kustas and Norman (1997) proposed the use of TIR measurements at multiple viewing angles to separate canopy and soil temperatures analytically, eliminating the need to parameterize the canopy transpiration through energy combination models. Indeed, using a similar approach but taking advantage of high-spatial resolution imagery, as those obtained from Unmanned Aerial Vehicles (UAVs), Nieto et al. (2019) proposed a dual-temperature TSEB (TSEB-2T), which directly retrieves canopy and soil temperatures from the TIR imagery and, as such, does not need to parameterize an a priori transpiration estimate.

Indeed, there is a growing interest from the research community to further develop methods that take advantage of UAVs to detect water stress at fine spatial scales to implement precision agriculture practices (Bellvert et al. 2016; de Castro et al. 2021; Nassar et al. 2020; Nieto et al. 2019). The use of high spatial resolution imagery can isolate the flux signals at the plant level and allow to better understand the spatial variability of crop stress within plot scales. In addition, using physically-based models to derive ET at fine resolution allows to not only assess the spatio-temporal variability of water stress but also to quantitatively assess the vine water demand giving potentially better recommendations for irrigation inputs required for different production goals. As such, this study aimed to evaluate the capabilities of different crop water stress metrics derived from process-based surface energy balance models and high-resolution UAV imagery to detect within-plot water stress over an experimental vineyard. In this study, both the original TSEB-PT and TSEB-2 T were implemented and were, as a first step, evaluated against eddy-covariance (EC) tower measurements located at the study site, to assess the general accuracy and expected uncertainties of these models at the EC footprint level. Subsequently, different crop water stress metrics, taking advantage of both modelled bulk and canopy fluxes, were computed and examined against in-situ vine physiological measurements, such as stomatal conductance (gst), leaf (Ψleaf) and stem (Ψstem) water potential, serving as in-situ vine water stress proxies at treatment/irrigation sector level. In fact, the discrimination of the transpiration signal from soil evaporation has potentially very high utility to viticulture since transpiration is linked to photosynthesis, and thus grape productivity, while soil evaporation may be seen as a loss of water from an agronomic point of view. This study allows us to better understand how these UAV and ET-based indicators relate to vine water stress and demand over a small area of interest, assessing its potential to support decision making processes with quantitative estimates of vine water demand to better implement the most appropriate irrigation regime depending on production objectives, such as applying deficit irrigation practices to control vine quality or to optimize water use efficiency and grape production.

Materials and methods

Study site

The study was implemented in a 0.5 ha vineyard (Petit Verdot variety) located at the ‘El Socorro’ experimental farm in central Spain (40.14 N 3.37 E, Belmonte de Tajo, Madrid, Spain; altitude 755 m.a.s.l.). The grapevines were planted following a 2 m inter-row spacing by 1.1 m inter-vine spacing, with rows oriented in the North–South direction. The area is characterized by a typical semi-arid continental Mediterranean climate with mean annual air temperatures of about 14 °C and average annual rainfall of 420 mm (Guerra et al. 2022). The soil is characterized as Calcic Haploxeralf (pH 8.4) with a clay-loam texture. The inter-rows were mowed with a flail mower (Guerra et al. 2022), leaving no cover crop understory. The grapevines were placed on a vertical shoot position (VSP) trellis system with the crop height reaching maximums of roughly 1.5 m. In 2021, a drip irrigation system was installed allowing to establish an experimental design of randomized blocks of three replications for three different irrigation treatments (Fig. 1). The three different irrigation treatments consisted on maintaining three different crop coefficient (Kc) compared to the reference ET (ET0) as calculated by the FAO56 Penman–Monteith method (Allen et al. 1998) using daily meteorological data from the nearby Chinchón weather station of the Spanish Agroclimatic information systems for irrigation (SIAR, https://eportal.mapa.gob.es/websiar/SeleccionParametrosMap.aspx?dst=1). The typical deficit irrigation practice of the region maintains Kc values at around 0.4 throughout the growing season (Romero et al. 2016; Rallo et al. 2021). Therefore to induce a large contrast and variability of vine water stress in the experimental study, we implemented treatments that imposed an extreme deficit irrigation (0.2Kc; 20% of ET0), typical deficit irrigation (0.4Kc; 40% of ET0) and highly irrigated practices (0.8Kc; 80% of ET0.). Irrigation began during the grapevine flowering (early June) and ended prior to harvest (mid-September) and were adjusted weekly over the three treatments taking into account the weekly accumulated rainfall and reference ET. This study was implemented over two vine growing seasons in 2022 and 2023, with an average annual irrigation input of 93, 194 and 371 mm for the 0.2Kc, 0.4Kc and 0.8Kc treatments, respectively.

Fig. 1
figure 1

Aerial view of the experimental scheme in ‘El Socorro’ farm with three different irrigation treatments and repetitions (0.2Kc in red, 0.4Kc in green and 0.8 Kc in blue) along with 27 permanent sampling points used to measure in-situ vine physiological measurements (orange circles) and location of eddy-covariance (EC) tower (red triangle) with the typical 2D footprint dimensions and weighted contribution to flux measurements where yellow-purple indicates a greater weight (estimated from Kljun et al. 2015 on 2022-08-30 11:00 UTC) directed towards the dominant southwest fetch. Orthomosaic taken from RGB camera camera (DJI P1) acquired on 2022-08-30

Full size image

Data and instrumentation

For both 2022 and 2023, field campaigns were performed every 2 weeks between June and September. During each campaign, a UAV flight was performed during midday conditions (~ 11–13 UTC) with a thermal infrared (TIR) and multispectral visible-to-near-infrared (VNIR) sensors, in addition to a high-resolution visible (RGB) camera (see Sect. “UAV imagery”). An eddy-covariance (EC) tower was installed on August 4th 2022 to sample continuous flux and meteorological measurements at high temporal frequency (see Sect. “Flux and energy balance data”). In this study, we solely made use of UAV campaigns with simultaneous measurements from the EC tower to apply and evaluate the TSEB modelling methods, resulting in a total of eight UAV campaigns (Table 1). During each flight overpass, in-situ physiological data of leaf (Ψleaf) and stem (Ψstem) water potential, along with leaf stomatal conductance (gst) were acquired to obtain a reference dataset of vine water stress during different phenological periods ranging from BBCH scale (Lorenz et al. 1995) of 77 (fruit development) to 85 (Ripening of berries) (see Sect. “In-situ physiological data”).

Table 1 Date, UAV takeoff time, meteorological conditions during flight time and vine phenological stage according to BBCH scale (mode of in-situ observations) for each field campaign performed at El Socorro experimental farm. Meteorological data were acquired from EC tower installed within experimental farm
Full size table

UAV imagery

A DJI Matrice-300 UAV (DJI Technology Co., Ltd, Shenzhen, China) was used to acquire VNIR, TIR and RGB imagery using the sensors Parrot Sequoia + (Parrot S.A., Paris, France), DJI’s Zenmuse H20T and DJI’s Zenmuse P1, respectively. The VNIR Sequoia + and TIR H20T acquired images simultaneous at 40 m above surface looking at nadir with 70 and 80% frontal and lateral overlap, resulting in a native pixel resolution of about 3–4 cm. The RGB images were acquired at 15 m above surface (also with 70 and 80% frontal and lateral overlap) to acquire a dense point cloud to do a 3D reconstruction through photogrammetric techniques.

UAV images were processed using OpenDroneMap (https://www.opendronemap.org/), an open-source drone processing software. Raw TIR H20T image tiles (i.e. in R-JPEG format) were first converted to single band radiometric temperatures using the open-source DJI Thermal SDK software (https://www.dji.com/downloads/softwares/dji-thermal-sdk). These individual temperature image tiles were then mosaicked together with OpenDroneMap. Congruently, multispectral images from Sequoia + were radiometrically calibrated using camera corrections, such as vignetting, black level and gain/exposure compensations, using the available routines developed for OpenDroneMap. (https://github.com/OpenDroneMap/ODM/blob/master/opendm/multispectral.py).

Flux and energy balance data

The EC and energy balance tower measurement system was located on the eastern edge of the experiment (Fig. 1) to have an adequate fetch from the dominant prevailing winds from the South-West. The tower was instrumented with an integrated open-path infrared gas analyzer and 3D Sonic anemometer Campbell ScientificFootnote 1 (IRGASON, Campbell Scientific, Logan, UT, USA) to measure ecosystem-level carbon and water gas exchanges. The raw data were sampled at a frequency of 20 Hz and recorded using a CR6 datalogger (Campbell Scientific, Logan, UT, USA). The Easyflux datalogger program (Easyflux-DL, Campbell Scientific, 2020) corrected the raw high-frequency data using the full suite of standard corrections and adjustments, including spike filtering, measurement quality control flags and applying correction for high/low frequency losses, to generate corrected half-hourly turbulent fluxes. More details of EC data post-processing are available in the EasyFlux-DL product manual (Campbell Scientific, 2020).

The full radiation budget, both shortwave and longwave, were measured using a four-component net radiometer (SN-500-SS, Apogee, Logan, UT, USA). Both EC and radiation instrumentation were installed on the tower at 3.3 m above ground level (agl) during the 2022 campaign and at 3.5 m agl during the 2023 campaign, over two times the maximum heights reached by the vine canopies. In addition, five soil heat flux plates (HFT-3, Radiation Energy Balance Systems, Bellevue, WA, USA) were installed along a cross-row southwest diagonal transect between the two vine rows following the dominant wind fetch (as shown in Fig. 1) each buried at 10 cm depth. This was done to characterize G conditions at different distances to the vine canopies but also takes into account the two outer rows of treatments 0.8Kc and 0.4Kc, which are the most dominant treatments within the flux footprint (Fig. 1). Five pairs of soil thermocouple probes (TCAV-L, Campbell Scientific, Logan, UT, USA) and five soil moisture sensors (CS655, Campbell Scientific, Logan, UT, USA) were co-located with each soil heat flux plate along the transect at depths of 2 and 8 and 5 cm depths, respectively for estimating storage term above the the heat flux plates.

Due to the well-known issue of the frequent lack of energy balance closure (EBC) from the EC system measurement technique, latent heat fluxes (LE) and sensible heat fluxes (H) were corrected, as described in Kustas et al. (2022), using the average of three typical closure methods: (1) assigning all residuals to LE, (2) assigning all residuals to H and (3) assigning residuals proportionally to LE and H maintaining the Bowen Ratio (i.e., H/LE). It is assumed that the ensemble mean comes closer to the ‘true’ observed flux values compared to any single correction method. This rationale is described and justified in Kustas et al. (2022) and Bambach et al. (2022). The energy balance closure (i.e. Rn-G vs H + LE) is shown in Fig. A1 for the entire sampling time of the EC measurements and for the specific time steps of the UAV overpasses, where the average closure ratio (H + LE/Rn-G) was 0.79 and 0.81, respectively.

In-situ physiological data

Midday leaf-level Ψleaf, Ψstem and gst were acquired simultaneous to UAV overpass over 3 vines per treatment and block, resulting in a total of 27 evenly spaced sampling vines (3 treatments × 3 blocks × 3 vines) (Fig. 1). This resulted in a total of 72 and 54 measurements within each treatment for Ψleaf/stem and gst, respectively, with gst only being sampled during the 2023 campaigns. Ψleaf and Ψstem were measured using a pressure chamber (Scholander et al. 1965), where a well-developed sunlit leaf was excised in each sampled vine for the Ψleaf sample while the Ψstem sample was covered using an opaque aluminum bag prior to excision (at least one hour before sampling). Simultaneous to this, gst was measured using the LI-600 porometer (LI-COR model 660, LI-COR Biosciences, Lincoln, NE, USA) by sampling six leaf samples per vine (three in the upper canopy and three in lower canopy). All three of these physiological measurements were acquired during midday conditions (between 11 and 13 UTC) simultaneous to UAV overpass.

TSEB-PT and TSEB-2 T models

Model descriptions

Remote sensing-based surface energy balance (SEB) models estimate LE as residual of the energy balance, relying on the ability of radiometric remote sensing information to estimate net radiation (Rn) and sensible heat flux (H), along with soil heat flux (G) usually as a proportion of Rn at the soil surface. There are different types of SEB models, broadly grouped between one-source models, which do not discriminate between soil and vegetation components over the surface, and multi-source models, which explicitly solves the energy balance over the different soil and vegetation sources. The main source of uncertainty in these models lies in the estimation of H, which is computed with the heat transport equation (Eq. 1).

$$H=\frac{\rho {C}_{p}\left({T}_{0}-{T}_{A}\right)}{{R}_{AH}}$$
(1)

where ρCp is the volumetric heat capacity of air (J m−3 K−1), T0 is the aerodynamic surface temperature (K), TA is the air temperature at a reference height (K) and RAH is the aerodynamic resistance to heat transport (s m−1). This heat transport equation is only satisfied when using the conceptual aerodynamic surface temperature at the canopy source-sink height (i.e. T0). However, the radiometric LST from remote sensing can differ significantly to T0, with their relationship not well establish and dependent on the surface and observation conditions (Colaizzi et al. 2004). One-source models therefore need to relate remotely sensed LST with T0, often parameterizing additional resistance terms (Kustas et al. 2016) or applying empirical contextual methods (i.e., Allen et al. 2007) to account for these differences. By contrast, two-source models consider soil and vegetation sources separately within their framework, which explicitly accommodates the major factors that influence the relationship between radiometric LST and T0, such as vegetation gap fraction and viewing angle. Indeed, the Two-Source Energy Balance (TSEB) model (Norman et al. 1995) is well established and has been shown to effectively simulate energy fluxes in vineyards, notably through the various scientific advances from the GRAPEX project (Kustas et al. 2018, 2019, 2022).

In TSEB, the energy balance is separated between a canopy (or vegetated) and soil sources (Eq. 2a, b), explicitly accounting for the aerodynamic coupling between the soil (subscript s) and vegetation canopy (subscript c) layers.

$$Rn,c=LEc+Hc$$
(2a)
$$Rn,s=LEs+Hs+G$$
(2b)

The radiative transfer through the canopy and soil (Rn,c and Rn, s) were estimated using an extinction coefficient approach as described in chapter 15 of Campbell and Norman (1998) and incorporated into TSEB by Kustas and Norman (1999). This is primarily dictated by the foliage density (i.e. LAI) and architecture of the vegetated canopy. In this case, we incorporated the clumping index taking into account the row structure of the grapevines, simulating them as simplified clumped rectangular canopies as described in Nieto et al. (2019). On the other hand, the ratio approach was used to model G, where it was estimated as 20% of the net radiation reaching the soil (Rn,s).

To decompose the energy balance, TSEB must separate the radiometric land surface temperature (LST) into its vegetation and soil components, assuming that the total temperature emitted by the surface is the combined emission of both vegetation and soil sources weighted by the fraction of vegetation canopy observed by the sensor (Eq. 3).

$$LST\left(\theta \right)={\left[f\left(\theta , LAI\right){T}_{c}^{4}+\left(1-f\left(\theta , LAI\right)\right){T}_{s}^{4}\right]}^{1}/4$$
(3)

where \(f\left(\theta \right)\) is the fraction of vegetation observed by the TIR sensor looking at a zenith angle \(\theta\), which also depends on LAI (Kustas and Norman 1999); Tc is the vegetation canopy temperature (K) and Ts is the soil substrate temperature (K).

Since both Tc and Ts are unknown a priori in Eq. 3, different approaches have been developed to estimate both component temperatures to solve the series of equations in TSEB. In this study, we make use of two approaches: the original modeling scheme based on a Priestley-Taylor (PT) initialization (TSEB-PT, Norman et al. 1995) and the contextual retrieval of both component temperatures using high resolution imagery Nieto et al. 2019).

TSEB-PT initializes the model by assuming, as a first guess, that the photosynthetically active part of the vegetation is transpiring at a potential rate (Eq. 4) and iteratively accounts for vegetation water stress until conservation of energy and temperature is achieved assuming daytime constraints and LST as a boundary condition.

$$LEc={\alpha }_{PT}{f}_{g}\frac{\Delta }{\Delta +\gamma }Rn,c$$
(4)

where \({\alpha }_{PT}\) is the Priestley–Taylor coefficient, initially set to 1.26 but automatically reduced in TSEB-PT until convergence is reached, fg is the fraction of vegetation that is photosynthetically active, Δ is the slope of the saturation vapour pressure versus temperature and γ is the psychrometric constant. For more details regarding the TSEB-PT model scheme, the reader is referred to Norman et al. (1995) and Kustas and Norman (1999).

For the TSEB-2 T implementation, Tc and Ts were directly retrieved using high resolution UAV imagery (see Sect. “TSEB implementation with UAV imagery”). This removes the need for a parametrization of the initial canopy transpiration of Eq. 4 and was originally proposed by Kustas and Norman (1997) using multiple TIR viewing angles. For more details regarding TSEB-2 T using UAV imagery, the reader is referred to Nieto et al. (2019a).

TSEB implementation with UAV imagery

The TSEB model runs were implemented by resampling the original UAV orthomosaics to 2 m pixel size, which is roughly the width of the grapevine rows. We assume that this spatial resolution is compatible with aerodynamic and radiation formulations applied within TSEB to be consistent with micro-meteorological lengths scales (Xia et al. 2016) and, therefore, contain both the vine canopy and soil interrow at the pixel level.

Ancillary canopy structural variables are required in TSEB to account for both radiation transmission in row crops and turbulent heat exchange between the soil, plant and the atmosphere (Parry et al. 2019, Raupach 1994). Canopy height (CH) was estimated by generating the digital terrain model (DTM) and the digital elevation model (DEM) using photogrammetric techniques from the overlapping UAV-based RGB images (i.e. CH = DEM – DTM). CH was then resampled by extracting the maximum value within the 2 × 2 m pixel window.

Vegetation fractional cover (fc) was estimated from the classification of vine pixels at 0.03 m using a supervised classification (random forest classification) and further limited to pixels with normalized-difference-vegetation-index (NDVI) greater than 0.3 (limiting edge effects) and CH greater than 0.5 m to avoid non-vine vegetation (i.e. weeds) present in the interrows. Then, fc was obtained from the fraction of vine pixels identified across the 2 × 2 m window. Similarly, Tc and Ts, needed for TSEB-2 T, were estimated by acquiring the mean of pixels classified as vine and soil, respectively within the 2 × 2 m window.

Finally, the leaf area index (LAI) at 2 m pixel resolution was estimated by developing a random forest regression model based on in-situ LAI measurements acquired at the sampling points shown in Fig. 1 using VNIR vegetation indices (OSAVI, Rondeaux et al. 1996 and reNDVI, Gitelson and Merzlyak 1994) and structural metrics (CH, fc) as predictor variables. Refer Fig. A1 to see the calibration and validation of the LAI model. During each campaign, in-situ LAI measurements were acquired over the 27 vine sampling points (Fig. 1) using the LAI-2200C (LI-COR, Biosciences, Lincoln, NE, USA) following the protocol developed by White et al. (2019) for vineyards. In this case, four measurements were performed at each of the 27 sampling point below the vine-row and then 1⁄4, 1⁄2 and 3⁄4 distance from vine-row to obtain an ‘ecosystem’-level (i.e. vine + interrow) LAI reference. These references values were used to train the empirical model as similarly implemented by Kang et al. (2022) (see Fig. A2 for the LAI model training and testing).

Crop water stress metrics

From the TSEB outputs, different crop water stress indices were estimated and applied in this study to evaluate which indices were more related to in-situ physiological measurements. Traditionally, the Crop Water Stress Index (CWSI), based on the relationship between actual and potential evapotranspiration (Jackson et al. 1981), has been widely used to detect water stress from thermal remote sensing imagery. However, as discussed in Nieto et al. (2022) other stress indicators can also be generated taking advantage that two-source models, such as TSEB, separating canopy and soil fluxes. Nieto et al. (2022) proposed the Crop Transpiration Stress Index (CTSI, Eq. 5a) and Crop Stomatal Stress Index (CSSI, Eq. 5b).

$$CTSI=\frac{LEc}{LEc,sw}$$
(5a)
$$CSSI=\frac{{g}_{s}}{{g}_{st,0}}$$
(5b)

where LEc,sw is the potential canopy latent heat flux computed by Shuttleworth-Wallace model (Shuttleworth and Wallace 1985) (described in appendix B) and gst,0 is the maximum stomatal conductance (see Nieto et al. 2022 for its derivation). Effective gst can be estimated from TSEB modelled fluxes and aerodynamic resistances as demonstrated by Nieto et al. (2022). This is done by deriving the effective resistance to water vapour diffusion exerted by all leaves (Rc, Eq. 6)

$${R}_{c}=\frac{{\rho }_{a}{C}_{p}\left({e}_{L}-{e}_{0}\right)}{LEc\gamma }-{R}_{x}$$
(6)

where \({\rho }_{a}\) is the air density, eL is water vapour pressure in the leaf, which is assumed saturated at leaf temperature Tc (Farquhar and Sharkey 1982), e0 is the vapour pressure of the air at the canopy interface, related to air vapour pressure (ea) measured at reference height as \({e}_{0}={e}_{a}+\frac{LE\gamma {R}_{a}}{{p}_{a}{C}_{p}}\) and Rx is the resistance to momentum and heat transport at the boundary layer of the canopy interface (see Norman et al. 1995 for details). As such, effective leaf-level gst is estimated by inverting Eq. 7 (Kustas et al. 2022)

$${R}_{c}=\frac{1}{{f}_{t}{f}_{g}LAI{g}_{st}}$$
(7)

where \({f}_{t}\) is either 1 or 2 depending on stomatal distribution in the leaves (1 for hypostomatous and 2 for amphistomatous leaves).

Along with these stress indices that take advantage of partitioned LE, more traditional indices were also computed such as the Crop Water Stress Index using the one-source or big-leaf Penman–Monteith formulation to estimate potential LE (Eq. 8a, CSWI-PM), using the two-source Shuttleworth-Wallace formulation to estimate potential LE (Eq. 8b, CWSI-SW) and the widely used evaporative stress index (Eq. 8c, ESI; Anderson et al. 2007), which uses the FAO56 reference ET (Allen et al. 1998) to normalize the actual ET. The Penman–Monteith and Shuttleworth-Wallace energy combination equations are described in Appendix B.

$$CWSI\text{-}PM=\frac{{\text{LE}}}{{LE}_{PM}}$$
(8a)
$$CWSI\text{-}SW=\frac{{\text{LE}}}{{LE}_{SW}}$$
(8b)
$$ESI=\frac{{\text{LE}}}{{LE}_{fao56}}$$
(8c)

Both LEPM and LESW use a minimum stomatal resistance dependant on the vapour pressure deficit (VPD) as derived and detailed in Nieto et al. (2022) and Kustas et al. (2022).

Model evaluation

Modelled fluxes from both TSEB-PT and TSEB-2T were evaluated against tower measurements using the eddy-covariance (EC) technique. The two-dimensional EC footprint areas were estimated following the method described in Kljun et al. (2015) during the time step of the image acquisition, including estimating the footprint probability density function. Winds were often light and conditions were strongly unstable resulting in a relatively small footprint (Fig. 1) given the EC measurement height to canopy height ratio of ~ 2. The mean of the instantaneous modelled fluxes (LE, H, Rn and G) of all pixels located within the footprint were weighted by the normalized probability density function and evaluated against EC tower measurements.

In addition, modelled gst based on TSEB-based LEc inversion were compared against in-situ gst measured with the LI-600 porometer over sampled vines (Sect. “In-situ physiological data”, Fig. 1). Model performance was evaluated based on the root-mean-square-error (RMSE, Eq. 9a), the relative RMSE (rRMSE, Eq. 9b), the mean bias (bias, Eq. 9c) and Pearson correlation coefficient (r, 9d).

$$RMSE=\sqrt{\frac{\sum {\left({Y}_{obs}-{Y}_{mod}\right)}^{2}}{N}}$$
(9a)
$$rRMSE=\frac{RMSE}{{Y}_{obs,max}-{Y}_{obs,min}}$$
(9b)
$$bias=\frac{\sum \left({Y}_{mod}-{Y}_{obs}\right)}{N}$$
(9c)
$$r=\frac{\sum \left(\left({Y}_{mod}-{Y}_{mod,mean}\right)\left({Y}_{obs}-{Y}_{obs,mean}\right)\right)}{\sqrt{\sum \left({\left({Y}_{mod}-{Y}_{mod,mean}\right)}^{2}\right)\sum \left({\left({Y}_{ivs}-{Y}_{obs,mean}\right)}^{2}\right)}}$$
(9d)
$$\rho =1-\frac{6\sum {d}_{i}^{2}}{N({N}^{2}-1)}$$
(9e)

where Yobs are the observed values, Ymod are the modelled values, N is the sample size and di represents the difference between the ranks of corresponding variables in the two datasets.

To understand how sensitive the modelled fluxes were to the irrigation treatments over the study area, a one-way analysis of variance (ANOVA) was performed to test whether the distribution of values between irrigation regimes were statistically different. In this case, we report both the F value, the ratio of variances where a high value indicates that the variance between treatment means was larger than the variance within treatment groups, and the p value, which quantifies the probability of obtaining such a F value if null hypothesis were true (i.e. that the distribution of values from the different treatments are the same). In addition, the different TSEB-based stress metrics were compared to in-situ physiological measurements and their relation was quantified through the Spearman rank correlation coefficient (\(\rho\), Eq. 9e), which is a non-parametric measure of the monotonic relationship of variables, to account for possible non-linear relationships.

Results

Evaluation of TSEB-PT and TSEB-2T

In Fig. 2, the instantaneous flux estimates from both TSEB-PT (Fig. 2a) and TSEB-2 T (Fig. 2b) were plotted against EC measurements over the study area. In general, both TSEB variants performed similarly, with error metrics within comparable ranges (Table 2), although LE (H) was more systematically overestimated (underestimated) with TSEB-PT compared to TSEB-2 T. H was generally well modelled by TSEB-2T (RMSE = 79 Wm−2; bias = 3 Wm−2), but was more consistently underestimated by TSEB-PT (bias = −49 Wm−2). By consequence, TSEB-PT modelled LE was more overestimated (RMSE = 85 Wm−2; bias = 67 Wm−2) compared to TSEB-2T (RMSE = 44 Wm−2; bias = 31 Wm−2), as well as being less correlated (r = 0.83 vs 0.94). In general, both models simulated Rn and G with low systematic errors. However, Rn was somewhat more overestimated by TSEB-2T (bias = 15 Wm−2) while having a greater negative G bias (−20 Wm−2), resulting in an overall larger overestimation of available energy (AE, Rn—G). This led to a modest systematic LE bias (31 Wm−2) for TSEB-2 T even though H was largely in line with EC measurements (bias = 3 Wm−2). For TSEB-PT, the AE was less overestimated but the larger H bias (−33 Wm−2) led to a more consistent LE overestimation (bias = 47 Wm−2). It should be noted that only eight UAV flights were used, therefore overall error metrics were sensitive to any significant discrepancies from one day. For example, TSEB-2 T H estimates from 2023-06-22 showed a larger bias compared to the other dates, increasing the scatter of errors. Indeed, while the average energy balance closure (i.e. [LE + H]/[Rn – G]) during the UAV overpass time steps was 0.81, the energy balance closure on 2023-06-22 was one of the lowest at 0.58. This may be a factor in the observed biases as the correction method has a more important weight leading to greater possible uncertainties in the observations than the model estimates. It also should be mentioned that the EC measurements are influenced by sources from both the 0.8Kc and 0.2Kc treatments, thus the model validation should be considered more qualitatively as an indicator that the models provide satisfactory flux estimates.

Fig. 2
figure 2

Evaluation of modelled fluxes from TSEB-PT (a) and TSEB-2 T (b) against eddy-covariance (EC) observations

Full size image
Table 2 Model performance indicators of root-mean-square-error (RMSE), relative RMSE (rRMSE), bias and pearson’s correlation coefficient (r) for TSEB-PT and TSEB-2T against tower measurements
Full size table

LE spatial distribution and irrigation treatment effects

In general, different LE distributions were detected over the study area as a consequence of the different irrigation regimes (Fig. 3). Modelled LE showed a distinct spatial distribution due to different irrigation regimes, with visually distinct spatial patterns observed in the map in Fig. 3 for both TSEB-PT and TSEB-2 T. This is supported quantitatively by the boxplots in Fig. 3 showing different distribution of values for each irrigation regime, where an ANOVA test revealed significant treatment effects for LE for both TSEB-PT (F value = 35.83; p value = 1.65e-14) and TSEB-2T (F value = 24.71; p value = 1.59e-10). Pairwise z-tests between each pair of treatments also revealed significantly different LE distributions for all pairs (p value < 0.001), except between 0.2 and 0.4 Kc treatment for TSEB-2T (p value = 0.16). By contrast, TSEB-PT modelled LE were significantly different between 0.2 and 0.4 Kc treatments (p value = 0.04), as well as all other treatment pairs (p value < 0.001).

Fig. 3
figure 3

TSEB-PT (top row) and TSEB-2T (bottom row) modelled latent heat flux (LE) on 2023-08-16 11:30 UTC over study area (left column) and corresponding boxplots for all dates grouped by 0.2 (red), 0.4 (green) and 0.8 (blue) Kc treatments (right column). The eddy-covariance (EC) tower is depicted by a black circle. The coordinates on X–Y axis are projected in UTM zone 30N with units in meters

Full size image

While LE estimates from TSEB-PT and TSEB-2T were relatively similar, much greater contrast was observed for the modelled LE partitioning (LEc/LE). As Fig. 4 shows, TSEB-PT’s median LEc/LE oscillated around 0.4 for all treatments, much lower than TSEB-2T’s LEc/LE, which treatment median ranged roughly between 0.7 and 0.9. This was due to the important differences in estimated canopy and soil temperatures between both TSEB variants (Fig. 5). In general, the canopy (soil) temperature used in TSEB-2T were roughly 4 degrees cooler (hotter) than those estimated from TSEB-PT (Fig. 5). By using LEc/LE estimates, TSEB-2T outputs were found to have more significant treatment effects (F value = 18.71; p value = 2.77e-08) compared to TSEB-PT (F value = 7.21; p value = 0.0009). TSEB-2T had significant differences for all pairwise comparisons (p value < 0.05) while TSEB-PT had generally higher p values for each pairwise comparison, including demonstrating a non-significant difference between 0.2 and 0.4 Kc treatments (p value = 0.12).

Fig. 4
figure 4

TSEB-PT (top row) and TSEB-2T (bottom row) modelled latent heat flux (LE) partitioning (LEc/LE) on 2023-08-16 11:30 UTC over study area (left column) and corresponding boxplots for all dates grouped by 0.2 (red), 0.4 (green) and 0.8 (blue) Kc treatments (right column). Eddy-covariance (EC) tower is depicted by a black circle. The coordinates on X–Y axis are projected in UTM zone 30N with units in meters

Full size image
Fig. 5
figure 5

Comparison between the estimated canopy (Tc, left panel) and soil (Ts, right panel) temperatures estimated by TSEB-PT and those applied in TSEB-2T over the 27 sampling points for all campaigns assessed

Full size image

Estimation of gst through TSEB inversion

Leaf stomatal conductance (gst) was estimated by inverting both TSEB-PT and TSEB-2T fluxes and were compared against in-situ leaf-level measurements (Fig. 6). Considering the scaling difficulty of comparing downscaled ecosystem-level fluxes (TSEB outputs) and upscaled leaf-level measurements (in-situ), modelled results agreed reasonably well, where values were within similar ranges to in-situ measurements with low bias (TSEB-PT bias = 0.1 mol m−2 s−1; TSEB-2T bias = 0.05 mol m−2 s−1). TSEB-2T achieved higher correlation (r = 0.35) than TSEB-PT but larger dispersion and errors (RMSE = 0.11 mol m−2 s−1; rRMSE = 46%). However, the largest biases for TSEB-2T were primarily concentrated during early and late phenological periods. By solely evaluating the results of the UAV flights during the peak growth period (2023-07-18, 2023-08-01 and 2023-08-16), as shown in Fig. 7, TSEB-2T errors decreased (RMSE = 0.05 mol m−2 s−1; rRMSE = 20%) and showed higher correlation (r = 0.59). By contrast, TSEB-PT, although with low errors and biases, did not capture as much variability during this peak growth period (r = 0.28). This result is likely due to TSEB-PT LEc/LE values having little change with treatment.

Fig. 6
figure 6

Evaluation of TSEB-PT (a) and TSEB-2T (b) inversion for stomatal conductance (gs) estimations against in-situ observations measured with handheld porometer (LI-600) across all sampling dates in 2023

Full size image
Fig. 7
figure 7

Evaluation of TSEB-PT (a) and TSEB-2T (b) inversion for stomatal conductance (gs) estimations against in-situ observations measured with handheld porometer (LI-600) during the peak vegetation growth period (2023-07-18, 2023-08-01 and 2023-08-16)

Full size image

Performance of different stress indices compared to in-situ measurements

Figures 8 and 9 illustrate the relationship between bulk crop water stress indices computed with TSEB-PT and TSEB-2T, respectively and in-situ Ψleaf and Ψstem measurements. Results follow very similar patterns with little difference between stress indices and between TSEB-2T and TSEB-PT. CWSI-PM was most related to Ψstem with both TSEB-PT (\(\rho\) = 0.56) and TSEB-2T (\(\rho\) = 0.57). In general, we observed that stress indices based on bulk fluxes tend to plateau for lowest Ψstem values (Ψstem < -1.0 Mpa), indicating that perhaps the UAV-based stress indicators were less able to capture the variability of highly stressed plants. Ψleaf was generally less related to stress indices (\(\rho\) < 0.4) with the highest score achieved by TSEB-2T with CWSI-SW (\(\rho\) = 0.4).

Fig. 8
figure 8

TSEB-PT crop water stress indexes (CWSI) using bulk fluxes against in-situ measurements of leaf (Ψleaf) and stem (Ψstem) water potential

Full size image
Fig. 9
figure 9

TSEB-2T crop water stress indexes (CWSI) using bulk fluxes against in-situ measurements of leaf (Ψleaf) and stem (Ψstem) water potential

Full size image

By contrast, when comparing the crop water stress indices computed with partitioned or canopy fluxes (i.e. LEc, CTSI, CSSI), large differences were observed between TSEB-PT and TSEB-2T (Fig. 10 vs Fig. 11). In general, TSEB-PT based stress indices based on canopy fluxes had low correlation with Ψleaf and Ψstem, with the highest correlation observed between with CTSI and Ψstem (\(\rho\) = 0.49). TSEB-2T stress indices generally achieved higher correlation with Ψleaf ( \(\rho\) > 0.4) and Ψstem (\(\rho\) > 0.6), where the highest correlation was observed between CSSI and Ψstem (\(\rho\) = 0.62).

Fig. 10
figure 10

TSEB-PT crop water stress indexes (CWSI) using partitioned canopy latent heat fluxes (LEc) against in-situ measurements of leaf (Ψleaf) and stem (Ψstem) water potential

Full size image
Fig. 11
figure 11

TSEB-2T crop water stress indexes (CWSI) using partitioned canopy latent heat fluxes (LEc) against in-situ measurements of leaf (Ψleaf) and stem (Ψstem) water potential

Full size image

Discussion

The results from this study indicated that TSEB, either through TSEB-PT or TSEB-2T, was capable of reliably estimating energy fluxes through UAV imagery with RMSE bounds within 25–80 W/m2 compared to EC measurements. These errors are within the expected uncertainties for the models and measurements (50–100 W/m2) and are within similar ranges compared to other studies in vineyards (Gao et al. 2023; Kustas, et al. 2022; Nieto et al. 2019, 2022). For instance, LE errors using UAVs and different TSEB model versions were reported by Gao et al. (2023) to be between 69–90 W/m2 for vineyards in California while Ortega-Farias et al. (2021) obtained an RMSE of 55 W/m2 using a two-source patch energy balance model in a Chilean vineyard site. Similar error ranges (i.e., roughly 60 W/m2) were also reported by Simpson et al. (2022) for a savanna ecosystem in Spain applying TSEB with over 25 UAV overpasses and evaluated using 6 different EC towers. Therefore, the results and energy flux uncertainties presented in this study were very much in agreement with similar studies available in the literature. However, the in-situ EC measurements are also expected to have their own uncertainties, especially since the tower footprint overlaps at least two different irrigation treatments leading to a certain level of heterogeneity, which may not be captured as effectively by the net radiometer or the soil heat flux plates. For example, while the five soil heat flux plates were installed cross sectionally between two treatments, this sampling design may not effectively capture the G heterogeneity present in the flux footprint, which must account for both irrigation treatment effects but also differences between the row crop and interrow since the drip irrigation applies the water input locally near grapevine. As such, there is likely additional uncertainties related to the estimation of the available energy compared to sampling over a uniform site, even though the estimated Rn and G compared rather well to tower measurements (Fig. 2, Table 2). Furthermore, the widely acknowledged energy balance closure issue (Bambach et al. 2022; Mauder et al. 2020) is another source of uncertainty and it is worth noting that the largest discrepancy between modelled and measured H occurred on the date with lowest energy balance closure (i.e. 0.58).

The different stress indicators estimated through bulk fluxes (i.e. CWSI-PM, CWSI-SW and ESI) related similarly to in-situ measurements, either applied with TSEB-PT or TSEB-2T. The CWSI-PM yielded slightly better results compared to CWSI-SW, showing little evidence of the advantage of using the SW model to compute potential LE (LEpot) compared to PM, even though its dual-source model structure should be more suited for clumped row vegetation structures such as grapevines. As Nieto et al. (2022) noted, this may be an issue of how LEpot is defined, where, in this study, we parameterized LEpot in the SW model as having a rather constant dry top soil but with enough water in the root-zone for vegetation to transpire at a potential rate. It is likely that the assumption of a very dry soil was not adequate for all the different UAV campaigns assessed here, with soil conditions being affected by seasonal changes and, in certain cases, by precipitation events occurring relatively close in time to the field campaigns, especially during the first campaign of 2023. Soil conditions not only affects the bulk LE (through soil evaporation) but also interacts with the canopy fluxes as it influences the evaporative demand in the canopy by transporting hot (or cool) and dry (or wet) air parcels from the soil towards the canopy-air interface (Kustas and Norman 1999).

The crop water stress indicators that took advantage of modelled canopy fluxes (CTSI and CSSI) showed large differences between TSEB-PT and TSEB-2T. The CTSI and CSSI estimated by TSEB-PT were much less correlated with in-situ measurements, with the highest correlation achieved by CSSI and Ψstem (⍴ = 0.49). On the other hand, the CTSI and CSSI estimated by TSEB-2T were much more correlated with in-situ measurements, with both yielding better results compared to the indices based on bulk fluxes (i.e., CWSI-PM, CWSI-SW or ESI). In fact, CSSI had the highest correlation out of all studied indices with both Ψleaf (⍴ = 0.43) and Ψstem (⍴ = 0.61). These correlations are largely in line with results presented in Bellvert et al. (2020) which reported an R2 ranging from 0.29 to 0.55 between a transpiration-based CWSI and Ψstem. Indeed, Bellvert et al. (2020) also showed that the relationship between Ψstem and canopy transpiration was stronger and more consistent compared to bulk ET in vineyards, as estimated through TSEB with both satellite and airborne imagery. These results, along with the results presented in Nieto et al. (2022) demonstrate the potential of using partitioned canopy fluxes to limit the effect of the soil signal on CWSI and to improve crop water stress detection using remote sensing.

In general, Ψstem was more correlated (⍴ > 0.5) to UAV ET-based stress indices compared to Ψleaf (⍴ < 0.4). The poor results with Ψleaf were also observed by Nieto et al. (2022), while several studies have highlighted the limitation of Ψleaf to robustly represent water stress (García-Tejera et al. 2021; Jones 2004). Clearly, one of the main limitations is that these measurements are highly sensitive to plant architecture with different management interventions, such as vine pruning, affecting the hydraulic conductivity from the root to leaf. As such, Ψleaf at different phenological periods may be more affected by structural changes than to water stress. For example, Bellvert et al. (2015) showed that the relationship between CWSI and Ψleaf was significantly affected by both the grapevine variety and phenological stage, requiring model calibration to account for these effects.

Several authors have suggested the use of gst rather than water potential for water stress detection as it accounts for crop physiology within atmospheric conditions, along with linking the processes of transpiration and photosynthesis (Jarvis 1986). Nieto et al. (2022) demonstrated TSEB’s effectiveness to estimate gst through flux inversion and tower-based inputs, which compared well with in-situ measurements of leaf gas exchange. In this study, we also show that TSEB-based gst estimated through UAV imagery were within similar error ranges (RMSD ~ 0.1 mol m−2 s−1) compared to in-situ leaf measurements, although we observed a lower correlation (r < 0.35). There was also a large difference between TSEB-PT and TSEB-2T, where gst estimated by TSEB-PT showed much less variability, likely affected from the apparent underestimation of LEc (Fig. 3). This was particularly apparent when solely evaluating the peak vine growth period (Fig. 5) where TSEB-2T very effectively estimated gst during this period (RMSE = 0.05 mol m−2 s−1; r = 0.59) compared to TSEB-PT (RMSE = 0.06 mol m−2 s−1; r = 0.28). The larger errors during the early and late phenological periods may be related to (i) lower fg·LAI values during these dates, that might cause arithmetic instability when inverting Eq. 7 and (ii) the greater difficulty to measure in-situ gst since fewer well-developed leaves are present or, especially for late phenological period, may be diseased or damaged. In fact, there is an implicit bias to sample well illuminated and healthy leaves, which also depends on the subjectivity of the operator, while the UAV imagery samples the entire top of the canopy, composed of leaves with varying conditions including being sunlit and shaded. Jones et al. (2002) demonstrated significant differences in gst between sunlit and shaded leaves, with sunlit leaves having, on average, double the gst values as shaded leaves. As such, there is an inherent difficulty in comparing top-down ecosystem-level fluxes (TSEB outputs) and leaf-level measurements (in-situ), which can be a factor that increases the scatter of results observed between the UAV imagery and the leaf-scale measurements.

Large differences in modelled LE partitioning (LEc/LE) between TSEB-PT and TSEB-2 T were apparent in this study, both by evaluating CTSI and CSSI against in-situ indicators (Fig. 9 vs Fig. 10) and by observing the spatial variability and sensitivity to treatment effects (Fig. 4). TSEB-PT showed on average a much lower LEc/LE (~ 0.4) compared to TSEB-2T (~ 0.8). Burchard-Levine et al. (2022) reported a LEc/LE of about 0.8 during the main grapevine growing season in California, USA, both by applying an adapted TSEB version (three-source energy balance model, 3SEB) and an EC-based partitioning method (Zahn et al. 2022). Kustas et al. (2019) also reported a monthly LEc/LE between 0.8 and 0.9 during the summer months as estimated from EC partitioning method (Scanlon and Kustas 2012) and showed that TSEB-PT, forced from tower-based inputs, largely underestimated LEc/LE with values below 0.6 during these periods. Using UAV imagery, Gao et al. (2023) also compared TSEB-PT and TSEB-2T in a vineyard and showed that LEc/LE was better depicted by TSEB-2T, especially when applying the so-called quantile temperature separation algorithm to better extract Tc from the canopy. By contrast, using tower-based measurements, Kool et al. (2021) showed that TSEB-PT was more reliable than TSEB-2T for LEc/LE retrievals. At the satellite level, when comparing the atmosphere–land exchange inverse model (ALEXI) model, based on TSEB, using both a PT and PM initialization, Knipper et al., (2022) also revealed that the PT initialization had greater uncertainty for LEc/LE compared to the PM initialization, even though bulk fluxes were well simulated. Indeed, our results, along with past studies, show that TSEB’s ET partitioning is sensitive to the initial transpiration parameterization, necessary to decompose LST into canopy and soil sources when the pixel is mixed by both components. Future work should concentrate on better understanding the effect of the different transpiration initialization, such as using a PM (Colaizzi et al. 2014) or SW (Kustas, et al. 2022) parameterization, and understand the respective sensitivity of input variables affecting LEc/LE estimations.

Conclusion

These results demonstrate the utility of physically-based models (i.e. TSEB) and UAV imagery to estimate ET and partitioned canopy fluxes, which can enhance the detection of vine water stress. Moreover, the results indicate satisfactory congruence of modelled energy fluxes compared to a reference EC tower. Modelled bulk fluxes were able to depict irrigation treatment effects and were somewhat correlated to in-situ measurements (more notably with Ψstem). However, partitioned flux indicators, by better isolating the vegetation signal and limiting soil/substrate effects, were more effective in detecting these irrigation treatment effects and had higher correlation to in-situ water stress proxies, including comparing well to in-situ gst measurements. Since TSEB-2T and TSEB-PT showed large differences in modelled LEc/LE, future work should concentrate on better evaluating the modelled ET partitioning. This includes testing different TSEB transpiration formulations and soil and canopy temperature partitioning methods, to better understand the effects of the different model initializations on LEc/LE estimates, especially under different agronomic (e.g. presence of cover crop) and climatic conditions. This includes better understanding the effect of advection present in many irrigated agricultural regions as explored by Kustas et al. (2022).