High-throughput field phenotyping reveals genetic variation in photosynthetic traits in durum wheat under drought

Chlorophyll fluorescence (ChlF) is a powerful non-invasive technique for probing photosynthesis. Although proposed as a method for drought tolerance screening, ChlF has not yet been fully adopted in physiological breeding, mainly due to limitations in high-throughput field phenotyping capabilities. The light-induced fluorescence transient (LIFT) sensor has recently been shown to reliably provide active ChlF data for rapid and remote characterisation of plant photosynthetic performance. We used the LIFT sensor to quantify photosynthesis traits across time in a large panel of durum wheat genotypes subjected to a progressive drought in replicated field trials over two growing seasons. The photosynthetic performance was measured at the canopy level by means of the operating efficiency of Photosystem II ( F 0 q = F 0 m ) and the kinetics of electron transport measured by reoxidation rates ( F 0 r1 and F 0 r2 ). Short-and long-term changes in ChlF traits were found in response to soil water availability and due to interactions with weather fluctuations. In mild drought, F 0 q = F 0 m and F 0 r2 were little affected, while F 0 r1 was consistently accelerated in water-limited compared to well-watered plants, increasingly so with rising vapour pressure deficit. This high-throughput approach allowed assessment of the native genetic diversity in ChlF traits while considering the diurnal dynamics of photosynthesis.

tions in high-throughput field phenotyping capabilities.The light-induced fluorescence transient (LIFT) sensor has recently been shown to reliably provide active ChlF data for rapid and remote characterisation of plant photosynthetic performance.We used the LIFT sensor to quantify photosynthesis traits across time in a large panel of durum wheat genotypes subjected to a progressive drought in replicated field trials over two growing seasons.The photosynthetic performance was measured at the canopy level by means of the operating efficiency of Photosystem II (F 0 q =F 0 m ) and the kinetics of electron transport measured by reoxidation rates (F 0 r1 and F 0 r2 ).Short-and long-term changes in ChlF traits were found in response to soil water availability and due to interactions with weather fluctuations.In mild drought, F 0 q =F 0 m and F 0 r2 were little affected, while F 0 r1 was consistently accelerated in water-limited compared to well-watered plants, increasingly so with rising vapour pressure deficit.This highthroughput approach allowed assessment of the native genetic diversity in ChlF traits while considering the diurnal dynamics of photosynthesis.
Attempts to exploit new molecular tools to their full potential, particularly to dissect the genetics of quantitative traits, such as yield and stress tolerance, are limited by our ability to quantify relevant traits with the necessary throughput (Araus & Cairns, 2014;Tuberosa, 2012).Integration of high-throughput phenotyping with genotyping may lead to a major breakthrough in our understanding of the fundamentals of complex adaptive physiological traits related to drought tolerance (Edmeades, McMaster, White, & Campos, 2004;Reynolds et al., 2020;van Eeuwijk et al., 2019).High-throughput phenotyping platforms (HTPPs) have been deployed to quantify ChlF and other traits to circumvent existing bottlenecks in phenotypic and genomic selection (e.g., Barbagallo, Oxborough, Pallett, & Baker, 2003;Chen et al., 2014;Flood et al., 2016;Humplík et al., 2015;Jansen et al., 2009;McAusland, Atkinson, Lawson, & Murchie, 2019;Tschiersch, Junker, Meyer, & Altmann, 2017;Wang et al., 2018).However, they are generally confined to operate under controlled or semicontrolled environments and restricted to detached leaves or small and medium-sized plants, and only a few are suitable for large plants (>1.50 m).HTPPs for measuring ChlF under natural fluctuating field conditions (e.g., the "Field Scanalyzer" reported by Virlet, Sabermanesh, Sadeghi-Tehran, & Hawkesford, 2017) are hitherto scarce.
By monitoring the kinetics of LIFT-based ETR over different timeframes beyond Q A , Keller, Vass, et al. (2019) developed the Q A reoxidation efficiency parameters (F 0 r1;r2 ) for photosynthesis phenotyping.Using the LIFT method for automated plant phenotyping under semi-field conditions, Keller, Matsubara, et al. (2019) demonstrated that the ChlF-based parameters not only facilitated the understanding of photosynthetic interactions with varying environmental factors but also identified differences between and within crop species.Nevertheless, to the best of our knowledge, the LIFT approach has not yet been applied to large-scale field phenotyping of ChlF traits under natural fluctuating growing conditions.
Here, we aimed to assess the inherent genetic variation in photosynthetic traits at the canopy level in a large durum wheat panel (>220 elite accessions) under progressive drought in the field.The photosynthetic performance was measured using the LIFT method by means of ChlF traits: F 0 q =F 0 m and the newly developed reoxidation efficiency parameters, F 0 r1 and F 0 r2 , which were slightly modified from Keller, Vass, et al. (2019).In particular, our study aimed (a)

| Plant material
In two growing seasons, we evaluated a set of elite durum wheat (Triticum turgidum L. ssp.durum Desf.) accessions, mainly cultivars and advanced lines, from the association mapping population "UNIBO-Durum Panel" assembled at the University of Bologna (UNIBO), Italy (see list of materials in Table S1).This panel contains a representative selection of the genetic diversity existing in the major improved durum wheat gene pools adapted to Mediterranean environments (Maccaferri et al., 2006(Maccaferri et al., , 2011)).The collection includes "founder genotypes" used extensively worldwide as parents in breeding programmes, as well as  224 accessions were planted on December 18, 2018.In both trials, genotypes were sown in 2-row plots, 3.5 m long with 0.76 m between rows, with an average seeding rate of 16.8 seeds per meter, which were laid out in a resolvable row-column design (Patterson & Williams, 1976;Piepho, Williams, & Michel, 2015) with two replicates (14 rows Â 18 columns per rep) in Y1, and with three replicates (14 rows Â 16 columns per rep) in Y2.Pre-plant granular nitrogen fertilizer at 112 kg/ha and phosphorus (P 2 O 5 ) at 56 kg/ha was incorporated into the soil.Fields were managed following the standard agricultural practices for the region and were regularly monitored to prevent damage from above-ground insect pests and pathogens.

| Experimental design
Sprinkler irrigation was used to germinate seeds and establish the crop, followed by subsurface drip irrigation as needed for optimal plant growth, once or twice a week.The pressurized subsurface drip irrigation system was installed before planting when one dripline with emitters spaced every 0.30 m was buried at ~0.10 m depth along each seed row.In Y1, the final irrigation event was on March 11, 2018 (i.e., 103 days after sowing, DAS) when ~50% of the genotypes had flag leaf sheaths opened (i.e., at growth stage [GS] 47; Zadoks et al., 1974).From this time point, the whole experiment was subjected to a progressive water deficit until 2-3 April 2018 when plants at the anthesis halfway stage (GS65, on average) were harvested to measure total above-ground biomass.By contrast, in Y2, on March 5, 2019 (77 DAS), when ~50% of the genotypes were at late tillering phase (GS26), well-watered (WW) and water-limited (WD) treatments were implemented and assigned to entire single seed rows in an alternating pattern across the field, such that the initial 2-row plots where split into two side by side subplots of one row each (i.e., a strip-plot type design with water treatment as one main-plot factor).Due to operational limitations, the water treatment could not be randomized, and this alternating pattern of WW and WD strips was assigned to keep a homogeneous field.Control subrows remained well watered by recurring subsurface drip irrigation, whereas water-stressed subrows experienced a progressive water deficit.Both water treatments were imposed until April 9, 2019 (112 DAS) at early anthesis (GS61, on average) when the experiment was terminated, and plots were harvested to determine total above-ground biomass.Table S2 summarizes both growing seasons.

| Meteorological data and soil moisture monitoring
Daily and hourly meteorological reports for both growing seasons were obtained from the Arizona Meteorological Network (AZMET; https:// cals.arizona.edu/azmet/06.htm).In addition, high temporal resolution meteorological data, particularly air temperature, relative humidity and photosynthetic photon flux density (PPFD), for the experimental site were recorded at 5-second intervals with an automated weather station (Clima Sensor US, Adolf Thies GmbH & Co. KG, Göttingen, Germany) and a quantum sensor (SQ-214, Apogee Instruments, Inc., Logan, UT, USA).These data were made available by the TERRA Phenotyping Reference Platform (TERRA-REF; https://terraref.org/).VPD was calculated as the difference between the saturation and actual vapour pressure (Allen, Pereira, Raes, & Smith, 1998).
In Y1, the soil volumetric water content (VWC) was monitored in and between seed rows with time-domain reflectometry (TDR) sensors (True TDR-315, Acclima, Inc., Meridian, ID, USA) installed at three locations within the experiment and at 1 cm, 10 cm and 50 cm depths at each location.In Y2, the TDR sensors (True TDR-310S, Acclima, Inc., Meridian, ID, USA) were installed in both WW and WD seed

| Leaf relative water content
The plant water status was monitored by leaf relative water content (RWC) as described by Mullan and Pietragalla (2012)

| Plant height and above-ground biomass
Plant height was manually measured with a metre stick as the distance from the soil surface to the spike's base or the uppermost level of leaves in the absence of the spike.Median height was measured in the central portion of the plots to reduce plot edge effects.In Y1, plant height was recorded only once at 122 DAS, whilst multiple measurements over time (82,93,101 and 107 DAS) were taken in Y2.
Plants were harvested before the ripening stage (Table S2), and therefore biomass data indicate the status at a point in time rather than direct estimates of final yields.In Y1, at the end of the experiment (125-126 DAS), plants within the 2-row plots were cut with a mechanical forage harvester (Carter Manufacturing Company, Inc.) for aboveground whole plot weights, while subsamples for moisture content [%] were hand-cut before mechanical harvesting for measurements of fresh weight and dry weight after oven-drying for 2-3 days.Total shoot dry matter yield (SDMY) was adjusted to 0% moisture and is reported as (kg/ha).In Y2, the experiment was ended at 112 DAS, and SDMY was obtained for two replicates by hand-cutting lengths of 0.914 m of plants from each single-row plot (WW and WD), which were bundled and placed into large driers until a constant weight was achieved.et al., 2005;Osmond et al., 2017).

| LIFT device and method
In our study, the Q A flash reproduced the FRRF 0.75ms protocol introduced by Keller, Vass, et al. (2019).According to this protocol, the SQ A phase (lasting ~0.75 ms) consists of a sequence of 300 subsaturating flashlets (1.6 μs pulse length) applied at 2.5 μs discrete intervals.The RQ A phase (lasting ~209 ms) consists of 127 flashlets (1.6 μs pulse length) with an initial interval between flashlets of 20 μs followed by exponential increments.The exponent factor increases from 1.025 to 1.05, linearly over the RQ A sequence's length to best cover the temporal dynamics of the fluorescence relaxation signal.At a distance of 0.60 m, the average excitation power for the SQ A phase was ~72,000 μmol photons m À2 s À1 in Y1 and ~55,000 μmol photons m À2 s À1 in Y2.Excitation power was measured at a 1% duty cycle using a 5-second calibration flash (a sequence of 50,000 flashlets with 1 μs pulse length at 100 μs intervals) measured by a quantum sensor (LI-190R, LI-COR, Inc., Lincoln, NE).
Apart from the ChlF sensor, the LIFT device is also equipped with a Vis micro-spectrometer (STS-VIS, Ocean Optics, Inc., Winter Park, FL) with an optical resolution of 1.5 nm (FWHM) for acquiring spectral bands between 400-800 nm (integration time 790 ms).Acquisition of spectral data was synchronized to be performed immediately after completion of each LIFT Q A flash.

| Field measurements
The LIFT instrument was installed to the front end of a cart (model based on White & Conley, 2013) in the vertical direction pointing downward (nadir) and above the plant canopy, perpendicular to the crop row (Figure S1).The distance from the LIFT lens to the median uppermost-canopy (i.e., the target area) was ~0.60 m, being regularly adjusted as plants grew.The blue LED light beam was ~30 mm in diameter at the focal point over the target area.The fluorescence channel's gain was adjusted at the beginning of each measurement day to maintain the raw fluorescence within the 2,000-20,000 signal range, aiming to operate with an optimal signal-to-noise ratio.
Regardless of the growing season, the day of measurement and the time of day, the cart's front end was always facing south to avoid selfshadowing over the target area.The cart was manually pushed across the plots at an average speed of 8 cm/s while a total of 20 independent measurements, each one carried out in a time frame of ~1 s, were acquired from every experimental unit per day of measurement.In Y1, the entire field trial (252 genotypes Â 2 reps) was measured at 0, 1, 2, 3, 4, 5, 6, 8, 10, 12 and 16 days after withholding water (DAWW).
The spectral data were calibrated by a 0.50 m Â 0.50 m white reference panel with 95% reflectance (Zenith Lite™ diffuse target, SphereOptics GmbH, Herrsching, Germany) horizontally placed at 0.60 m in front of the LIFT apparatus (Figure S1b).Over a field phenotyping day, white reference measurements were regularly taken every 36 plots (~30-min interval) in Y1 and 32 plots (~25-min interval) in Y2.Dark reference measurements were acquired in a dark room with the LIFT lens covered with a dark cloth.

| Fluorescence parameters for field phenotyping
Based on a slight modification of the procedure reported by Keller, Vass, et al. (2019), we derived the photosynthetic traits from the LIFT-measured ChlF transients using the Q A flash protocol.The PSII operating efficiency (F 0 q =F 0 m ) from light-adapted plants was estimated as , where F 0 is the ChlF yield of the first flashlet and F 0 m is the maximum ChlF yield observed between flashlets 298th and 302nd inclusive (Figure S4).
The Q A ¯reoxidation efficiency trait (Fr 0 ) is typically estimated by integrating the ChlF yield curve at a specific time interval and using the normalized integral area as a proxy for the slope of the ChlF transient to assess the kinetics of the relaxation phase (Keller, Matsubara, et al., 2019;Keller, Vass, et al., 2019).Here, however, we directly assessed these kinetics by fitting log-log regression models for the time intervals of interest, namely t 1 from 0.82 to 1.44 ms (i.e., from 303rd to 320th flashlets) and t 2 from 1.56 to 8.08 ms (i.e., from 321st to 360th flashlets), where both variables, independent (time) and dependent (ChlF yield), were log-transformed.According to the power-law relationship f x ð Þ ¼ αx β , the slope and the constant of a straight line from a log-log model equal β and logα, respectively (Marquet et al., 2005).Therefore, the efficiency of electron transport up to ~0.65 ms after reducing Q A in light-adapted plants (1) was estimated as the slope β of the log-log regression fitted within the t 1 interval, while the efficiency of electron transport up to ~6.64 ms after Fr1 0 (i.e., Fr2) was equal to the slope β of the log-log model fitted within the t 2 interval (Figure S4).The time intervals t 1 and t 2 approximate the time frame in which electron transfer from Q A to the PQ pool, and to some extent from the PQ pool to PSI, respectively (de Wijn & van Gorkom, 2001;Govindjee, 2004;Keller, Vass, et al., 2019;Kolber et al., 1998;Osmond et al., 2017Osmond et al., , 2019;;Stirbet & Govindjee, 2011).These time intervals are also supported by various in silico models (Lazár, 2003;Lazár & Jablonský, 2009;Stirbet & Strasser, 1995;Xin, Yang, & Zhu, 2013;Zhu et al., 2005).Moreover, in our study, the time intervals for estimating t 1 and t 2 remained the same regardless of genotype, water treatment and/or growing season.
The phenotypic plasticity of a given genotype for a ChlF trait in Y1 was based on the overall drought-induced relative change (%Δ), being estimated as ChlF final ÀChlF initial ChlFinitial Â 100 , where ChlF initial is the mean value of genotype i for a ChlF trait (F 0 q =F 0 m , Fr1 0 or F 0 r2 ) at 0 DAWW (non-stress), and ChlF final is the mean value of genotype i for a ChlF trait at 16 DAWW (severe stress).

| Spectral reflectance
The calibrated reflectance was estimated by normalising the target spectrum against the dark and white references.Thus, plant canopy reflectance [%] at 1 nm interval from 400 to 800 nm was calculated as DNrawÀDNdark DNwhiteÀDNdark h i , where DN raw is the raw digital value of the target, DN dark is the dark reference measurement and DN white is the white reflectance measurement (Bruning, Berger, Lewis, Liu, & Garnett, 2020).As red edge-based vegetation indices have been shown as good indicators of leaf area index (LAI), leaf and canopy chlorophyll content and plant water content (Bruning et al., 2020;Dong et al., 2019;Filella & Peñuelas, 1994;le Maire, François, & Dufrêne, 2004;Liu et al., 2004;Liu, Miller, Haboudane, & Pattey, 2004;Mutanga & Skidmore, 2007), we derived the Vogelmann red edge index (VOGREI) from the calibrated spectrum reflectance as the ratio λ 740 nm λ 720 nm À Á (Vogelmann, Rock, & Moss, 1993).

| Data cleaning
We

| Statistical analysis
A linear mixed model (LMM) approach was used to analyse the resolvable row-column designs with repeated measures for both Y1 and Y2.
Single-stage analysis models were applied to partition variance components and to estimate genotypic effects for all traits based on "Best Linear Unbiased Prediction" (BLUP) (Robinson, 1991).Variance components for each trait were estimated by residual maximum likelihood (REML) using the Average Information (AI) algorithm with sparse matrix methods (Gilmour, Thompson, & Cullis, 1995), as implemented in GENSTAT (VSN International, 2019).For random terms, the variance components were set to be positive-constrained.Details regarding the baseline models (BL), with covariates (BL Cov ) and spatiotemporal modelling (BL CovþSTM ), as well as the model selection with LMM are in Appendix B (see after the reference list).The adjusted coefficient of determination (R 2 adj )-like statistic for the final "best" LMM was estimated based on the average semivariance approach (Ω ASV β ), as proposed by Piepho (2019).Conditional F test statistic was used to test fixed effects.The Fisher-Hayter procedure, a modified LSD (MLSD) test using the Studentized Range statistic (Hayter, 1986), was used to perform pairwise comparisons between adjusted means.
The impact of adding covariates and modelling the spatiotemporal correlations was evaluated by means of relative efficiency (RE) in terms of the error size.Thus, RE was used to assess the improvement in precision of the alternative models over the BL models (i.e., models without covariates and/or spatiotemporal covariance structures) for Y1 and Y2, respectively.The RE [%] was calculated as suggested by Qiao, Basford, DeLacy, and Cooper (2000) and can be defined as SEDBL SEDAT Â 100 , where SED is the REML-based average standard error of the difference between genotype means for the baseline model (SED BL ) and for the alternative models (SED AT ).The higher the RE estimate, the better the precision of the field evaluation of genotypes.

| Heritability and trait correlation estimation
Broad-sense heritability on an entry-mean basis (H 2 ), or repeatability, of a trait for a single time point was estimated according to Cullis, Smith, and Coombes (2006) as where υ BLUP ΔÁÁ is the mean variance of a difference of two BLUPs for the genotypic effect and σ 2 g is the genotypic variance.Bivariate LMM (see details in Piepho, 2018;Piepho & Möhring, 2011) were used to estimate genetic correlations (r g ) between each pair of traits (e.g., between ChlF traits and SDMY or ΔRWC) in each time point.Details regarding these models are in Appendix B. Coefficients of phenotypic correlation r p between traits over time were estimated by Pearson's coefficients of correlation between the BLUEs ("Best Linear Unbiased Estimators"; Appendix B) of genotypes.

| Covariates and spatiotemporal modelling
Before covering our key findings, here we report the overall outputs of model selection with LMM.The model fit for the ChlF traits considerably improved after accounting for the biological and experimental sources of variation, as well as for the spatiotemporal correlations among neighbouring plots (Table S3).All models fitted to ChlF traits for Y1 and Y2 are reported in Tables S4-S9, respectively.The addition of fixed regression coefficients on top of the baseline models (i.e., BL Cov ) to accommodate differences in plant height (i.e., RelF), in plant growth and development (iZDS or ZDS), in canopy structure and leaf pigments (i.e., VOGREI), along with fluctuating environmental factors which were recorded at a 5-s interval (e.g., PPFD and VPD), promoted a net gain in precision, in terms of relative efficiency (RE) estimates, by 25.5, 28.9 and 23.4%, on average, for F 0 q =F 0 m , F 0 r1 and F 0 r2 , respectively.Models with the further addition of spatiotemporal trends (BL CovþSTM ) resulted in smaller but meaningful gains in precision relative to BL Cov , yielding total final improvements of 29.6, 36.5 and 27.3%, on average, for F 0 q =F 0 m , F 0 r1 and F 0 r2 , respectively.These final "best" fit models (BL CovþSTM , as indicated in Tables S4-S9), whose coefficients of determination (Ω ASV β ) ranged from .49 to .73 (Table S3), served as the basis for all results hereafter reported.The conditional F test statistics for fixed effects are found in Table S10.

| Weather conditions and drought severity
Overall, south-central Arizona's climate conditions were quite distinct between the two growing seasons.According to the National Centres for Environmental Information (NOAA, 2020), the 6-month period (November-April) in the 2017/2018 season (Y1) was characterized as the warmest and the driest on record for a 126-year period (1895-2020), whereas the 2018/2019 season (Y2) was the 53rd warmest and the 75th driest for the same period.The average temperature, precipitation and Palmer Z Index, as a measure of short-term drought severity, for the 6-month period (Nov-Apr) in Y1 were 16.6 C (+3.0 C anomaly compared to the 1901-2000 mean), 41.7 mm (À99.1 mm anomaly) and À2.28 (severe drought; À2.39 anomaly), against 13.9 C, 147.3 mm and 0.19 in Y2, overall a near-normal season (NOAA, 2020).
Mean meteorological data for the time period when LIFT data were recorded in both growing seasons are found in Figure S5.The higher atmospheric water demand in Y1 led to a faster and more acute reduction in soil moisture compared to Y2 (Figure S6).After withholding water, θ PWP at 10 cm depth was reached in roughly 3 and 17 days in Y1 and Y2, respectively.From this time point until the last day of field measurements using the LIFT sensor, the soil VWC dropped ~27.1% up to 16 DAWW in Y1, and only ~11.2% up to 25 DAWW in Y2.The WD rows in Y2 were, on average, 31.1% drier at 10 cm depth compared to the WW counterpart rows at D1, even though soil VWC was still slightly above (~8.5%) the θ PWP (Figure S6).At the D2 and D3 time points, WD rows were, on average, 43.1 and 45.5% drier than WW rows, respectively.The soil VWC for WD rows was around and slightly below (~10.0%) the θ PWP in D2 and D3, respectively (Figure S6).Overall, soil moisture in WW rows at 10 cm depth remained at 68.5% (standard deviation, SD = 7.2) of the θ FC over time in Y2.

| Effects of water stress on leaf RWC and above-ground biomass
In Y1, when the drought was severe, the mean relative change in leaf RWC (ΔRWC) from 0 to 15 DAWW was À25.3%Δ (Figure S7a).High variability among the 252 genotypes was observed (p < .001),where ΔRWC for the least and the most dehydrated genotypes were À14.9%Δ (standard error, SE = 1.98) and À44.1%Δ (SE = 2.98), respectively.In Y2, when the drought was milder, even after imposing 21 days of water-limiting conditions, leaf RWC of WD plants decreased, on average, only by À1.83%Δ (SE = 0.13) compared to WW plants (Figure S7b).Despite this mild stress and a minor drop in RWC in Y2, the water treatment had a major effect on the total shoot dry matter yield (SDMY), F(1, 27.2) = 176, p < .001,where WD plants produced, on average, 4,815 kg/ha (SE = 52.4),and WW plants produced 6,020 kg/ha (SE = 65.6).Although SDMY greatly varied among genotypes (p < .001) in Y2, the genotype-by-water treatment interaction was not significant (p = .258).In Y1, the mean SDMY at the end of the severe stress was 5,161 kg/ha, varying from 4,343 kg/ha (SE = 146) to 5,801 kg/ha (SE = 267) among genotypes (p < .001; Figure S7a).However, because the progressive drought was imposed throughout the entire experiment, a comparison with a WW condition within the same growing season was not possible.

| Effects of drought stress on ChlF traits
The increasing severity of drought stress condition in Y1 resulted in a slow but steady reduction in F 0 q =F 0 m and in both reoxidation efficiency r1 and F 0 r2 (Figure 1a,b).At 16 DAWW, F 0 q =F 0 m , F 0 r1 and F 0 r2 significantly decreased by À14.4,À17.8 and À20.3%, respectively, relative to 0 DAWW (Figure 1c).This down-regulation of ChlF traits was well aligned with the combined effect of progressive soil moisture dry-down at 10 cm and 50 cm depths (Figure S8).Indeed, the soil VWC at 50 cm depth strongly correlated with overall daily means for ChlF traits (Figure 2).Although linearly slowing down over time, F 0 r1 tended to accelerate immediately after the onset of drought (up to 3 DAWW) before decelerating in the long term (Figures 1c and 2b).
The overall main effect of water treatment in Y2, averaged across time, was minor but significant for F 0 q =F 0 m , F(1, 13.3) = 54.7,p ≤ .001,where WD plants (mean, M = 0.507, SE = 0.0008) had slightly higher values than WW plants (M = 0.497, SE = 0.0008).In addition, F 0 r1 for WD plants (M = À0.368,SE = 0.001) was faster than for WW plants (M = À0.347,SE = 0.001), F(1, 13.0) = 205, p < .001.However, F 0 r2 for both treatments, WD (M = À0.212,SE = 0.0006) and WW (M = À0.210,SE = 0.0006), performed alike, F(1, 13.4) = 1.73, p = .211.The water treatment-by-time interaction was significant for F 0 q =F 0 m , F(2, 111) = 5.99, p = .003and for F 0 r1 , F(2, 111) = 29.1,p < .001,but of minor effect for F 0 r2 , F(2, 110) = 2.01, p = .139.At time point D1, when the soil in the WD rows was only marginally dry, F 0 q =F 0 m for WD plants was +1.08% relative to control plants (Figure S9).At D2 and D3, with further depletion of soil VWC, stressed plants showed higher F 0 q =F 0 m compared to non-stressed plants, +2.72% and +2.58%, respectively (Figure S9).Likewise, F 0 r1 for WD plants was faster than for WW plants throughout the season in the order of +3.10%, +7.50% and +7.86% for D1, D2 and D3, respectively (Figure S9).These drought-induced trends for F 0 r1 in Y2 (milder drought) were similar to those observed at the onset of the water-limiting conditions in Y1, particularly around 3 DAWW (Figure 1c).Independently of the environment, the three ChlF traits operated in a highly coordinated manner, which was observed in Y1 (Figure 3) across multiple time points with varying ambient conditions, either above or below ground.Daily means for F 0 r1 and F 0 q =F 0 m (Figure 3a) were strongly correlated, indicating that the faster or slower the electron flow from Q A towards PQ pool, the higher or lower the PSII operating efficiency.Moreover, F 0 r2 and F 0 q =F 0 m (Figure 3b) were even better correlated, suggesting that the faster or slower the electron flow from PQ pool towards PSI, the higher or lower the PSII operating efficiency.Interestingly, this correlation was even stronger than the relationship between both reoxidation processes, F 0 r1 and F 0 r2 (Figure 3c).

| Phenotypic plasticity and variability across subpopulations
The relationships between overall drought-induced percentage changes (%Δ), as a measure of phenotypic plasticity and initial (nonstress) or final (severe stress) values for ChlF traits were evaluated for the subgroups of genotypes (Figure 4).For this assessment, we A weak negative correlation was found between the initial F 0 q =F 0 m values for genotypes under well-watered conditions at 0 DAWW and their respective %Δ after experiencing severe drought up to 16 DAWW (Figure 4a).On the other hand, moderate positive correlations were found for F 0 r1 (Figure 4b) and F 0 r2 (Figure 4c), whose initial values explained (by means of R 2 ) nearly 20 and 45%, respectively, of variability in %Δ throughout genotypes.Nevertheless, there was a strong positive correlation between the final F 0 q =F 0 m values for genotypes at 16 DAWW and their corresponding overall drought-induced percentage changes (Figure 4d), which explained around 70% of the variability.Similarly, there was a very strong correlation (negative) for final F 0 r1 values (Figure 4e), which explained almost 90% of %Δ variability.Final F 0 r2 values (Figure 4f) were moderately and negatively correlated with their respective %Δ, explaining roughly 30% of the variability.In short, based on Figure 4a,b, it is possible to state that F 0 q =F 0 m and F 0 r1 from non-stressed plants inform little about the a posteriori effect of drought across genotypes.On the other hand, F 0 q =F 0 m and F 0 r1 from severely stressed plants can better indicate those genotypes most and least affected by drought (i.e., with the highest and the lowest %Δ), Figure 4d,e.Notably, F 0 r2 from both non-stressed and severely stressed plants may serve as indicators of the potential magnitude of drought effect across genotypes, but interpretations require caution since opposite relationships occurred (Figure 4c,f).It is noteworthy that genotypes grouped at the tail end of %Δ within each ChlF trait (i.e., the two extreme groups of 15 genotypes ranked at the top and bottom according to drought-induced relative changes in F 0 q =F 0 m , F 0 r1 or F 0 r2 ) also had contrasting means for leaf ΔRWC, but equal means for SDMY (Table S11).
The variation among subpopulations alone accounted for approximately 7.5 (p < .001),8.0 (p < .001)and 16.7% (p < .001) of the overall observed variation in the drought-induced %Δ for F 0 q =F 0 m , F 0 r1 and F 0 r2 , respectively.The subgroup S8 had the lowest overall %Δ in ChlF traits (Table S12), suggesting that genotypes from S8 were less prone to reduce photosynthetic activities even after severe drought stress.

| Changes of heritability over time
Varying temporal dynamics of genotypic effects were observed for all ChlF traits during progressively increasing drought severity in Y1.The broad-sense heritability (H 2 ; Figure 5) for F 0 q =F 0 m increased from 0.55 to 0.64 between 0 to 5 DAWW (i.e., from non-stress to moderate drought) and then gradually decreased to 0.46 in severe drought at 16 DAWW.H 2 for F 0 r1 marginally increased to 0.59 within the first two DAWW, remained stable up to five DAWW, and then continuously decreased to 0.40 up till 16 DAWW (Figure 5).H 2 for F 0 r2 , likewise, had starting values of 0.61, peaked at 0.68 in 3 DAWW and decreased to a minimum of 0.40 (Figure 5).The unknown (residual) environmental effects increased from moderate to severe stress conditions, which could explain the decline in heritabilities (i.e., the decreasing relative contribution of genetic effects to phenotypic variance).The overall BLUP of genotypic values averaged across time for Y1 are reported in Figure S10.
For Y2, on average, H 2 for F 0 q =F 0 m , F 0 r1 and F 0 r2 were 0.65, 0.74 and 0.64, respectively, regardless of time (drought severity) and water treatment (WD and WW).Neither the genotype-by-treatment interaction effect nor the three-way interaction of genotype, treatment and time for ChlF traits were significant (p > .10)under the mild drought conditions in Y2. Figure S11 shows the overall BLUP of genotypic values, irrespective of water treatment and time.
Values of H 2 for SDMY and ΔRWC were 0.63 and 0.80, respectively, under the severe drought conditions in Y1.Moreover, H 2 for SDMY was 0.45 for both WD and WW plants in Y2 since the genotype-by-water treatment interaction effect was not significant (p = 0.26).

| Trait-trait genetic and phenotypic correlations
Bivariate LMMs for testing the genetic overlap between traits (see Methods) were fitted to investigate the trait-trait genetic correlations (r g ) over time between ChlF traits and SDMY or ΔRWC.The correlation patterns changed according to the drought severity (Figure 6).The r g between F 0 q =F 0 m and SDMY (Figure 6a) decreased during the steady progression of drought, with the highest positive correlation at 1 DAWW (0.65, SE = 0.14, p < .001),and a weak negative correlation at 16 DAWW (À0.21,SE = 0.34, p = .274).Both F 0 r1 and F 0 r2 (Figure 6a) showed similar r g patterns with final SDMY over time, and they mirrored that displayed by F 0 q =F 0 m .During the mild drought in Y2, the r g between ChlF traits and final SDMY (Figure S12) were, on average, 0.70, À0.45 and À0.65 for F 0 q =F 0 m , F 0 r1 and F 0 r2 , respectively, which were similar to those correlations observed at the onset of waterlimiting conditions in Y1, particularly between 1 and 5 DAWW.

Time-varying genetic correlations were also evident between
ChlF traits and the relative change in leaf RWC (Figure 6b).Within the first week after withholding water, r g between F 0 q =F 0 m and ΔRWC was, on average, À0.36, and increased to 0.35 (SE = 0.28, p = .128)under severe stress at 16 DAWW.On the other hand, an opposite trend was seen for both reoxidation traits, F 0 r1 and F 0 r2 , whose r g with ΔRWC were, on average, 0.36 and 0.31, respectively, throughout the first week after imposing drought, and then shifted to À0.33 (SE = 0.22, p = .109)and À0.34 (SE = 0.33, p = .162),respectively, at 16 DAWW.
Trait-trait phenotypic correlations (r p ) over time between ChlF traits and SDMY (Figure 7a) or ΔRWC (Figure 7b) were estimated by Pearson's correlation coefficients.Similar to the genetic correlations (Figure 6), the phenotypic correlation patterns dynamically changed according to the drought severity.Indeed, all results for r p exhibited analogous trends to those described for the trait-trait r g , but the magnitude of the phenotypic associations was lower, especially between ChlF traits and SDMY.In Y1, r p between SDMY and ΔRWC was À0.17 (p = .008).

| Environmental responses and diurnal courses of ChlF traits
Both PPFD and air VPD were major fluctuating environmental factors driving nonlinear changes in the ChlF traits from light-adapted plants within a measurement day.In fact, their interaction effect (logPPFD Á logVPD; i.e., both variables were log-transformed) was especially important for accounting for the variations observed in the ChlF traits under severe drought, as in Y1 (Table S10).Increasing light intensity from 150 to 2,500 μmol m À2 s À1 induced a continuous but nonlinear reduction in F 0 q =F 0 m (Figure S13a) and a deceleration in F 0 r2 (Figure S13c).Furthermore, such effects were even more pronounced when VPD increased from 1.5 to 2.5 kPa.F 0 r1 (Figure S13b) accelerated under higher VPD, especially at low light intensity (<~450 μmol m À2 s À1 ), whereas this effect was negligible when PPFD was >~900 μmol m À2 s À1 .Neither the PPFD-by-VPD interaction nor its interaction with water treatments were significant (p > .10)for explaining the variations in the ChlF traits in Y2.Accordingly, these effects were dropped from the final fitted models.Nevertheless, the single main effects of logPPFD and logVPD were still important in Y2 and did show similar trends as in Y1 (Table S10).Notably, solely for F 0 r1 trait (Table S10), the logVPD-by-water treatment interaction effect was significant, indicating that the faster rate of F 0 r1 induced by increasing VPD occurred even more rapidly in WD than in WW plants (Figure S13d).
Over the course of the day, ChlF traits exhibited clear patterns which were consistent across multiple days and growing seasons, and generally followed diurnal courses of PPFD and VPD (Figures 8 and   S14).The diurnal course of F 0 q =F 0 m (Figures 8b and S14b) showed a local minimum around midday at the peak of incoming sunlight.Relative to WW conditions (Figure 8b), WD plants tended to have slightly higher F 0 q =F 0 m values towards the afternoon, when PPFD was decreasing but VPD was at the highest.The diurnal course of F 0 r2 (Figures 8d  and S14d) mirrored that displayed by F 0 q =F 0 m , showing a local maximum (i.e., slowest rate) around midday.However, there were no evident changes between diurnal patterns of WD and WW plants (Figure 8d).By contrast, F 0 r1 (Figures 8c and S14c) was stable early morning to midday and then decreased linearly (i.e., gradually accelerating the rates of Q A ¯reoxidation) towards the afternoon.Although stressed plants always had faster rates of F 0 r1 throughout the day (Figure 8c), especially in the afternoon under increasing VPD and decreasing PPFD, the shape of the diurnal patterns of WW and WD plants were comparable.

| DISCUSSION
Changes in LIFT-measured ChlF traits of light-adapted durum wheat plants subjected to progressive soil drying were induced in both the Days after withholding water F I G U R E 5 Dynamics of broad-sense heritability on an entrymean basis (H 2 ) for operating efficiency of PSII (F 0 q =F 0 m ) and both reoxidation processes, F 0 r1 and F 0 r2 , from light-adapted durum wheat plants under progressive drought stress, from 0 to 16 days after withholding water (DAWW), in Y1.The progression of soil volumetric water content (VWC) at 10 cm depth over time is also displayed.Small arrows on the right-hand side approximate to the soil permanent wilting point (θ PWP ) and the soil field capacity (θ FC ) at 10 cm depth [Colour figure can be viewed at wileyonlinelibrary.com]F I G U R E 6 Mean trait-trait genetic correlations (r g ) ± SE, n = 252 genotypes, from 0 to 16 days after withholding water, between each chlorophyll fluorescence trait (F 0 q =F 0 m , F 0 r1 and F 0 r2 ) and (a) the total shoot dry matter yield (SDMY) at the end of the stress period, and (b) the relative change in leaf relative water content (ΔRWC) as a consequence of the severe drought stress imposed in Y1 [Colour figure can be viewed at wileyonlinelibrary.com] short and long term.In fact, a reduced photosynthetic activity, estimated by the LIFT parameters (F 0 q =F 0 m , F 0 r1 and F 0 r2 ), was only observed under a persistent moderate to severe drought (Figure 1c).It is worthy of note, though, that the absence of a contiguous control treatment (WW) in Y1 may have limited our ability to isolate small changes caused solely by drought, particularly when it was not severe.Indeed, the photosynthetic machinery, especially PSII photochemistry, is known to be relatively resilient to water stress (Flexas et al., 2009;Flexas, Escalona, & Medrano, 1998;Havaux, 1992;Kaiser, 1987).Such a resilience, however, has been shown to be more pronounced for ChlF traits measured in dark-adapted plants as the maximum quantum efficiency of PSII photochemistry (F v =F m ), rather than in light-adapted plants as F 0 q =F 0 m (Athar & Ashraf, 2005;Lu & Zhang, 1999;Zivcak et al., 2014).Therefore, these findings indicate that light-adapted ChlF traits might be physiologically preferable for assessing the effects of environmental stressors.Concurrent electron acceptor sinks (i.e., photosynthetic carbon reduction and carbon oxidation) may explain the high resilience of ChlF traits, particularly F 0 q =F 0 m and F 0 r2 , even after 10 days of waterlimiting conditions in Y1 (Figure 1c) and also the fact that no declines relative to WW plants were found under mild drought in Y2.Stomatal ( g s ) and mesophyll ( g m ) conductance are key CO 2 diffusion components that regulate leaf transpiration efficiency, playing pivotal roles in plant acclimation to drought (Flexas et al., 2009;Ouyang, Struik, Yin, & Yang, 2017).In the short term, at the onset of water-limiting conditions, stomatal closure is induced to reduce water loss and thereby CO 2 availability, leading to increased photorespiration (Cornic, 2000;Lawlor, 2002).In C 3 plants under mild drought, the O 2 uptake via photorespiratory activity can almost entirely replace the lower CO 2 availability as an electron acceptor pathway (Cornic & Fresneau, 2002).Drought-stressed tomato plants, for instance, doubled electron dissipation through photorespiration relative to nonstressed plants (Haupt-Herting & Fock, 2002).This repartitioning of light energy, or energy balancing network (Walker, Kramer, Fisher, & Fu, 2020), may result in minor changes in ChlF-based traits in mild stress.Nevertheless, if drought progresses, LET might be electron sink-limited, as both photorespiration and the Calvin-Benson cycle can be repressed, causing impairment of ribulose bisphosphate (RuBP) regeneration and adenosine triphosphate (ATP) synthesis (Flexas & Medrano, 2002;Tezara, Mitchell, Driscoll, & Lawlor, 1999).Ultimately, whole photosynthetic electron transport activity will be downregulated (Cornic & Fresneau, 2002;Haupt-Herting & Fock, 2002;Medrano, Escalona, Bota, Gulías, & Flexas, 2002;Schöttler & T oth, 2014).Hence, these considerations can explain the decrease in F 0 q =F 0 m co-occurring with reduced F 0 r1 and F 0 r2 in response to long-term severe drought stress observed in Y1.
The very strong linear relationship between F 0 q =F 0 m and F 0 r2 (i.e., the kinetics of electron transport from PQ pool towards PSI; are known to be at the core of plant acclimation to long-term drought stress (Kohzuma et al., 2009).Therefore, a slower F 0 r2 might suggest a deceleration of electrons through Cytb 6 f with simultaneous decreasing in F 0 q =F 0 m due to a higher NPQ.Regardless of drought, LIFT-measured ChlF traits displayed similar diurnal temporal patterns (Figures 8 and S14), which demonstrate the high level of inherent regulation of the photosynthetic apparatus under fluctuating growing conditions, particularly through, but not limited to, NPQ.Similar patterns have also been observed in various other plant species growing in open fields (Pieruschka et al., 2010(Pieruschka et al., , 2014;;Raesch et al., 2014).Instantaneous light intensity and temperature have been reported by Keller, Matsubara, et al. (2019) as the key drivers of such dynamics, in agreement with our findings.Actually, because air VPD is strongly correlated with air temperature (Figure S15; Gates, Zolnier, & Buxton, 1998;Yuan et al., 2019) and includes air relative humidity, we investigated the impacts of VPD instead.It is well known that adjustments in the PSII/PSI stoichiometry are crucial to optimize the quantum efficiency of photosynthesis under fluctuating environment (Chow, Melis, & Anderson, 1990;Külheim, Ågren, & Jansson, 2002).
Recently, Grieco et al. (2020) reported short-and long-term photosynthetic adjustments to drought and fluctuating light in wheat by changing NPQ relaxation and the PSII-LHCII phosphorylation pattern combined with a fine-tuning in protein stoichiometry.Indeed, the efficiency of electron transport under changing ambient conditions is highly dependent on the tight co-ordination among the several electron carriers between PSII and PSI, whose intricate regulatory processes occur at different time scales and multiple sites (see details in Dietz, 2015;Horton, 2012;Kono & Terashima, 2014;Rochaix, 2011;Schöttler & T oth, 2014;Tikkanen et al., 2012;Walters, 2005).
The sustained faster F 0 r1 (i.e., the kinetics of electron transport from Q A towards PQ pool) in Y2 relative to control (Figure 8c) was the most remarkable effect of mild drought.This suggests that plants could sense a subtle shortage of soil moisture, quickly modulate their photosynthetic electron transport and potentially trigger responses to either acclimate or cope with reduced water availability.Interestingly, increasing atmospheric VPD intensified this response (Figure S13d).
We hypothesize that stomata responses combined with alternative electron flows apart from the LET, such as water-water cycle (WWC), cyclic electron flow around PSI (CEF) and/or chlororespiration mediated by the plastid terminal oxidase (PTOX) (Cruz et al., 2005;Kono & Terashima, 2014), might explain a faster F 0 r1 at the onset of drought.Takahashi, Milward, Fan, Chow, and Badger (2009) demonstrated that CEF enhances the pmf and helps to alleviate photoinhibition by either suppressing photodamage to PSII via a q E -independent mechanism or preventing the inhibition of the repair of photodamaged PSII via a q Edependent mechanism.It has long been recognized that CEF enhances under drought (Golding, Finazzi, & Johnson, 2004;Golding & Johnson, 2003;Zivcak et al., 2014).In addition, PTOX mediates the electron transfer from plastoquinol (PQH 2 ) to reduce O 2 to H 2 O via a non-electrogenic process (Shirao et al., 2013), potentially acting as a safety valve by protecting the PQ pool and mediating physiological responses (Krieger-Liszkay & Feilke, 2016;McDonald et al., 2011;Nawrocki, Tourasse, Taly, Rappaport, & Wollman, 2015).Indeed, both CEF and PTOX are dependent on the redox state of the PQ pool, a vital component of photosynthesis with multiple functions, including photoprotection and stress tolerance (Havaux, 2020).Remarkably, Wang et al. (2016) recently demonstrated the role of the PQ pool over-reduction as a mechanism of chloroplast-mediated stomatal closure.Moreover, it is known that rising VPD increases atmospheric demand for water, leading to stomatal closure (Franks, Cowan, & Farquhar, 1997;Massmann, Gentine, & Lin, 2019).Besides VPD, diurnal and seasonal stomata kinetics are also driven by combined effects of temperature, irradiance and soil moisture (Matthews, Vialet-Chabrand, & Lawson, 2018;McAusland et al., 2016;Neukam, Böttcher, & Kage, 2016;Sack & Holbrook, 2006).Altogether, early changes in F 0 r1 may appear as acclimation responses to the onset of water-limiting conditions, which might promote photoprotection, even when drought stress effects are not obvious at the whole-plant level.Our results show dynamic fluctuations in broad-sense heritability for ChlF traits (Figure 5) due to changes in the relative contribution of genetic variance over time.Even after correcting the ChlF traits for biological (e.g., plant height, phenology and canopy reflectance) and environmental (e.g., PPFD and VPD) variations, strong significant differences between genotypes were still found, demonstrating that there was substantial genetic variability that could not be explained by those covariates alone.Araus, Amaro, Voltas, Nakkoul, and Nachit (1998) have similarly reported genetic variability for ChlF traits in durum wheat under field conditions, where phenology was also considered.In our data, the highest H 2 for F 0 q =F 0 m , F 0 r1 and F 0 r2 were observed in mild drought but with gradual reductions when the soil became drier.It is noteworthy that H 2 values estimated in Y2 (milder drought) were of similar orders of magnitude as in Y1 at the onset of drought stress (i.e., when the soil VWC of both seasons were comparable).Using a high-throughput image phenotyping approach, Chen et al. ( 2014) also observed dynamic changes in heritability over time for fluorescence-based traits in barley under drought, where H 2 similarly decreased during progressive stress.Time-varying H 2 for F 0 q =F 0 m in Arabidopsis growing in fluctuating light has also been reported by Flood et al. (2016).It has been argued that the dynamic change of heritability over time is due to changes in the magnitude of genotype and environment effects, as well as their interaction (Visscher, Hill, & Wray, 2008).Fluctuations in H 2 for a trait can be challenging for plant breeding programmes, particularly in drought-prone environments, where a lower heritability under severe stress could negatively impact the effectiveness of selection.
The changing genetic and phenotypic correlations between ChlF traits and above-ground biomass yield (Figures 6a and 7a) or ΔRWC (Figures 6b and 7b) during soil drying might suggest that multiple water use strategies are in place to cope with water deficit.In mild drought, genotypes with high photosynthetic activity tended to have both high biomass yield and high dehydration, altogether indicating a high transpiration rate.As the opposite behaviour was also true (i.e., low photosynthetic activity with low biomass yield and low dehydration), the identification of water savers and spenders (Nakhforoosh, Bodewein, Fiorani, & Bodner, 2016) may be somewhat facilitated by ChlF values, at least under mild stress.Nonetheless, such correlations were weaker, or even shifted directions, in severe drought.These circumstances were probably due to other traits related to drought tolerance which may also affect ChlF responses, including stay-green (delayed senescence), osmotic adjustment and antioxidant defence (Chen et al., 2017;Christopher, Christopher, Borrell, Fletcher, & Chenu, 2016;Farooq, Hussain, & Siddique, 2014).
It is also known that plants under drought stress can adapt by altering biomass partitioning among roots and grain development (Davies & Zhang, 1991;Fang et al., 2017).However, it was beyond the scope of our study to evaluate root dynamics and possible changes due to drought.
Apart from the high level of phenotypic plasticity for ChlF traits among genotypes, a crossover interaction was noticed (Figure 4), especially for F 0 r1 and F 0 r2 , which reinforces the roles of genotype-bytime (i.e., drought severity) effect and the genetic complexity of plant responses to drought.In other words, genotypes with high photosynthetic activity (high F 0 q =F 0 m and fast F 0 r1 and F 0 r2 ) in non-limiting environments will likely perform worse when grown under very poor conditions compared to those with low photosynthetic activity, and vice versa.A crossover effect has also been reported for grain yield (Araus, Slafer, Reynolds, & Royo, 2002;Araus, Slafer, Royo, & Serret, 2008;Cooper, Stucker, DeLacy, & Harch, 1997) and is a "source of frustration" (Blum, 2005) to plant breeding for drought stress adaptation.Properly setting the target environment, therefore, seems to be essential for accurate comprehension of and good use of the genetic variation in ChlF traits.

| CONCLUSIONS
We here demonstrated that short-and long-term changes in ChlF traits induced by progressive drought were rapidly and non-invasively monitored at canopy level in durum wheat under field conditions.
Integrating LIFT-measured ChlF traits with high temporal resolution environmental data facilitated the assessment of genotype-byenvironment interaction effects under drought stress.Simultaneous statistical modelling of spatial patterns and temporal trends combined with time-varying covariates helped to improve the precision and interpretation of experiments under changing ambient conditions.Strong significant differences in ChlF traits were found among genotypes, demonstrating that there is substantial genetic variability available for breeding programmes to select for drought-adaptive traits.This also suggests that the LIFT method can enable genome-wide association studies (GWAS) for dissecting the QTLome of photosynthetic traits and assess the effects on yield associated with the relevant quantitative trait loci (QTLs).At an unprecedented scale, our high-throughput approach for field phenotyping ChlF traits not only allowed for estimation of genetic effects over time in a large durum wheat panel but also shed light on the diurnal dynamics of the photosynthetic apparatus, leveraging the ability to dissect complex physiological traits.We propose that plant ecophysiology studies and physiological plant breeding benefit from more flexible and versatile methods to measure photosynthesis and related traits, enabling knowledge of the mechanisms of drought adaption under natural plant stand and agricultural field conditions alike.

ACKNOWLEDGMENTS
where T is time points (i.e., days after withholding water), the repeated factor.On top of the Model (B2), covariates were also included, and the full baseline model (BL Cov ) for Y1 was defined as varying covariates.Model (B3) was used to fit F 0 q =F 0 m , F 0 r1 and F 0 r2 : Besides the spatial modelling of trends along rows and columns, temporal correlation structures (Littell, Pendergast, & Natarajan, 2000;Payne et al., 2019), such as compound symmetry (CS), banded Toeplitz (BAND), power (POW), ante-dependence (ANTE), unstructured (UN) or general correlation (COR), with equal or unequal variances were also fitted to the residuals to accommodate trends over time due to multiple observations on the same experimental unit.
Modelling of serial correlation was also extended to G Á T in order to assess genetic correlations for the same trait across time, allowing for heterogeneity of genetic variances.For estimating BLUEs for each time point, G=T were fitted as fixed in the Model (B3).
In Y2, a LMM for each phenotype trait measured in each time point after imposing water treatment was defined as where TRT is the water treatment and SUB the subrows.Model (B4) was used to fit SDMY, which was natural log-transformed.For estimating BLUEs within each water treatment, G=TRT were fitted as fixed in the model (B4).
For assessing the differences among levels of drought severity, repeated measures were incorporated into the Model (B4).Therefore, the BL model, which considers the entire observed data in Y2 in one stage at the level of individual plots, was defined as where T is time points after imposing water treatment (i.e., the levels of drought severity), the repeated factor.Covariates were also added as fixed in the Model (B5) to account for the fluctuating environment, and so the BL Cov model for Y2 was defined as where ZDS is the growth stage in the Zadoks scale.RelF, ZDS, VOGREI, PPFD and VPD were time-varying covariates.Model (B6) was used to fit F 0 q =F 0 m , F 0 r1 and F 0 r2 in Y2.Modelling of spatiotemporal (STM) correlations was also performed for the residuals only.For estimating BLUEs within each water treatment and across time points, G=T and G Á TRT ð Þ =T were fitted as fixed in the Model (B6).
In all models, covariates were mean centred, except RelF which was centred to one (i.e., target area at a distance of 0.60 m).PPFD and VPD were also log-transformed and fitted to BL Cov models to evaluate whether a nonlinear relationship with the ChlF traits would improve model fit.
Comparison between candidate models was assessed by the REML-likelihood ratio test (REMLRT), provided that the two models being compared were nested and had the same fixed effects model (Galwey, 2014).Otherwise, when models were non-nested, yet with the same fixed effects, the Akaike Information Criterion (AIC) was used to assess their goodness-of-fit, judged by the "smaller the better" form of the criterion (Cheng, Edwards, Maldonado-Molina, Komro, & Muller, 2010).
Bivariate LMM for estimating trait-trait genetic correlations as the response vector of observed phenotype for the trait k (k ¼ 1, 2), the bivariate model for a single time point was defined as where TRAIT stands for the traits of interest.Model (B7) was used to assess the correlation between SDMY and ΔRWC in Y1.Covariates were also accommodated to the bivariate models for assessing the genetic correlations between each ChlF trait and SDMY or ΔRWC in Y1 as where is a quantitative Z variable, being set to Z = 0 when TRAIT = SDMY or ΔRWC, and to Z ¼ 1 when TRAIT ¼ F 0 q =F 0 m , F 0 r1 or F 0 r2 .Coefficients of genotypic correlation between pairs of traits were defined as where σ g 1 g 2 is the genetic covariance between two traits, σ 2 g 1 and σ 2 g 2 are the genotypic variances of both traits under analysis; such variance-covariance components were estimated through the random G Á TRAIT effect in the Models (B7) or (B8).The REML estimate of ρ g is denoted as r g .The REMLRT was used to estimate the significance of the genetic correlations by comparing the model with varying genetic covariance between the two traits and the model with the genetic covariance fixed to zero.
to estimate relevant population parameters (genotypic variation and heritability) for the LIFT-measured ChlF traits; (b) to evaluate the phenotypic plasticity of ChlF-related traits in response to drought; (c) to assess correlations between ChlF traits and relative change in leaf relative water content (ΔRWC), or total shoot dry matter yield (SDMY); and (d) to assess the effects of a fluctuating environment, in terms of light intensity and vapour pressure deficit (VPD), and their interactions with varying soil moisture, on ChlF traits.
accessions bred and released by CIMMYT (the International Maize and Wheat Improvement Centre), ICARDA (the International Centre for Agricultural Research in the Dry Areas), INRAE (the French National Institute for Agriculture, Food and Environment), IRTA (the Spanish Institute of Agriculture and Food Research and Technology) and by public breeding programmes in Italy, in the Northern Great Plains of the USA and Canada (North Dakota, Montana, Saskatchewan and Alberta), and materials from the Southwestern USA, namely "Desert Durum ® ", which are co-owned by the Arizona Grain Research and Promotion Council and the California Wheat Commission.Recently, Condorelli et al. (2018) reported the occurrence of a strong population genetic structure in the UNIBO-Durum Panel, identifying eight distinct subpopulations, despite a considerable admixture mostly among germplasm from ICARDA, CIMMYT and Italy.These subgroups were also considered for further assessment in our study.

Field
experiments were conducted at the Maricopa Agricultural Centre of the University of Arizona (33.07454N, 111.97494W, elevation 360 m) in Maricopa, AZ, on a Casa Grande sandy loam soil (fine-loamy, mixed, superactive, hyperthermic Typic Natrargids) under a semi-arid low desert climate.In the first growing season 2017/2018 (Year 1 = Y1), a total of 252 accessions were planted on November 28, 2017, while in the subsequent season, 2018/2019 (Year 2 = Y2), rows at three locations within the experimental field at 2, 10, 25 and 50 cm depths and at 15 and 35 cm depths between seed rows.Additional soil sensors were installed between rows at 15 cm depth for measuring the soil matric potential (Tensiomark, ecoTech GmbH, Bonn, Germany).All soil sensors recorded data at 15-min intervals throughout the entire growing seasons.Based on characterisations of the soil hydraulic and physical properties of the experimental site by Prof. Dr Markus Tuller at the University of Arizona (Tucson, AZ) under the TERRA-REF project, the permanent wilting point (θ PWP ) and the field capacity (θ FC ) at 10-15 cm depth, based on the van Genuchten (1980) model, corresponded to approximately 0.110 and 0.282 m 3 /m 3 , respectively.
16:40 hours local Mountain Standard Time (MST).In Y2, aiming to achieve three different levels of drought severity between WW and WD (D1, D2 and D3), field data were collected at three time points after imposing water treatment.However, due to the size of the trial (224 genotypes Â 2 treatments Â 3 replicates), only one replicate per day was operationally possible between 08:45-15:55 hr MST, on average.Thus, three consecutive days were required to phenotype the entire experiment with three replicates: D1 was taken between 12 and 14 days after imposing water treatment, D2 between 17 and 19 days and D3 between 23 and 25 days.Further field operational details related to the LIFT data collection are reported in Figures S2 collected a total of 110,880 and 80,640 ChlF transients (and spectral data) over time in Y1 and Y2, respectively.The LIFT sensor relies on its artificial excitation light source to induce ChlF emission from the target canopy, and a ChlF transient, such as in Figure S4, is only possible in the presence of living photosynthetic tissues.However, due to the highly fluctuating environment, particularly solar irradiation and winds, shifting of the leaves, off-target measurements (e.g., soil), and/or technical constraints, low-quality data can occur.Hence, a data cleaning pipeline was defined and implemented in the R environment (R Core Team, 2020).Further details about the criteria for data cleaning are in Appendix A (see after the reference list).After this step, the remaining data points, 100,947 (91%) and 77,946 (97%) ChlF transients in Y1 and Y2, respectively, were averaged, resulting in one value per trait per plot per time of measurement (N = 5,544 data points per trait in Y1; and N = 4,032 data points per trait in Y2).The raw data and the processed and cleaned datasets for both growing seasons are publicly accessible (https://doi.org/10.5281/zenodo.4305673).

F
I G U R E 1 LIFT-measured chlorophyll fluorescence (ChlF) traits from light-adapted durum wheat plants in response to progressive drought stress, from 0 to 16 days after withholding water (DAWW), in Y1.(a) Operating efficiency of PSII (F 0 q =F 0 m ); and (b) efficiency of electron transport up to ⁓0.65 ms after reducing Q A (i.e., F 0 r1 ; the kinetics of electron transfer from Q A to PQ pool), and up to ⁓6.64 ms after F 0 r1 (i.e., F 0 r2 ; the kinetics of electron transfer from PQ pool to PSI).Values are means, averaged across genotypes, with ±95% confidence intervals (CIs), n = 252.(c) Percentage changes (%Δ) in ChlF traits over time.Mean relative changes with ±95% CIs are shown with respect to the baseline value at 0 DAWW.Pairwise comparisons between time points and baseline within each ChlF trait were performed by the Fisher-Hayter procedure.The levels of significance are indicated by * (p ≤ .05),** (p ≤ .01),and *** (p ≤ .001),otherwise blank if p > .05.Dashed arrow indicates when the soil permanent wilting point (θ PWP = 0.11 m 3 /m 3 ) at 10 cm depth was reached traits, F 0 grouped genotypes into subpopulations (S) based on the genetic structure in the UNIBO-Durum Panel reported by Condorelli et al. (2018), as follows: S1 includes Mediterranean and North African germplasm; S2 includes cultivars bred for dryland areas at ICARDA (Syria) in the early 1970s; S3 includes mainly IRTA (Spain) and INRAE (Morocco) accessions bred in early 1970s, and CIMMYT and ICARDA accessions selected for temperate areas; S4 contains predominantly high-yielding materials for temperate zones from ICARDA, and some Italian accessions from the 1970s; S5 comprises materials derived from broadly adapted (photoperiod-insensitive) CIMMYT germplasm released between the late 1970s and the early 1980s; S6 includes Italian accessions from the mid-1970s; S7 includes mostly high-yielding CIMMYT genotypes released from the late 1980s and the early 1990s; S8 contains American (North Dakota), Canadian, French and Australian genotypes; and finally, mostly due to a significant exchange of genetic resources among international breeding programmes, an admixture subgroup is composed mainly by ICARDA, CIMMYT and Italian materials.
Relationships between LIFT-measured chlorophyll fluorescence traits from light-adapted durum wheat plants and the soil volumetric water content (VWC) at 50 cm depth in Y1.(a) Operating efficiency of PSII (F 0 q =F 0 m ); and both reoxidation processes, (b) F 0 r1 and (c) F 0 r2 .Values are means, averaged across genotypes, with ±95% confidence intervals, n = 252.Nearby points, the numerical label indicates the timing (in days) after withholding water (DAWW).The colour schemes indicate drought progression over time, from non-stress (darker colours) to severe stress (lighter colours) [Colour figure can be viewed at wileyonlinelibrary.com] Relationships among LIFT-measured chlorophyll fluorescence traits from light-adapted durum wheat plants under progressive drought stress in Y1.(a) Relationship between F 0 r1 (the kinetics of electron transport from Q A to PQ pool) and operating efficiency of PSII (F 0 q =F 0 m ); (b) relationship between F 0 r2 (the kinetics of electron transport from PQ pool to PSI) and F 0 q =F 0 m ; and (c) relationship between both reoxidation processes, F 0 r1 and F 0 r2 .Values are means, averaged across genotypes, with ±95% confidence intervals, n = 252.Nearby points, the numerical label indicates the timing (in days) after withholding water (DAWW).The colour scale for points is based on the soil volumetric water content (VWC) at 10 cm depth, and the point sizes are based on the soil VWC at 50 cm depth [Colour figure can be viewed at wileyonlinelibrary.com]R 2 =.03, p =.003 /F m ' (a.u.) − [0 DAWW] Overall percentage change (%∆) .28,p <.001 −0.15 −0.14 −0.13 −0.12 −0.11 −0.10 F r2 ' (a.u.) − [16 DAWW] Relationships across durum wheat genotypes between the overall drought-induced percentage changes (%Δ; i.e., the size of the absolute change between 0 to 16 DAWW in comparison to the reference value at 0 DAWW) in chlorophyll fluorescence (ChlF) traits and their initial (a) F 0 q =F 0 m , (b) F 0 r1 , and (c) F 0 r2 values from non-stressed plants at 0 days after withholding water (DAWW), or their final (d) F 0 q =F 0 m , (e) F 0 r1 , and (f) F 0 r2 values from severely stressed plants at 16 DAWW in Y1.Genotypes were assembled into subgroups (S) according to the population genetic structure reported by Condorelli et al. (2018) [Colour figure can be viewed at wileyonlinelibrary.com]

Figure 3b )FF
Figure3b) supports the mechanism of photosynthetic control of electron transfer when metabolism is repressed under environmental stresses to prevent photodamage in both PSII and PSI.According toKanazawa et al. (2017), ATP synthase activity decreases in limiting CO 2 , slowing proton efflux from the thylakoid lumen and, consequently, increasing proton motive force (pmf ) across the thylakoid membrane.As a consequence, a more acidic lumen can concomitantly (a) trigger the energy-dependent (q E ) non-photochemical quenching (NPQ), which thermally dissipates the surplus of absorbed light energy from the light-harvesting complexes (LHCs) to prevent over-excitation of PSII, as well as (b) slow down the electron transfer through the cytochrome b 6 f complex (Cytb 6 f ), which prevents over-reduction of PSI electron acceptors(Kanazawa et al., 2017;Tikhonov, 2013).These mechanisms of photoprotection governed by ATP synthase activity

F
I G U R E 8 Diurnal course of (a) photosynthetic photon flux density (PPFD) and air vapour pressure deficit (VPD), (b) operating efficiency of PSII (F 0 q =F 0 m ) and both reoxidation processes, (c) F 0 r1 and (d) F 0 r2 , for well-watered (WW) and water-limited (WD) durum wheat plants in light-adapted conditions in Y2.Values are M ± SD, averaged across days of field phenotyping, n = 9.The local time zone is Mountain Standard Time (MST).The discontinuity around midday relates to a daily operational break during the LIFT data collection (see details in Figure S3) [Colour figure can be viewed at wileyonlinelibrary.com] This research was performed partly within the German Plant Phenotyping Network (DPPN), which was funded by the German Federal Ministry of Education and Research (BMBF; Project Identification Number 031A053), and partly within the TERRA-REF experiment, funded by the Advanced Research Projects Agency-Energy (ARPA-E), U.S. Department of Energy (DOE), under Award Number DE-AR0000594 with additional support from the U.S. Department of Agriculture (USDA), Agricultural Research Service (ARS), under agreement number 58-2020-6-019 with the University of Arizona.The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency where iZDS is the initial growth stage in the Zadoks scale measured 2 days before withholding water, VOGREI the Vogelmann red edge index, PPFD the photosynthetic photon flux density, VPD the vapour pressure deficit and RelF is the relative deviation of the target area from the focal point of the LIFT light beam set at 0.60 m, calculated as LIFT height ÀPH 60 h i , where LIFT height is the distance from the soil surface to the LIFT lens [cm] and PH is the plant height [cm].RelF and iZDS were time-constant covariates, whilst VOGREI, PPFD and VPD were time-