Multiplex^® (FORCE-A, Orsay, France, patent pending) is a hand-held, multi-parametric fluorescence sensor based on light-emitting-diode (LED) excitation and filtered-photodiode detection that is designed to work in the field under daylight on leaves, fruits and vegetables (Figure 1).

A block diagram of Multiplex^® functions is shown in Figure 2. The present version of the sensor has a yellow (Y) filter at the third emission channel (it may also have a blue filter for blue-green fluorescence, BGF, according to FORCE-A). The sensor has six UV-light sources (LED-matrices) at 375 nm protected by DUG11 filters (Schott, Mainz, Germany) and it has three, Red-Blue-Green (RGB) LED-matrices emitting lights at 470 nm (blue, B), 516 nm (green, G) and 635 nm (red-orange, R) protected by a 650-nm short-pass filter (Edmund Scientific, United Kingdom) (Figure 1). LEDs are pulsed sequentially at 476 Hz with 20 μs per flash.

There are three, synchronised, photodiode detectors for fluorescence recording: yellow (YF), red (RF) and far-red (FRF), which are defined by the 590NB10, 678WB22 and 750WB65 interference filters (Intor, Socorro, NM USA), respectively. In addition, the RF channel has a 3-mm RG665 red glass filter (Schott, Mainz, Germany) and the FRF channel has a 3-mm RG9 far-red glass filter (Schott, Mainz, Germany). The sensor illuminates an 8-cm-diameter surface (50 cm²) at a 10-cm distance from the source and detector, where the cluster or berry plates were positioned for measurements, which lasted less than a second per cluster. Each measurement consisted of a train of 500 flashes of four colours (UV, B, G and R). The sensor calculates a set of chosen ratios after each series of four-color flashes. The mean and standard deviation of the 500 measurements for the 12 signals (Table 1) and 10 ratios are recoded on a SD card and displayed on the sensor’s screen [Figure 1(b)]. The Multiplex^® was used in the field under daylight and was also used indoors on berries. Two versions of the Multiplex were used that differed only in their design and ergonomics. Multiplex^® 2 [Figure 1(b)] was used for calibration, laboratory measurements and in the Fort Chabrol vineyard, and Multiplex^® 3 [Figure 1(c)] was used in commercial vineyard blocks.

Because the fluorescence screening method used in the Multiplex^® was described in detail in [16], here we will only identify and describe the nomenclature of the fluorescence indices provided in the commercial Multiplex^® sensors used in the present study. The decadic logarithm of the ratio of far-red fluorescence (FRF) excited at two different wavelengths [red FRF_R and green FRF_G (Table 1)] is called ANTH_RG because it is proportional to skin anthocyanin content: (1) ANTH _ RG = log ( FRF _ R / FRF _ G )

We also calculated two equivalent indices based on the combination of blue and red excitation and blue and green excitation (not recorded on the sensor) that were tested in this study: (2) ANTH _ RB = log ( FRF _ R / FRF _ B ) (3) ANTH _ BG = log ( FRF _ B / FRF _ G )

Ben Ghozlen et al. [19] recently proposed the use of a log-transformed version of an inverted, single-signal FRF_R: (4) FERARI = log ( 1 / FRF _ R ) = − log ( FRF _ R )

They found that the log-transformed version had a good positive correlation with skin anthocyanin content. This index is provided in the commercial Multiplex^® under the abbreviation FERARI (Fluorescence Excitation Ratio Anthocyanin Relative Index, used hereafter).

The emission ratio (SFR_R) is linked to the chlorophyll content of leaves [30,31] and grape berries [19]. It is a simple chlorophyll fluorescence ratio (SFR) of far-red emission (FRF, 735 nm) divided by red emission (FR, 685 nm) under red excitation: (5) SFR _ R = FRF _ R / RF _ R

Due to the overlap of the chlorophyll absorption and emission spectrum, re-absorption occurs at shorter wavelengths (RF) but not at longer wavelengths (FRF) [30,31]. Therefore, SFR increases with increasing sample chlorophyll content. It should be noted that the inverse ratio, RF/FRF, is also often used in the literature (cf. [32]). The latter will decrease with increases in chlorophyll content.

Data were treated, transformed, statistically analysed, fitted and plotted using a combination of software: Excel 2003 (Microsoft, USA), Statistica 6 (StatSoft, Tulsa, OK USA) and Igor Pro 6.02 (WaveMetrics, USA). Model solving and computations were performed with Mathematica 4 (Wolfram Research, Champaign, IL, USA).

In addition to the intrinsic linearity of the response of Multiplex^® detectors guaranteed by the producer, we have tested the linearity of the ANTH indices themselves using standards of known absorbance. Figure 3 shows that ANTH_RG deviates substantially from the y = x line above the absorbance of 0.9 (more than 10%). ANTH_RB, corresponding to the ratio of red to blue excitation, is more linear and has less than 10% deviation up to the absorbance of 2. However, none of the ANTH values acquired in this study were larger than 0.9; we can thus consider them all having a linear response. The FERARI index, which is just a log transformation of FRF_R, was linearly related to the absorbance of the standards over two orders of magnitude (therefore, not presented in Figure 3). The addition of aqueous solution of chlorophyll-protein complexes revealed the detection limit of the sensor’s FRF_R signal to be 0.7 μg chlorophyll L⁻¹ (3.5 ng chlorophyll cm⁻²) and the sensitivity to be 1.86 mV μg⁻¹ chlorophyll (1 mV per 2.7 ng chlorophyll cm⁻²).

Table 2 shows the various sources of variability and the precision of the Multiplex^® measurements. Repeatability was calculated from 30 consecutive measurements on the blue standard at 25 °C. For repeatability, we present the signal both in millivolts and as the standardised signal because the latter was obtained using the same blue standard. In addition, presenting percentage standard deviation (%SD) for the standardised ANTH_RG index is mathematically meaningless because it is a logarithm of a signal ratio; therefore, its mean is equal to zero in the absence of anthocyanins. The precision of the Multiplex^® was good (better than 1%) and the reproducibility, assessed using the same type of measurement but during the entire season (n = 11) at a temperature range from 27 to 28 °C, was satisfactory (better than 2%). Temperature affected the signals (1% per °C) more than the ratios (0.25% per °C, Table 2).

In the most recent version of the Multiplex^® (the Multiplex^® 3 that we used in the second part of this study in the field), all of the signals were corrected for temperature in the sensor itself. The major source of variability was the distance of measurement. A 30 to 40% variation may be induced depending on the berry type for a 30% deviation from the standard distance of measurement (Table 2). As expected, a single signal (FRF_R) and its transformation as a FERARI index (cf. below) were influenced much more by the distance of measurement than an index based on either fluorescence emission (SFR_R) or excitation (ANTH_RG) ratio (Table 2), decreasing to only 2 to 4% depending on the berry type. Two conclusions can be drawn from the data in Table 2. First, a single measurement per cluster is sufficient. Repeated successive measurements would only increase the variability due to the induction of the variable chlorophyll fluorescence [26] because the latter depends on environmental conditions. Second, the FRF_R value for green berries was 40% larger than for the blue standard (1.373 vs. 1) and ANTH_RG was larger than zero (0.072). These results indicate that green berries are more fluorescent than the blue standard and they are less excited in the green than the standard (compared to red light excitation). Because this absorption might vary with chlorophyll content and berry structure, we did not correct for it (cf. negative value in Figures below).

We first present the kinetics of Multiplex indices produced by the sensor recorded on 40 marked clusters during the whole 2008 maturation season (Figure 5).

The kinetics were similar to the one obtained in 2007 with the Multiplex^® prototype [19], but here the sampling was more frequent (twice a week) and lasted 45 days longer, until mid October (DOY 284), when the harvest had been finished in all of Champagne (CIVC, personal communication). This timeframe allowed us to see that the FERARI index continued to increase until the last day. However, there was a change in slope from DOY 260 [Figure 5(a)] corresponding to the date from which ANTH_RG remained constant [Figure 5(b)]. DOY 260 corresponded to the date when all berries of all clusters were coloured red in Pinot Noir and Pinot Meunier (end of véraison, stage BBCH 85 [35]). Therefore, there were two types of potential influences on anthocyanin-related optical signals: the influence of the proportion of red to green berries and the effect of anthocyanin screening in red berries. We will analyse these two influences independently in the next sections.

The chlorophyll-related SFR index, whose calibration is beyond the scope of this paper, decreased steadily in all cultivars during grape maturation. Although technical maturity, estimated on 2-kg samples of grape from the same block, showed large fluctuations (Figure 6) probably caused by a sampling problem, a good correlation can be seen between SFR and both sugar and total acidity (r² = 0.85 and r² = 0.85, respectively) (Figure 7).

This characteristic could be especially useful for the non-destructive, optical monitoring of white grape cultivars devoid of anthocyanins. In contrast, ANTH_RG increased steadily during the whole season [Figure 5(b)], primarily due to the loss of chlorophyll but also due to changes in berry optical properties (becoming more translucent over time) [16]. Thus, ANTH_RG could also be a useful indicator for white grape cultivars. It is interesting to note that the fluorescence emission index SFR, thanks to its independence of excitation screening, was strikingly similar for both red and white grape cultivars. The behaviour of anthocyanin-related indices in Chardonnay [Figure 5(a) vs. Figure 5(b)], which is devoid of anthocyanins, showed that FERARI is noisier than ANTH_RG, displaying larger variations and larger confidence intervals, as was expected from the above-described tests on signal sensitivity.

In previous publications on the application of the chlorophyll fluorescence screening method for grapes [16,18,29], the simple influence of green berries on the overall optical signal was not analysed. This type of influence on the chlorophyll fluorescence screening method is illustrated in Figure 8. Two populations of berries were mixed, including both red berries chosen individually with the Multiplex^® to have the same anthocyanin content and green berries devoid of anthocyanins.

Thus, if we name p the proportion of red berries, FRF signals should be proportional to the sum of the fluorescence of green berries having “naked” chlorophyll, 1 − p, and red berries, p, in which chlorophyll is screened (Figure S2). For green excitation, this behaviour can be described by the equation: (6) FRF _ G = γ G [ ( 1 − p ) + ( p   T G ) ]where T_G is the apparent transmittance of the skin for green light before reaching the chlorophyll of red berries and γ_G is a constant. The transmittance for green berries (1 − p) is assumed to be 1. Indeed, all three signals (FRF_B, FRF_G and FRF_R) declined linearly with the increasing presence of red berries (Figure 8, insert). This trend occurred during véraison at the level of each cluster or for samples of 200 berries from the vineyard. For all three wavelengths, this simple model showed a very large coefficient of determination, r² = 0.99, between the observed and predicted values, so p could be estimated with an error between 1 and 4%. The root mean square error (RMSE) for p estimation was 0.029, 0.037 and 0.028 for FRF_R, FRF_G and FRF_B, respectively.

ANTH indices were less affected than FERARI, as expected, but they changed significantly above p = 0.5 (equivalent to the half véraison stage) [5] (Figure 8). Combining Equations (1) and (6) and using the corresponding suffixes, R, G and B for red, green and blue excitation light, respectively, the equation for the effect of p on ANTH indices will be: (7) ANTH _ RG = log ( γ RG [ ( 1 − p ) + p   T R ] / [ ( 1 − p ) + p   T G ] ) (8) ANTH _ RB = log ( γ RB [ ( 1 − p ) + p   T R ] / [ ( 1 − p ) + p   T B ] )and for FERARI from Equations (4) and (6): (9) FERARI = − log ( γ R [ ( 1 − p ) + ( p   T R ) ] )

The γ_RG, γ_RB and γ_R are constants, and T_R and T_B are the apparent transmittances of berry skins before chlorophyll is reached. The fits of these equations to experimental data were very good, with a standard error smaller for ANTH_RG (0.0084) than for ANTH_RB (0.019) and FERARI (0.047) (Figure 8).

The effect of anthocyanin screening in red berries now needed to be quantified and expressed in usual skin anthocyanins content units. Towards this goal, we analysed a series of berry samples chosen for their increasing redness with the Multiplex^®, followed by the total extraction of the skins as done in [16–19]. The decrease in FRF signals that depends on the increase in anthocyanins in the skin, without the effect of the presence of green berries, is presented in Figure 9.

The chlorophyll fluorescence screening method postulates [27] that the FRF_G signal should be attenuated by the presence of anthocyanins in accordance with the Beer-Lambert law: (10) FRF _ G = γ G 10 − ɛ G   anthocyaninswhere ɛ_G is the apparent absorptivity of anthocyanins in the green, and γ_G is a constant. Therefore, the inversed and log-transformed FRF signal should be linearly related to anthocyanins content. From the insert in Figure 9, where normalised (division by the constant γ_G), transformed signals are presented, it can be seen that this is not the case. Although the order of the absorptivity coefficients for anthocyanins was preserved (for malvidine-3-glucoside 24,590, 12,290 and 70 M⁻¹ cm⁻¹, for 516, 470 and 635 nm, respectively), the relationship was not linear and the absorptivity decreased with anthocyanin content. The transformed red-excited FRF signal, −log(FRF_R), which is the FERARI index, increased much more rapidly than expected from its in vitro absorptivity coefficient (anthocyanins do not absorb red light), but this signal was the closest to a linear dependence. In contrast, the low and constantly changing in vivo absorptivity of anthocyanins, obtained from transformed green and blue excited FRF (Figure 9, insert), can probably be explained by the overlap between anthocyanins and chlorophyll in the grape skin [16]. Indeed, unlike the situation in leaves, where a good separation exists between the proximal layer of epidermal cells and a distal layer of chloroplasts along the pericline cell wall of the first layer of mesophyll cells [36], in grape berry skins the first hypodermal cell layer containing anthocyanins already contains some chlorophyll. Therefore, the chlorophyll fluorescence screening method, which works very well for flavonols in leaf epidermises and probably also for flavonols in the berry epidermis [37], must depend on the small difference of 10 to 20 μm between the maximum in anthocyanins and chlorophyll distribution, the former preceding the latter in the berry hypodermis [16]. As discussed in [16], the decrease in the absorptivity of anthocyanins in vivo can also be due to the increase in skin pH during maturation. All these observations explain why the best fit of the data was obtained with a negative exponential function (11) − log ( FRF _ G ) = A [ 1 − exp ( − a G   anthocyanins ) ]where A and a_G are constants of the fit. The constant A is the maximal value at infinite anthocyanin content, so we kept it the same for all three signals. This function can also be used to define the FRF’s decay with increasing anthocyanins. From Equations (10) and (11), we obtain the following power and exponential decay function for FRF_G: (12) FRF _ G = γ G 10 A [ exp ( − a G   anthocyanins ) − 1 ]which not only fits the signal best (Figure 9) but can now be used to derive the response function for ANTH_RG from the two signals decay functions and Equation (1): (13) ANTH _ RG = A [ exp ( − a R   anthocyanins ) − exp ( − a G   anthocyanins ) ] + log ( γ RG )

Thus, in Figure 10(b), the ANTH indices measured during calibration were fitted with the respective forms of Equation (13). Using Equation (13) and the fit parameters of the calibration, A, a_R, a_B and a_G (Figure 9, insert and Figure 10(b)), we could simulate the response curves for ANTH_RG, ANTH_RB and ANTH_BG for the full range of anthocyanin content expected for any grape cultivar (0 to 1 mg cm⁻², equivalent to 0 to 4 mg g⁻¹, Figure 11). One can see that the ANTH indices have two ranges of response to anthocyanin content, which increase in one range and decrease in the other, separated by a maximum (Figure 11). Response curves in Figure 11 can be used to derive graphically (numerically) anthocyanin content from any index value. The first range is delimited by the maxima at 0.2, 0.27 and 0.16 mg cm⁻² of skin anthocyanin content for ANTH_RG, ANTH_RB and ANTH_BG, respectively. In addition, for the first range of the response curve, we derived polynomial functions by fitting the inversed response curves anthocyanins vs. Multiplex^® indices. Therefore, the following fourth-order polynomials can be proposed for ANTH_RG until a first limit of 0.16 mg cm⁻² (equivalent to 0.6 mg g⁻¹) (to avoid the flat range around the maximum) (14) anthocyanins = 0.0567   ANTH _ RG + 0.688   ANTH _ RG 2 − 2.05   ANTH _ RG 3 + 2.28   ANTH _ RG 4and for ANTH_RB until a first limit of 0.2 mg cm⁻² (equivalent to 0.8 mg g⁻¹) (15) anthocyanins = 0.181   ANTH _ RB + 2.33   ANTH _ RB 2 − 12.3   ANTH _ RB 3 + 26.1   ANTH _ RB 4to estimate skin anthocyanin content from Multiplex^® indices. The RMSE for anthocyanin estimation was very similar, 16 μg cm⁻² (20%) and 16 μg cm⁻² (16%), with an r² of 0.90 and 0.94 for ANTH_RG and ANTH_RB, respectively. For Pinot Noir and Pinot Meunier, which have quite different skin characteristics [38], individual fitting of the response curves [Equation (13)] was not significantly different for ANTH_RG and ANTH_RB (data not shown) [cf. Figure 10(b)].

In contrast, the FERARI index could be used in the entire range up to 0.45 mg cm⁻² (1.8 mg g⁻¹) [cf. Figures 10(a) and 11] but with a maximum error up to 13% and an r² of 0.96 for both PM and PN (RMSE = 29 and 21 μg cm⁻², respectively). The differences in the slope of the FERARI calibration curves for the two varieties are trivial and can be explained by the larger average berry size for Pinot Meunier (2.25 g) than for Pinot Noir (2.05 g) later in the season. Pinot Meunier berries will give a larger fluorescence signal because they will be closer to the Multiplex^®’s detectors. In the future, we recommend the use of a measurement geometry that will keep the proximal side of the berry (and of the cluster) at a constant distance from the detector. FERARI data can then be transformed into skin anthocyanin content based on the surface (mg cm⁻²) by using the inverse of the linear fit of Figure 10 (16) anthocyanins = FERARI / 2.9when considering both cultivars together with an r² of 0.93 and an RMSE of 42 μg cm⁻² (17%). The latter figure was much larger than for individual cultivars or for ANTH indices, but the index was validated for a much wider range (0 to 0.45 mg cm⁻²–1.8 mg g⁻¹) (Table 3).

For ANTH_RG, different limits have been described previously using three different fluorometric devices: (1) using a spectrofluorometer with a limit at 300 nmol cm⁻² skin equivalent to 0.15 mg cm⁻² skin and 0.55 mg g⁻¹ berry [16], (2) using a fluorescence imager (0.25 mg cm⁻²–1 mg g⁻¹) [17], or (3) using the “leaf clip” Dualex-ANTH on peeled skins (0.3 mg cm⁻²–1.2 mg g⁻¹) [18]. The limit attained here of 0.45 mg cm⁻² (1.8 mg g⁻¹) with the FERARI index based on a weakly absorbed wavelength is sufficiently high for the viticultural practice in Champagne (cf. below) and probably other regions. The limit of 0.2 mg cm⁻² (0.8 mg g⁻¹) attained using the first range of ANTH_RB is well adapted for a large number of table grapes cultivars but not for all winegrapes. Several red-winegrape cultivars, such as Nebbiolo (<0.8 mg g⁻¹), have lower anthocyanin contents than Pinot Noir, even at full maturity [4]. For these cultivars, both the ANTH and FERARI indices are fully applicable using the first range of the response curve (the rising part in Figure 10). Many other cultivars (Cabernet Sauvignon, Merlot, Syrah) have even higher anthocyanin content values than Pinot Noir, so they will be in the second range of the response curve where ANTH decreases with increasing anthocyanins (Figure 10). However, ANTH indices could also be used in this second range to quantify anthocyanins using a polynomial function, but they will have to be calibrated with the larger skin anthocyanin content of appropriate cultivars in the future. It should be mentioned here that many other red fruits (apples, pears) have one order or magnitude less anthocyanins in their skins [20,39], so the first range of the ANTH response curve is well adapted to them.

We can now address the question of a combined P and A model to deconvolute both the proportion of red berries and anthocyanin content from the kinetics of Multiplex^® indices recoded in Figure 5. A combination of the basic equations of the two models, Equations (6) and (12): (17) FRF _ G = γ G   { ( 1 − p ) + p   10 A [ exp ( − a G   anthocyanins ) − 1 ] }could not produce a viable solution. Therefore, we decided to fit the FRF decays (Figure 9) by a function that can be inverted and can still satisfy the quality of fits obtained by Equation 12 (tested by comparing the Chi², not shown). That new function was a hyperbolic function of the type f(x) = a/(b + x^1.1). If we include this empirical hyperbolic function for each FRF signal, using the usual suffixes R, G and B to corresponding to Equation 6 of the P-model, we obtain the following system of Equations: (18) FRF _ R = γ R [ ( 1 − p ) + p   α R / ( β R + anthocyanin 1 . 1 ) ] (19) FRF _ G = γ G [ ( 1 − p ) + p   α G / ( β G + anthocyanin 1 . 1 ) ] (20) FRF _ B = γ B [ ( 1 − p ) + p   α B / ( β B + anthocyanin 1 . 1 ) ]with the six parameters, α and β coming from the calibration of the A-model (Figure 9). The constants γ were the values of the signals of green berries. For each date and cultivar (40 PM and 40 PN clusters), pairs of p and anthocyanins were calculated using any of the possible three pairs of excitation wavelengths [Equations (18) to (20)]. The solution for the red and green pair is shown in Figure 12(b). The solutions for the other two pairs were similar (cf. supplementary data, Figure S3). The proportion of red berries was fitted by a sigmoid function [Figure 12(b)]. The half-véraison, characterised by p = 0.5 (50% red berries), occurred at DOY 233 for Pinot Meunier and at DOY 241 for Pinot Noir. The confidence interval for this estimation was ±1 day. We would have made an error of seven days (a whole week) if we had based the estimation only on the direct ANTH_RG index. The model estimates the half-véraison date more accurately; therefore, the harvest date can be forecast more reliably [5]. The period from the first red berry in a cluster (véraison start; stage BBCH 81) to fully red clusters (hereafter DOY 260; véraison end; stage BBCH 85) can last more than 20 days (Figure 5). Therefore, it is very difficult to define half-véraison with precision by simple visual observation. This growth stage is important and often used to adjust viticultural practice (pest treatments, cluster thinning, addition of growth regulators). In addition, the deconvolution of p and anthocyanins (Figure 12) shows that the anthocyanin increase was rather linear during the season (r² = 0.99 and r² = 0.98 from DOY 232 and DOY 238 on for PM and PN, respectively) (Figure 12), unlike the increase indicated by the ANTH and even the FERARI index (Figure 5).

The model was validated on 200-berry samples from commercial blocks for which photographs and visual estimation of p were available (Figure S2). The combined P and A model was applied to twenty independent samples. A RMSE of 0.036 (7%) was obtained for the p estimation. The following characteristics of the linear regression (not shown) between the observed and estimated p were found: slope = 0.947 ± 0.0378, intercept = 0.049 ± 0.018, r² = 0.988 and p < 0.0001 (extremely significant). Of the six parameters of the model (for two wavelengths), the two γ constants seem responsible for most of the uncertainty (data not shown). These constants depend on the instrument function (intensity of excitation light, sensitivity of detectors) that is almost eliminated by the use of the blue standard but also on chlorophyll content [16], on chlorophyll fluorescence yield (the sample should be under similar light and temperature conditions) and on the distance of measurement (Table 2), which should change as little as possible between the calibration and sample measurements. To that aim, we propose that grids or windows be used in the future to fix the measuring distance.

The decrease in skin chlorophyll content during maturation, attested by the decrease of SFR (Figure 5), was not taken into account. Our attempt to include this component in the model failed (not shown), possibly because it might be smaller than, or compensated by, the chlorophyll-overlap effect. Indeed, as explained above, the accumulation of anthocyanins produces a double exponential attenuation: the first from anthocyanins absorption and the second from the decreased chlorophyll excitation due to an anthocyanins-chlorophyll overlap [16]. Our goal was to produce a practical model to obtain p by an analytical, or at least a numerical, inversion. It was successfully tested for p and so it is useable in this respect. In addition, the model calculates anthocyanin content, but it was not tested for anthocyanin content explicitly because Fort Chabrol data were followed non-destructively and the 200-berry samples extracted by the standard wet chemistry method yielded incomplete extraction (cf. below).

Armed with calibrated indices and a model for the deconvolution of the contribution of green berries, we could address an example of practical large-scale application of the Multiplex^®.