Leaf optical properties reflect variation in photosynthetic metabolism and its sensitivity to temperature.

Researchers from a number of disciplines have long sought the ability to estimate the functional attributes of plant canopies, such as photosynthetic capacity, using remotely sensed data. To date, however, this goal has not been fully realized. In this study, fresh-leaf reflectance spectroscopy (λ=450-2500 nm) and a partial least-squares regression (PLSR) analysis were used to estimate key determinants of photosynthetic capacity-namely the maximum rates of RuBP carboxylation (V(cmax)) and regeneration (J(max))-measured with standard gas exchange techniques on leaves of trembling aspen and eastern cottonwood trees. The trees were grown across an array of glasshouse temperature regimes. The PLSR models yielded accurate and precise estimates of V(cmax) and J(max) within and across species and glasshouse temperatures. These predictions were developed using unique contributions from different spectral regions. Most of the wavelengths selected were correlated with known absorption features related to leaf water content, nitrogen concentration, internal structure, and/or photosynthetic enzymes. In a field application of our PLSR models, spectral reflectance data effectively captured the short-term temperature sensitivities of V(cmax) and J(max) in aspen foliage. These findings highlight a promising strategy for developing remote sensing methods to characterize dynamic, environmentally sensitive aspects of canopy photosynthetic metabolism at broad scales.

Introduction

Using an array of technologies, researchers from a number of disciplines continue to pursue methods for remotely estimating the biochemical, structural, and physiological traits of plant leaves and canopies based on their optical properties (Wessman ; Martin and Aber, 1997; Smith ; Biewer ). Thus far, target foliar traits have included concentrations of nitrogen (Nmass, Bolster ; Gillon ), lignin, cellulose (Wessman ; Martin and Aber, 1997; Kokaly and Clark, 1999; Petisco ), and photosynthetic pigments (Richardson ; Gitelson ; Moorthy ) as well as water content (Sims and Gamon, 2003; Stimson ; Cheng ). In addition, leaf isotopic ratios (δ13C and δ 15N; Richardson and Reeves, 2005; Wang ; Kleinebecker ), specific leaf area (SLA; Asner and Martin, 2008), and leaf mass per area (LMA; Asner ; Doughty ) have been successfully estimated using leaf optical properties. To date, however, remote sensing of leaf functional attributes, such as photosynthetic metabolism, has not progressed as rapidly (Grace ).

Much of the effort to relate optical remote sensing data to photosynthetic status and ecosystem function has focused on the use of the photochemical reflectance index (PRI, Gamon , 1997). The PRI, which provides a linkage with photosystem II (PSII) efficiency by tracking the variation in xanthophyll cycle pigments, has been successfully used to assess photosynthetic functioning across a range of vegetation types (Peñuelas , 1998; Gamon ; Nichol ; Stylinski ; Guo and Trotter, 2004; Fuentes ; Drolet ; Hilker ; Middleton ) and responses of plants to environmental stress (Dobrowski ; Suarez ; Gray ).

A parallel avenue of research has focused on the detection of vegetation chlorophyll fluorescence (CF) to exploit its relationship with photosynthetic functioning (Freedman ; Louis ; van der Tol ; Damm ). CF has been detected through passive monitoring of solar-induced steady-state fluorescence (Zarco-Tejada ; Dobrowski ; Grace ; Campbell ) and active laser-induced methods (Ananyev ). Although remote sensing of CF remains experimental owing to a variety of technical issues (Grace ; Coops ), the recent creation of regional and global CF maps using space-borne observations (Guanter ; Joiner ) highlights the technique’s potential.

The results of an effort to further explore the links between the photosynthetic and optical properties of tree leaves are summarized here. Specifically, in a glasshouse study of trembling aspen (Populus tremuloides) and eastern cottonwood (Populus deltoides) trees grown under different temperature regimes, it was assessed whether climate-mediated variation in leaf photosynthetic metabolism could be effectively estimated using data from visible and near-infrared reflectance spectroscopy (VIS/NIRS). With respect to photosynthetic metabolism, the focus was on two important parameters, the maximum rate at which ribulose bisphosphate (RuBP) is carboxylated (Vcmax) and regenerated (Jmax). With robust estimates of these two traits, the photosynthetic performance of a given leaf can be predicted using a widely adopted biochemical model (Farquhar ; Farquhar and von Caemmerer, 1982). This model has been effectively scaled to the canopy level (Medlyn ; Thum ), and its use in canopy and ecosystem process models is increasingly common.

Vcmax and Jmax vary substantially across plant species, functional groups, and growth environments (Wullschleger, 1993; Kattge ). Moreover, both parameters are very sensitive to short-term (i.e. seconds to hours) dynamics in leaf temperature (Medlyn ; Kattge and Knorr, 2007). The accuracy and credibility of outputs from process models would increase considerably if a feasible methodology were developed for remotely sensing canopy photosynthetic capacity and its temperature sensitivity across broad scales.

The principal objectives of our study were (i) to develop spectroscopic models for estimating Vcmax and Jmax based on leaf data collected across a wide range of growth temperature regimes; and (ii) to assess the effectiveness of those models in estimating the short-term temperature sensitivity of Vcmax and Jmax in the field. The second objective was addressed with measurements of leaf traits in field-grown aspen trees.

Materials and methods

Glasshouse treatments and experimental design

The relationships were tested between leaf photosynthetic and optical properties in climate-controlled glasshouses at the University of Wisconsin–Madison Biotron. Aspen and cottonwood germinants were reared in flats until they reached a height of about 20 cm. Seedlings were then transplanted into 4-litre pots and transferred to the Biotron glasshouses, where they were grown for 8 weeks under one of three different temperature regimes, with fixed day/night air temperatures of 30/23, 25/18, and 20/13 °C. These thermal regimes were chosen to span ranges in air temperature observed during the growing season along a latitudinal transect for a study that examined the physiological factors limiting geographic distributions of temperate and boreal tree species (Dillaway and Kruger, 2010). Each regime was replicated in two glasshouses.

Measurements of leaf gas exchange

During the fourth and eighth weeks of growth in the Biotron glasshouses, leaf gas exchange was measured on a total of 8–10 trees per species in each treatment (4–5 per species and glasshouse, n=78) using a LI-6400 portable photosynthesis system (Li-Cor Biosciences, Lincoln, NE, USA). All measurements were conducted on the youngest, fully expanded leaf of each tree. Leaves were measured under high light intensities (photosynthetic photon flux=1800 μmol m−2 s−1, provided by a red-blue LED array) at several CO2 partial pressures (pCO2) ranging from 7.5 to 120 Pa. Photosynthesis (A) was assessed first at a cuvette reference pCO2 of 40 Pa, and then again after each of three step-wise decreases in pCO2 (i.e. at 25, 15, and 7.5 Pa). Photosynthesis was then measured at 60, 90, and 120 Pa CO2, respectively. Cuvette reference pCO2 was controlled using the LI-6400 CO2 injector system. The potentially confounding influences of diffusion leaks across the cuvette gasket on gas exchange calculations were taken into account by applying the manufacturer’s equation to determine the gasket diffusion coefficient (Anonymous, 2005).

At a given pCO2, leaves were allowed to acclimate to cuvette conditions for 2–5 min, depending on when photosynthetic rate stabilized. Vapour pressure deficit between leaf and air in the cuvette ranged from 0.95–2.01 kPa. For each leaf, the photosynthetic response to pCO2 was assessed at the daytime air temperature of each treatment (20, 25, and 30 °C), which, owing to concerns about IRGA signal stability, was maintained through the manipulation of cuvette rather than leaf temperature. Thus, because leaf temperature was not directly controlled, it ranged from 20.40–21.80, 25.06–26.14, and 30.05–30.40 °C at the target temperatures of 20, 25, and 30 °C, respectively. Measurements were conducted throughout the day, as long as stomatal conductance remained comparatively high (within 25% of the daily maximum).

Measurement of leaf optical properties

Reflectance was measured for both species on the same leaves assessed for gas exchange using a high-spectral-resolution ASD FieldSpec® 3 Full-Range (350–2500 nm) spectroradiometer (Analytical Spectral Devices, Boulder, CO, USA). All measurements occurred on the leaf adaxial surface using a leaf-clip assembly attached to a plant probe with an internal, calibrated light source. The relative reflectance of each leaf was determined from the measurement of leaf radiance divided by the radiance of a 99.9% reflective white standard (Spectralon, Labsphere Inc., North Dutton, NH, USA). Reflectance was measured on ten different areas of each leaf lamina, and the resulting ten spectra were averaged to determine mean optical properties for each leaf. Each measurement required less than 5 s. Measures of leaf optical properties and gas exchange occurred within 24 h of one another.

Leaf nitrogen and leaf mass per area

Immediately following gas exchange and optical assessments, leaves were harvested and measured for projected area using a LI-3100 leaf area meter (Li-Cor Biosciences, Lincoln, NE, USA), oven-dried to a constant mass at 70 °C, and weighed. These data were combined to calculate leaf mass per area (Marea, g m−2). Dried leaves were then finely ground and analysed for nitrogen concentration using an Elementar Vario Macro CHN analyser (Elementar Analysensyteme GmbH, Hanau, Germany).

Estimation of leaf photosynthetic traits

Relationships between photosynthesis (A) and intercellular pCO2 (Ci) were used to estimate maximum rates of RuBP carboxylation (Vcmax) and regeneration (Jmax) at a given leaf temperature with a curve-fitting method that minimized the sums of squares for error resulting from comparisons of observed versus estimated A (Long and Bernacchi, 2003). Michaelis–Menten constants for CO2 (Kc) and oxygen (Ko), and photosynthetic (CO2) compensation point (Γ*) were calculated with formulae from Long and Bernacchi (2003). Vcmax was estimated from the lower portion of the A–Ci curve (where Ci <30 Pa), and Jmax was estimated from the upper portion of the curve (Ci >60 Pa). It was noted that this approach does not account for the influence of mesophyll conductance on estimates of Vcmax and Jmax (Dillaway and Kruger, 2010), and thus our reported values are based on intercellular as opposed to chloroplastic pCO2.

It was assumed that leaf temperature closely tracked air temperature in the glasshouses, and thus, because the two differed somewhat during gas exchange measurements, temperature-response models were used to estimate Vcmax and Jmax for each leaf at its respective glasshouse daytime temperature (i.e. 20, 25, or 30 °C). Temperature-response models were generated for each species based on Vcmax and Jmax data pooled across the three temperature treatments (data not shown). Owing to its exponential form, the temperature sensitivity of Vcmax was modelled using an Arrhenius equation (Hikosaka ), whereas the asymptotic temperature response of Jmax was characterized with a ‘peak’ model (Kattge and Knorr, 2007). The resulting models produced unbiased estimates of Vcmax and Jmax in the cases of both species and all measurement treatments. Specifically, the slopes and intercepts of relationships between observed and predicted Vcmax and Jmax did not differ significantly from 1 and 0, respectively (data not shown). Implicit in this approach was the assumption that photosynthetic metabolism of glasshouse tree foliage did not, to any appreciable extent, acclimate to the different thermal regimes, which would be consistent with our observations in the field (Dillaway and Kruger, 2010).

Generation of predictive models from leaf reflectance spectra

Predictive models of metabolic, biochemical, and morphological traits based on leaf optical properties were examined using partial least-squares regression (PLSR) analysis (Wold ; Geladi and Kowalski, 1986; Wolter ). While PLSR has not been widely embraced in ecology (Carrascal ), its use in remote sensing research has increased in recent years (Smith ; Townsend ; Ollinger and Smith, 2005; Martin ; Wolter ). This is because PLSR is useful in situations of high predictor collinearity and/or the predictor variables are equal to or higher than the number of observations, which is often the case in spectroscopic and/or remote sensing research. In addition, models developed using PLSR are much more robust than classical regression in that the calibrated model parameters do not vary greatly given different calibration subsets from a population of observations (i.e. high parameter stability; Geladi and Kowalski, 1986).

A standard PLSR approach for spectral-chemical analysis utilizes the continuous, full-spectrum data (Asner and Martin, 2008; Doughty ) or a pre-determined spectral subset (Bolster ; Richardson and Reeves, 2005). The spectral loadings (or regression coefficients), which directly relate the target leaf attributes to corresponding spectral features, are generated through the development of a smaller set of orthogonal linear latent components which are obtained through the decomposition of the model variables and the optimization of the covariance structure in the data (Wold ; Geladi and Kowalski, 1986; Wolter ). In this study, the choice was made to incorporate an automatic variable selection method similar to Wolter for predictor dimensionality reduction and model optimization. This allowed us to build more parsimonious models and investigate regions of the spectrum that were important for the prediction of each target variable.

The dimensionality reduction involved a two-stage selection of predictor variables (i.e. spectral wavelengths) whereby the variables and number of latent components chosen are those which minimized the model PRESS statistic (Wolter ), followed by a second selection technique minimizing the Root Mean Square Error (RMSE), once the model PRESS is minimized. This enabled the continued elimination of redundancy in the spectral data until the model RMSE reached a minimum. In this second step, the iterative-cross validation PRESS statistic (Asner and Martin, 2008; Wolter ) was used to select the optimal number of components for each variable subset. The remaining variables and components with the RMSE and PRESS minimized are chosen for the final model. This was done independently for each variable of interest (i.e. Vcmax, Jmax, Nmass, and Marea). This resulted in PLSR models with only salient variables and components for dependent variable prediction.

The PLSR analysis and variable selection was carried out using the PLSREGRESS function in Matlab (Mathworks, Natick, MA, USA) and a set of custom functions for the variable selection. The entire spectrum from 450–2500 nm was sub-sampled by retaining every fifth wavelength prior to running the analyses to decrease the computation time for variable and component selection. Once the optimal models were chosen, a final PLS-PRESS analysis was carried out to calculate the spectral loadings and model diagnostics.

As additional model verification, a 100× cross-validation of our Jmax and Vcmax models was performed using a random 70/30% split of the data for model calibration and testing. This was done by using the final set of wavelengths and components selected for each variable during the two-stage PLSR modelling to generate new PLSR estimations of the observations left out of each iteration (i.e. 30% validation), based on the remaining data left for calibration (i.e. 70% calibration). The results from this analysis were used to examine model and data stability. From this, the distribution of error resulting from the multiple permutations of the data is reported.

Estimating Vcmax and Jmax with leaf optical data from field-grown aspen

To determine whether the results from our glasshouse PLSR analysis captured short-term (seconds to hours) temperature-dependent variation in leaf metabolic traits rather than–or in addition to–longer-term (days to weeks) photosynthetic acclimation of foliage to variation in growth conditions, leaf properties were examined on five field-grown trembling aspen trees. Measurements were made during an 8 h period on two different days across a range of ambient air temperatures. On 15 July 2009, under clear skies, gas exchange and temperature were measured (on the adaxial and abaxial surface using an Agri-Therm III infrared thermometer, Everest Interscience, Tucson, AZ, USA) on one sunlit leaf per tree in the morning and again in the afternoon. This protocol allowed us to obtain a 5–13 °C span in the temperatures at which a particular leaf was measured.

On a different set of aspen trees, gas exchange and temperature were again measured on one sunlit leaf per tree in the morning and afternoon of 16 June 2010. Here, in addition to measuring leaf temperature just prior to spectral measurements, it was monitored inside the spectroradiometer leaf-clip assembly using a fine-wire thermocouple. This allowed us to test the assumption that spectral measurements did not cause large perturbations of leaf temperature. On average, leaf temperature in the leaf-clip assembly was 0.15 °C (±0.55 °C) higher than the corresponding ambient value.

For these experiments, leaf gas exchange and spectral reflectance were measured in the same fashion as in the glasshouse study, with two exceptions: (i) owing to time constraints, gas exchange measures in July 2009 were confined to the initial portion of the A–Ci curve (cuvette pCO2 <40 Pa), affording only the estimation of Vcmax, while those in June 2010 were confined to cuvette pCO2 values >60 Pa, affording only the estimation of Jmax, and (ii) the spectroradiometer was used in its backpack configuration in order to collect leaf spectral data in situ. Using these data, the accuracy and precision with which the glasshouse-based PLSR models predicted short-term, temperature-mediated variation in Vcmax and Jmax were examined on an independent set of data under different conditions.

Results

Variation in target leaf traits among glasshouse temperature treatments

In the glasshouses, leaf nitrogen concentration (Nmass), leaf mass per area (Marea), and maximum rates of RuBP carboxylation (Vcmax) and regeneration (Jmax) varied considerably across temperature regimes and/or species (Fig. 1). Trait responses to growth temperature were fairly consistent across species, except that, as temperature rose, Nmass increased in aspen and decreased in cottonwood. Nmass and Marea were generally higher in cottonwood than in aspen, and for both species Marea decreased in the warmest glasshouses. Vcmax increased exponentially with temperature, while Jmax exhibited a less pronounced, asymptotic response. At a given temperature, averages for Vcmax and Jmax were higher in cottonwood than in aspen.

Fig. 1.

Variation in nutritional, morphological, and metabolic leaf traits for trembling aspen (white boxes) and eastern cottonwood (grey boxes) measured in the Biotron facility, summarized by night-time/daytime glasshouse temperatures. Traits are nitrogen concentration (Nmass, %), leaf mass per area (Marea, g m−2), and maximum rates of RuBP carboxylation (Vcmax, μmol m−2 s−1) and regeneration (Jmax, μmol m−2 s−1). The box plots display the median for each trait by group (dark horizontal line), the interquartile range (IRQ, boxes), the range (whiskers), and the extreme observations (black dots).

Relationships between target leaf traits and optical properties in the glasshouse

Leaf reflectance varied substantially within and across temperature treatments (Fig. 2a), with a 45%, 20%, and 37% range in reflectance in the visible (450–700 nm), near-infrared (NIR, 700–1300 nm) and shortwave infrared (SWIR, 1500–2500 nm) regions, respectively. The relative range in reflectance for individual wavelengths was at a minimum in the NIR (15% at 772 nm) and peaked in the SWIR (59% at 2494 nm). Target leaf traits were variably correlated (r, 0.25–0.75) with leaf reflectance at wavelengths broadly distributed across the full spectrum (i.e. 450—2500 nm; Fig. 2b). In general, positive correlations were observed in the blue (450–495 nm) and red-edge (650–680 nm) regions, whereas moderate to strong negative correlations occurred in the green (505–570 nm) and red (620–650 nm) regions, and across the NIR and SWIR (Fig. 2b, 2c).

Fig. 2.

(a) Mean, ±1 standard deviation, and minimum and maximum leaf reflectance for the pooled aspen and cottonwood seedlings grown in the Biotron facility. Correlation coefficients showing the strength of relationships between spectral wavelengths and each of the four target leaf traits across the full spectrum (b) and for only the visible spectrum (c) Traits are nitrogen concentration (Nmass, %), leaf mass per area (Marea, g m−2), and maximum rates of RuBP carboxylation (Vcmax, μmol m−2 s−1) and regeneration (Jmax, μmol m−2 s−1). (This figure is available in colour at JXB online.)

PLSR models based on glasshouse data

PLSR analysis yielded accurate and precise empirical predictions of target leaf traits based on leaf reflectance spectra (Table 1; Fig. 3). All models possessed a high coefficient of determination (r2), while their root mean square error (RMSE) values averaged 8% of the mean. Through automated variable selection, a set of 13 wavelengths was found that described nearly 90% of the variation in Vcmax across leaves and temperature treatments, while 44 wavelengths provided an effective model of Jmax (Table 1). Both the Vcmax and Jmax models involved broadly similar portions within the visible and short-wave spectral regions, although the loadings were generally larger in the Vcmax model (Fig. 4; see Supplementary data and Tables S1–S4 at JXB online). In contrast to that for Vcmax, the model for Jmax included wavelengths in the NIR related to leaf water content and the green reflectance peak. All models incorporated wavelengths in the visible spectrum (i.e. 450–700 nm), with all but Marea including the chlorophyll absorption regions (i.e. ∼430–460 nm and 640–670 nm). In addition, wavelength regions included in the Jmax and Vcmax models overlapped somewhat with those selected for the Nmass model, while Jmax further displayed some similarities with Marea in the NIR and SWIR (Fig. 4).

Table 1.

Summary of two-stage PLSR modelling results for target leaf traits, including nitrogen concentration (Nmass, %), leaf mass per area (Marea, g m−2), and maximum rates of RuBP carboxylation (Vcmax, μmol m−2 s−1) and regeneration (Jmax, μmol m−2 s−1).

Variable	Number of wavelengthsa	Number of componentsb	R2	RMSE
Nmass	13	11	0.89	0.31
Marea	11	8	0.95	3.69
Vcmax	13	10	0.89	15.4
Jmax	44	13	0.93	18.67

Number of wavelengths selected in the final PLS models.

Number of PLSR components used to generate the wavelength coefficients in the final models.

Fig. 3.

The observed versus predicted values from the final PLS leave-one-out (LOO) cross-validation procedure for glasshouse leaf nitrogen concentration (Nmass, %), leaf mass per area (Marea, g m−2), maximum rates of RuBP carboxylation (Vcmax, μmol m−2 s−1) and regeneration (Jmax, μmol m−2 s−1). Note that the colour scale for Nmass (a) depicts corresponding variation in Marea, while those for Marea (b), Vcmax (c), and Jmax (d) depict variation in Nmass. Each plot has a total of 53 observations and the symbols correspond to the three temperate regimes.

Fig. 4.

Final distribution of the wavelengths selected in each two-stage PLSR model for leaf nitrogen concentration (Nmass, %), leaf mass per area (Marea, g m−2), and maximum rates of RuBP carboxylation (Vcmax, μmol m−2 s−1) and regeneration (Jmax, μmol m−2 s−1). The colour of each vertical bar represents the magnitude of the standardized variable loading (see colour map). Note that each vertical bar represents only one wavelength and the bar width is exaggerated for display purposes only. The final number of wavelengths selected is 13 for Nmass, 11 for Marea, 13 for Vcmax, and 44 for Jmax (see Table 1; see Supplementary data and Tables S1–S4 at JXB online).

As an additional test of the model stability and robustness for the metabolic parameters (i.e. Vcmax and Jmax), a jackknife test of the glasshouse calibrated PLSR models was performed. The 100× cross-validation analysis showed that the models performed consistently across multiple permutations of the data, with both Vcmax (median R2=0.78, median RMSE=16.2) and Jmax (median R2=0.77, median RMSE=20.1) having prediction errors that did not significantly change from the original leave-one-out cross-validation (Fig. 5 shows the R2 and RMSE histograms for the 100× resampling of the glasshouse models).

Fig. 5.

Results of the 100X jackknife resampling of the PLSR glasshouse models (Fig. 3) for Vcmax (a, b) and Jmax (c, d). The Vcmax and Jmax models had a median R2 of 0.78 and 0.77, respectively, while the median model RMSE was 16.2 for Vcmax and 20.1 for Jmax.

Relationships among target leaf traits in the glasshouse

The relationships among target leaf traits, within and across temperature regimes, were examined in an effort to clarify the nature of our PLSR models. In particular, an attempt was made to determine whether predictive models of Vcmax and Jmax might have arisen primarily as a result of the commonly observed dependence of photosynthetic capacity on leaf N status (Field, 1983). To round out this assessment, leaf nitrogen content (Narea, the product of Marea and Nmass) was included in the correlation matrix. Traits were positively correlated with one another in at least one—and often all three—temperature regimes (Fig. 6). In several relationships, however, growth temperature significantly affected the slope and/or intercept. For example, photosynthetic traits were positively correlated with leaf N status (either Nmass or Narea) within a given temperature regime (r ≥0.45, P ≤0.027), but, especially in the case of Vcmax, the correlations deteriorated when data were pooled across temperatures (Fig. 6). Particularly in relationships among Vcmax, Jmax, and Narea, this deterioration resulted from a marked separation of the trend at 30 °C from those at 25 °C and 20 °C (Fig. 4), brought about by the differential temperature responses of each leaf trait (or, in the case of Narea, its components Nmass and Marea) illustrated in Fig. 1.

Fig. 6.

Relationships among nutritional, morphological, and metabolic leaf traits, including nitrogen concentration (Nmass, %), leaf mass per area (Marea, g m−2), and maximum rates of RuBP carboxylation (Vcmax, μmol m−2 s−1) and regeneration (Jmax, μmol m−2 s−1). Note that leaf nitrogen content (Narea, g m−2) —the product of Marea and Nmass—is also included in this analysis. Pearson correlation coefficients (r) are presented by temperature treatment and across all data pooled.

Predicting the temperature dependence of Vcmax and Jmax in field-grown aspen

In July 2009, averages for leaf temperature and Vcmax increased by 7.6 °C (±3.3 °C) and 59.5 μmol m−2 s−1 (±30.4 μmol m−2 s−1), respectively, between morning and afternoon on foliage from the five field-grown aspen trees (Fig. 7a). Using the spectral Vcmax model developed with the glasshouse data, it was found that the observed Vcmax variation across leaves and temperatures could be predicted with reasonable accuracy and precision (r2=0.86, RMSE=10.2 μmol m−2 s−1) based solely on the corresponding leaf spectral data (Fig. 7c).

Fig. 7.

Relationships between observed Vcmax or Jmax and leaf temperature (a, b), and observed versus PLSR-predicted Vcmax or Jmax (c, d), for field-grown aspen trees. Predicted values are derived using only spectral reflectance data in conjunction with the glasshouse PLSR models (see Supplementary Tables S3 and S4 at JXB online), while the observed Vcmax and Jmax data are derived from gas exchange (A–Ci) analyses.

In June 2010, on a different set of field-grown aspen, leaf temperature increased by an average of 3.5 °C (±1.3 °C) from morning to afternoon (Fig. 7b), and corresponding Jmax responses to temperature were variable, showing no clear temperature sensitivity overall. Nevertheless, the spectral Jmax model yielded precise estimates (r2=0.93, RMSE=8.2 μmol m−2 s−1) based solely the corresponding leaf spectral data (Fig. 7d).

Discussion

Our success in estimating leaf morphology and nitrogen status with spectroscopic models is consistent with previous studies using PLSR approaches (Bolster ; Asner and Martin, 2008; Doughty ) and other methods (Wessman ; McLellan ; Kokaly and Clark, 1999). Recently, spectroscopic data combined with PLSR modelling has been used to estimate levels of various other leaf biochemical and nutritional constituents (Gillon ; Richardson and Reeves, 2005; Petisco ). For example, Asner and Martin (2008) found that spectroscopy could be used to estimate leaf concentrations of chlorophylls (i.e. Chl a and Chl b), water, carotenoids, and phosphorus. Other studies have provided empirical evidence that concentrations of structural compounds such as lignin and cellulose, as well as tissue 13C and 15N isotopes, can be predicted effectively using the combined spectroscopic PLSR approach (Bolster ; Brinkmann ; Richardson and Reeves, 2005; Petisco ; Kleinebecker ).

At present, there is only one other study in the literature (Doughty ) that relates the photosynthetic parameters Jmax and Vcmax to full-spectrum leaf optical properties (within a 5 °C temperature range). The authors found weak to moderate predictive power for the Vcmax (r2=0.39, RMSE=36 μmol m−2 s−1) and Jmax models (r2=0.52, RMSE=39 μmol m−2 s−1), in contrast to models of the other variables of interest (e.g. Amax, LMA, leaf N). This was attributed to error propagation in the estimation of Vcmax and Jmax (Doughty ). On the other hand, Stylinski found a close relationship between Jmax and the narrow-band photochemical reflectance index (PRI), an optical indicator of the xanthophyll cycle pigments (Gamon ; 1997), in foliage of pubescent oak (Quercus pubescens) trees. They related this correlation to the down-regulation of electron transport capacity associated with an increase in non-photochemical quenching (NPQ, Demmig-Adams and Adams, 1996, 2006) provided by the xanthophyll cycle pigments.

In addition, Wang observed that an in situ broadband simple ratio (SR), based on infrared and photosynthetically active radiation (PAR) reflectance (i.e. 400–700 nm), was a good predictor of the Vcmax in Japanese beech (Fagus crenata) forests. However, the relationship varied significantly among the three study sites along an elevation gradient, limiting the generality of the results (Wang ). As such, a logical continuation of our research effort is an assessment of the utility of our PLSR models across other plant species, regions, and growth environments.

Perhaps the most novel outcome of this study is the apparent ability of spectral reflectance data to capture the short-term temperature sensitivity of Vcmax and Jmax. In particular, our predictive algorithms based on leaf optical properties (Fig. 3) collapsed substantial species- and temperature-mediated variation in Vcmax and Jmax (Fig. 1) into a single trend (Fig. 3c, d). This finding, along with the pronounced temperature-mediated variation observed in certain leaf-trait relationships (Fig. 6), indicate that the derived Vcmax and Jmax PLSR models are not simply scalars of other leaf traits, or of one another. The fact that our glasshouse-based PLSR models for Vcmax and Jmax also performed fairly well when applied to an independent data set of leaves from field-grown trees (Fig. 7c, d) underscores the considerable potential for remote sensing of environmentally sensitive traits governing photosynthetic metabolism in forest canopies.

An examination of the relationship between reflectance and other leaf traits (Fig. 2) highlighted that (i) leaf physiological traits are correlated with leaf optical properties, but the strength of these correlations varies across the VIS/NIR/SWIR regions; and (ii) there are regions of highly collinear wavelengths that allow for reduction of dimensionality in the predictors, potentially without penalty to overall model performance. Although PLSR can handle datasets with high data dimensionality, such as spectral data, appropriate variable selection techniques can enhance the results of PLSR and provide more parsimonious models (Martens and Martens, 2000; Lestander ; Schmidtlein ; Li ; Wolter ; Chun and Keles, 2010; Feilhauer ).

Use of full-spectrum data can often result in PLSR models with spectral loadings that do not contribute significantly to the prediction (i.e. are near 0) and may negatively affect results (Martens and Martens, 2000; Chun and Keles, 2010). This is important for issues in scaling from the field to broad-scale remote sensing applications where the number of available wavelengths is limited and knowledge of the key spectral regions in predicting the component of interest is important for sensor design. The full-spectrum data was simplified by iteratively removing wavelengths with low predictive power and producing a set of final models with the most significant predictors for each leaf trait (Fig. 4). While there are a variety of PLSR variable selection techniques including genetic algorithm PLS (GA-PLS; Leardi, 2000), interval PLS (iPLS; Norgaard, ), sparse PLS (SPLS; Chun and Keles, 2010), and backward selection techniques based on coefficient stability (Martens and Martens, 2000; Feilhauer et al., 2011) the method presented here significantly reduced the spectrum (Table 2; Fig. 4) and provided consistent results for each variable (Fig. 3).

Table 2.

The number of wavebands selected for each variable during the two-stage PLSR modelling within the visible (VIS), near-infrared (NIR), short-wave 1 (SWIR1), and short-wave 2 (SWIR2) spectral regions

Variable	VIS 400–700 nm	NIR 700–1300 nm	SWIR1 1300–1900 nm	SWIR2 1900–2500 nm
Nmass	4	1	3	5
Marea	2	8	6	7
Vcmax	4	1	4	4
Jmax	7	3	14	20

Reviewing the location of selected wavelegnths in our PLSR models (Fig. 4), it was found that many fell within spectral regions associated with leaf characteristics such as internal structure, water content, and chlorophylls. For example, the NIR spectrum of a leaf varies in concert with leaf water content and structure (e.g. thickness and or density), which are related to Marea (Jacquemoud and Baret, 1990; Fourty and Baret, 1997; Ceccato ). Specifically, leaves require more dry material, structurally, in order to hold higher water content. Related to this, there are two well-known water absorption features centred at around 970 nm and 1200 nm (Elvidge, 1990; Penuelas ; Sims and Gamon, 2003; Kokaly ), where both the Jmax and Marea models contain selected wavelengths (Fig. 4). With the exception of the model for Marea, all PLSR models contained wavelengths with relatively high spectral loadings in regions with known sensitivities to variations in leaf nitrogen, which occurs primarily in mesophyll proteins and chlorophylls. Nitrogen typically comprises 6.5% (by weight) of the primary light-harvesting molecules and 30–50% of green leaf N is allocated to the protein ribulose-1,5 bisphosphate carboxylase-oxygenase (Rubisco; Elvidge, 1990). In total, roughly 70% of leaf N is invested in compounds that support carbon fixation (Field, 1983; Evans, 1989) and leaf N status is often strongly associated with net photosynthetic capacity (Amax; Field and Mooney, 1986; Evans, 1989; Reich ). Notably, Rubisco has several relatively broad spectral absorption features in the NIR and SWIR centred at 1.5, 1.68, 1.74, 1.94, 2.05, 2.17, 2.29, and 2.47 μm (Elvidge, 1990) that are located in close proximity to several of the wavelength regions selected for leaf Nmass, Vcmax, and Jmax (Fig. 4). For example, the Vcmax model had wavelengths selected at 1.51, 1.68, 1.76, 1.94, 2.21, and 2.49 μm (Fig. 4: see Supplementary data and Tables S1–S4 at JXB online). These considerations further highlight and reinforce plausible linkages between leaf photosynthetic metabolism and spectral reflectance data.

Remote-sensing approaches offer the potential to estimate the landscape- to regional-scale carbon, water, and energy fluxes, as well as other aspects of terrestrial ecosystem function (Carter, 1998; Rahman ; Townsend ; Fuentes ; Asner and Martin, 2008; Hashimoto ). As a consequence, there has been much research focused on the development of methods to relate remotely sensed observations, from the shortwave (i.e. 0.3–3 μm) through the mid-infrared and thermal (i.e. 8–15 μm) wavelengths, to the photosynthetic functioning of vegetation (Sellers ; Gamon ; Carter, 1998; Zarco-Tejada ; Grace ; Anderson ; Hilker ; Sims ). The spectroscopic technique presented in this study complements previous remote sensing methods that utilize vegetation indices (Gamon ; Fuentes ; Sims ), fluorescence observations (Louis ; Damm ; Joiner ) or light-use efficiency (LUE; Monteith, 1972, 1977) approaches (Rahman ; Asner ; Hilker ) to estimate vegetation carbon fluxes. In this study, hyperspectral data were used to empirically estimate parameters that provide a mechanistic link to the biochemistry of carbon assimilation (Farquhar ; Sellers ). The method described here can potentially provide rapid and accurate assessments of key metabolic properties at the leaf level. In addition, this approach offers the opportunity to enhance or validate other methods through the integration with long-term monitoring networks such as FLUXNET (Baldocchi ) and SpecNet (Gamon ). SpecNet, in particular, is designed to explore the linkages between optical remote-sensing data and key parameters governing the exchange of CO2 and water between vegetation and the atmosphere.

The methods presented here and by others (Doughty ) may also provide the basis for regional estimation of photosynthetic metabolism using imaging spectrometers such as the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS; Green ) as well as future instruments such as the Hyperspectral Infrared Imager [HyspIRI, a two sensor platform having a spectrometer (400–2500 nm) and an 8-band multi-spectral thermal instrument]. Utilizing such instruments, the potential exists to map parameters such as Vcmax and Jmax while providing empirical and broad-scale observations that can be used to test for photosynthetic thermal acclimation in plants across large climatic gradients (Dillaway and Kruger, 2010; Gunderson ). These observations could further be used to improve the parameterization of regional as well as dynamic global vegetation models (DGVMs; Kucharik ; Sitch ; Krinner ; Alton ) that rely on the Farquhar–von Caemmerer–Berry model of photosynthesis (Farquhar ; Farquhar and Sharkey, 1982).

Supplementary data

Supplementary data can be found at JXB online.

Further details of the four PLSR models, including the wavelengths selected and corresponding regression coefficients, are provided. This information can be used to derive estimates of the four traits based on the linear summation of the reflectance values at each wavelength, multiplied by the corresponding regression coefficients, and the intercept value.

Supplementary Table S1. Summary of the leaf nitrogen PLSR model wavelengths, regression coefficients, and jackknife statistics.

Supplementary Table S2. Summary of the leaf mass per area PLSR model wavelengths, regression coefficients, and jackknife statistics.

Supplementary Table S3. Summary of the Vcmax PLSR model wavelengths, regression coefficients, and jackknife statistics.

Supplementary Table S4. Summary of the Jmax PLSR model wavelengths, regression coefficients, and jackknife statistics.