Tracking the phenology of photosynthesis using carotenoid-sensitive and near-infrared reflectance vegetation indices in a temperate evergreen and mixed deciduous forest

Summary

• Photosynthetic phenology is an important indicator of annual gross primary productivity (GPP). Assessing photosynthetic phenology remotely is difficult for evergreen conifers as they remain green year-round. Carotenoid-based vegetation indices such as the photochemical reflectance index (PRI) and chlorophyll/carotenoid index (CCI) are promising tools to remotely track the invisible phenology of photosynthesis by assessing carotenoid pigment dynamics. PRI, CCI and the near-infrared reflectance of vegetation (NIR_V) index may act as proxies of photosynthetic efficiency (ɛ), an important parameter in light-use efficiency models, or direct proxies of photosynthesis.
• To understand the physiological mechanisms reflected by PRI and CCI and the ability of vegetation indices to act as proxies of photosynthetic activity for estimating GPP, we measured leaf pigment composition, PRI, CCI, NIR_V and photosynthetic activity at the leaf and canopy scales over 2 years in an evergreen and mixed deciduous forest.
• PRI and CCI captured the large seasonal carotenoid/chlorophyll ratio changes and good relationships were observed between PRI–ɛ and CCI–photosynthesis and NIR_V–photosynthesis.
• PRI-, CCI- and NIR_V-based models effectively tracked observed seasonal GPP. We propose that carotenoid-based and near-infrared reflectance vegetation indices may provide useful proxies of photosynthetic activity and can improve remote sensing-based models of GPP in evergreen and deciduous forests.

Introduction

Northern forests have an important role in the global carbon (C) cycle and in modulating global climate through the exchange of C, energy and water between the land and atmosphere (Bonan, 2008; Richardson et al., 2013). The seasonal dynamics of C uptake characteristics, or photosynthetic phenology, are sensitive to photoperiod and climate (Fu et al., 2017; Flynn & Wolkovich, 2018). Global warming is leading to shifts in photosynthetic phenology generally towards a longer growing season via advanced spring phenology and delayed autumn phenology, which has important implications for annual C uptake (Peñuelas & Filella, 2001; Cleland et al., 2007; Piao et al., 2007; Peñuelas et al., 2009; Jeong et al., 2011; Gill et al., 2015). In addition, the extent of the shift in phenology differs among species, suggesting the need for a more precise assessment of photosynthetic phenology to capture site-specific heterogeneity to improve quantification of the terrestrial C budget (Morin et al., 2009; Tang et al., 2016).

The eddy covariance (EC) technique provides a powerful means to assess ecosystem C, water and energy fluxes for quantifying C uptake or gross primary productivity (GPP) (Baldocchi et al., 2001). However, it is limited to the flux footprint of the EC system (Baldocchi, 2003). Remote sensing provides another means for surface observations that can extrapolate beyond the flux footprint of EC systems to larger regions for global coverage (Gamon, 2015). Remote sensing largely captures the spectral reflectance properties of leaves and canopies, which may be related to plant function (Ustin et al., 2009). Therefore, the use of remotely sensed vegetation indices as proxies of plant function can be exploited to parameterize models for estimating GPP. The light-use efficiency (LUE) model is widely used to estimate plant productivity, including GPP, because of its simplicity and parameterization by meteorological data and remotely sensed products (Running et al., 2004). The LUE model is based on the concept that GPP is directly related to the amount of absorbed photosynthetically active radiation (APAR) and the photosynthetic efficiency (ɛ), which reflects the ability to utilize APAR for photosynthesis (Eqn 1) (Monteith, 1977; Medlyn, 1998). APAR can be further defined as the amount of incoming photosynthetically active radiation (PAR) and the fraction of PAR absorbed by vegetation (ƒ_APAR) (Eqn 2).

(Eqn 1)

(Eqn 2)

The ƒ_APAR parameter is sensitive to physiological processes of the foliage, including photosynthetic pigment composition and chloroplast movement, but also structural features such as leaf position and angle, as well as overall canopy structure and incoming PAR (Björkman & Demmig-Adams, 1994; Kasahara et al., 2002). ƒ_APAR is closely related to canopy greenness and can be assessed by remotely sensed vegetation indices like the normalized difference vegetation index (NDVI) and the enhanced vegetation index (EVI), which are calculated from visible and near-infrared wavebands (Sellers, 1987; Baret & Guyot, 1991; Carlson & Ripley, 1997; Myneni et al., 2002). For deciduous and annual vegetation, APAR generally represents photosynthetic phenology well, owing to the close relationship between canopy greenness and photosynthetic activity (Balzarolo et al., 2016; Junker & Ensminger, 2016b; Yu et al., 2017). However, evergreen conifers present a challenge for tracking photosynthetic phenology with APAR because they retain their needles and show little variation in foliage over the course of the year and hence canopy greenness remains largely unchanged (Hmimina et al., 2013; Wu et al., 2014). Thus, using the LUE model according to Eqn 1, ɛ reflects the seasonal regulation of photosynthetic efficiency in evergreen conifers, and is the parameter that represents photosynthetic phenology (Garbulsky et al., 2011; Fréchette et al., 2015, 2016).

The direct estimation of ɛ has been challenging because it represents the fraction of APAR used for photosynthesis. A common indirect approach to estimate ɛ is based on a fixed biome-specific maximum constant (ɛ_max) downscaled via realized ɛ (ɛ_realized) using temperature and water availability limitations for photosynthesis from meteorological data (Running et al., 2004; Yuan et al., 2007). However, as a result of this broadscale approach, site-specific heterogeneity of ɛ may be poorly represented (Drolet et al., 2008; Garbulsky et al., 2010; D'Odorico et al., 2014). Therefore, estimating ɛ directly from vegetation via remote sensing may improve the LUE model (Coops et al., 2010; Hilker et al., 2011).

A remote sensing approach to assess ɛ is based on carotenoid pigment composition via the photochemical reflectance index (PRI) (Garbulsky et al., 2011; Peñuelas et al., 2011). Carotenoid pigments are involved in regulating photosynthetic processes and are important modulators of photoprotective nonphotochemical quenching (NPQ). Carotenoid content and composition also vary in response to environmental conditions (Demmig-Adams & Adams, 1996). Over diurnal periods, PRI captures the photoprotective conversion of the xanthophyll cycle pigment violaxanthin into antheraxanthin and zeaxanthin in response to light to facilitate the distribution of absorbed light energy and therefore ɛ (Gamon et al., 1992; Peñuelas et al., 1995). Specifically, PRI reflects the conversion of the xanthophyll cycle pigments based on changes in spectral reflectance at 531 nm in comparison to a reference waveband at 570 nm. However, the focus on the xanthophyll cycle as a proxy of ɛ has caused some misinterpretation of PRI at the seasonal timescale. Seasonal PRI is largely driven by the ratio of carotenoid to chlorophyll pigments (Car : Chl) rather than the xanthophyll cycle in evergreen conifers (Busch et al., 2009; Fréchette et al., 2015; Wong & Gamon, 2015b) and deciduous trees (Sims & Gamon, 2002; Junker & Ensminger, 2016b). Owing to the role of Car : Chl in regulating photosynthetic activity, the seasonal variation of the PRI signal is associated with the seasonal variation in photosynthetic activity and a sustained component of NPQ. This suggests that PRI may provide a reliable proxy of photosynthetic phenology (Busch et al., 2009; Porcar-Castell et al., 2012; Hmimina et al., 2015; Wong & Gamon, 2015a; Fréchette et al., 2016; Wong et al., 2019). Together, these studies suggest that the remote assessment of carotenoid pigments may provide a useful proxy of ɛ at both diurnal and seasonal timescales.

The chlorophyll/carotenoid index (CCI) is another indicator of Car : Chl and therefore photosynthetic phenology. CCI is an analog of PRI, which assesses Car : Chl at 531 nm and uses the 645 nm waveband as a reference, instead of the 570 nm waveband used to estimate PRI (Gamon et al., 2016). This shift of the reference waveband from 570 to 645 nm results in an increased sensitivity of CCI for Chl pigments compared with PRI (Wong et al., 2019). Originally, Gamon et al. (2016) demonstrated that CCI at the canopy scale and derived from the satellite-based moderate resolution imaging spectroradiometer (MODIS) sensor effectively tracks photosynthetic phenology in evergreen conifers at both leaf and canopy scales. Springer et al. (2017) observed that CCI is effective at tracking photosynthetic phenology in both evergreen and deciduous trees at the canopy scale. Because CCI is sensitive to both physiological and structural features that are associated with photosynthetic phenology across plant functional types, Springer et al. (2017) recently suggested that CCI could potentially bypass the LUE model as a direct proxy of GPP.

Taken together, there is considerable potential for using carotenoid-sensitive vegetation indices to track photosynthetic phenology. Because CCI was introduced only recently, it has not been used in any comparative studies as yet. By contrast, the relationship between PRI and ɛ has been evaluated in two meta-analyses (Garbulsky et al., 2011; Zhang et al., 2016). Both studies clearly demonstrate that PRI is a good proxy of ɛ at different spatial and temporal scales and across plant functional types. However, both studies also highlighted the need for a better understanding of the mechanisms that contribute to the variation of this vegetation index across spatial scales. For example, canopy-scale PRI is known to be sensitive to canopy structure, solar angle and background signals, which may influence the spectral reflectance signal reflected by PRI (Barton & North, 2001). In addition, Gitelson et al. (2017a) showed at the leaf scale in crops and deciduous trees that seasonal PRI is controlled by Car : Chl, while at the canopy scale in crops, Gitelson et al. (2017b) observed that PRI is influenced by canopy structure and Chl content, which weakens the relationship between PRI and Car : Chl. These confounding effects require further clarification to elucidate the physiological mechanisms driving canopy-scale PRI and CCI variation, which often remains uncertain. Together, this warrants further work to validate the mechanisms influencing the PRI and CCI signals at the canopy scale by incorporating photosynthetic pigment quantification and leaf-scale validation measurements (Grace et al., 2007; Gamon, 2015).

A newly derived index, the near-infrared reflectance of vegetation (NIR_V), provides an alternative remote sensing approach via MODIS to assess GPP (Badgley et al., 2017). NIR_V is based on NDVI multiplied by near-infrared (NIR) reflectance to represent light capture of vegetation accounting for leaf area, orientation and canopy structure and minimalizes the influence of background soil contamination (Badgley et al., 2017, 2019). MODIS NIR_V evaluation at FLUXNET validation sites reveals a good prediction of monthly and annual GPP (Badgley et al., 2017, 2019). However, the evaluation of MODIS NIR_V to assess photosynthetic phenology at higher frequencies remains unclear. In grasslands, Wang et al. (2020) reported a stronger CCI relationship to GPP than to NIR_V, potentially because of the sensitivity of NIR_V to vegetation development and the influence of seasonal hysteresis in vegetation greenness and photosynthetic activity. Therefore, there is a need to evaluate the ability of NIR_V to track seasonal photosynthetic activity, especially in evergreen vegetation, which shows subtle seasonal changes in vegetation greenness, which typically limits the use of greenness sensitive indices like NDVI (Hmimina et al., 2013; Wu et al., 2014).

In order to overcome current limitations in our ability to assess the phenology of photosynthesis and GPP through remote sensing approaches, we compared four vegetation indices (NDVI, PRI, CCI, NIR_V) with photosynthetic activity at the leaf and canopy scales in a temperate evergreen and a mixed deciduous forest stand over a period of 2 years. Our goals were to assess the relationship of photosynthetic pigments and NDVI, PRI and CCI; to assess how the NDVI, PRI, CCI, and NIR_V reflect the phenology of photosynthesis; and to predict GPP based on models parameterized by a combination of NDVI (as a proxy of ƒ_APAR) and PRI (as a proxy of ɛ) or solely from CCI or NIR_V.

Materials and Methods

Site description

This study was conducted at two long-term C monitoring forest stands at the Turkey Point Flux Station in Ontario, Canada from March 2015 to the end of 2016. The evergreen forest stand (TP39; 42°42′36.6″N, 80°21′26.6″W) is dominated by 80-yr-old eastern white pine (Pinus strobus L.). Stand density is 421 ± 166 trees ha⁻¹ and mean canopy height is 20 m. A 28-m-high flux tower was established in 2003, which was extended to 32 m in May 2016. The mixed deciduous forest stand (TPD; 42°38′07.2″N, 80°33′27.8″W) is dominated by white oak (Quercus alba L.) and red maple (Acer rubrum L.). Other species include American beech (Fagus grandifolia Ehrh.), black oak (Q. velutina Lam.), red oak (Q. rubra L.), white ash (Fraxinus americana L.) and eastern white pine. Stand density at the deciduous site is 504 ± 18 trees ha⁻¹ and mean canopy height is 25.7 m. A 35.4-m-high flux tower was established in 2012.

Carbon flux and climatic measurements

Both forest stands are equipped with a closed-path EC system and a meteorological station located at the top of the flux towers. Flux data were filtered, de-spiked, gap-filled and processed as described in Arain & Restrepo-Coupe (2005) following standard algorithms to provide half-hourly fluxes and estimated EC GPP (GPP_observed), and climatic parameters. Daily total GPP_observed was summed from the half-hourly data. EC ƒ_APAR (ƒ_APAR(canopy)) was estimated using above- and below-canopy PAR sensors. A pair of above-canopy PAR sensors were located at the top of the flux towers, one facing upwards to capture incoming PAR (PAR_downwelling) and the other facing downwards to capture outgoing PAR from the canopy (PAR_reflected). A below-canopy PAR sensor was located 2 m above the ground facing upwards to capture PAR transmitted through the canopy to the ground (PAR_transmitted). There was no below-canopy, downward-facing PAR sensor for soil reflectance so we assumed soil to be black with no light reflectance, as our forest stands had a closed canopy during the growing season (Widlowski, 2010). Using the half-hourly GPP_observed and PAR data, we calculated ƒ_APAR(canopy) and canopy-scale LUE (ɛ_canopy) as:

(Eqn 3)

(Eqn 4)

Midday ƒ_APAR(canopy) and ɛ_canopy were averaged from the half-hourly data from 12:00 to 15:00 h bordering solar noon to capture midday values where PAR was assumed to be maximized for comparison with optical measurements.

Spectral measurements

Canopy-scale spectral reflectance was obtained for four narrow wavebands for the calculation of NDVI and PRI using a set of spectral reflectance sensors (SRS; Meter Group Inc., Pullman, WA, USA). The SRS sensors were set up at the top of the flux towers and data were collected every 1 min, with 5 min average values stored on a datalogger (EM50G; Meter Group Inc.). Each site was equipped with three target NDVI and PRI sensors with a field of view of 20° for spot canopy signal, and two reference NDVI and PRI sensors with a field of view of 180° to provide reference values of sky irradiance. Each SRS sensor measured two bands, 630 and 800 nm for the NDVI sensors, and 532 and 570 nm for the PRI sensors. The target sensors faced east, south and west away from the flux tower and pointed downwards towards the canopy at a 45° angle. The reference sensors were mounted c. 2 m from the target sensors and elevated an extra 2 m to ensure they were not shaded from other structures on the flux tower.

To calculate reflectance from the target sensors, radiance values for each waveband of the target sensors were normalized by the average of the corresponding wavebands’ radiance values from the reference sensors. In addition, a white panel calibration was performed every 2–3 months in order to cross-calibrate the target sensors according to Gamon et al. (2015). For the white panel calibration, a Teflon white reference panel was placed under each target sensor covering the entire field of view for a minimum of 15 min, while the reference sensors sampled sky irradiance. This procedure yielded a cross-calibration ratio by dividing the target sensor data of the white reference panel by the reference sensor for each waveband and sensor pair. Reflectance data were calibrated by dividing the reflectance data for each waveband by the respective cross-calibration ratio. With the calibrated reflectance values, NDVI and PRI were calculated using data from the respective NDVI and PRI sensors, CCI was calculated using wavebands from both NDVI and PRI sensors, and NIR_V was calculated by multiplying NDVI by a NIR waveband (800 nm) from the NDVI sensor. The canopy-scale NDVI (NDVI_canopy), PRI (PRI_canopy), CCI (CCI_canopy) and NIR_V (NIR_Vcanopy) were calculated as:

(Eqn 5)

(Eqn 6)

(Eqn 7)

(Eqn 8)

where R indicates reflectance at the specific wavelengths in nm. The 5 min data were averaged to 30 min aggregates to match flux data. Daily midday NDVI_canopy, PRI_canopy and CCI_canopy were expressed as the mean of the half-hourly data obtained between 12:00 and 15:00 h to represent solar noon and to minimize the effects of the diurnal variation in solar angle.

Leaf-scale spectral reflectance measurements were used to ground truth NDVI_canopy, PRI_canopy and CCI_canopy. Leaf-scale spectral reflectance measurements were obtained using a portable spectrometer (UniSpec-SC; PP Systems, Amesbury, MA, USA) equipped with a bifurcated fiber-optic (UNI410; PP Systems) and a needle leaf clip (UNI501; PP Systems) to measure leaf surface reflectance at a fixed angle of 60° (0.6 mm diameter spot size). Measurements were completed between 12:00 and 13:00 h on sunlit leaves near the top of the canopy at 1 to 2 wk intervals during the spring and autumn, and at 4 to 5 wk intervals during the summer and winter. At each site, three trees per species were selected that had branches that could be accessed from the flux tower. From each selected tree, we chose one sunlit branch at the top of the canopy. From each branch, we obtained 10 separate measurements from different leaves in order to account for heterogeneity within the canopy. Only the dominant species were measured: eastern white pine at the evergreen forest, and white oak and red maple at the mixed deciduous forest. For pine, only second-year needles were measured. Before leaf reflectance measurements, a dark current correction scan was measured to correct for dark current instrument noise. This was followed by a reference scan on a white reference standard (Spectralon, Labsphere, North Sutton, NH, USA) and finally a leaf target scan for leaf radiance. Reflectance was calculated by dividing leaf radiance by the white reference standard scan. Leaf-scale NDVI (NDVI_leaf), PRI (PRI_leaf), CCI (CCI_leaf), NIR_V (NIR_Vleaf) were calculated using Eqns 5–8.

Leaf-scale gas exchange

Leaf-scale gas exchange was measured using a portable photosynthesis system (Li-6400; Li-Cor, Lincoln, NE, USA). A bundle of conifer needles or a single deciduous leaf which remained attached to the tree were placed inside the gas exchange chamber (6400-40; Li-Cor). The same three trees from the leaf-scale spectral reflectance measurements were used. Three sets of leaves were measured per tree. The leaf chamber was set to a flow rate of 400 ml min⁻¹, c. 55% relative humidity and a CO₂ concentration of 400 ppm. Temperature was set to match ambient temperature, which ranged from −1 to 30°C in winter and summer, respectively. Air vapor pressure deficit (VPD) ranged from 0.2 to 1.3 kPa from winter to summer, respectively. Leaves were dark-acclimated for 20 min inside the leaf chamber. Maximum CO₂ assimilation rate (A_max) was measured at 1500 µmol m⁻² s⁻¹ photosynthetic photon flux density (PPFD) once steady state was achieved (usually between 7 to 10 min). Leaf-scale LUE (ɛ_leaf) was calculated as A_max/1500 µmol m⁻² s⁻¹ PPFD. After measurements, leaves were collected and for the conifers, the needle leaf surface area was measured using WinSEEDLE package (Regent Instruments Inc., Québec City, QC, Canada), while the deciduous leaf surface area was measured using ImageJ (https://imagej.net/).

Pigment analysis

Leaf samples for pigment analysis were collected following leaf-scale spectral reflectance measurements from the same location on the same branch. Samples were flash-frozen in liquid nitrogen and later transferred to a −80°C freezer. Three batches of leaf material were collected per tree. Each batch was immersed in liquid nitrogen and ground to a fine powder using a mortar and pestle. Pigments were extracted in 98% methanol buffered with 2% 0.5 M ammonium acetate according to Junker & Ensminger (2016a). Pigments were separated and quantified using a high-performance liquid chromatography (HPLC) system (1260 Infinity; Agilent Technologies, Santa Clara, CA, USA) according to Junker & Ensminger (2016a). The HPLC system was equipped with a reverse-phase C₃₀ column (YMC Carotenoid; YMC Co. Ltd, Kyoto, Japan) The mobile phase consisted of a gradient of methanol, water buffered with 0.2% ammonium acetate and methyl tert-butyl ether at a flow rate of 1 ml min⁻¹. Pigments were detected at 450 and 656 nm. Chla and Chlb were quantified using standards from Sigma-Aldrich, and antheraxanthin, β-carotene, lutein, neoxanthin, violaxanthin and zeaxanthin were quantified using standards from DHI Lab Products (Hørsholm, Denmark). Total Chl content was expressed as the sum of Chla and Chlb on a FW basis (µmol g⁻¹). The ratio of Car : Chl was expressed as total Car, including violaxanthin, antheraxanthin, zeaxanthin, neoxanthin, lutein, α-carotene and β-carotene, divided by Chl content (mmol mol⁻¹).

Evaluation of different LUE models for the prediction of GPP

The phenology of photosynthesis based on daily GPP was modeled using three different approaches parameterized by midday NDVI_canopy and PRI_canopy, CCI_canopy or NIR_Vcanopy and compared with daily total GPP_observed from the EC measurements. To calibrate and test our models, we randomly divided the entire canopy-scale dataset into a model calibration dataset consisting of two-thirds of the data and a model testing dataset consisting of one-third of the data. We used a LUE model parameterized by NDVI and PRI, and solely driven CCI and NIR_V models using GPP_observed for the calibration and scaling of spectral reflectance data which are described in the following:

1. PRI model. Here we used PAR, NDVI_canopy as a proxy of ƒ_APAR and a normalized PRI_canopy from 0 to 1 (PRI_normalized) as a proxy of ɛ directly in the PRI model to estimate an uncalibrated GPP:

(Eqn 9)

The uncalibrated GPP was used to calculate a linear regression equation with GPP_observed using the model calibration dataset. For model testing, we used the model testing dataset to estimate an uncalibrated GPP using Eqn 9 and empirically estimated GPP (GPP_PRI) using the linear regression obtained from the model calibration dataset.

1. CCI model. Here we found a nonlinear relationship between CCI_canopy and GPP_observed using the model calibration dataset. Using this regression model, we estimated GPP (GPP_CCI) directly with a CCI function (ƒ(_CCI)) with the model testing dataset:

(Eqn 10)

1. NIR_V model. Here we found a linear relationship between NIR_Vcanopy and GPP_observed using the model calibration dataset. Using this regression model, we estimated GPP (GPP_NIRv) directly with a NIR_V function (ƒ(_NIRv)) with the model testing dataset:

(Eqn 11)

The linear and nonlinear regressions used in these models were determined separately for the evergreen and the mixed deciduous forest as a result of different observed relationships. The estimated GPP from all models were cleaned by removing all time points where PAR was < 50 µmol m⁻² s⁻¹ and when precipitation occurred, because rainfall events have confounding effects on the spectral reflectance signal. In addition, when daily average air temperature was < 0°C, the estimated GPP was set to zero.

Results

Tracking photosynthetic phenology with NDVI, PRI, CCI and NIR_V

Photosynthetic phenology in the evergreen and mixed deciduous forests were revealed by the seasonal variation in GPP_observed. In both forests, no C uptake occurred during the winter season indicated by zero GPP_observed values (Fig. 1). In the evergreen forest, C uptake began in April in 2015 and March in 2016, increasing quickly until July, indicated by peak GPP_observed values of about 12 g C m⁻² d⁻¹ for both years. Following this peak of C uptake, GPP_observed declined during the autumn and had ceased by the end of December. In the mixed deciduous forest, the growing season was considerably shorter than the evergreen forest with GPP_observed beginning in May, which quickly increased until July to about 12 g C m⁻² d⁻¹ and declined during the autumn where GPP_observed had ceased by the end of October (Fig. 1).

Fig. 1

Seasonal dynamics in an evergreen (left) and mixed deciduous (right) forest at Turkey Point, ON, from March 2015 to January 2017. (a–h) Eddy covariance gross primary productivity (GPP_observed; black line), leaf-scale (symbols) and canopy-scale (red lines) normalized difference vegetation index (NDVI; a, b), photochemical reflectance index (PRI; c, d), chlorophyll/carotenoid index (CCI; e, f) and near-infrared reflectance of vegetation (NIR_V; g, h). GPP_observed and canopy-scale vegetation indices show the 5 d running average. Leaf-scale measurements show the dominant species from each forest stand: pine (green dots), maple (blue triangles) and oak (cyan squares). Error bars denote the SE of the mean (n = 3). Alternating gray and white bars represent periods of 2 months.

The seasonal variation of GPP_observed was not equally well represented by midday NDVI, PRI, CCI and NIR_V in the evergreen and mixed deciduous forests (Fig. 1). NDVI revealed limited ability to track photosynthetic phenology in the evergreen forest indicated by the nearly absent seasonal variation of NDVI_canopy and NDVI_leaf (Fig. 1a). By contrast, in the mixed deciduous forest, both NDVI_leaf and NDVI_canopy were able to track photosynthetic phenology as revealed by a distinct increase in NDVI during spring and a decline NDVI during the autumn (Fig. 1b). PRI revealed a good ability to track photosynthetic phenology in both the evergreen and mixed deciduous forests indicated by the seasonal variation of both PRI_leaf and PRI_canopy (Fig. 1c,d). Similar to GPP_observed, PRI increased during the spring and declined during the autumn. In the mixed deciduous forest, PRI_canopy exhibited large variations of PRI values during the autumn (Fig. 1d). We note that the autumn decline of PRI showed a slight lag in timing compared with GPP_observed (Fig. 1c,d). CCI revealed a similar pattern to PRI and appears more coupled in timing with GPP_observed, and hence reflected photosynthetic phenology in both evergreen and mixed deciduous forests for both CCI_leaf and CCI_canopy (Fig. 1e,f). NIR_Vcanopy revealed a good ability to track photosynthetic phenology in both the evergreen and mixed deciduous forests (Fig. 1g,h). We note that in the evergreen forest, the recovery of NIR_V begins earlier than that of GPP_observed (Fig. 1g).

Sensitivity of NDVI, PRI and CCI to seasonal Chl and Car : Chl dynamics

To better understand how pigment composition contributes to variation in the vegetation indices at different scales, we evaluated the relationship of leaf- and canopy-scale NDVI, PRI and CCI with Chl and Car : Chl based on linear regression analysis (Fig. 2). There was no relationship between NDVI_leaf and Chl and Car : Chl in pine, while for both deciduous trees, we observed significant relationships (Fig. 2a,b). Interestingly for maple, NDVI_leaf had a higher regression coefficient with Chl than with Car : Chl, whereas for oak, the opposite pattern was observed (Fig. 2b). The relationship between NDVI_canopy and Chl and Car : Chl was significant for both evergreen and deciduous trees, although Chl generally had higher correlation coefficients (Fig. 2c,d). Comparing the spatial scales of NDVI, we observed a significant relationship between NDVI_leaf and NDVI_canopy only for the deciduous trees (Fig. 2e).

Fig. 2

Linear regressions between leaf- and canopy-scale normalized difference vegetation index (NDVI; a–e), photochemical reflectance index (PRI; f–j), and chlorophyll/carotenoid index (CCI; k–o) with Chl content and carotenoid : chlorophyll (Car : Chl) and between leaf- and canopy-scale vegetation index for pine (green circles), maple (blue triangles) and oak (cyan squares) from Turkey Point, ON, from March 2015 to January 2017. Leaf spectral reflectance data represent measurements obtained from three trees per species (10 measurements per tree). Pigment data reflect measurements obtained from three trees per species (three batches of leaf material per tree). No symbol, P > 0.05; *, P < 0.05; **, P < 0.01; ***P < 0.001.

The relationship between PRI_leaf and Chl and Car : Chl was significant for both evergreen and deciduous trees with regression coefficients higher between PRI_leaf and Car : Chl for pine, whereas for the deciduous trees, PRI_leaf had stronger relationships with Chl (Fig. 2f,g). Like PRI_leaf, PRI_canopy was significantly correlated to both Chl and Car : Chl, although the regression coefficients were generally weaker compared with PRI_leaf, except for maple between PRI_canopy and Chl (Fig. 2h,i). The direct comparison of PRI_leaf and PRI_canopy indicated significant relationships for both evergreen and deciduous trees with similar regression coefficients (Fig. 2j).

CCI_leaf and CCI_canopy exhibited a significant relationship with Chl and Car : Chl for both evergreen and deciduous trees (Fig. 2k–n). At the leaf scale, stronger regression coefficients between CCI_leaf and Car : Chl were observed for pine and oak, while for maple, CCI_leaf was more strongly related to Chl (Fig. 2k,l). At the canopy scale, CCI_canopy generally had stronger regression coefficients with Chl than Car : Chl (Fig. 2m,n). Comparing CCI from leaf and canopy scales, we found a significant regression coefficient between CCI_leaf and CCI_canopy for evergreen and strongly significant regression coefficients for the deciduous trees (Fig. 2o).

NDVI, PRI, CCI and NIR_V as proxies of ƒ_APAR, ɛ and photosynthesis

NDVI_canopy was strongly related with ƒ_APAR(canopy) for both evergreen and mixed deciduous forests (Fig. 3). However, the linear regression line differed between the evergreen and mixed deciduous forests as a result of limited variation of ƒ_APAR(canopy) in the evergreen forest (Fig. 3). PRI was strongly correlated with ɛ for all species at both leaf and canopy scales based on the regression coefficients; however, the canopy-scale regression coefficients were much lower compared with those at the leaf-scale (Fig. 4a,b). To explore the role of CCI in the LUE model, we evaluated the relationship of CCI with ɛ and photosynthesis. At the leaf scale, CCI_leaf was a good predictor of both ɛ_leaf and A_max for all species, with a linear relationship for pine and nonlinear relationships for the deciduous maple and oak (Fig. 4c). At the canopy scale, CCI_canopy was a significant predictor of both ɛ_canopy and GPP_observed in both forests (Fig. 4d). The regression coefficients were generally higher between CCI_canopy and GPP_observed than with ɛ_canopy (Fig. 4c). NIR_Vleaf exhibited a good relationship with A_max only for the deciduous species and had no relationship with the pine (Fig. 4g). At the canopy scale, NIR_Vcanopy had significant relationships with GPP_observed for both evergreen and mixed deciduous forests (Fig. 4h).

Fig. 3

Linear regression of canopy-scale fraction of absorbed photosynthetically active radiation (f_APAR) and normalized difference vegetation index (NDVI_canopy) in an evergreen needleleaf forest (ENF; black) and mixed deciduous broadleaf forest (DBF; red) at Turkey Point, ON, from March 2015 to January 2017. Canopy spectral reflectance data represent measurements obtained from three spectral reflectance sensors per forest stand. All relationships are statistically significant at P < 0.001.

Fig. 4

Regressions of leaf- (left) and canopy-scale (right) photosynthetic efficiency (ɛ) and photochemical reflectance index (PRI; a, b), ɛ and chlorophyll/carotenoid index (CCI; c, d), maximum assimilation rate (A_max) or eddy covariance gross primary productivity (GPP_observed) and CCI (e, f), A_max or GPP_observed and near-infrared reflectance of vegetation (NIR_V; g, h) for pine (green circles) of the evergreen forest (ENF; black circle), maple (blue triangles) and oak (cyan squares) of the mixed deciduous broadleaf forest (DBF; red triangles) at Turkey Point, ON, from March 2015 to January 2017. Leaf spectral reflectance data represent measurements obtained from three trees per species (10 measurements per tree). Canopy spectral reflectance data represent measurements obtained from three spectral reflectance sensors per forest stand. No symbol, P > 0.05; *, P < 0.001.

Estimation and validation of GPP and phenology using models driven by vegetation indices

The seasonal dynamics of GPP_observed were represented well by the PRI, CCI and NIR_V models (Fig. 5). Linear regression analysis for model performance demonstrated strong model predictions of GPP_observed, indicated by the significant regression coefficients (Fig. 6). Interestingly, the PRI model performed slightly better than the CCI and NIR_V model in the evergreen forest, whereas the CCI and NIR_V models performed slightly better than the PRI model in the mixed deciduous forest based on the regression coefficients, root mean square error and bias analysis (Fig. 6).

Fig. 5

The 5 d running means (lines) of observed and estimated daily gross primary productivity (GPP; symbols) using three modeling approaches based on a normalized difference vegetation index (NDVI) and photochemical reflectance index (PRI) model (GPP_PRI), a chlorophyll/carotenoid index (CCI) model (GPP_CCI) and a near-infrared reflectance of vegetation (NIR_V) model (GPP_NIRv) at an evergreen (a) and mixed deciduous (b) forest at Turkey Point, ON, from March 2015 to January 2017. Symbols represent the daily estimate of GPP from each model.

Fig. 6

Linear regressions of estimated gross primary productivity (GPP) using the three model approaches of a normalized difference vegetation index (NDVI) and photochemical reflectance (PRI) model (GPP_PRI; circles), a chlorophyll/carotenoid index (CCI) model (GPP_CCI; triangles) and a near-infrared reflectance of vegetation (NIR_V) model (GPP_NIRv; squares) and eddy covariance GPP (GPP_observed) at an evergreen (a) and mixed deciduous (b) forest at Turkey Point, ON, from March 2015 to January 2017. All relationships are statistically significant at P < 0.001.

Discussion

In this study, we evaluated the ability of four vegetation indices (NDVI, PRI, CCI and NIR_V) to track the phenology of photosynthesis in an evergreen and a mixed deciduous forest, with the goal of using these vegetation indices to develop a robust and simplified model for estimating canopy-scale GPP. We demonstrate that leaf- and canopy-scale NDVI represents photosynthetic phenology only in the mixed deciduous forest, while PRI, CCI and NIR_V track photosynthetic phenology in both the evergreen and mixed deciduous forests. Furthermore, we show that NDVI, PRI, CCI and NIR_V perform as suitable proxies of ƒ_APAR, ɛ and photosynthesis, respectively. As proxies of these photosynthetic parameters, we successfully estimate daily GPP using LUE, CCI and NIR_V models.

Seasonal variation of photosynthetic pigments reflected by NDVI, PRI and CCI

Photosynthetic phenology includes the regulation of photosynthetic activity, involving seasonal adjustments of photosynthetic pigment composition and pool sizes, which may be assessed using spectral reflectance vegetation indices (Blackburn, 2007; Ustin et al., 2009). Deciduous trees undergo large seasonal variation in leaf greenness, driven by the synthesis and degradation of Chl content during the spring and autumn, respectively. This seasonal variation in leaf greenness and Chl is generally assessed by NDVI (Tucker, 1979; Sellers, 1987; Baret & Guyot, 1991; Gamon et al., 1995; Carlson & Ripley, 1997). Our data confirm this strong relationship between NDVI and Chl at both leaf and canopy scales in the deciduous trees (Fig. 2a,c). However, for the evergreen conifer, a relationship between NDVI and Chl was only observed at the canopy scale, which was much weaker than for maple as a result of limited seasonal variation in Chl (Fig. 2c). Therefore, as a result of the limited seasonality of Chl and vegetation greenness, NDVI is a relatively weak proxy for the photosynthetic phenology of evergreen conifers (Gamon et al., 1995; Nagai et al., 2012; Hmimina et al., 2013).

For evergreen conifers, photosynthetic phenology is probably better represented by the Car : Chl ratio, which is involved in regulating photosynthetic activity and photoprotection via enhanced Car : Chl during overwintering and lower Car : Chl during the growing season (Adams et al., 2002; Demmig-Adams & Adams, 2006). This seasonal variation of Car : Chl may be assessed with PRI (Filella et al., 2009; Garrity et al., 2011; Wong & Gamon, 2015b). Our data confirm a relationship between PRI and Car : Chl at leaf and canopy scales for both evergreen and deciduous trees (Fig. 2g,i). However, a relationship between PRI and Chl is also observed and, specifically for the deciduous trees, the regression coefficient is much higher (Fig. 2f,h). This suggests that PRI tracks the degradation of Chl more strongly than the change in Car : Chl in deciduous vegetation (Nakaji et al., 2006; Gitelson et al., 2017b).

Like PRI, CCI may assess Car : Chl to track seasonal photosynthetic phenology in northern deciduous and evergreen trees (Gamon et al., 2016; Springer et al., 2017). However, much like PRI, CCI has a relationship with both Car : Chl and Chl in which the CCI–Car : Chl relationship is stronger for the evergreen conifer, whereas the CCI–Chl relationship is stronger for the deciduous trees at both leaf and canopy scales (Fig. 2k–n). The relationship between CCI and both Car : Chl and Chl suggests that CCI may reflect the regulation of photosynthetic activity via light harvesting and photoprotective processes. This relatively new finding provides a mechanistic support to the work of Springer et al. (2017). Consequently, CCI is probably effective at tracking photosynthetic phenology in both evergreen and deciduous trees.

NDVI, PRI, CCI and NIR_V as proxies of ƒ_APAR, ɛ and photosynthesis

In LUE models, NDVI is commonly used as a proxy of ƒ_APAR because of a strong linear relationship (Baret & Guyot, 1991; Myneni & Williams, 1994; Running et al., 2004). Our data confirm a linear relationship between NDVI_canopy and ƒ_APAR(canopy) in both forests (Fig. 3). However, the relationships differed largely between forests where seasonal variations of NDVI_canopy and ƒ_APAR(canopy) were large in the mixed deciduous forest, whereas the evergreen forest showed limited seasonal variation of ƒ_APAR(canopy) as needles are retained. In the mixed deciduous forest, we also observed some decoupling between NDVI_canopy and ƒ_APAR(canopy) (Fig. 3), which is probably a result of highly heterogenous canopy structure and cover during early spring and late autumn because of different stages of leaf growth and senescence between trees and species in the mixed forest (Jenkins et al., 2007). This may result in a partially open canopy introducing contributions of reflectance from understory vegetation and leaf litter (Widlowski, 2010; D'Odorico et al., 2014). In addition, our estimations of ƒ_APAR(canopy) were based on single-point PAR data, which may not adequately represent the variation of light transmittance within the forest stand potentially leading to further bias when canopy cover is strongly heterogenous.

Photochemical reflectance index is a good proxy of ɛ across spatial scales and different plant functional types; however, heterogeneity has been shown to make scaling from leaf to canopy difficult (Garbulsky et al., 2011; Gamon, 2015; Zhang et al., 2016). This is largely a result of leaf heterogeneity in such parameters as leaf angle and position, which is absent at the leaf scale. At the canopy scale, the signal of PRI may not only be influenced by leaf angle and position, but also sun angle (Barton & North, 2001). Our data confirm a good relationship between PRI and ɛ at both leaf (Fig. 4a) and canopy scales (Fig. 4b) with suitable extrapolation from the leaf- to canopy-scale PRI (Fig. 2j). However, the canopy-scale PRI_canopy and ɛ_canopy regression coefficient was much lower than at the leaf scale, probably because of factors such as leaf angle and position, canopy structure, shading and sun angle which were shown to influence the PRI_canopy signal (Barton & North, 2001; Hernández-Clemente et al., 2011; Gitelson et al., 2017b).

One of the goals of this study was to evaluate the relationship between CCI and ɛ as a result of its similarity to PRI, and the relationship between CCI and photosynthesis as was suggested by Springer et al. (2017). This work demonstrates that CCI can act as a reliable proxy for ɛ and photosynthesis at both leaf (Fig. 4c) and canopy scales (Fig. 4d). Interestingly, the regression coefficients between CCI and photosynthesis were higher than with ɛ (Fig. 4d,f), indicating that CCI may perform better as a direct proxy of photosynthesis. Surprisingly, the regression coefficients between leaf- and canopy-scale CCI with photosynthesis (Fig. 4e,f) were similar, indicating strong extrapolation further supported by the relationships between CCI_leaf and CCI_canopy (Fig. 2o). With CCI becoming more readily available from satellite platforms (Gamon et al., 2016), CCI can provide a powerful and robust tool for assessing seasonal variation and phenology of photosynthesis in both evergreen and mixed deciduous forests.

The newly described NIR_V has been suggested as a direct proxy of photosynthesis based on satellite NIR_V validation at FLUXNET validation sites (Badgley et al., 2017, 2019). Here, we demonstrate that canopy-scale NIR_V is a good proxy of GPP at the daily timescale in both evergreen and mixed deciduous forests (Fig. 4h). However, at the leaf scale, NIR_V was unable to track seasonal A_max in pine (Fig. 4g). This may be a result of the limited seasonal variation of leaf greenness, which NIR_V reflects, whereas at the canopy scale, NIR_V may reflect the light capture of vegetation influenced by a combination of vegetation greenness, leaf area, orientation and canopy structure (Badgley et al., 2017, 2019), resulting in the detection of seasonal GPP in the evergreen forest. This suggests that NIR_V at the canopy scale may provide another approach to assess GPP based on NIR reflectance.

Performance of models for the estimation of GPP

We present three simple GPP models using NDVI as a proxy of ƒ_APAR and PRI as a proxy of ɛ in a LUE model, CCI as a direct proxy of GPP in the CCI model, and NIR_V as a direct proxy of GPP in the NIR_V model. All three models performed well at tracking seasonal GPP in the evergreen and mixed deciduous forests (Figs 5, 6). Interestingly, the CCI and NIR_V models displayed similar performance in assessing GPP, indicating two alternative remote sensing approaches based on Car : Chl via CCI and light capture via NIR_V for tracking photosynthetic phenology. Comparing the performance of the PRI, CCI and NIR_V models between forest stands, the PRI model, which includes NDVI and PAR, performed slightly better than the standalone CCI and NIR_V models in the evergreen forest, whereas in the mixed deciduous forest, the CCI and NIR_V models performed slightly better (Fig. 6). The weaker performance of the CCI and NIR_V models in the evergreen forest may be a result of differences in timing between the recovery of photosynthesis and pigment content (Wong & Gamon, 2015a). For CCI, the relationship between Car : Chl and photosynthetic activity may decouple during early spring recovery (Fréchette et al., 2015; Wong & Gamon, 2015a). In the case of NIR_V, the phenology of vegetation greenness via Chl content and photosynthesis may be decoupled (D’Odorico et al., 2015). The PRI model may be less affected by this decoupling during spring, as PRI was complemented by NDVI and PAR in parameterizing the LUE model, which may correct for some of the mismatched timing during spring recovery. This suggests that the CCI and NIR_V models may be improved with additional driving parameters. However, using CCI and NIR_V as sole model drivers, they have the added benefit of being completely driven by remote sensing products without the need for meteorological data. In the mixed deciduous forest, the potential for decoupling is probably smaller, as photosynthetic pigments, canopy greenness and photosynthetic activity are highly synchronized as a result of the rapid rate of spring recovery (Sims et al., 2006; Toomey et al., 2015). Together, these findings demonstrate the ability of greenness and carotenoid-sensitive vegetation indices for modeling GPP, and potentially complementary roles of NDVI and PRI in the LUE model for improved assessment of evergreen forest GPP.

Conclusions

This study demonstrates the ability of greenness and carotenoid-sensitive vegetation indices to remotely detect photosynthetic phenology in evergreen and mixed deciduous forests at the leaf and canopy scales. The seasonal variation of photosynthetic pigment composition in Car : Chl is captured accurately by carotenoid-sensitive vegetation indices such as PRI and CCI, which can therefore act as proxies of ɛ and photosynthesis, respectively, at both leaf and canopy scales. We further demonstrate that the relatively simple LUE model parameterized by NDVI and PRI, and even models solely based on the recently described CCI and NIR_V, can provide accurate estimates of GPP and hence represent promising tools for capturing GPP with remote sensing-based models. With satellite-based PRI and CCI products becoming more readily available, in addition to the newly described NIR_V, remote sensing of vegetation spectral reflectance promises powerful applications for assessing photosynthetic phenology in evergreen- and deciduous-dominated ecosystems remotely and continuously at large scales.

Acknowledgements

This work was supported by funding to IE from NSERC (grant RGPIN-2015-06514), CFI (grant 27330), and the Ontario Ministry of Research and Innovation (grant ER10-07-015); funding for flux measurements provided by the NSERC Discovery, Global Water Futures (GWF) program and the Ontario Ministry of Environment, Conservation and Parks (MOECP) grants to MAA; and The Ann Oaks Doctoral Scholarship awarded by the Canadian Society of Plant Biologists (CSPB) to CYSW. We would like to thank Yazad Bhathena and Dhyani Patel for assistance in pigment analysis and field data collection; Felix Chan, Eric Beamesderfer and Myroslava Khomik for support and advice at Turkey Point; and Steve Garrity, METER Group (formerly Decagon Devices) and SpecNet for providing the SRS sensors used throughout the study.

Author contributions

CYSW and IE developed the initial framework and objectives of the study. CYSW contributed to the field data collection. CYSW analyzed the data, with meteorological and EC data contributions by MAA and the phenology fitted double logistics function by PD. CYSW and IE wrote the manuscript with contributions from all other authors.