A Broadband Green-Red Vegetation Index for Monitoring Gross Primary Production Phenology

The chlorophyll/carotenoid index (CCI) is increasingly used for remotely tracking the phenology of photosynthesis. However, CCI is restricted to few satellites incorporating the 531 nm band. This study reveals that the Moderate Resolution Imaging Spectroradiometer (MODIS) broadband green reflectance (band 4) is significantly correlated with this xanthophyll-sensitive narrowband (band 11) (R2 = 0:98, p < 0:001), and consequently, the broadband green-red vegetation index GRVI—computed with MODIS band 1 and band 4—is significantly correlated with CCI—computed with MODIS band 1 and band 11 (R2 = 0:97, p < 0:001). GRVI and CCI performed similarly in extracting phenological metrics of the dates of the start and end of the season (EOS) when evaluated with gross primary production (GPP) measurements from eddy covariance towers. For EOS extraction of evergreen needleleaf forest, GRVI even overperformed solar-induced chlorophyll fluorescence which is seen as a direct proxy of plant photosynthesis. This study opens the door for GPP and photosynthetic phenology monitoring from a wide set of sensors with broadbands in the green and red spectral regions.


Introduction
Terrestrial gross primary production (GPP), the total amount of carbon dioxide (CO 2 ) assimilated by plants by photosynthesis, is the most variable and uncertain flux in the global carbon cycle [1]. Accurate characterization of the spatiotemporal dynamics of GPP is crucial for improving our understanding of the responses and feedbacks of vegetation to climate change.
Remote sensing provides a feasible way to track GPP dynamics at a large scale, but it is still not straightforward to achieve spatiotemporally continuous monitoring with high reliability [1]. The recently emerging satellite-recorded solar-induced chlorophyll fluorescence (SIF) opened a new avenue to directly track GPP, considering its mechanistical link with plant photosynthesis [2]. However, the temporal frequency and spatial resolution of current SIF satellite products are still very limited [3,4]. Therefore, primary productivity models based on simple light-use efficiency (LUE) considerations still prevail in GPP community. Primary productivity models represent GPP as the product of absorbed photosynthetically active radiation (APAR) and LUE [5]. This LUE paradigm provides a robust and simple framework for calculating GPP from satellite observations [6]. APAR is closely related with the normalized difference vegetation index (NDVI), a measure of vegetation green biomass [7]. The key to estimating GPP using an LUE model is the determination of LUE. LUE is commonly parameterized by identifying the maximum LUE for each biome and then downregulating based on stress conditions, expressed using climatic variables, e.g., vapor-pressure deficit, temperature, and soil-moisture concentration [8,9]. This parameterization involves meteorological data and preassigned maximum LUE, which are both uncertainty-prone. Therefore, LUE parameterization is one of the main sources of uncertainty in estimates of GPP [6,10,11].
The photochemical reflectance index (PRI) provides a promising way to determine LUE directly from satellite measurements [12]. The underlying mechanisms of PRI for representing LUE vary with the timescale: PRI represents both the diurnal activity of the xanthophyll cycle and seasonal changes in chlorophyll/carotenoid pigment ratios [13].
Although CCI provides a reliable tool to monitor GPP dynamics, it can be calculated from a very few sensors (e. g., MODIS), because most of the current running optical satellites lack the xanthophyll-sensitive narrowband. In addition, chlorophyll/carotenoid ratio retrieval is very challenging because of the high atmospheric contamination [27]. Many studies revealed atmospheric correction would not improve and even reduce the performance of CCI [17,18,28,29]. We hypothesized that the xanthophyll-sensitive narrowband MODIS band 11 (ranging from 526 to 536 nm) and the broadband in the green and MODIS band 4 (ranging from 545 to 565 nm) are highly correlated. If this hypothesis holds, the applicability of CCI would substantially improve, considering the ready availability of broadband green reflectances from existing multispectral sensors.
In fact, the normalized difference between broadband green and red reflectances has long been proposed [30] and was later named as green-red vegetation index (GRVI) [31]. However, its potential in capturing GPP dynamics may be underevaluated. GRVI has been reported to outperform the commonly used NDVI and enhanced vegetation index (EVI) in tracking plant photosynthesis [32,33].
This study is aimed at evaluating the validity of GRVI, as an alternative to CCI, for monitoring gross primary production and vegetation phenology. Specific scientific questions include the following: (1) whether the MODIS-derived GRVI and CCI highly correlate with each other and (2) how well does GRVI track the dynamics of GPP, especially for the end of season when a temporal lag between NDVI and GPP often occurs.

Vegetation Indices. CCI can be calculated from MODIS reflectance data as
where B 1 and B 11 are the surface reflectances in MODIS band 1 (620-670 nm) and band 11 (526-536 nm), respectively. The underlying spectroscopic mechanism of CCI is the foliar spectra in the green-red region, which is mainly controlled by pigmentation ( Figure 1). The reflectance for summer acclimated leaves is much higher at the green than the red band. Winter acclimated leaves, which often have a lower chlorophyll/carotenoid pigment ratio [34], are characterized by higher reflectance except around 531 nm (MODIS band 11), leading to lower CCI values (Equation (1)). CCI is therefore a reliable indicator of the temporal variation of the chlorophyll/carotenoid pigment ratio, which indicates photosynthetic downregulation during autumn and winter [7,20,21,24,26].
GRVI was established based on the contrasting reflectances at green and red bands. GRVI can be calculated from MODIS band 1 (B1) and band 4 (545-565 nm) (B4), as Although the GRVI and CCI are mechanistically different, they may be similar at satellite scale because the reflectance change caused by varied chlorophyll/carotenoid pigment ratio is quite subtle, compared with other confounding factors, e.g., atmospheric disturbance [27]. Therefore, we inferred that bands 4 and 11 may exhibit high consistence, considering their close spectral distance.
For comparison, the commonly used NDVI and the EVI, as well as the solar-induced chlorophyll fluorescence (SIF), were also employed in this study. NDVI and EVI were calculated from the MODIS reflectance data in bands B1 (620-670 nm), B2 (841-8756 nm), and B3 (459-479 nm), as SIF was extracted from GOSIF dataset with a resolution of 0.05°and 8 days. GOSIF used a machine learning method to predict Orbiting Carbon Observatory-2 observations with MODIS EVI and meteorological data (specifically, photosynthetically active radiation, vapor pressure deficit, and air temperature) as explanatory variables [35].
2.2. Data. In situ GPP time series from the FLUXNET-2015 dataset [36] were used as a benchmark to assess the performance of GRVI in tracking GPP dynamics. Daily GPP calculated from a nighttime method [37] was used. We selected 2 Journal of Remote Sensing 27 grasslands, 37 evergreen needleleaf forests, and 21 deciduous broadleaf forests. Detailed information for each site is presented in Table S1. Sites were selected based on the criteria: (1) they were in the Northern Hemisphere at latitudes > 30°, and (2) they had at least three years of concomitant in situ GPP and MODIS observations for 2001-2020. We used the daily MCD19A1 Version 6 [38] to calculate CCI, GRVI, NDVI, and EVI, for two main reasons: (1) it uses an adaptive time series and a spatial analysis, the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm, to derive atmospheric aerosol concentrations and surface reflectances without empirical assumptions to obtain a more accurate surface reflectance, and (2) reflectances at bands 1, 2, 3, 4, and 11 are all delivered, allowing the comparison between CCI, GRVI, NDVI, and EVI. All the bands were resampled to 1 km.
Data contaminated by clouds, snow, or a high aerosol optical depth were excluded. MCD19A1 data with viewing zenith angles > 40°were also excluded to minimize the anisotropic effects of the reflectance [19,25].

Statistical Analysis.
We firstly compared the reflectances at MODIS bands 11 and 4 and also compared the, respectively, derived CCI vs. GRVI over all the selected 115,916 reflectance samples.
We evaluated the performance of CCI, GRVI, NDVI, EVI, and SIF for tracking GPP and retrieving phenological metrics as compared to FLUXNET GPP. The maximumseparation (MS) method was adopted to extract the dates of the start of season (SOS) and end of season (EOS) [39]. MS is a variant of the threshold method and applies a moving window that estimates the ratio of observations that exceed a threshold (50%, in this study) before and after the central day. SOS/EOS is the day of the year when the difference between the ratios before and after the central day are minimum/maximal. Details for MS are provided in our previous study [39]. SOS/EOS was estimated from CCI, GRVI, NDVI, EVI, SIF, and FLUXNET GPP. SOS and EOS from GPP were seen as the references of starting and ending days of photosynthetically active season, respectively. Note that, before phenology extraction, the daily vegetation indices and GPP were resampled to 8-day temporal resolution with a maximum-value composting method to smooth the noise in the time series.
US-Oho, CA-TP3, and US-IB2, were selected as examples of deciduous broadleaf forest (DBF), evergreen needleleaf forest (ENF), and grassland (GRA), respectively, to illustrate the seasonal dynamics of GPP and the capacity of different vegetation indices to capture its dynamics ( Figure 3). During spring season, GPP and all the vegetation indices increased rapidly and nearly at the same time, as reflected by the narrow range of variation of SOS values especially for grassland (Figure 3(c)). Contrarily, the decrease during the senescence process was gradual. The satellite vegetation indices showed a systematic positive temporal lag difference (i.e., a delay in phenology) compared to ground GPP measurements. NDVI showed the highest positive lag in autumn phenology compared with GPP, implying that photosynthesis shuts down even when plants still have high green biomass. CCI and GRVI were very similar 3 Journal of Remote Sensing throughout the growing season, but GRVI lagged slightly compared with the CCI, especially for the forests (Figures 3(a) and 3(b)).
Compared with the GPP-derived SOS, those from the satellite indices all exhibit earlier estimates over DBF (Figure 4), with the largest and least bias happen for NDVI   Figure 4(a)) and SIF (Figure 4(g)), respectively. Over the three selected vegetation types, SOS for ENF showed the highest uncertainty. For example, R 2 between NDVI and GPP-derived SOS was 0.11. The best estimate was from SIF, with a R 2 of 0.58. For GRA, due to its simple structure, no obvious discrepancy was found for the SOS estimates     Journal of Remote Sensing from the five indices. Generally, GRVI performed comparable with CCI in SOS estimation, and they both show similar performance as EVI. However, closer inspection reveals that GRVI and CCI both obtained better accuracy than EVI for ENF, for which CCI was originally designed [7,20]. EOS dates estimated using NDVI were considerably delayed compared with the reference dates using GPP for all selected vegetation types (Figures 5(a)-5(c) for DBF, ENF, and GRA, respectively). The highest bias was observed in ENF sites (43.03 days) ( Figure 5(b)). Compared with NDVI, the lag of the estimated EOS from other indices was remarkably mitigated; e.g., the bias for ENF was reduced to 6.28 and 9.63 d for CCI and GRVI, respectively. GRVI performed comparably than CCI over all site-years, with only slightly more delayed estimates: 2.56 (20.15-17.59), 3.35 (9.63-6.28), and 0.62 (12.5-11.88) d for DBF, ENF, and GRA, respectively. As expected, SIF performed very well for every vegetation type. Yet, CCI and GRVI showed slightly better performances, in terms of R 2 and RMSE, than SIF for EOS retrieval over ENF.

Discussion
This study demonstrated that at the satellite scale the MODIS narrowband CCI can be reproduced with the broadband GRVI. GRVI performed satisfactorily in tracking GPP dynamic over DBF, ENF, and GRA vegetation types.
The physiological chlorophyll/carotenoid index (CCI) was originally designed for ENF [7] but also well represents GPP variation for other vegetation types, e.g., DBF [23] and GRA [25]. The advantage of CCI to track GPP mainly lies in the extraction of the end of the photosynthetically active season (EOS) [26]. For deciduous plants (including DBF and GRA), senescence follows a physiological timetable: (1) reallocation of foliar nutrients, (2) degradation of chlorophyll, (3) foliar coloration, and (4) foliar abscission. The reallocation of foliar nutrients is difficult to detect with remote sensing but is strongly associated with the degradation of chlorophyll (with decreasing chlorophyll/carotenoid ratio) [40]. As for evergreen plants, the absorbed light during late autumn cannot be fully exploited for carbon uptake because of environmental stress (e.g., low temperature) [41]. Such an imbalance between light energy supply and utilization activates a mechanism which adjusts leaf pigment pools and dissipates excess energy by increasing the carotenoid/ chlorophyll ratio [13,42]. Therefore, CCI is more capable of capture the downregulation of plant photosynthesis than other vegetation indices more sensitive to green biomass [26].
Previous studies revealed that NDVI was mainly determined by vegetation structure and EVI is relatively more controlled by leaf coloration [43]. This explains why the observed positive temporal lag phenomenon in EOS extraction was shorter for EVI (with less bias compared with GPP EOS) than for NDVI. Compared with NDVI and EVI, GRVI was capable of capturing the reflectance contrast between the green hump and red valley (see Figure 1) which are mainly determined by leaf chlorophyll content. Leaf chlorophyll content was found a good proxy of maximum photo-synthetic rate ( [44]). GRVI therefore performed better than NDVI and EVI to extract photosynthetic phenology, especially for EOS. SIF, a direct "proxy" of plant photosynthesis [2], was also employed as a reference to assess the performance of GRVI in GPP tracking. Our results reveal that GRVI derived from MODIS data performs comparable with SIF derived from GOSIF data. GOSIF has, however, a lower spatial resolution and revisit frequency (0.05°every 8-day temporal frequency) than MCD19A1 data (1 km spatial resolution and 1-day temporal frequency). The current existing SIF datasets are indeed characterized by low spatial resolution and temporal frequency, so the GRVI here defined provides a useful complementary or even alternative tool for high spatiotemporal GPP monitoring.
We found that MODIS bands 11 (narrow green band, sensitive to xanthophyll interconversion) and 4 (wide green band, signaling leaf chlorophyll content) were highly correlated at the seasonal scale (Figure 2(a)), so CCI (Equation (1)) can be safely reproduced by GRVI (Equation (2)). The direct comparison between CCI and GRVI (Figure 2(b)), the comparison with FLUXNET GPP, and the validation of phenology metrics supported our hypothesis of equivalence between CCI and GRVI for monitoring GPP and phenology. Considering that band 4 has higher spatial resolution (500 m) than band 11 (1 km), MODIS GRVI is capable of characterizing GPP seasonality at smaller granularity than CCI. Furthermore, the calculation of GRVI is easy to transfer to other sensors, because the green and red broadbands are readily available on most existing multispectral sensors. Note that the xanthophyll-sensitive narrowband (~531 nm) was removed in the VIIRS, the successor of MODIS [45]. The application of MODIS CCI, however, can be continuously persisted by GRVI from VIIRS.
GRVI can quantify the maximum leaf photosynthetic rate ( [32,33]); therefore, performs satisfactory in tracking GPP dynamics, even for the evergreen forests [32,33,46]. Other indices were also reported having the potential to capture the pigment variation and deserved further assessment. For example, Penuelas et al. [16] found that the red and blue bands were the best combination to capture the variation of chlorophyll/carotenoid pigment ratios (Normalized Difference Pigment Index ðNDPIÞ = ðred-blueÞ/ ðred + blueÞ and Structural Independent Pigment Index ðSIPIÞ = ðIR-blueÞ/ðIR-redÞ). The blue band, however, was highly sensitive to atmospheric distortion, so the feasibility of NDPI or SIPI calculated from satellite observations remains unknown. Red edge bands were also found highly sensitive to pigment content, especially to chlorophyll content [47]. Therefore, red edge-based vegetation indices, e. g., MERIS Terrestrial Chlorophyll Index (MTCI) and OLCI Terrestrial Chlorophyll Index (OTCI) [48], were also worth assessment in terms of photosynthetic phenology extraction.
This study demonstrated that MODIS CCI could be reproduced by the GRVI. GRVI can be directly transferred to other satellites considering the readily availability of broadband green and red bands. We will assess the application of GRVI to Landsat 8, Sentinel-2, and other commonly used optical satellites in future studies. Future studies should 7 Journal of Remote Sensing focus on assessing the feasibility of GRVI to retrieve chlorophyll/carotenoid ratio also as compared with MODIS CCI and ground data. Residual atmospheric contamination effects on CCI also require further attention [27]. Finally, we will also test other band combinations and investigate red-edge and blue bands from Sentinel-2 and Sentinel-3 in a future study.

Conclusions
This study demonstrated that MODIS-derived CCI, originally calculated with the support of the narrow xanthophyll-sensitive (~531 nm) band, could be safely reproduced by wide green and red bands, GRVI (R 2 = 0:97 ), for monitoring GPP and phenology. The comparison with FLUXNET GPP showed that the broadband GRVI performed comparably with the original CCI in tracking the dynamics of GPP and for extracting the dates of the start and end of the photosynthetically active season. GRVI provides a powerful and robust tool for monitoring the temporal variation of photosynthesis activity from a wide set of sensors with broadbands in the green and red spectral regions.

Data Availability
The data that support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
The authors declare that there is no conflict of interest regarding the publication of this article.