Greenness indices from digital cameras predict the timing and seasonal dynamics of canopy scale photosynthesis ‐

The proliferation of digital cameras co located with eddy covariance instrumentation provides new ‐ opportunities to better understand the relationship between canopy phenology and the seasonality of canopy photosynthesis. In this paper we analyze the abilities and limitations of canopy color metrics measured by digital repeat photography to track seasonal canopy development and photosynthesis, determine phenological transition dates, and estimate intra annual and interannual variability in canopy photosynthesis. We used 59 ‐ site years of camera imagery and net ecosystem exchange measurements from 17 towers spanning three plant ‐ functional types (deciduous broadleaf forest, evergreen needleleaf forest, and grassland/crops) to derive color indices and estimate gross primary productivity (GPP). GPP was strongly correlated with greenness derived from camera imagery in all three plant functional types. Specifically, the beginning of the photosynthetic period in deciduous broadleaf forest and grassland/crops and the end of the photosynthetic period in grassland/crops were both correlated with changes in greenness; changes in redness were correlated with the end of the photosynthetic period in deciduous broadleaf forest. However, it was not possible to accurately identify the beginning or ending of the photosynthetic period using camera greenness in evergreen needleleaf forest. At deciduous broadleaf sites, anomalies in integrated greenness and total GPP were significantly correlated up to 60 days after the mean onset date for the start of spring. More generally, results from this work demonstrate that digital repeat photography can be used to quantify both the duration of the photosynthetically active period as well as total GPP in deciduous broadleaf forest and grassland/crops, but that new and different approaches are required before comparable results can be achieved in evergreen needleleaf forest.


Introduction
Climate change impacts on vegetation phenology have been widely documented across a range of biomes and plant functional types (Richardson et al. 2013). In particular, long term records of leaf and fower phenology in temperate and boreal forest indicate that spring onset is occurring earlier ( While a large number of studies have identified widespread patterns of change, the impacts of changes in phenology on ecosystem function and feedbacks to the climate system remain poorly understood and quantified (Richardson et al. 2013). For example, multisite comparisons show that growing season length is positively correlated with net ecosystem productivity (NEP; Churkina et al. 2005, Baldocchi 2008), but spatial patterns observed across sites are not identical to temporal patterns at individual sites, which are driven primarily by interannual variability in weather (Richardson et al. 2010). Warmer springs and longer growing seasons have been shown to increase annual carbon uptake in boreal deciduous forest (Barr et al. 2004, mixed temperate forest (Dragoni et al. 2011), and evergreen needleleaf forest (Richardson et al. b, 2010. In subalpine forest, on the other hand, longer growing seasons can lead to lower NEP if warmer temperatures (Sacks et al. 2007) or shallower spring snowpacks (Hu et al. 2010) reduce soil moisture sufficiently to create drought conditions. Similarly, drought conditions in grassland can also shorten the growing season length, thereby lowering annual NEP (Flanagan and Adkinson 2011).
Because phenology is a key regulator of ecosystem function, substantial effort has recently been devoted to expanding networks that track seasonal vegetation dynamics (Morisette et al. 2009). Methods to monitor phenology fall into two broad categories: visual observations and remote sensing. Visual observations provide the oldest and longest running phenology records in existence (e.g., (250-500 m) that encompass considerable landscape heterogeneity within each pixel. An additional weakness is the relatively low temporal resolution of some space borne remote sensing instruments. While coarse spatial resolution sensors such as MODIS provide observations with repeat intervals of 1-2 days, moderate spatial resolution sensors such as Landsat provide a revisit frequency of 16 days, a relatively long interval for capturing rapid changes during seasonal transition periods. In both cases, persistent cloud cover can significantly reduce the frequency of usable observations, which can substantially decrease the utility of spaceborne remote sensing for observing and characterizing the timing of key phenological transitions.
Digital repeat photography, a form of near surface remote sensing, provides data at higher temporal frequency and finer spatial scale than satellite remote sensing . Specifically, digital repeat photography can provide imagery that is nearly continuous in time, rarely obscured by clouds, and robust to variation in illumination conditions . Exploiting this, color indices derived from digital Color indices derived from digital repeat photography have also been correlated with canopy photosynthesis in deciduous broadleaf forest (Richardson et al. 2007b, Ahrends et al. 2009, Mizunuma et al. 2012), grasslands (Migliavacca et al. 2011), and desert shrublands (Kurc and Benton 2010). However, each of these studies was limited to one or two sites and it is unclear how well results from these efforts generalize within and across PFTs at regional to continental scales. Further, a large proportion of previous studies have focused on temperate deciduous forest. Not only does the relationship between annual carbon exchange and the length of the carbon uptake period vary substantially across PFTs (e.g., Richardson et al. 2010), but relationships among camera based color metrics, phenology, and carbon exchange remain under studied in -ecosystems and PFTs outside of deciduous broadleaf forest (Richardson et al. 2013). Hence, there is a need for improved understanding regarding how canopy photosynthesis is linked to canopy phenology across and within PFTs, and by extension, the role of digital repeat photography for studying these relationships.
With these issues in mind, our objective in this study was to perform a systematic analysis of digital repeat photography as a tool for understanding the relationship between canopy phenology and canopy photosynthesis, both within and among multiple PFTs. To this end, the specific questions guiding this study were:

1.
Can camera derived color indices be used to monitor the seasonality of GPP within and across multiple -PFTs?

2.
How does the relationship between canopy phenology and GPP vary within and across PFTs?

3.
What is the relationship between dynamics in greenness measured from digital camera imagery and key phenophase transitions in different PFTs?

4.
Can interannual variation in annual GPP be estimated using camera derived color indices? - To address these questions, we used data from the PhenoCam network of co located cameras and eddy covariance towers to assess the relationship between canopy phenology and the seasonality of photosynthesis. Our study, conducted across a range of PFTs, provides the most comprehensive analysis of canopy development and photosynthesis using digital repeat photography to date, and provides useful new understanding regarding the ability of camera derived color indices to track the seasonality of GPP across space and time.

Study sites
The study spanned 13 geographically distinct research sites, including 17 fux towers in total (Table 1; Appendix A). We used all possible sites that were members of both the PhenoCam 22 and the AmeriFlux 23 or Canadian Carbon Program 24 networks. In addition, we included four towers, managed by the University of Illinois (UI), that were not members of either network. Each site was dominated by one of three PFTs: deciduous broadleaf forest (DBF), evergreen needleleaf forest (ENF), and grassland/crops (GRS; Table 1). The Groundhog site in Ontario is most accurately described as mixed ENF/DBF; here, we group it with ENF sites because conifer species are dominant. Together, measurements from these sites comprised 59 site years of concurrent fux and camera data, with 26, 11, and 22 site years in DBF, ENF, and GRS PFTs, respectively. Most sites had 2-5 years of data. One notable exception, however, is the ARM site in Oklahoma, where data were collected nearly continuously from 2003 to 2011. One of the UI sites featured a crop rotation from maize to soybean in the second year (out of two), which caused significant changes in the magnitude of carbon fuxes. To address this, we treat the two site years (2009 vs. 2010) as separate sites: UI Maize and UI Soy.
-  (Table 1). To minimize the impact of variation in scene illumination (e.g., clouds and aerosols), auto white/color balance was turned off, and exposure adjustment for each camera was set to automatic mode. Note, however, that Vaira was an exception in this regard. To correct for variability induced by auto color balancing at this site, we used a gray reference panel in the camera field of view (e.g., Jacobs et al. 2009). Images were either archived by the site investigator or automatically transferred to the PhenoCam server via file transfer protocol (FTP). Time series were first visually inspected for camera shifts and changes in field of view. Noting these changes, we processed the image archives to extract regions of interest (ROI) that encompassed all portions of the full canopy within the foreground (Fig. 1). At Vaira, the ROI was restricted to the grass portion of the image, excluding distant oak trees from analysis. To quantify canopy greenness, we calculated the green chromatic coordinate (GCC), which is widely used to monitor canopy development and where DN is the digital number and R, G, and B denote the red, green and blue channels, respectively. For completeness, we also calculated the Excess Green (ExG) index which has been shown to be less noisy than GCC in some coniferous canopies ). To characterize canopy coloration in fall, the red chromatic coordinate (RCC) was calculated using the same form as Eq. 1, substituting DNR in the numerator. , we calculated the 90th percentile of GCC, ExG, and RCC values for three day moving windows, yielding up to 122 observations each year. Only photos taken during daylight hours (0600-1800 local time) were included, and any images with underexposed ROIs (which we defined as <15% color saturation, or DN < 39, in any band) were excluded. We did not exclude photos due to poor weather conditions or snow, as the 90th percentile filter successfully removed these ). To eliminate any residual noise we removed GCC or ExG values that exceeded ±2 standard deviations of the mean within 27 day windows. To account for changes in camera settings or shifts in camera fields of view, GCC, RCC, and ExG values were manually screened and rescaled (as needed) to preserve a smooth and continuous time series at each site.
We used nonlinear least squares regression to fit logistic functions to GCC, RCC, and ExG time series, which were then used to estimate phenophase transition dates from DBF and GRS sites (e.g., Fisher and Mustard

2007, Richardson et al. 2009 a). For GRS sites, we used separate logistic functions in spring (s) and fall (f)
where t is the day of year and the remaining terms are empirically estimated coefficients. For DBF sites, we used the modified logistic function presented by Elmore et al. (2012), which includes an additional parameter (a2) that accounts for "summer greendown" that is widely observed in DBF greenness time series (Keenan et al. 2014) Note that in Eq. 4, a1 + b denotes the early summer maximum GCC, while the minimum summer GCC value preceding fall coloring is given by (b − a2 × t). Coefficients in Eqs. 3 and 4were estimated using the Levenberg-Marquardt method. Early analysis indicated that the logistic function provided a poor representation of GCC dynamics at many ENF sites; a separate method was needed to explore links between GCC and GPP seasonality in evergreen sites.
Hence, we calculated splines along GCC curves and examined correlations between dates at which a range of GCC thresholds (5-75% of seasonal amplitude, in 5% intervals) were reached, and dates at which a similar range of GPP thresholds were reached.

Eddy covariance data
To assess the ability of camera based indices to capture seasonal dynamics in carbon fuxes, we compared color indices with estimates of GPP derived from eddy covariance measurements.
To do this we used 30 minute non gap filled net ecosystem exchange (NEE) data to estimate GPP, except at the ---Harvard Forest and Morgan Monroe sites, where only hourly data were available. NEE was partitioned into GPP (micromoles of CO2 per square meter per second) using the Q10 method (Raich and Schlesinger 1992) where Rref is a scaling parameter, Q10 is the temperature sensitivity of ecosystem respiration (Reco), and Tref (=10°) is the base temperature where Reco = Rref. Friction velocity (u*) filtering was used to remove nocturnal NEE measurements when there was insufficient turbulence using site specific u* values. The Q10 function was estimated independently for every site year, yielding 30 minute estimates of R -eco and GPP. When available, we compared our GPP estimates with estimates provided by site investigators. Results from this comparison showed that the estimates were in close agreement (mean R 2 = 0.95; range: 0.91-0.98).
To make the GPP data comparable to the camera based color indices, we calculated the mean daily integrated --GPP (grams of carbon per square meter per day) across the three day periods over which the camera data were processed. In addition, we also calculated mean daytime instantaneous fux rates (calculated across all daytime hours, defined as photosynthetic photon fux density (PPFD) ≥ 5 μmol·m −2 ·s −1 ), as well as estimates of the light saturated rate of photosynthesis ( -Amax, measured as micromoles of CO2 per square meter per day), which was derived by fitting a Michaelis Menten light response function to the high frequency (hourly or half --hourly) fux measurements. The use of these alternative metrics did not change our interpretation (see Results). To allow comparison at annual time scales, we calculated annual GPP sums, using the same Q10 method as above, but including gap filled NEE. When gap filled NEE data were not provided by site -investigators, we used an online tool that implements standardized gap filling methods (Reichstein et al. 2005). 25 To evaluate GCC as a predictor of photosynthesis, daily GPP was regressed against three day GCC for each tower site. We also regressed the mean daytime instantaneous fux rate (GPP30; averaged over equivalent three day periods) against GCC, which allowed us to assess this relationship independent of day length.
-Goodness of fit was based on the coefficient of determination ( --R 2 ), calculated using linear and quadratic functions at a significance level of 0.05.
A key goal of this analysis was to assess how well dynamics in GCC capture changes in photosynthetic activity corresponding to phenological transitions. For example, one question we examined was, "Does start of spring, estimated by GCC, correspond to the first day of photosynthesis (GPP > 0 g C·m −2 ·d −1 ) in spring?" To compare relative photosynthetic capacity across sites, we fit smoothing splines to the daily GPP time series for each of the six DBF sites and calculated the percentage of maximum annual fux (maximum daily GPP within a given year = 100%) at 1% intervals along the estimated splines. These data were then pooled, providing a composite DBF data set of 19 site years. Using phenophase transition dates (start of spring, middle of spring, middle of fall, end of fall) extracted from the GCC and RCC time series, we performed geometric mean regression between camera derived dates and a range of fux amplitudes (1-90%). Goodness of fit was evaluated using --the coefficient of determination and the slope of the regression. Bias was quantified using the mean deviation, and accuracy was evaluated using the root mean square deviation (RMSD) between transition dates estimated from GCC data and transition dates estimated from GPP data.
To explore these relationships at the GRS sites, we pooled data from the four UI sites and performed a parallel analysis. The ARM site in Oklahoma was excluded because both the fux data and the camera data included mixtures of differing phenological patterns associated with multiple crop cycles. We also excluded the Vaira site because it is characterized by asynchronous seasonality (winter active vs. summer active elsewhere) relative to the rest of the sites in our analysis of transition dates. To compare the timing of maximum greenness (GCC90%) and carbon fux (GPP90%), we determined the dates when each metric reached 90% of the maximum annual value at each site using only complete site years. - Because the rates of spring increase and fall decrease in daily GPP or GCC can vary between years (see Richardson et al. 2010), dates corresponding to the start and end of the growing season may not fully characterize patterns of interannual variability in phenology.
To assess this, we tested the hypothesis that during the spring or fall transition periods time integrated GCC values provide more information about anomalies in GPP than start of season or end of season dates estimated from GCC time series. To do this, we ---first re scaled the GCC and GPP data to account for differences across sites in the magnitude of carbon fuxes and canopy greenness. This provided normalized time series of daily GPP and GCC, both on a scale from 0 to 1.
We then fit splines to the normalized GPP and GCC values over 60 day periods following the earliest start of spring and preceding the latest end of fall, and calculated the integral under each spline curve using numerical approximation. These integrals were then converted to anomalies relative to each site level mean and used to calculate linear correlations between integrated GCC anomalies and integrated GPP anomalies. To determine whether integrated GCC values provide greater explanatory power than discrete dates such as the start of spring, we compared these results with linear correlations between phenophase transition date anomalies and integrated GPP anomalies. Lastly, we tested whether spring and fall greenness anomalies were correlated with integrated annual GPP anomalies via multiple linear regression, using spring and fall normalized integrated GCC anomalies as independent variables and annual GPP anomalies as the response variable.

Canopy development and photosynthesis; patterns among plant functional types
Time series of GCC and daily GPP ( Fig. 2; Appendix B) demonstrate broadly consistent relationships within each of the three PFTs, with some notable exceptions. DBF and GRS sites exhibited clear seasonality in both GCC and GPP, with high values during the photosynthetically active season and low values during the inactive season.
GRS sites exhibited shorter but well defined growing seasons compared to those in DBF ( - Fig. 2c). In ENF sites, the annual cycle in GCC was roughly sinusoidal, with a relatively short period of minimum values in winter (Fig.   2b). Relationships between GCC and GPP in both the active and dormant seasons were phase shifted, with spring increases in GCC preceding those in GPP, and autumn decreases in GCC lagging behind GPP.  throughout the year. There was also a wide range in GPP among PFTs owing to differences in ecosystem productivity arising from factors such as species composition, leaf area, and local climate. Across all sites and PFTs, daily GPP values showed strong seasonal patterns, but there was substantial day to day variation caused -by changes in short term environmental conditions (e.g., clouds, vapor pressure deficits, and soil moisture) that limit short term productivity, and by extension, decrease correlation between GPP and GCC on short (i.e., hours to days) timescales. Canopy development and photosynthesis; patterns within plant functional types DBF sites exhibited two primary modes of variation in GCC during the photosynthetically active season. First, over the course of two or three weeks in late spring, GCC tended to exhibit a distinct late spring "green peak" that was not observed in either ENF or GRS. Second, following this peak, GCC tended to gradually decline over roughly three months, leading to a decrease in GCC of ∼30% relative to the seasonal amplitude. At the onset of leaf coloration, GCC tended to decrease rapidly, leading into the annual winter minimum. Daily GPP, by contrast, increased more slowly throughout the spring, reaching its maximum value 2-4 weeks after the GCC peak. And, whereas GCC remained high during the summer months, daily GPP tended to decline almost immediately after its peak, well in advance of the fall decline in GCC.
As we noted previously, daily GPP exhibited substantial day to day variability in all PFTs. At the Missouri Ozarks -site in 2007, however, daily GPP decreased sharply in July, nearly two months before the autumn decrease in GCC, likely in response to moisture stress (Yang et al. 2010). Otherwise, covariance between daily GPP and GCC for DBF sites was generally strong overall (R 2 = 0.50-0.79; Table 2; Fig. 3a; Appendix C) and tended to be linear at lower values of GCC. At higher values of GCC, however, there was little or no relationship between daily GPP and GCC for most DBF sites, which refects the fact that daily GPP during midsummer is controlled by day to -day variation in weather that does not affect canopy greenness on short time scales. Correlations between daily GPP and GCC were comparable with those between GCC and GPP30( Table 2), indicating that GCC-GPP relationships are robust and independent of seasonal changes in day length.  Table 2 for coefficients of determination). All years of data are featured in each subplot. ENF sites were characterized by unique patterns of seasonality in GCC and GPP. Most notably, the period associated with minimum GCC values during winter dormancy was short lived. At most ENF sites GCC continued to decline into early winter, even when daily GPP was near zero, before rising again in late winter well in advance of the spring onset of photosynthesis. This pattern was not observed at the Wind River site, which was photosynthetically active throughout almost the whole year (Appendix B). Among all ENF sites, the summertime peak in GCC occurred close to the peak in daily GPP. Overall, correlations between daily GPP and GCC were almost as strong (R 2 = 0.53-0.76; Table 2; Appendix C) as those for DBF sites. As with DBF, correlation between GCC and GPP30 were comparable with those between GCC and daily GPP ( Table 2).
For all but one GRS site, correlations between daily GPP and GCC were high (R 2 = 0.80-0.90; Table 2; Appendix C), and the relationship was linear. Similar to the ENF sites, GCC at GRS sites exhibited a short summer plateau.
At the UI Switchgrass and UI Prairie sites, GCC was modestly phase shifted, with GCC leading daily GPP in spring and lagging daily GPP in fall. Covariance between GPP and GCC at the ARM Oklahoma site, where the growing season extends well beyond that at most other sites, was substantially higher between GCC and GPP30than between GCC and daily GPP ( Table 2).
For DBF and GRS, relationships between GPP and ExG were similar to those observed for GCC (Table 2). At ENF sites, correlations between ExG and GPP were marginally higher than those between GCC and GPP, but the magnitude of these differences was site specific. At Wind River, in particular, ExG accounted for ∼15% more variance in daily GPP than GCC because of the greater stability (less day to day noise) in ExG. Similar (but less -pronounced) increases were also observed at Chibougamau, Howland, and Niwot.

Camera and fux based phenophase transitions -
Using a combination of greenness (GCC) and redness (RCC) indices, digital repeat photography facilitated accurate determinations of the start and end of the photosynthetic period for DBF and GRS. In ENF sites, however, the lack of a discernible winter baseline prevented accurate estimation of the start and end of canopy photosynthesis. In ENF and GRS, GCC provided a relatively accurate estimation of the date of maximum photosynthesis; however, the relationship in ENF was statistically insignificant. In the section below we elaborate on these themes, discussing four camera based phenology metrics (start of spring, middle of spring, middle of fall, and end of fall) and their relationship with the seasonality of GPP.
At DBF sites, camera derived spring and fall phenophase transition dates successfully captured spatiotemporal variability in the beginning and end of the photosynthetic period. Start of spring, estimated using Eq. 4 fit to the GCC time series, was most highly correlated with the day of year corresponding to when fux amplitudes were between 24% and 30% of maximum GPP (R 2 = 0.62; n = 17). Mean deviation (MD) and RMSD between start of spring from GCC and GPP was smallest at 20% and 24% of GPP amplitudes, respectively (Fig. 4a).
Results were even stronger (Fig. 4b) for the "middle of spring" (the date on which 50% of the seasonal amplitude in GCC was reached), which corresponded to 30-40% of the spring amplitude in GPP (R 2 = 0.82). In contrast, GCC was a relatively poor predictor of the date of maximum photosynthesis (GPP90%) in DBF sites, with the date of GCC90% consistently preceding the date of GPP90% by more than three weeks, on average. Note, however, that the magnitude of this bias was disproportionately infuenced by one site year (Harvard Forest in -2010), in which a late summer increase in GPP delayed the 90% threshold significantly (Fig. 5a).  Correlations between the date at which 50% of the seasonal amplitudes in GCC and GPP were reached in fall was relatively weak (R 2 = 0.43; Fig. 4d). Similarly, correspondence between GCC and GPP derived end of fall -dates was also weak. Canopy redness (RCC), rather than greenness, provided the best indicator of the end of the photosynthetically active period, with the date of peak RCC strongly correlated to the date when GPP amplitude reached 14% (R 2 = 0.69; Fig. 4c).
At GRS sites, GCC provided more information about seasonal dynamics in photosynthesis during spring than in fall. GCC was a good indicator of the beginning of the photosynthetically active period, with high correlation between both the start and middle of spring derived from GCC time series and the date corresponding to a wide range of amplitudes in spring GPP (Fig. 6a, b). Relative to GPP90%, GCC90% was less biased at GRS sites than at DBF sites (Fig. 5c). Similar to patterns observed in spring, the timing of both the middle and end of fall from GCC showed significant (but lower relative to spring) correlations across a broad range of GPP amplitudes (Fig.   6c). At ENF sites, GCC typically started to increase prior to the onset of the growing season, when GPP was still zero, and continued to decrease late in the year after GPP had returned to zero. Thus, at both the start and end of the growing season, significant variations in GCC occur that are not associated with dynamics in GPP.
Indeed, correlations between the timing of changes in GCC and GPP across a wide range of spring and fall amplitude thresholds (5-75% of seasonal amplitude, in 5% intervals) were statistically insignificant at P ≤ 0.05.
It would appear, therefore, that camera based GCC time series cannot be used to predict the beginning or end of the photosynthetically active period for ENF sites. It is worth noting, however, that GCC did provide a rough indication of the date of maximum GPP. While the correlation between the dates on which 90% of the spring amplitudes in GCC and GPP were reached was statistically insignificant (R 2 = 0.32, P = 0.11), the mean bias (across all site years) was less than one day (0.3 d ± 10 d). -

Integrated GCC and GPP
In the final element of our analysis, we investigated whether spring and fall time integrated sums of daily GCC provide additional or complementary information regarding interannual variation in GPP relative to phenophase transition dates estimated from the GCC time series. To do this, we first focused on the Barlett Forest site and calculated springtime integrated daily GPP and GCC from 2006 to 2012 (Fig. 7). Starting on day of year (DOY) 115 (selected to precede the earliest observed green up day, DOY 118), we integrated both GCC and GPP over successively longer time segments at five day increments (e.g., DOY 115-120, 115-125, etc.).
-(DOY 1 is 1 January.) Results from this analysis showed that springtime integrated GCC anomalies were strongly and significantly correlated with integrated GPP anomalies for up to 30 days (R 2 = 0. 56  We then extended this analysis to include all DBF and three of the four Illinois GRS sites (we excluded the UI Maize and UI Soy sites, for which only a single year of data was available). For the DBF sites, we found moderate correlation (as high as R 2 = 0.49; n = 19 after 30 days) between normalized GCC integral anomalies and normalized GPP integral anomalies up to 60 days after green up ( - Fig. 8). Over this period, anomalies of up to 158 g C/m 2 (Harvard), or ∼8% of the annual total GPP, were observed. Correlations based on time integrals extending beyond 60 days after the earliest green up were not statistically significant. In contrast to results at -Bartlett Forest, start and middle of spring transition date anomalies were more highly correlated with --normalized GPP integral anomalies (R 2 = 0.71 and 0.60 at 20 and 30 days, respectively). In fall, correlations between GCC integral anomalies and GPP anomalies were not statistically significant, whereas end of fall -transition date anomalies were weakly correlated with normalized GPP integral anomalies (R 2 = 0.  At GRS sites we found strong correlation between normalized GCC anomalies and normalized GPP anomalies up to 60 days after the start of spring (R 2 = 0.97; n = 6; Fig. 9) and during the period 20-50 days preceding the end of the growing season (R 2 = 0.83). While these results are promising, it is important to note that the sample size is small (n = 6) and each tower is represented by only two site years. As at DBF sites, information related to interannual variation in fall GPP from time integrated GCC values was comparable to that provided from transition dates, but provided less information related to spring GPP variations. For example, correlation of start of spring with spring GPP anomalies was lower than that for GCC anomalies ( --R 2 = 0.85), while correlation of end of fall with fall GPP anomalies ( --R 2 = 0.81) was equivalent to that of the integrated GCC anomaly.

Figure 9
Open in figure viewer PowerPoint Regression of normalized GPP integrated sums (dimensionless) vs. GCC integrated sums (dimensionless) during the first 60 days following green up (in 10 day increments) for the GRS sites, UI --Miscanthus (black circles), UI Prairie (white circles) and UI Switchgrass (gray circles).

Canopy development and photosynthesis
Results from this study demonstrate that canopy greenness is correlated with rates of photosynthesis in both forest and grassland. Consistent with results from previous studies, canopy greenness and GPP were correlated across DBF sites (Richardson et al. 2007a, Ahrends et al. 2009). For reasons that are unclear, we found a stronger relationship between greenness and photosynthesis in grassland than Migliavacca et al. (2011). At ENF sites, our results are consistent with those obtained by  and showed moderate to strong correlation between canopy greenness and GPP across all of the sites we examined. This was particularly true for ExG, suggesting that camera based modeling of GPP in ENF should be based on this index.
-Specifically, ExG was less sensitive than GCC to variation in illumination conditions. Thus, ExG appears to minimize the impact of shadows, which are prominent and highly variable in conifer canopies.
Our analysis also revealed several limitations of canopy greenness as a predictor of GPP. For example, there was a pronounced peak in GCC at the end of spring in DBF sites (also noted by Mizunuma et al. 2012) that preceded the peak in GPP by several weeks. Peak GCC is caused by seasonal variation in foliage pigments (e.g., Sims and Gamon 2002) and is accentuated by the oblique viewing angle used by the cameras in this study (Keenan et al. 2014). As a result, GCC90% tended to occur several weeks before GPP90%. Data from the Missouri Ozarks site also demonstrated limitations of GCC during drought conditions when photosynthesis was reduced by moisture stress, but canopy color was unaffected. As a result, GPP and GCC became decoupled as GPP dropped rapidly while GCC remained high (Appendix B). Even though ENF sites exhibited well defined seasonality in greenness, GCC was only weakly correlated to GPP at these sites. Conifers undergo seasonal changes in chlorophyll content, with winter minima ∼40% lower than summer maxima (Billow et al. 1994, Ottander et al. 1995. Hence, seasonal variation in chlorophyll concentrations at sites with long winters (Chibougamau, Groundhog, Niwot) may be driving observed patterns in canopy greenness, even during the non photosynthetic period ( - Fig. 2; Appendix C).

Phenophase transitions and integrated GCC-GPP
Start of spring and end of fall, determined based on GCC and RCC, provided biased estimates for the beginning and cessation of the photosynthetically active period (i.e., GPP > 0 g C·m −2 ·d −1 ). In deciduous broadleaf forest and grassland/crops sites, the MD and RMSD for the start of spring were lowest for GPP values between 20% and 26% of the spring amplitude, while for end of fall, deviations were lowest for GPP at 14-16% of the fall amplitude. Local maxima in the change in curvature rate, which is used to identify the start of spring and end of fall (Zhang et al. 2003), occurs above wintertime minimum values, when GCC reach ∼10% and 90% of the amplitude of Eqs. 3 and 4, respectively. In DBF, an additional source of disagreement between the timing of GCC and GPP is early season photosynthesis from sub dominant evergreen trees, which can increase -ecosystem GPP well before leaf emergence in deciduous trees. Likewise, leaf level studies indicate long periods (50-80 days) between green up and maximum -photosynthesis (Reich et al. 1991, Bassow and Bazzaz 1998, Morecroft et al. 2003. Thus, it is perhaps not surprising that we found that changes in GCC tend to lead changes in GPP in both spring and autumn in DBF.
A particularly important conclusion from this work is that repeat digital photography not only allows us to identify when photosynthesis begins and ends, but also helps us estimate how much of an impact phenological variability has on seasonal and annual carbon budgets. Using an independent measure of canopy phenology, we showed how changes in the timing of green leaf phenology in the spring and fall affects cumulative photosynthesis. Among DBF sites, we also found significant correlation between combined spring-fall GCC anomalies and anomalies in annual GPP.

Impacts and future work
By examining relationships between camera derived metrics of greenness and GPP across a large set of sites spanning multiple years and three plant functional types, this research provides an improved foundation for using digital repeat photography to model the impact of phenological dynamics on the carbon cycle of terrestrial ecosystems. Key contributions of this study are (1) demonstration of relatively general relationships between GPP and GCC, and (2) quantification of spatiotemporal variability in canopy development and GPP among and across three major PFTs. More generally, results from this study highlight the role that cameras can play in refining and calibrating phenological subroutines in Earth System models, which vary widely in their representation of green leaf phenology (e.g., Richardson et al. 2012). The Community Land Model, for instance, includes seven PFTs (Bonan et al. 2002), four of which were represented in our study: deciduous broadleaf forest (DBF), coniferous evergreen forest (ENF), grasses (GRS), and crops (GRS). Our study did not include broadleaf evergreen forest, and deciduous and evergreen shrubs, and we are not aware of any studies that have compared camera based phenology and carbon fuxes in broadleaf evergreen forest (but see Doughty and Goulden 2008 for radiometry based phenology). However, given the major role of humid tropical forest in the global carbon cycle, there is a clear need for camera based studies in this biome.
-Although our study was focused on canopy scale phenology, digital repeat photography also has significant potential as a tool for bridging the gap between canopy to landscape scale processes and organismal level --observations of leafing and fowering phenology. Digital repeat photography can also play an essential role in calibrate models of terrestrial phenology are likely to increase. Exploiting this, future work will explore how such camera networks can be used to characterize spatiotemporal variability in phenology and determine the environmental drivers (e.g., temperature, precipitation, photoperiod, snow cover) that regulate canopy development and senescence at regional to continental scales.

Conclusions
In this study, we demonstrate the strengths and limitations of camera based canopy greenness for monitoring the phenology of photosynthesis in three PFTs: deciduous broadleaf forest, evergreen needleleaf forest and grassland/crops. We encountered key differences among PFTs in the relationship between canopy development, expressed as greenness, and the seasonality of carbon fuxes. These differences were also evident in the detection of discrete phenophase transitions. Canopy greenness proved effective at detecting the beginning and end of the photosynthetically active period in GRS sites. In DBF sites, greenness was effective for detecting the beginning of the photosynthetic period, whereas redness was most effective for detecting the end. A key finding of this study was that integrated GCC was significantly correlated with total GPP during the first 30-60 days following green up, in both DBF and GRS. In some cases, integrated GCC was a better predictor of summed spring/fall GPP than discrete transitions dates. Further, in DBF there was a moderate correlation between combined spring-fall GCC anomalies and the annual GPP integral anomalies, indicating significant seasonal control of shifts in phenology on ecosystem productivity. Camera data thus provide a valuable and independent means by which ecosystem scale phenology can be characterized (cf. phenological metrics derived from CO2 fuxes themselves, as in Richardson et al. 2010). Finally, our results suggest that digital repeat photography may be used to estimate interannual variability in GPP resulting from phenological variability with greater accuracy than many existing ecosystem process models provide (Keenan et al. 2012).