The relationship between carbon dioxide uptake and canopy colour from two camera systems in a deciduous forest in southern England


Correspondence author.


  1. Carbon dioxide flux measurements using the eddy covariance (EC) methodology have helped researchers to develop models of ecosystem carbon balance. However, making reliable predictions of carbon fluxes is not straightforward due to phenological changes and possible abiotic/biotic stresses that profoundly influence tree functioning.
  2. To assess the influence of canopy phenological state on CO2 flux, we installed two different digital camera systems at different viewing angles (an outdoor webcam with a near-horizontal view and a commercial ‘fisheye’ digital camera with a downward view) on a flux measurement tower in southern England and tracked the visual change of the canopy in this oak-dominated (Quercus robur L.) forest over two growing seasons.
  3. Changes in the setting of the camera's white balance substantially affected the quality of the webcam images. However, the timing of the onset of greening and senescence was, nevertheless, detectable for the individual trees as well as the overall canopy for both years. The greening-up date assessed from the downward images from a hemispherical lens was ∼5 days earlier than from the horizontal-view images, because of ground vegetation development (not visible in the horizontal view).
  4. The effects of a late air frost in 2010 were evident in the canopy greenness, and these led to reductions in daily gross primary productivity (GPP). The cameras recorded differences between individual tree crowns, showing their different responses to the late frost.
  5. A major new finding from this work is the strong relationship between GPP and Hue, which was stronger than the relationship between GPP and NDVI.


Atmospheric CO2 concentration is strongly influenced by the seasonality of vegetation (Keeling, Chin & Whorf 1996). It is increasingly important to understand how forests will react to climate change. CO2 flux measurements using a global network of eddy covariance (EC) systems have provided us with knowledge of the carbon balance of forests. Although the seasonal patterns of carbon flux can be explained in general terms by vegetation physiology, soil and environmental factors, it is harder to understand the interannual variability due to annual variations in the duration of leafing and in the amount of leaves, caused by weather conditions and occasional but often severe biotic stresses that influence the physiological status of trees. Baldocchi et al. (2005) proposed that cameras should be installed to record the state of canopies. The relatively low cost of consumer-level digital cameras has made them an attractive proposition for ‘near-surface remote sensing’ due to the high temporal and spatial resolutions (Richardson et al. 2007, 2009; Wingate et al. 2008; Morisette et al. 2009). The camera network has expanded in North America (Sonnentag et al. 2012), Asia (Nagai et al. 2011b) and Europe (Wingate et al. 2011), and there is now a need to explore the relationships between the recorded images and the productivity and carbon fluxes of the forest.

The timing of biological events and the length of the growing season influence the productivity of all forest ecosystems, in particular deciduous forests (Baldocchi 2008). The study of the timing of biological events such as leaf appearance, flowering and leaf shedding (abscission) is termed phenology. Phenological observations are one of the simplest ways in which the response of species to global warming can be monitored (Menzel 2002). Phenological events are in themselves easy to record, but phenological work has traditionally required field visits with frequent and regular observations of the same organisms at the same location (Sparks, Jeffree & Jeffree 2000). The recent approach of using sensors from satellites provides quantitative data (Myneni et al. 1997; Zhou et al. 2001; Stockli & Vidale 2004). However, the time resolution is not high, and observations are often contaminated by clouds. Ground-based cameras allow us to record both qualitative (traditional) and quantitative measures of phenology: images may be assessed visually (Ahrends et al. 2008) or the time course of vegetation indices can be analysed quantitatively from digital values of red, green and blue (RGB) taken from image files (Richardson et al. 2007). The low cost of image capture has encouraged widespread use of cameras in ecological investigations (Ide & Oguma 2010). This also opens up new possibilities: huge sets of data from numerous public outdoor webcams may in the future be accessed to survey phenology at a large geographical scale (Jacobs et al. 2009; Graham et al. 2010). Vegetation indices based on the spectral reflectance such as the normalised difference vegetation index (NDVI) allow us to quantify phenology using remotely sensed data. However, analogous indices obtained using well-established colour analysis of data are clearly also possible. While these colours indices (Table 1) should not be confused with those based on the wavebands of electromagnetic radiation (e.g. NDVI), they may, nevertheless, provide a useful additional source of information.

Table 1. Vegetation indices used and/or discussed in this study
Vegetation indexFormulationReference
  1. The terms Sred, Sgreen and Sblue are chromatic coordinates sensu Gillespie, Kahle and Walker (1987); nred, ngreen and nblue are the digital numbers of red, green and blue (RGB), respectively; nred/255, ngreen/255, ngreen/255; Imax is the maximum value of r, g and b, and Imin is minimum value of r, g and b.

Strength of red
display math
Gillespie, Kahle and Walker (1987)
Strength of green
display math
Strength of blue
display math
display math
Joblove & Greenberg (1978)
Green Excess Index
display math
Woebbecke et al. (1995)

Early studies used manual or video cameras to record seasonal changes in leaf canopies (e.g. Nakakita 1990; Kawashima & Nakatani 1998). In recent years, however, time-lapse camera systems have enabled quantitative analysis of sequential images for phenological monitoring. Two distinct technologies have been used to provide ‘near-surface remote sensing’ of canopy phenology: (i) real-time webcams connected to a computer network (e.g. Richardson et al. 2007, 2009; Jacobs et al. 2008, Graham et al. 2010), and (ii) commercial digital cameras (e.g. Ahrends et al. et al. 2008, 2009; Maeda et al. 2008). Here, we aim to investigate the suitability and use of different camera systems. In the case of images from cameras, two vegetation indices are often used: the strength of green relative to the total of RGB is used to detect foliage phenology (e.g. Ahrends et al. 2008), while the Green Excess Index (Woebbecke et al. 1995) has also been useful (GEI or 2G_RBi, for example, Richardson et al. 2009; Migliavacca et al. 2011). Conceptually, this ‘strength’ is a representation of colour on three axes, representing RGB referred to as a ‘chromatic coordinates’ (Gillespie, Kahle & Walker 1987). The start of the growing season derived from these indices coincided well with the start of the carbon uptake period measured by EC, defined as when gross primary productivity (GPP) is >0 (Ahrends et al. 2008). However, GPP is influenced not only by the state of the canopy but also by the temperature, humidity and photosynthetic capacity, and a model considering these factors is needed to decide which vegetation index is a better proxy of GPP (Migliavacca et al. 2011). Graham et al. (2009), Mizunuma et al. (2011) and, very recently, Saitoh et al. (2012) have suggested using an alternative colour system based on the Hue, Saturation and Light aspects of the image (HSL), which is a three-dimensional coordinate system devised by Smith (1978). Hue is an angular scale, 0–360°, expressing colours as perceived by the average human eye, and thus, it is not directly related to spectral distribution. In a previous study, we found Hue detected leaf flushing, performing rather better than other vegetation indices (Mizunuma et al. 2011). This representation of colours was advanced also by Liu & Moore (1990) because of its ability to deal with colour in images containing both sunlit and shaded regions. As colour of leaves depends on their elemental content (e.g. Ollinger et al. 2008; Asner, Martin & Bin Suhaili 2012; Hufkens et al. 2012b), vegetation indices derived from digital images have the potential to allow detection of nutrient deficiencies and effects of biotic and abiotic stresses on canopy processes.

Although cameras are not calibrated instruments, the results of recent studies using eleven different cameras in a deciduous forest in North America showed that the choice of cameras and image formats made small differences in the detection of phenological change (Sonnentag et al. 2012). However, viewing angles of cameras are often horizontal, unlike the viewing angles of satellite sensors (Hufkens et al. 2012a). In this case, the lack of information on understorey and ground vegetation phenology may cause problems in matching phenological timings with the satellite-based observations. Hemispherical (‘fisheye’) lenses were pioneered 50 years ago (Anderson 1964) to measure canopy structure and the penetration of photosynthetically active radiation rather than phenology (see Maeda et al. 2008; Nagai et al. 2011a, 2012). Hemispherical lenses mounted on downward-viewing cameras give the user choice in selection of regions of interest within the circular image.

The first aim of this study is to compare the phenological data recorded by two different camera systems mounted above an oak woodland: (i) a networked digital webcam with a standard lens and a horizontal view and (ii) a commercial digital camera with a hemispherical lens facing downward. The second aim is to investigate any relationship between the colour of the canopy and the recorded fluxes of carbon dioxide.

Materials and methods

Study site

The Straits Inclosure (51°7′ N, 0°51′ W) is located within the Alice Holt Research Forest at an altitude of 80 m and is approximately 60 km south-west of London. The measurement site is affiliated to the FLUXNET/CarboEurope IP network and is included in several other monitoring and research projects. It has a long-term forest health observation plot within the European Network programme (ICP Forests), and it is a UK Environmental Change Network (ECN) site ( The surrounding landscape consists of mixed lowland woodland with arable and pasture agricultural land. The site is managed by Forest Research, an agency of the UK Government's Forestry Commission.

Early maps show that the western part of the Straits Inclosure was partly wooded and the east was under agricultural management in 1787, but the whole area was completely planted with oak in the 1820s. It was subsequently re-planted during the 1930s and is now a relatively homogeneous 90-ha woodland block, managed as a commercial lowland oak forest. The main tree species is Quercus robur L., and other species, mostly ash (Fraxinus excelsior L.) but including Qpetraea (Mattuschka) Liebl. and Q. cerris L., make up c. 10% of the tree cover. The understorey is dominated by hazel (Corylus avellana L.) and hawthorn (Crataegus monogyna Jacq.) (Pitman & Broadmeadow 2001). The soil is a surface-water gley (Pyatt 1982) with a depth of 80 cm to the C horizon of the Cretaceous clay (soil series Denchworth and Wykeham). The pH is 4·6 and 4·8 in the organic and mineral horizons, respectively. Mean top height and diameter at breast height (DBH) were 20·5 m and 29·0 cm, respectively, in 1999 at a density of 495 trees per hectare resulting in an estimated basal area of 23·4 m2 ha−1.

The climate of the region is oceanic, relatively mild and wet. The UK Meteorological Office-affiliated weather station located at the Research Station, approximately 1·8 km from the monitoring site, provides long-term data. In the 30-year period of 1961 through 1990, the mean annual precipitation was 779 mm and the mean annual air temperature was 9·5°C.

Microclimate measurements and canopy state

Environmental measurements recorded at the flux site include the following: wind speed and direction (model WAA151; Vaisala, Helsinki, Finland), wet and dry bulb air temperature (model DTS-5; ELE International, Loveland, CO, USA), above- and below-canopy solar irradiance (tube solarimeter; Delta-T Devices, Cambridge, UK), global solar radiation (model CM2; Kipp & Zonen B.V., Delft, the Netherlands) and net radiation (model DRN-301; ELE International). The fraction of absorbed photosynthetically active radiation (FPAR) was estimated from the midday average solar irradiance above (Iabove) and below the canopy (Ibelow). To take account of the different absorption of solar and PAR wavebands by the canopy, the formula given by Monteith (1993) was used:

display math(eqn 1)

where the exponent 1·35 accounts for the mean effect of the different absorptivities in the PAR waveband and for total solar radiation. The attenuation of radiation in the canopy follows an exponential function (Jones 1992).

display math(eqn 2)

where PAI is Plant Area Index, and k is an attenuation coefficient of total solar radiation, so PAI can be evaluated as:

display math(eqn 3)

Stem area was estimated using the mean of the midday average of attenuation of solar irradiance between January and March (Iabove0 and Ibelow0, respectively), and Leaf Area Index (LAI) was obtained by subtracting stem area from PAI.

display math(eqn 4)

The attenuation coefficient k for total solar radiation was obtained from the attenuation of PAR using previous in situ measurements (Broadmeadow et al. 2000, kPAR=0·4) using the expression:

display math(eqn 5)

LAI was also assessed using litterfall traps located within a 20-m radius of the flux tower. Canopy litterfall (leaves, twigs, frass, acorns and residual fraction) was collected in three cone-shaped traps held above the ground vegetation at a height of 1·5 m, each with a collecting surface area of 0·33 m2. Small cloth bags attached to the traps were collected every 2 weeks during the summer and autumn and subsequently sorted into their constituents. Leaf surface area was measured using a leaf area meter (model MK2, Delta-T Devices, Cambridge, UK), and peak leaf areas were back-calculated from cumulative litterfall (ICP Forests, 2004).

Eddy covariance measurements and data processing

The turbulent vertical fluxes of energy (sensible and latent heat), momentum, carbon dioxide and water vapour were measured continuously at a height of 28 m using the EC technique (Moncrieff et al. 1997). The exact details of the system installed at the Straits Inclosure have been published recently (Wilkinson et al. 2012). GPP was calculated as the differences between ecosystem respiration and net ecosystem exchange (NEE, negative indicates net uptake). The GPP term is hereafter called ‘actual GPP’ to distinguish it from modelled GPP. As small fluctuations in positive GPP were observed in winter, the carbon uptake period was determined using an arbitrary GPP threshold of 0·2 gC m−2 day−1 for a 5-day average; the start of the uptake period was defined as the days on which the average exceeded the threshold for three consecutive days. Photosynthetically active radiation (PAR) was estimated as 0·47 times the measured incident global solar radiation (Biggs et al. 1970), and incident radiation use efficiency (RUE) was calculated as:

display math(eqn 6)

Digital camera systems and settings

Two different digital camera systems (a commercial webcam ‘NetCam’ and a custom-made digital fisheye camera set-up, Automatic-capturing Digital Fisheye Camera (ADFC), Table S1) were mounted in separate weatherproof enclosures at the top of the EC tower at a height of 26 m. The NetCam SC 5MP (StarDot Technologies, Buena Park, CA, USA) was mounted so that it was facing slightly downward in a south-westerly direction, providing a field of view 47° in the horizontal plane. This camera was equipped with a 4–8 mm, 90°–47° wide varifocal auto-iris lens. Exposure was set to auto, and image quality was set to 1296 × 960 pixels. Colour balance was initially set to auto but changed to fixed with settings of red 256, green 180 and blue 256 on 12 March 2010. The camera was networked to a PC, and image capture was controlled using the manufacturer's DRV software. Images were taken half-hourly between 06:00 and 18:30 GMT, although the subsequent analysis was restricted to the 12:00–13:30 GMT images to minimise effects of low solar angles. The ADFC is a standard fixed-view camera, incorporated in a system designed by the Phenological Eyes Network (PEN,, Tsuchida et al. 2005; Nishida 2007), which consists of a compact digital camera (COOLPIX 4500; Nikon Corporation, Tokyo, Japan), a circular fisheye attachment lens (FC-E8, Nikon Corporation, field of view 180°), a waterproof housing and a control system with a PC running the open software photopc ( It was mounted horizontally facing downward, recording images every hour from 05:00 to 21:00 GMT. As for the NetCam images, only the 12:00–13:30 GMT images were used here. Exposure was set to auto, and digital images were stored as compressed JPEG file (resolution 2272 × 1704 pixels, three channels of 8 bit RGB colour information, and a camera setting of fine image quality). The white balance was set to auto in 2009 and changed to fixed with the sunny setting on 6 April 2010.

Image analysis

ImageJ (National Institutes of Health, Bethesda, Maryland, USA; version 1.42q) was used to process the images, and the regions of interest (ROI) were fixed (Fig. 1). The crowns of two oak trees were ROI in the images taken by both cameras. We extracted RGB digital numbers from the images and calculated the strength of each colour channel (Sred, Sgreen, Sblue) and Hue (equations are listed in Table 1). The strength of a channel is the ratio of the digital numbers of each channel to the total digital numbers of RGB channels, and Hue is one of the dimensions of the HSL colour scheme (see Mizunuma et al. 2011). The change of white balance setting after the first year made a difference to the magnitude of RGB digital numbers. We screened the images using Hue values, which are sensitive to wet conditions where values tend to shift towards the range of blue (180°–270°). By analysing images under hazy and snowy conditions and images with droplets on the lens, a threshold was established for the elimination of the images in which Hue values were larger than 140°. Dates of phenological events were determined by visual inspection of the graph of the annual time course.

Figure 1.

Example images and the analysed regions of interest (ROI) for: (a and c) canopy and (b and d) individual trees. (a) and (b) were taken by the NetCam and (c) and (d) were taken by Automatic-capturing Digital Fisheye Camera (ADFC).

Modelling GPP

Daily vegetation indices calculated from canopy-scale images were used to model GPP using the equation for MODIS GPP and NPP products (MOD17A2/A3, Heinsch et al. 2003):

display math(eqn 7)

where ε is the radiation conversion efficiency and FPAR can be derived from measured light interception (eqn (eqn 1)). Using an approximation detailed in Running et al. (2000), and because NDVI generally assumes positive values, FPAR can be replaced by the NDVI:

display math(eqn 8)

The conversion efficiency ε was calculated using simple linear ramp functions of the daily minimum temperature (Tmin) and the daylight average vapour pressure deficit (VPD).

display math(eqn 9)

The ramp functions (f) enable ε to be reduced to take account of the sensitivity of photosynthesis to temperature and humidity. We used the observed Tmin, VPD and maximum RUE for εmax and obtained the parameters of the ramp functions for deciduous broadleaf forest (DBF) from the Biome Properties Look-Up Table (BPLUT) for the MODIS product of vegetation production (Heinsch et al. 2003). To compare with satellite-based data, we obtained the MODIS Land Product Subsets for the Straits Inclosure (Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC) 2010); NDVI (MOD13Q1 and MYD13Q1, 16-day composite, 250-m resolution) and selected those for which product quality was flagged good. For deciduous canopies, it is possible to normalise the Sgreen and Hue vegetation indices derived from images, using the maximum and minimum values in each year (VInorm)

display math(eqn 10)

and replacing the NDVI with the analogous VInorm.

display math(eqn 11)

Statistical analysis was performed using SigmaPlot for Windows (Systat Software, Chicago, Illinois, USA; version 11.0).


Productivity of the forest and the seasonal trend

The GPP in 2010 was 17% lower than in 2009 (Table 2), and both were substantially less than the 12-year average annual value (2034 gC m−2 year−1, Wilkinson et al. 2012). In both years, an outbreak of defoliation by caterpillars of winter moth (Operophtera brumata L.) was observed in May and early June, and, noticeably, an infection with oak mildew (Erysiphe alphitoides Griffon & Maublanc 1912) followed on some oak trees. The dry weight of insect frass collected in litter traps was approximately 6·5% of that of leaves in both years. Figure 2 shows the daily variations in air temperature, precipitation, global radiation, GPP, RUE and LAI estimated from light interception at the site in 2009 and 2010. In 2009, the increase in GPP started on day 103, while the GPP increase in 2010 started 2 weeks later on day 118 with the slower increase in spring temperature. The number of days in excess of 5°C in minimum temperature until the uptake started was similar in the 2 years (23 and 21 days, respectively). A cold spell occurred in the first half of May 2010, and the minimum air temperature dropped to −2°C on day 132, damaging leaves, although the effects were hardly observed in GPP and LAI. The highest daily GPP in 2009 was 18·3 gC m−2 day−1 on day 165 and that in 2010 was 15·0 gC m−2 day−1 on day 181 (Fig. 2c). Autumn 2009 was mild, and the first record of air temperature below zero was 1 day after the carbon uptake finished on day 288. In comparison, there were cold days in mid-October 2010, and the minimum temperature of the day 294 was −3·9°C, although carbon uptake continued until day 307. LAI suggested that the trees kept leaves at the end of the carbon uptake period in 2009, while a sharp decline in LAI was observed before the end of the carbon uptake period in 2010. Although the data from the MODIS product were sparse, the increase in NDVI started earlier than that of LAI, particularly in 2009 (Fig. 2e).

Table 2. Summary of the climate conditions, the CO2 flux measurements and litter collection in the two study years at the Straits Inclosure, Alice Holt Forest, south-east England
 Whole yearCarbon uptake period
Climate conditions
Precipitation (mm year−1)937747310358
Average temperature (°C)10·29·314·114·0
Maximum temperature (°C)29·028·829·028·8
Minimum temperature (°C)−9·8−9·6−1·1−3·9
Global solar radiation (MJ m−2 year−1)3949385630332761
Carbon flux
Carbon uptake period (day)  186190
Carbon uptake start (day of year)  103118
Carbon uptake end (day of year)  288307
GPP (gC m−2 year−1)1824150616841396
NEE (gC m−2 year−1)−359−296−619−532
RUE (gC MJ−1)0·980·831·181·08
Litter collection
Maximum LAI (m2 m−2)4·44·2  
Frass in litter traps (g m−2)29·323·5  
Figure 2.

Seasonal variations in (a) air temperature, (b) precipitation and global solar radiation, (c) GPP, (d) RUE and (e) LAI estimated from light interception and the MODIS NDVI. Dotted lines indicate the start and end of carbon uptake period determined from GPP.

Canopy observations from two cameras

The colour signals extracted from images taken with the NetCam were influenced by the white balance setting (Fig. 3a), while those from the ADFC showed distinctive seasonal patterns throughout the 2 years (Fig. 3b). The images taken by the NetCam with auto balance showed high values in red and small variations in green throughout the year, which made detecting changes in the vegetation difficult. In 2010, the change of white balance from auto to fixed (dash-dot line) substantially improved the stability of red and blue values in the NetCam images and increased the difference in Sgreen and Hue between winter and summer, making the transition more evident (Fig. 3a). The images taken by the ADFC were vibrant with either auto or fixed setting, and the stability of vegetation indices was not affected, although the red values with auto setting were lower and the green values were higher than those with fixed setting. Some of the scatter was caused by changes in sky conditions (Table S2); in sunlit conditions, the normalised Sgreen was higher but Hue was lower than in cloudy conditions for both cameras. For green canopies, the difference between images in cloudy and sunlit conditions was larger with the ADFC camera than that for the NetCam. After the autumn frost in 2010, Sgreen and Hue from both cameras decreased sharply showing the same trend as the LAI and MODIS NDVI (Fig. 2e).

Figure 3.

Seasonal variations in strength of red, green and blue (RGB) and Hue for the canopy images taken by: (a) NetCam and (b) Automatic-capturing Digital Fisheye Camera (ADFC). Dotted lines indicate the start and end of carbon uptake period determined from GPP. The camera setting of both cameras for white balance changed on the days indicated by dash-dot line. Dashed line indicates the date of late frost event.

The indices for the individual trees overlapping in the NetCam and the ADFC images showed that both camera systems detected the same seasonal trends for the same tree (the results for Sgreen are shown in Fig. 4). In 2009, Sgreen of tree 1 increased on day 102, peaked on day 127 and after a gradual decline dropped to the baseline on day 300. In 2010, the Sgreen of the same tree increased from day 117 and showed a double peak on day 140 and 178 indicating damage and recovery from the late frost event. Although the magnitudes were different for both camera systems and settings, their trends coincided. Mildew was observed on a tree in the ADFC images from day 200 in 2009. By visual inspection, the damaged crown showed distinctively grey among the other green crowns. The values in Sgreen and Hue also showed a sharp decrease (not shown).

Figure 4.

Seasonal variations in the Sgreen for two oak trees in both Automatic-capturing Digital Fisheye Camera (ADFC) and NetCam images. Dashed line indicates the date of late frost event.

Seasonal relationships between canopy colour, actual GPP and LAI

In both years and with both camera systems, when the GPP started to increase, the canopy-scale Sgreen and Hue also increased (open circles in Fig. S1). The sharp peaks of Sgreen came 30 days earlier than the broader peaks of GPP, while the less rapid increase of Hue in spring with a much broader plateau was a better match to the GPP pattern. However, Hue dropped much later than GPP in the autumn in both years. The match of the pattern of Hue with that of canopy LAI estimated from light interception was good through the summer and autumn (Fig. S2), although Hue increased before canopy LAI in the spring.

Using canopy colour to model daily GPP

The modelled GPP estimated from measured FPAR was correlated with the actual GPP (Fig. 5a), although it overestimated considerably particularly in 2010 (the coefficient of determination R2 in 2009 and 2010 was 0·68 and 0·60, respectively). The GPP values estimated by the MODIS NDVI correlated less well (Fig. 5b, R2 in 2009 and 2010 were 0·36 and 0·48, respectively), and the standard error of the slope was much higher (Table S3). Using the normalised vegetation indices from ADFC images (eqn (eqn 10) and eqn (eqn 11)), the model using Sgreen overestimated GPP over many days in the spring and early summer (Fig. 5c) showing slopes similar to those computed with MODIS NDVI values (R2 in 2009 and 2010 were 0·40 and 0·46, respectively). The GPP estimated using the Hue showed a very good correlation with the actual GPP (R2 in 2009 and 2010 was 0·78 and 0·73, respectively), with no significant difference between 2 years (Fig. 5d), although there was a consistent overestimation of approximately 15%. Modelling GPP with normalised vegetation index data from the NetCam canopy images in 2010 showed similar results (Table S3), although the slope of the relationship was closer to one.

Figure 5.

Comparison of actual GPP in the carbon uptake period with: (a) GPP estimated from FPAR, (b) GPP estimated from MODIS NDVI product, (c) Sgreen and (d) Hue from Automatic-capturing Digital Fisheye Camera (ADFC) images. Dashed–dotted lines indicate linear regression in 2009, dashed lines indicate linear regression in 2010, and dotted lines indicate y = x.


Observations on leaf canopies using different camera systems

We collected images of a deciduous forest from the top of a flux tower using two different camera systems. They had different fields of view (NetCam, 45°; ADFC, 180°) and were oriented differently. The NetCam was fixed facing south-west, that is, the direction of the prevailing wind and the predominant direction of the flux footprint. However, in the SW orientation, the camera faced the sun and received the forward-scattered signal, which may be less useful than the back-scattered signal that predominately contains reflected rather than transmitted radiation. The choice might therefore be between (i) a rotating NetCam and (ii) a hemispherical lens that offers the choice of viewing angle. The fisheye lens provides images with a distorted geometry, which limits the image areas available for analysis, and most data were collected from vegetation immediately below the camera (Fig. 1). However, this viewing angle has the advantage of detecting the earlier greening of the ground flora in the spring (Fig. 3). Although the NetCam covers more distance, and with little distortion, the sideways-looking images taken by the NetCam are still limited in the number of crowns recorded and do not capture the understorey and ground vegetation. However, and reassuringly, when the same tree crowns were examined (Fig. 4), image analysis from both cameras detected the same phenological timing despite the angle of view differences. The possibility of setting NetCams in a fixed position and with a narrow field of view to focus on individual buds is closer to the traditional requirement of phenologists, which is to capture with fine resolution the state of the forest canopy and to permit the recording of bud break and abscission dates for individual species.

The white balance setting affected the quality of images as described by Richardson et al. (2009), which emphasises the importance of protocol standardisation (e.g. PhenoCam However, the effect of changing the white balance differed between the webcam and the digital camera, and uncertainties still remain when making observations using non-calibrated commercial camera systems. We checked the colour responses of the NetCam using standard colour charts. With the ‘auto’ colour balance setting of the NetCam, colours were represented satisfactorily, but with the colour balance ‘fixed’ as described in the methods section, we found that green colours were faithfully reproduced, while yellow was poorly represented. Although calibration panels have been proposed as a solution for different light conditions, Ahrends et al. (2009) pointed out the difficulty of using them, as images from their calibration panel saturated in bright light. The distance from objects can influence colour values (Richardson et al. 2009; Mizunuma et al. 2011), and a possible degradation of camera sensitivity over time has been suggested (Ide & Oguma 2010). Notwithstanding these difficulties, we found that Sgreen and Hue showed similar seasonal patterns despite differences between cameras and radiation geometries; however, Sred and Sblue varied. Overall, supporting the previous studies, the results here show the robustness of camera observations for recording canopy phenology.

Camera monitoring of biotic and abiotic stress events

It is evident that continuous camera records of the canopy state can also help us detect the impact of biotic and abiotic stresses (Hufkens et al. 2012b). In this study, although the litterfall observations suggested substantial loss of leaf mass from trees due to insect herbivory, it was not possible to detect the defoliation from the image analysis, partly because the canopy was still developing at this time, and defoliation did not cause any separate colour change. However, visual inspection of the leaves in the NetCam images did show leaf damage, although this would be hard to quantify. In photographic standards for assessing the proportion of foliage damage (Innes 1990), the sample photograph for ten percentage foliage loss is hardly distinguishable from healthy crowns; however, once severe crown loss occurs, the damage would be detected by canopy image analysis. In contrast, the grey typical of the oak mildew attack changed the colour of the foliage of affected trees and was detectable in the images. Therefore, the impact of other stresses such as lack of nitrogen, saltwater spray or inundation, that change leaf colour, would also be detectable from such automatic canopy image capture. In addition, the spatial information in images can help identify differences between trees in the impacts or responses, as shown in the response to frost damage (Fig. 4). Some early flushing trees were damaged and then recovered; some trees avoided the stress due to their late bud break. Observations using camera systems would be useful to track the response of trees in different varieties and provenances to abiotic stresses and also extreme events such as typhoons (Ide & Oguma 2010).

Carbon flux and canopy colour

The timing of the spring rise and autumn decline in CO2 uptake approximately followed the seasonal pattern in Sgreen and Hue, suggesting that indices based on these attributes may be used to estimate the seasonality of GPP. As the rate of photosynthesis depends not only on the state of the canopy but also on the solar radiation input, it would be surprising to obtain a good agreement between the GPP and any index of colour or of leaf amount alone. In this case, Sgreen reaches a maximum before the peak of GPP, while Hue peaks somewhat later than GPP. This early Sgreen peak coincides with leaf expansion, when the cuticle may not be fully developed and the leaves have not yet been coated by deposition of particles from the atmosphere or colonised by microflora. Later, when leaves are fully expanded and solar radiation input is higher, the maximum photosynthetic performance occurs. We have shown that the peak Hue value shows when the leaves are in the mature ‘dark green’. At some study sites, the peak of Sgreen coincided with the GPP peak (e.g. Ide et al. 2011; Migliavacca et al. 2011), but in others, as here, GPP lagged behind the peak of Sgreen (e.g. Ahrends et al. 2009; this study). Analysing images of three deciduous tree species, Nagai et al. (2011a) reported that the peak of Sgreen occurred earlier than both the peak of LAI and the peak of leaf chlorophyll content. In the first case, the rapid growth of LAI may have shortened the lag. Some physiological studies revealed that the development of full photosynthetic capacity took more than 50 days after the bud break (Morecroft, Stokes & Morison 2003; Muraoka et al. 2010). Compared with the sharp increase in Sgreen, the values of Hue changed from the region of yellowish green to the region of dark green gradually. This suggests a possible relationship with the slow development of leaf pigments.

It is surprising that the image-derived vegetation indices can be better predictors of GPP than FPAR itself. This may be due to the low spatial sampling intensity of below-canopy radiation employed here to derive FPAR. However, even extensive sampling of below-canopy radiation using solarimeters or broadband sensors would not detect the change in leaf reflectance caused by physiological changes, pigment changes and colonisation of oak leaves by fungi. In this study, the Sgreen showed a rapid increase at the time of leaf flushing, and the peak often preceded the peak of LAI. On the contrary, the gradual increase of Hue paralleled the trends of LAI. Moreover, Hue does not show the pronounced early peak with a drop off in summer that Sgreen and other indices based on the green channel values do (Mizunuma et al. 2011). Furthermore, using camera image–derived vegetation indices to model daily GPP from incident light and conversion efficiency showed that Hue was a considerably better proxy for fractional light interception than Sgreen (Fig. 5). This agrees with the good match that Hue gives to the pattern of LAI, although it should be noted that the light interception measurement, from which LAI is derived, was only at one location at the flux site, and not across the whole of the CO2 flux footprint. Of all the expressions of canopy colour, Hue is the one that seems to have the greatest utility, providing a means to calculate GPP which worked well for both years of study. Further research at other sites is needed to explore whether such a good relationship always exists, as it may be site and time specific. Of particular concern is the necessity to scale the VI values to the maximum and minimum observed for the site. While some researchers may consider this an unacceptable degree of empiricism, we believe it offers a useful approach worth further assessment.


At this deciduous woodland site, there was only a moderate relationship between NDVI from MODIS and the actual GPP over 2 years. However, vegetation colour indices calculated from digital camera images showed better correlation with GPP. In particular, the Hue parameter was an excellent predictor of GPP over 2 years. We have shown that the digital camera is an important aid in monitoring canopy condition and physiology especially when supplementary data from non-imagery devices are available, including various broad- and narrowband sensors (e.g. Garrity, Vierling & Bickford 2010; Ryu et al. 2010). It has the advantage of revealing a quantifiable image of the canopy as well as a set of derived reflectance indices. Thus, it enables researchers to link reflectance change to ‘real’ phenology and also reveals such occurrences as attack by disease and damage by storms.


The ADFC camera installation in Alice Holt was supported by a grant from the UK–Japan 2008 Collaborative Project Grant Award of the British Embassy, Tokyo, and the British Council to commemorate the 150th anniversary of official diplomatic relations between Japan and the UK and by funding from the European Commission Integrated Carbon Observing System (ICOS). We are indebted to the Phenological Eyes Network (PEN) for their support with the ADFC systems, in particular, Shin Nagai, Takahisa Maeda and Kenlo N. Nasahara. We are grateful to Lisa Wingate and Jérôme Ogée for their valuable discussions. The purchase of the NetCam was partly funded through the Forestry Commission and the FutMon: Forest Monitoring for the Future project, a European Commission Life+ co-financed project. The Straits Inclosure forest C & GHG balance measurement project is funded by the Forestry Commission. The authors thank the editor and the three anonymous reviewers for their many valuable comments.