The comparison of several colour indices for the photographic recording of canopy phenology of Fagus crenata Blume in eastern Japan

Abstract

Background: To understand how forests and woodland respond to global climate change, phenological observations are being made at a number of sites worldwide. Recently, digital cameras have been deployed as part of the existing network of ecosystem CO2 flux towers to provide a time-series of canopy images, and various numerical indices have so far been used by different authors.Aims: To identify which are the most effective colour indices to calculate from the signals extracted from digital cameras, in order to provide recommendations to the scientific community.Methods: Sample images of a Japanese beech (Fagus crenata) forest on Mt. Tsukuba (Japan) were used to define and calculate 12 colour signals and vegetation indices.Results: Although the strength of green signal and green excess index were reliable indicators for estimating foliage growth period, the indices were susceptible to low-visibility weather conditions and distance from the camera. Hue provided a robust metric, showing much less scatter during the vegetative period and a good indication of spring bud break. The bud break dates derived from the indices were slightly earlier than those assessed by visual observation, while the abscission dates were later.Conclusions: We propose that of all the candidate colour indices, hue is the most promising for the detection of bud break as it was least affected by atmospheric conditions.

Full text

PLEASE SCROLL DOWN FOR ARTICLE
This article was downloaded by:
[Mizunuma, Toshie]
On:
10 June 2011
Access details:
Access Details: [subscription number 938549930]
Publisher
Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-
41 Mortimer Street, London W1T 3JH, UK
Plant Ecology & Diversity
Publication details, including instructions for authors and subscription information:
http://www.informaworld.com/smpp/title~content=t793409773
The comparison of several colour indices for the photographic recording of
canopy phenology of <i>Fagus crenata</i> Blume in eastern Japan
Toshie Mizunumaa; Tomokazu Koyanagib; Maurizio Mencuccinia; Kenlo N. Nasaharab; Lisa Wingatec;
John Gracea
a School of GeoSciences, University of Edinburgh, Edinburgh, UK b Institute of Agricultural and Forest
Engineering, University of Tsukuba, Tsukuba, Japan c School of the Biological Sciences, University of
Cambridge, Cambridge, UK
Accepted uncorrected manuscript posted online: 04 March 2011
First published on: 07 April 2011
To cite this Article Mizunuma, Toshie , Koyanagi, Tomokazu , Mencuccini, Maurizio , Nasahara, Kenlo N. , Wingate, Lisa
and Grace, John(2011) 'The comparison of several colour indices for the photographic recording of canopy phenology of
<i>Fagus crenata</i> Blume in eastern Japan', Plant Ecology & Diversity, 4: 1, 67 — 77, First published on: 07 April 2011
(iFirst)
To link to this Article: DOI: 10.1080/17550874.2011.563759
URL: http://dx.doi.org/10.1080/17550874.2011.563759
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article may be used for research, teaching and private study purposes. Any substantial or
systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or
distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss,
actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.

Plant Ecology & Diversity
Vol. 4, No. 1, March 2011, 67–77
The comparison of several colour indices for the photographic recording of canopy phenology of
Fa g u s c r e n a t a Blume in eastern Japan
Toshie Mizunumaa*, Tomokazu Koyanagib, Maurizio Mencuccinia, Kenlo N. Nasaharab, Lisa Wingatecand John Gracea
aSchool of GeoSciences, University of Edinburgh, Edinburgh, UK; bInstitute of Agricultural and Forest Engineering, University of
Tsukuba, Tsukuba, Japan; cSchool of the Biological Sciences, University of Cambridge, Cambridge, UK
(Received 24 June 2010; final version received 14 February 2011)
Background: To understand how forests and woodland respond to global climate change, phenological observations are
being made at a number of sites worldwide. Recently, digital cameras have been deployed as part of the existing network of
ecosystem CO2flux towers to provide a time-series of canopy images, and various numerical indices have so far been used
by different authors.
Aims: To identify which are the most effective colour indices to calculate from the signals extracted from digital cameras, in
order to provide recommendations to the scientific community.
Methods: Sample images of a Japanese beech (Fagus crenata) forest on Mt. Tsukuba (Japan) were used to define and calculate
12 colour signals and vegetation indices.
Results: Although the strength of green signal and green excess index were reliable indicators for estimating foliage growth
period, the indices were susceptible to low-visibility weather conditions and distance from the camera. Hue provided a robust
metric, showing much less scatter during the vegetative period and a good indication of spring bud break. The bud break
dates derived from the indices were slightly earlier than those assessed by visual observation, while the abscission dates were
later.
Conclusions: We propose that of all the candidate colour indices, hue is the most promising for the detection of bud break as
it was least affected by atmospheric conditions.
Keywords: canopy phenology; deciduous broadleaved forest; digital image; HSL; RGB
Introduction
To understand how vegetation responds to climate warm-
ing, remote surveillance techniques have been established
to record continuously the long-term patterns of phenol-
ogy, including bud break, flowering and leaf abscission,
and relate these phenological events to the physiological
activity of the canopy (Richardson et al. 2007; Wingate
et al. 2008; Morisette et al. 2009). Ultimately, we expect
this will be achieved by using satellites; however comple-
mentary ‘ground truth’ approaches such as citizen-based
observations (e.g. the GLOBE programme in the UK,
http://www.globe.org.uk) or digital camera networks
(Wingate et al. 2008; Jacobs et al. 2009; Graham et al.
2010) are also essential to verify phenological patterns.
Recently, there have been attempts to analyse the colour
signals, red, green and blue (RGB), obtained from commer-
cially available digital camera images across many different
ecosystems (Richardson et al. 2007, 2009; Ahrends et al.
2008, 2009; Maeda et al. 2008; Jacobs et al. 2009; Ide
and Oguma 2010; Graham et al. 2010). So far, there is no
consensus on how these signals should be combined to
resolve adequately the key phenological events, such as
bud break and leaf abscission. There are many possible
indices based on the RGB signals. Maeda and Gamo (2004)
have patented a system that uses separate RGB indices in
a video system. Green fraction is sometimes used to detect
*Corresponding author. Email: Toshie.Mizunuma@ed.ac.uk
foliage phenology (e.g. Ahrends et al. 2008, 2009), while
red fraction has an advantage to detect autumn colouration
(Richardson et al. 2009). Others meanwhile have devised
vegetation indices analogous to the familiar Normalised
Difference Vegetation Index (NDVI) retrieved from satel-
lite spectral data. Normalised difference between red and
blue, which has a good correlation with the chlorophyll
content of crop plants (Kawashima and Nakatani 1998),
has been used as a weather filter of outdoor images (Ide
and Oguma 2010). The seasonal pattern in green excess
index (GEI) was correlated with data of gross primary pro-
ductivity of a forest (Richardson et al. 2009). Of particular
interest is the attempt to use the HSL (Hue, Saturation,
Lightness) colour system as the basis of an alternative
index. The HSL concept of describing colour by its inten-
sity and brightness in three-dimensional space was devised
by Smith (1978) as an early contribution to computer
graphics, and is said to be similar to the colour-recognition
system of the human eye. Previously, the HSL system has
been applied to phenological research to process spectral
data (Kodani et al. 2002), but more recently Graham et al.
(2009) used it to detect bud break and measure leaf area
expansion in Rhododendron occidentale Torr. & A. Gray.
Using viewpoints within the canopy, Graham et al. (2009)
found the hue-based colour system to be more useful than
the RGB colour system for their site.
ISSN 1755-0874 print/ISSN 1755-1668 online
© 2011 Botanical Society of Scotland and Taylor & Francis
DOI: 10.1080/17550874.2011.563759
http://www.informaworld.com
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

68 T. Mizunuma et al.
However, more tests are now required to evaluate the
applicability of these diverse approaches, and to assess
which of them is best related to phenological processes
and ideally to the seasonal cycle of forest CO2fluxes. We
present a detailed study of a particular case, using a num-
ber of indices obtained from a time-series of images. We
ask the general question: what is the best way to repre-
sent colour signals as an index? We examine recent data
of Fagus crenata Blume from a case study in a temperate
forest in Japan, collected by the Phenological Eyes Network
(PEN, http://www.pheno-eye.org/index_e.html; Tsuchida
et al. 2005; Nishida 2007).
Materials and methods
Site
Mt. Tsukuba is located 20 km north-east of Tsukuba city,
Japan. A meteorological observation station was installed
in January 2006 on the top of the west peak of Mt. Tsukuba
(36◦1330 N, 140◦552.8 E, 868 m above sea level)
by the University of Tsukuba (Hayashi et al. 2006). From
2007 onwards, a digital camera was mounted on a rooftop
terrace by the PEN to observe the seasonal change of adja-
cent trees. The mountain is covered by mixed forests of
Japanese beech (Fagus crenata), Japanese mountain maple
(Acer palmatum Thunberg) and a variety of conifers such
as Japanese cedar (Cryptomeria japonica D. Don) and
Japanese red pine (Pinus densiflora Siebold et Zucc). In
2008, the mean annual temperature was 9.9 ◦C, the mean
temperature in August was 20.5 ◦C and the annual precip-
itation was 1382 mm; the annual mean temperature and
the annual amount of precipitation were within the nor-
mal range, with low rainfall in July and heavy rain in
August.
Digital camera images
A commercially available compact digital camera (Coolpix
4500, 4.0 million pixels, Nikon Corp., Japan) was installed
within a waterproof housing at 12 m height, pointing
toward north-west, using the camera’s original lens (focal
length 7.85–32 mm) according to the protocol of PEN
(Tsuchida et al; 2005, Nishida 2007), and focusing on a
target F. crenata tree in a dense forest on the western
slope of Mt. Tsukuba. Canopy images of the same scene
were taken at 12:00 local time and stored as a compressed
JPEG file (resolution 2272 ×1704 pixels, three chan-
nels of 8-bit RGB colour information, and ‘fine’ image
quality by the setting of camera). Images were captured
in automatic aperture and exposure mode and white bal-
ance was fixed to ‘cloudy’. This study was based on the
images captured in 2008. Images for the periods from
Day-of-Year (DOY) 63 to 96, from DOY 210 to 219 and
from DOY 362 to 366 were missing due to system fail-
ure. A total of 315 days of images were incorporated in our
analysis.
Image analysis
A customised MATLAB program (7.7.0, R2008b; The
MathWorks) developed by AD Richardson and JP Jenkins
(cf. Richardson et al. 2007) was modified to process the
image files. Sample images and the region of interest (ROI)
used for this analysis are shown in Figure 1. The ROI was
set to three separate regions: (1) a prominent large beech
tree, well in focus; (2) a beech canopy to the left and behind
the main tree; and (3) a distant beech tree near the boundary
of the forest and the sky (Figure 1). Images were classi-
fied into three groups determined by weather conditions;
(1) images under high solar radiation conditions; (2) images
under low solar radiation conditions; and (3) low-visibility
images corresponding to days with haze and rain drops on
the window of the water-proof housing. Kawashima and
Nakatani (1998) reported different distribution patterns in
frequency of pixel values between images taken on a clear
day and those taken on a cloudy day. Although other studies
often excluded low-visibility images, we analysed them to
examine the sensitivity of colour indices to wet conditions.
Images taken during high and low solar radiation condi-
tions were identified using the data of downward shortwave
radiation measured every 10 min (EKO/Photo diode). A
threshold of 833.3 W m−2(0.5 MJ m−2per 10 min) was
used to separate high-radiation days from low-radiation
days. Low-visibility images were identified ‘by eye’.
Information from the 8-bit colour digital images
consisted of three values, each a digital number between
0 and 255, representing red, green and blue. We extracted
RGB values for each pixel of each image, from which the
mean of all pixel values for a given ROI was calculated.
A number of indices (equations are listed in Table 1)
were calculated to explore the suitability of each index for
observing changes in canopy phenology. This list of indices
is not exhaustive, but it includes those indices which have
been debated at recent phenology meetings. The relative
Figure 1. Example image (DOY =176). The white squares on
the image indicate the regions of interest. The camera was point-
ing north-west, and the overall field of view contains about 1 ha
of forest. The foreground is about 20 m away.
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

Comparison of colour indices of Fagus images 69
Table 1. Definitions of the signals and the derived vegetation indices from the digital camera images. See also Figure S1, available
online.
Signals and indices Formulation References
Strength of signal in RGB channel Schannel =nchannel/(nred +ngreen +nblu e) Richardson et al. (2007)
HSL signals
(Double hexagonal pyramid model)
Hue (degree)
Takagi and Shimoda (1991),
Joblove and Greenberg (1978)
H=(b−r)/((Imax −Imin)×60) +120 if g=Imax
H=(r−g)/((Imax −Imin)×60) +240 if b=Imax
H=(g−b)/((Imax −Imin)×60) +360 if g<b
H=(g−b)/((Imax −Imin)×60) otherwise
Saturation S=(Imax −Imin)/(Imax +Imin ) if lightness ≤0.5
S=(Imax −Imin)/(2 −(Imax +Imin )) if lightness >0.5
S=0ifImax =Imin
Lightness L =(Imax +Imin)/2
Monochrome luminance M=0.2989r+0.5870g+0.1140bSMPTE (2004)
Vegetation indices
Green Excess Index GEI =2n
green−(nred +nblu e) Woebbeche et al. (1995)∗,
Richardson et al. (2007)
Red Excess Index REI =1.4 nred−ng reen Meyer and Camargo Neto (2008)
Green Red Vegetation Index GRVI =(ngreen −nred)/(ngreen +nred ) Falkowski et al. (2005),
Kawashima and Nakatani (1998)
Green Blue Vegetation Index GBVI =(ngreen −nblue )/(ngreen +nblue ) Kawashima and Nakatani (1998)
Red Blue Vegetation Index RBVI =(nred−nbl ue)/(nred +nbl ue) Kawashima and Nakatani (1998)
Note: nred, ngreen, nblue are the values of red, green and blue respectively and nchannel is one of these; r=nred /255, g=ngreen /255, b=ngreen /255; Imax
is maximum value of r,gand band Imin is minimum value of r,gand b.
∗Cited in Meyer and Camargo Neto (2008).
strength of the signals in each channel was expressed as the
ratio of the values of each colour relative to the total values
of all three colours. In order to compare different colour
systems, the values of the RGB channels were converted
into hue, saturation, and lightness for the HSL colour
system (Joblove and Greenberg 1978) and monochrome
luminance following the National Television System
Committee (NTSC) standard (SMPTE, 2004). RGB and
HSL colour spaces used are explained diagrammatically
in Figure S1 of the supplementary materials (available on
the supplementary content tab of the article’s online page
at http://dx.doi.org/10.1080/17550874.2011.563759).
Classical observations of phenology made by human
observers are normally used to define particular phenolog-
ical phases (e.g. data from the International Phenological
Gardens in Europe). We compared the various dates of
phenological events obtained from visual inspection and
from the indices obtained from the digital camera images.
To compare the seasonal patterns of indices between the
different regions of interest, we excluded the data of low-
visibility images and calculated the 5-day running averages.
The dates of bud break and abscission were estimated by
detecting an abrupt increase or decrease in the indices.
Individual images were observed by eye; the day on which
leaves first appeared in each ROI was defined as the day
of bud break, and the day all leaves disappeared in each
ROI was defined as the day of abscission (sample images
are shown in Figure S2, available online). The dates for
the indices were manually determined by reading graphs;
the day that the trend started to change from winter state
was used for the bud break estimation, and the day that the
trend returned to the winter state was used for the abscis-
sion. Most indices showed a peak or trough in spring and
autumn, with some deviation from the dates of bud break
and abscission. We estimated peaks of spring and autumn
for each index using maximum or minimum values of the
running average data.
Results
Seasonal pattern of RGB values and indices
We first report the basic data from the regions of inter-
est delineated in Figure 1. The red, green, and blue values
extracted from image files for a region of the dominant
beech tree are shown in Figure 2. Both red and blue values
decreased sharply when leaf-flushing started. Thereafter
the red values stabilised over the summer period until a
dramatic increase associated with autumn leaf colouration
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

70 T. Mizunuma et al.
Figure 2. Red, green and blue (RGB) values extracted from
images in 2008 for one of the regions of interest, Tree1a, for a cen-
tred beech tree. In the graphs of values, solid circles indicate data
from images captured when the downward shortwave radiation
was low, open circles indicate images under high-radiation condi-
tions, and crosses indicate low-visibility images obscured by haze,
rain and/or water droplets (low-visibility images were selected
manually after visual inspection). The dotted line on DOY 122
indicates the visually selected date for budburst determined using
the images on the left, and DOY 318 indicates the selected date
for abscission using the images on the right.
was observed. The blue values increased slightly during
midsummer but declined again in autumn. In contrast, the
green values remained comparatively constant throughout
the year, with the exception of a slight decrease detected
in autumn. The values for low-visibility images showed
greater dispersion; in particular, the blue values were sus-
ceptible to atmospheric visibility in the foliage period.
The seasonal changes in 12 colour indices from the
dominant beech tree derived from the values are shown
in Figure 3. The meaning of the indices is defined math-
ematically in Table 1. The strength of the red channel
was stable over most of the year except during autumn,
when a peak was observed for autumn colouring and there-
after decreased with abscission (Figure 3A). The strength
of the green signal declined gradually over the summer
and reached the level of dormancy values after a small
negative inflection (Figure 3B). The strength of the blue
signal showed a negative bimodal distribution with inverted
peaks in spring and autumn (Figure 3C). The dispersion by
low visibility clearly is seen in the strength of green as well
as the strength of blue.
The three signals obtained for the HSL colour sys-
tem showed quite different seasonal patterns. Hue was the
least variable from day-to-day of the three (Figure 3D). It
dramatically increased after day 120, which was consis-
tent with the change observed in the green signal strength
shown in Figure 3B. Thereafter, while the strength of the
green signal gradually decreased over the season, the hue
signal continued to increase during the leaf-flushing stage
until day 130; thereafter it declined gradually during the
remainder of the season (Figure 3D). Hue was affected
by precipitation, but overall it was less ‘noisy’ than other
indices. In winter, the hue was affected by snow cover even
on clear days. The saturation signal was more suscepti-
ble to visibility than that of the hue (Figure 3E), but its
seasonal pattern depicted two obvious maxima that indi-
cated the peaks of leaf-flushing and autumn colouration.
Saturation values tended to be lower during hazy con-
ditions. Lightness values also displayed large day-to-day
variations, with maximal values occurring during the dor-
mant period (Figure 3F). Monochrome luminance (Figure
3G) showed a similar trend to that of the lightness. The dif-
ference between obscure and clear images was hard to dis-
tinguish for luminance. Although luminance was expected
to show the high density of leaves in the green period, it
showed relatively little seasonal variation and was quite
noisy.
The Green Excess Index (GEI) effectively detected bud
break, showing a dramatic increase at the spring onset;
thereafter a continuous decline was observed throughout
the season, with a slightly steeper drop in GEI observed
during the autumn senescence period (Figure 3H). The
Red Excess Index (REI) decreased in the leaf-flushing
phase and showed a peak in autumn (Figure 3I). Over the
season the Green Red Vegetation Index (GRVI) flipped
from negative to positive at leaf-flushing, and showed a
sharp decrease back to negative values during leaf shed-
ding (Figure 3J). Interestingly, the most negative values
occurred prior to leaf abscission and were more negative
than dormancy values. Both the Green Blue Vegetation
Index (GBVI) and the Red Blue Vegetation Index (RBVI)
had a bimodal distribution in spring and autumn showing
trends that were similar to those observed for the saturation
signal (Figure 3K, L). The dispersion by low visibility is
seen in GEI, GBVI and RBVI.
Low-visibility conditions added noise to most indices
except strength of red, hue and GRVI, while the differ-
ence between the high and low-light conditions was not
significant. To quantify fluctuation, we calculated the first
derivative for the moving average index excluding low-
visibility images (Figure S4, available online). Hue showed
the least day-to-day variability, followed by strength of
green. Lightness and monochrome luminance showed a
day-to-day variability throughout the year, which suggested
the difficulty to detect signals of seasonal change.
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

Comparison of colour indices of Fagus images 71
Figure 3. Seasonal variations in the colour indices calculated (listed in Table 1) in 2008 for Tree1a. Legend as in Figure 2.
Spatial variation in seasonal pattern of indices and
phenological dates
We combined data of all ROIs into one graph, to illustrate
how much the indices varied when the tree examined is in a
non-optimal position in the image, being partially shaded or
so far away as to be affected by haze (see Figure 1). To make
the data easier to observe on the graph, the 5-day running
averages were plotted rather than the daily data shown in
the previous figure (Figure 3). The different ROIs showed
similar trends to each other (Figure 4). Hue showed the least
scatter. A study in North American deciduous forests found
that the seasonal patterns in strength of red, green and blue
at more distant stands were less pronounced (Richardson
et al. 2009). The same trend appeared in the patterns of all
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

72 T. Mizunuma et al.
Figure 4. Seasonal variations in the colour indices calculated (listed in Table 1) in 2008 for all regions of interest extracted from images
under high-radiation conditions. Hue values over 180◦were converted into negative from 0◦(e.g. 350◦->−10◦).
indices; the signals from Tree1 were the strongest followed
by Tree2 and Tree3 corresponding to the distance from the
camera. Green fraction was significantly influenced by the
distance, while hue was less sensitive to the distance.
The estimated phenological dates derived from the sea-
sonal patterns of the indices are presented in Figure 5,
with the dates of bud break and abscission obtained from
visual inspection of the images. Most of the bud break
signals derived from images occurred before the visual
identification, while most of the abscission signals were
later (Figure 5A, B). Consequently, the foliage periods visu-
ally determined tended to be shorter than those from images
(Figure 5C). Most indices showed an agreement on the
spring peak date, while the peaks of the hue, the REI and
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

Comparison of colour indices of Fagus images 73
Figure 5. Phenological dates estimated by vegetation indices for the five regions of interest. Foliage periods are the numbers of days
between bud break and abscission. Dates determined visually from images are also shown for bud break, abscission and foliage period.
the GRVI occurred later (Figure 5D). Autumn peaks were
more variable between indices (Figure 5E).
Discussion
There are essentially three approaches to measuring pheno-
logical changes in vegetation. The first, and most widely
reported, involves direct human observation of dates of
bud break, flowering and leaf fall. Analysis of large data
sets, usually involving several species at a single location
over many years, has enabled generalisations to be made
about the relationship between these phenological stages
and weather patterns (Menzel 2000; Sparks et al. 2000;
Chmielewski and Rötzer 2001). A disadvantage of this
approach is that it relies on a somewhat subjective decision
of the stages by the observers, and can be influenced by
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

74 T. Mizunuma et al.
the ‘outliers’, i.e. those individual plants which happen to
be early or late in their phenological behaviour. Moreover,
it is labour intensive, although there appears to be no
shortage of willing volunteers who are able to carry out
daily observations, particularly on garden plants (Hepper
2003).
The second approach uses remotely sensed data from
satellites or towers to detect changes in colour (Myneni
et al. 1997; Zhou et al. 2001; Stockli and Vidale
2004). There are disadvantages in this approach too: long
time-series of data are hard to acquire because of the
limited life span of a satellite-borne sensor; in addition,
satellite data are not collected when the landscape is cloud
covered. In some respects it is an advantage that the satel-
lite ‘sees’ a large amount of landscape (spatial resolution is
from a few square metres to many square kilometres), as a
large sample integrates over many thousands of plants, but
species mixtures can affect signal interpretation. Sensors
mounted on flux towers, above the canopy, overcome many
of the objections that apply to satellite data. Such sensors
may be broad-band, like NDVI sensors (Miller et al. 2001),
or they may be hyperspectral radiometers (e.g. Motohka
et al. 2009; Nagai et al. 2010a).
The third approach, described in this paper, is rela-
tively new, and involves digital photography. The approach
has recently acquired popularity because it can be obtained
from public internet-connected cameras (Jacobs et al. 2009;
Ide and Oguma 2010; Graham et al. 2010) and, impor-
tantly, it is associated with the rapidly expanding network of
flux towers (Baldocchi et al. 2005; Richardson et al. 2007;
Wingate et al. 2008; Ahrends et al. 2009). Thus phenol-
ogy can now be linked to photosynthetic and respiratory
CO2exchange, hence providing a deeper understanding of
the controls on the terrestrial carbon cycle. The signals we
present are rich in information, indicating not only the clas-
sical phenological stages but also subtle changes during
the growing season, probably associated with changes in
leaf pigments, leaf area and pathogen infection. Moreover,
the signals from the camera can be stored indefinitely, pro-
viding an archive for subsequent analysis. However, there
remains some confusion about which signals are most
suitable for analysis, and in the current paper we have
attempted to identify the most robust indices. A prime con-
cern is filtering of effects caused by the changing light
conditions, especially those associated with bad weather.
We found that ‘hue’ provided a more robust metric than
those used by most other research groups who based their
indices on simply the strength of the green or red sig-
nal. It showed much less scatter during the summer and a
very good indication of spring bud break (Figure 3). Hue
varied between trees by only 25%, whereas all the other
indices were either much more variable or they did not
change with bud break (Figure 4). It is clear that a sys-
tem such as HSL is theoretically capable of defining any
colour as perceived by the human eye. It can cope with
variations in brightness and tone, which a simple expres-
sion of R, G and B, or a combination of these signals, can
never do. In canopy phenology there are situations of light
and shade, and the haze caused by scattering of short-wave
radiation, which provide much of the information content
perceived by humans. It is clear that hue represents the pure
colour that leaves and flower present, whilst ‘saturation’
and ‘level’ are influenced by non-biological variables such
as radiation flux and direction, and scattering by aerosols.
The use of ‘hue’ has, however, been overlooked in the anal-
ysis of remotely sensed images, in favour of indices such as
NDVI, which was initially developed for satellite sensors
with only limited capability of resolving the electromag-
netic spectrum. Nowadays, with hyperspectral sensors and
colour photography, this traditional approach may not be
appropriate.
There was some disagreement between the phenolog-
ical dates measured by the indices versus purely visual
observation of the images (Figure 5). The bud break dates
assessed by the indices were earlier and the abscission dates
assessed by the indices later than those assessed by visual
observation, and this can have a significant effect on esti-
mates of canopy duration. The trends coincide with the
results of a similar study using green fraction of beech
(Fagus sylvatica L.) trees in two European forests (Ahrends
et al. 2008, 2009). Abscission dates are recognised as being
harder to judge than spring signals even by visual inspec-
tion, as highlighted by Menzel et al. (2006). Presumably
this reflects the complexity of the processes whereby leaves
are shed, as wind and location-on-tree may be important
factors. These inaccuracies in visual assessment may also
partly explain the greater departures of abscission dates
compared with bud break dates of Figure 5.
All indices showed seasonal behaviour. The strength of
the green signal was, not surprisingly, a very reliable indi-
cator of bud break and the abscission of leaves. Hue, despite
its complex and conditional definition (Table 1), was
equally good at estimating canopy duration. Immediately
after bud break, and immediately before abscission, some
of the indices showed very clear maxima. In the case of
the green peak after unfolding, this indicates that all of
the leaves were fully developed. The gradual decline fol-
lowing the peak (in green indices) may be associated with
the maturation of the leaf cuticle and perhaps the pro-
gressive ‘weathering’ and ‘contamination’ of leaf surfaces,
potentially including the colonisation of leaf surfaces by
micro-organisms (Saunders 1971) and possibly the devel-
opment of negative water potentials in summer (Zakaluk
and Sri Ranjan 2008). In crop species, it has been shown
that the decline in green is associated with an increase
in chlorophyll, as a high optical thickness of chlorophyll
makes the leaves appear darker in the green waveband
(Kawashima and Nakatani 1998). In the case of the red peak
before abscission, we presume that the carotenoid pigments
become more apparent as the chlorophyll breaks down (e.g.
Archetti et al. 2009).
Ecological phenomena relating to the canopy, such as
leaf area expansion, leaf shedding, herbivory, disease and
tree health can be seen in these images. Monochrome
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

Comparison of colour indices of Fagus images 75
hemispherical images taken from ground-based locations
were commonly used for measuring leaf area index (van
Gardingen et al. 1999). However, monochrome luminance
did not show any evidence of the increase of leaf density
in this study. Although disease and insect damage were not
recorded, it is expected that a decrease in the numbers of
leaves will cause a decline in the green strength and GEI,
and that any change in leaf colour will have an effect on the
hue.
Richardson et al. (2009) reported that the change of
white balance on the camera setting from ‘auto’ to ‘fixed’
reduced the day-to-day variability in the green strength.
Although we used fixed white balance (cloudy), the signals
were still noisy even for images captured on only cloudy
days, which were less affected by direct sunlight and precip-
itation. The red and green values were less variable in com-
parison with the blue values. Consequently, hue and GRVI
were not susceptible to weather conditions. It is, however,
interesting to note that the distribution of pixel values for
each ROI showed some systematic differences depending
on the illumination conditions and the season. For exam-
ple, under high illumination conditions, the distributions of
the values of RGB clustered clearly around two separate
distributions for summer and autumn/winter, respectively
(Figure S3, available online). This resulted in very differ-
ent means for the two periods but very similar standard
deviations around the distributions. Images taken under
low illumination conditions appeared more variable season-
ally, both in their means and in their standard deviations,
with some very narrow and some very wide distributions
(Figure S3, available online). Kawashima and Nakatani
(1998) reported a good correlation between chlorophyll
content and the index derived from red and blue signals
for crops in cloudy conditions. However, their reported
condition only applied to a fraction of the growing sea-
son and they did not extend their findings to the rest of
the year. In addition, we think the red- and green-based
indices would be better for a long-term, outdoor observa-
tion programme, because of the vulnerability of the blue
signal to radiation scattered from aerosol optical effects. It
is undeniable that the index without the blue signal loses
not only the optical effects on blue but also the important
colour attribute of leaves. Previous studies (Nakakita 1996;
Graham et al. 2009) claimed the advantage of using hue.
In this study, hue showed robustness in optical conditions
in the vegetative period, which is useful for observing leaf
condition.
There are some important lessons to be learned from
the present work before an automated camera system could
be reliably used. First, the use of a standard reference card
in the field of view would be a useful aid for quality con-
trol. The card would have a black and white target area
and also a representation of hues (a colour chart). It might
also incorporate a light sensor, to keep track of changes in
illuminance of the card. Second, it is clear that camera man-
ufacturers are continuously improving the technology of
their product to achieve high-quality photographs, and that
different brands and models of camera may not therefore
produce the same result. Researchers therefore need to co-
operate to share experiences and where possible to make
comparisons of camera and lens types. Third, it is notable
that the result from any one camera depends to a large
extent on the camera settings, and that most ‘convenience’
settings apply a number of corrections. It may be preferable
to disable such corrections, including white balance, so that
consistent data are obtained.
Further work is needed to relate the colour fluctuations
to the pigment system and the physiological activity of the
leaves. It may be some time before colour information may
be used with weather data to reliably estimate canopy pho-
tosynthesis. However, concentrating available cameras at
sites where CO2flux is being measured by eddy covariance
will enable predictive equations to be developed (Nagai
et al. 2010b).
We conclude that the data from this inexpensive digital
camera may be used to detect phenological events with ade-
quate precision, and that other more gradual events which
elude the human observer are also detected by the camera.
Further work will be required to relate these changes to the
pigment concentrations in the canopy, and ultimately to the
physiological activity. We propose that of all the candidate
colour indices, hue is the most promising for the detec-
tion of bud break as it was less affected by atmospheric
conditions (e.g. for distant trees seen through bluish haze).
Acknowledgements
This work was supported by 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
KAKENHI (19688012; Grant-in-Aid for Young Scientists (A))
of the Japan Society for the Promotion of Science. Salary
for TM was made available through a Jim Gray Seed Trust
awarded to LW by Microsoft Research. LW was supported by
the Natural Environment Research Council Advanced Fellowship
award NE/G014418/1. Yousei Hayashi is thanked for the mete-
orological data collected at Mt. Tsukuba. We thank Andrew
Richardson and his team for allowing us to use their MATLAB
code. We are indebted to the Phenological Eyes Network (PEN)
for their support with the camera systems.
Notes on contributors
Toshie Mizunuma is a Ph.D. student at the University of
Edinburgh, studying the seasonal change of forest canopy and
their relevance for the global carbon cycle. She received an M.Sc.
in Forest GeoSciences from the University of Edinburgh and an
M.Sc. in Environmental Sciences from the University of East
Anglia. Her main research interests are phenology of plants in
boreal and temperate ecosystems and the observation methods
using optical equipment.
Tomokazu Koyanagi participated in this research as a Masters
student in the Institute of Agricultural and Forest Engineering,
University of Tsukuba with main interests in vegetation changes
in East Asia using remote sensing.
Maurizio Mencuccini is a Reader in the School of Geosciences
at the University of Edinburgh. His research interests include
net ecosystem productivity of forest ecosystems; functional
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

76 T. Mizunuma et al.
implications of architectural scaling in plants; tree–water rela-
tions; carbon stocks and fluxes; tree development and ageing;
photosynthesis.
Kenlo Nishida Nasahara is an Associate Professor, in the
Institute of Agricultural and Forest Engineering, University of
Tsukuba. His research interests focus on remote sensing of ter-
restrial ecosystems in relation to evapotranspiration and carbon
budget.
Lisa Wingate is a Natural Environment Research Council
Advanced Fellow in the Plant Sciences Department at the
University of Cambridge. Her research interests include under-
standing the response of tree growth to changes in climate,
constraining photosynthesis and respiration at large scales using
the carbon and oxygen isotope composition of CO2exchanged
between ecosystems and the atmosphere, and other ‘forest synthe-
sis’ activities, including linking phenology to plant function and
carbon allocation using webcam networks.
John Grace is Emeritus Professor of Environmental Biology,
in the School of GeoSciences at the University of Edinburgh.
His research interests include plant ecophysiology, atmosphere–
biosphere interactions and the long-term effects of global change
on ecosystem functioning, including carbon sequestration by
terrestrial ecosystems.
References
Ahrends HA, Brugger R, Stockli R, Schenk J, Michna P, Jeanneret
F, Wanner H, Eugster W. 2008. Quantitative phenological
observations of a mixed beech forest in northern Switzerland
with digital photography. Journal of Geophysical Research
113:G04004.
Ahrends HE, Etzold S, Kutsch WL, Stoeckli R, Bruegger R,
Jeanneret F, Wanner H, Buchmann N, Eugster W. 2009. Tree
phenology and carbon dioxide fluxes: use of digital photog-
raphy for process-based interpretation at the ecosystem scale.
Climate Research 39:261–274.
Archetti M, Döring TF, Hagen SB, Hughes NM, Leather SR, Lee
DW, Lev-Yadun S, Manetas Y, Ougham HJ, Schaberg PG,
et al. 2009. Unravelling the evolution of autumn colours: an
interdisciplinary approach. Trends in Ecology and Evolution
24:166–173.
Baldocchi DD, Black TA, Curtis PS, Falge E, Fuentes JD, Granier
A, Gu L, Knohl A, Pilegaard K, Schmid HP, et al. 2005
Predicting the onset of net carbon uptake by deciduous
forests with soil temperature and climate data: a synthesis of
FLUXNET data. International Journal of Biometeorology 49:
377–387.
Chmielewski FM, Rötzer T. 2001. Response of tree phenology
to climate change across Europe. Agricultural and Forest
Meteorology 108:101–112.
Falkowski MJ, Gessler PE, Morgan P, Hudak AT, Smith AMS.
2005. Characterizing and mapping forest fire fuels using
ASTER imagery and gradient modeling. Forest Ecology and
Management 217:129–146.
Graham EA, Yuen EM, Robertson GF, Kaiser WJ, Hamilton
MP, Rundel PW. 2009. Bud burst and leaf area expansion
measured with a novel mobile camera system and simple
color thresholding. Environmental and Experimental Botany
65:238–244.
Graham EA, Riordan EC, Yuen EM, Estrin D, Rundel PW.
2010. Public internet-connected cameras used as a cross-
continental ground-based plant phenology monitoring sys-
tem. Global Change Biology doi: 10.1111/j.1365-2486.2010.
02164.x.
Hayashi Y, Research Group for Intramural Project. 2006.
Meteorological Observation Station at the summit of Mt.
Tsukuba. Tsukuba Geoenvironmental Sciences 2:19–24.
Hepper FN. 2003. Phenological records of English garden plants
in Leeds (Yorkshire) and Richmond (Surrey) from 1946 to
2002. An analysis relating to global warming. Biodiversity
and Conservation 12:2503–2520.
Ide R, Oguma H. 2010. Use of digital cameras for phenological
observations. Ecological Informatics 5: 339–347.
Jacobs N, Burgin W, Fridrich N, Abrams A, Miskell K,
Braswell B, Richardson A, Pless R. 2009. The Global
Network of Outdoor Webcams: Properties and Applications.
In ACM International Conference on Advances in Geographic
Information Systems (SIGSPATIAL GIS). Proceedings
November 2009.
Joblove GH, Greenberg D. 1978. Color spaces for computer
graphics. Computers Graphics 12:20–27.
Kawashima S, Nakatani M. 1998. An algorithm for estimating
growth and development on the pathway of floral develop-
ment, chlorophyll content in leaves using a video camera.
Annals of Botany 81:49–54.
Kodani E, Awaya Y, Tanaka K, Matsumura N. 2002. Seasonal
patterns of canopy structure, biochemistry and spectral
reflectance in a broad-leaved deciduous Fagus crenata
canopy. Forest Ecology and Management 167: 233–249.
Maeda T, Gamo M. 2004. Japan Patent Number 2004-4280823
(2004.02.24).
Maeda T, Gamo M, Kondo H, Panuthai S, Ishida A, Nagai
S, Okamoto S. 2008. Leaf phenology detected by fixed
view camera images in a tropical seasonal forest at Mae
Klong, Thailand. In FORTROP II International Conference.
Proceedings November 2008.
Menzel A. 2000. Trends in phenological phases in Europe
between 1951 and 1996. International Journal of
Biometeorology 44:76–81.
Menzel A, Sparks TH, Estrella N, Koch E, Aasa A, Ahas R,
AlmKubler K, Bissolli P, Braslavska O, Briede A, et al. 2006.
European phenological response to climate change matches
the warming pattern. Global Change Biology 12:1969–1976.
Meyer GE, Camargo Neto J. 2008. Verification of color vegetation
indices for automated crop imaging applications. Computers
and Electronics in Agriculture 63:282–293.
Miller PCH, Lane AG, Wheeler HC. 2001. Matching spray appli-
cations to canopy characteristics in cereal crops. Pesticide
Outlook 12:100–102.
Morisette JT, Richardson AD, Knapp AK, Fisher JI, Graham,
EA, Abatzoglou J, Wilson BE, Breshears DD, Henebry GM,
Hanes JM, et al. 2009. Tracking the rhythm of the sea-
sons in the face of global change: phenological research in
the 21st century. Frontiers in Ecology and the Environment
7:253–260.
Motohka T, Nasahara KN, Miyata A, Mano M, Tsuchida S.
2009. Evaluation of optical satellite remote sensing for rice
paddy phenology in monsoon Asia using a continuous in situ
dataset. International Journal of Remote Sensing 30:4343–
4357.
Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR. 1997.
Increased plant growth in the northern high latitudes from
1981 to 1991. Nature 386:698–702.
Nagai S, Nasahara KN, Muraoka H, Akiyama T, Tsuchida
S. 2010a. Field experiments to test the use of the
normalized-difference vegetation index for phenology detec-
tion. Agricultural and Forest Meteorology 150:152–160.
Nagai S, Saigusa N, Muraoka H, Nasahara KN. 2010b. What
makes the satellite-based EVI-GPP relationship unclear in
a deciduous broad-leaved forest? Ecological Research 25:
359–365.
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011

Comparison of colour indices of Fagus images 77
Nakakita O. 1996. Color of forest. Japanese Forestry Society
18:52 (in Japanese).
Nishida K. 2007. Phenological Eyes Network (PEN) – a vali-
dation network for remote sensing of the terrestrial ecosys-
tems. AsiaFlux Newsletter 21:9-13 (available online at
http://www.asiaflux.net/newsletter.html).
Richardson AD, Jenkins JP, Braswell BH, Hollinger DY, Ollinger
SV, Smith M. 2007. Use of digital webcam images to track
spring green-up in a deciduous broadleaf forest. Oecologia
152:323–334.
Richardson AD, Braswell BH, Hollinger DY, Jenkins JP, Ollinger
SV. 2009. Near-surface remote sensing of spatial and
temporal variation in canopy phenology. Ecological
Applications 19:1417–1428.
Saunders PJW. 1971. Modification of the leaf surface and its envi-
ronment by pollution. In: Preece TF, Dickinson CH, editors.
Ecology of leaf surface micro-organisms. London: Academic
Press.
Smith AR. 1978. Color Gamet Transformation Pairs. Computer
Graphics 12:12–19.
SMPTE, Society of Motion Picture and Television Engineers.
2004. SMPTE 170M Television Composite Analog Video
Signal NTSC for Studio Applications SMPTE 170M
Television – Composite Analog Video Signal – NTSC for
Studio Applications.
Sparks TH, Jeffree EP, Jefree CE. 2000. An examination
of relationships between flowering times and temperature
at the national scale using long-term phenological record
from the UK. International Journal of Biometeorology
44: 82–87.
Stockli R, Vidale PL. 2004. European plant phenology
and climate as seen in a 20-year AVHRR land–surface
parameter dataset. International Journal of Remote Sensing
25:3303−3330.
Takagi M, Shimoda Y. 1995. Handbook of image analysis. Tokyo:
Tokyo University Press (in Japanese).
Tsuchida S, Nishida K, Iwao K, Kawato W, Oguma H, Iwasaki
A. 2005. Phenological eyes network for validation of remote
sensing data. Journal of Remote Sensing Society Japan
25:282–288 (in Japanese with English summary).
van Gardingen PR, Jackson GE, Hernandez-Daumas S, Russell
G, Sharp L. 1999. Leaf area index estimates obtained
for clumped canopies using hemispherical photography.
Agricultural and Forest Meteorology 94:243–257.
Wingate L, Richardson A, Weltzin JF, Nasahara KN, Grace J.
2008. Keeping an eye on the carbon balance: linking canopy
development and net ecosystem exchange using a webcam
network. FluxLetter 1:14–17.
Zakaluk R, Sri Ranjan R. 2008. Predicting the leaf water potential
of potato plants using RGB reflectance. Canadian Biosciences
Engineering 50:7.1–7.12.
Zhou L, Tucker CJ, Kaufmann RK, Slayback D, Shabanov NV,
Myneni RB. 2001. Variations in northern vegetation activ-
ity inferred from satellite data of vegetation index during
1981 to 1999. Journal of Geophysical Research – Atmosphere
106:20069–20083.
Downloaded By: [Mizunuma, Toshie] At: 10:40 10 June 2011