Open Journal of Forestry
2013. Vol.3, No.2, 70-74
Published Online April 2013 in SciRes (http://www.scirp.org/journal/ojf) http://dx.doi.org/10.4236/ojf.2013.32012
Copyright © 2013 SciRes.
70
New Methods to Quantify Canopy Structure of Leafless
Boreal Birch Forest from Hemispherical Photographs
T. D. Reid*, R. L. H. Essery
School of Geosciences, The University of Edinburgh, Edinburgh, UK
Email: *tim.reid@ed.ac.uk
Received November 30th, 2012; revised January 21st, 2013; accepted February 19th, 2013
Copyright © 2013 T. D. Reid, R. L. H. Essery. This is an open access article distributed under the Creative
Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium,
provided the original work is properly cited.
Hemispherical photography has been used for many years to measure the physical characteristics of for-
ests, but most related image processing work has focused on leafy canopies or conifers. The boreal forest
contains large areas of deciduous trees that remain leafless for over half the year, influencing surface al-
bedo and snow dynamics. Hemispherical photographs of these sparse, twiggy canopies are difficult to
acquire and analyze due to bright bark and reflections from snow. This Note presents new methods for
producing binary images from hemispherical photographs of a leafless boreal birch forest. Firstly, a
thresholding method based on differences between colour panes provides a quick way to remove bright
sunlit patches on vegetation. Secondly, an algorithm for joining up fragmented pieces of tree after thresh-
olding ensures a continuous canopy. These methods reduce the estimated hemispherical sky view fraction
by up to 6% and 3%, respectively. Although the processing remains subjective to some degree, these tools
help to standardize analysis and allow the use of some photographs that might have previously been con-
sidered unsuitable for scientific purposes.
Keywords: Hemispherical Photography; Image Processing; Leafless Canopies; Boreal Forest; Snow;
Abisko
Introduction
Forest canopies strongly affect the radiation balance at the
Earth’s surface. One well-established technique for quantifying
this effect is to take hemispherical photographs (hemiphotos)
looking upwards from the forest floor. Hemiphotos can be ana-
lyzed to estimate forest parameters such as sky view fraction
and plant area index van Gardingen 1999, but the first impor-
tant step involves applying thresholds of brightness to produce
a binary map of black (trees/horizon) and white (sky) pixels.
The thresholding is usually done manually, and has been criti-
cised for being subjective, prompting the development of
automatic methods Nobis 2005, Ishida 2004. Sky conditions are
also important, and overcast skies are preferable for their ho-
mogeneous quality. Zhang 2005 suggest measuring a reference
exposure for the open sky before lengthening exposure time to
increase contrast between sky and vegetation, while Pueschel
2012 suggest taking five photographs on different exposure
settings. However, such complicated approaches are not always
practical in remote, challenging field environments. It is bene-
ficial to take photos quickly, with minimal equipment and sim-
ple protocols that can be performed by different operators, in-
cluding students.
Moreover, most papers on hemiphoto analysis have focussed
on leafed deciduous trees or needleleaf conifers. The motive-
tion for this Note arose during fieldwork in March 2011 in an
area of leafless mountain birch forest in Abisko National Park,
Sweden. Hemiphotos taken in this region present considerable
challenges; homogeneous overcast conditions are rare or ac-
companied by blizzard conditions, and the high albedo snow
cover, low solar elevations and white birch bark cause consid-
erable light scattering.
This Note describes new techniques that can aid processing
of hemiphotos of sparse, leafless Arctic canopies. Section 3 de-
scribes a method of manipulating the colour panes of a hemi-
photo to identify sunlit parts of the canopy that are too bright to
be distinguished from the sky through a simple threshold. Sec-
tion 4 describes a branch-joining algorithm that acts to repair
hemiphotos in which branches with heterogeneous colour char-
acteristics have become broken-up during the thresholding pro-
cess. These new concepts are easy to use, and remove some
subjectivity from the processing of hemiphotos. They would
provide useful additions to popular existing software packages
such as Gap Light Analyzer, or GLA Frazer 1999, which don’t
tend to include special settings or functions for leafless cano-
pies.
Hemiphoto Acquisition and Thresholding
All photographs were taken using a Nikon Coolpix 4300
digital camera with a Nikon FC-E8 fisheye converter lens on an
automatic exposure setting, under the mountain birch canopy
around 3 km south of Abisko village (68.32˚ N, 18.83˚ E). Im-
ages were saved as JPEGs with resolution that gave a hemi-
*Corresponding author.
T. D. REID, R. L. H. ESSERY
spherical view of radius 1704 pixels. Algorithms were encoded
using IDL (Interactive Data Language©, 2011 Exelis Visual In-
formation Solutions). Canopy metrics-hemisphere-averaged sky
view fraction
s
v and effective plant area index were cal-
culated by first obtaining values of gap fraction,
Pv
, as the
ratio of sky pixels to total pixels in nine concentric 10 bands
of elevation angle
in the hemisphere.
s
v was calculated
by summing the weighted individual v
values according to
Essery 2008, and was calculated from logarithms of P v
according to van Gardingen 1999.
The thresholding of images in this paper is all done manually.
The automated threshold method of Nobis 2005 was tested on
Abisko hemiphotos, but it resulted in binary images that looked
far from the reality; large parts of trees were lost, perhaps be-
cause their software was developed for leafed canopies and has
difficulty detecting sky-tree edges in the leafless canopy.
Colour Pane Manipulation
Previous studies have highlighted the benefit of selecting the
blue colour pane of an RGB image to separate canopy and sky,
because blue is enhanced in the sky and reduced on trees Frazer
1999, Nobis 2005. The method presented here expands on this
concept, and that of Normalized Difference Vegetation Index
(NDVI), which indicates the presence of vegetation via the dif-
ference in intensities at certain spectral ranges Kriegler 1969.
Figure 1 shows that a similar argument can be applied to
mountain birch forest. A spectrophotometer (Analytical Spec-
tral Devices, Inc. (ASD)) was used to measure reflectivity
spectra of snow, white bark and darker copper-coloured bark in
the area where hemiphotos were taken in 2011 (Richardson and
Sandells, unpublished); these are similar to spectra measured by
Figure 1.
Reflectivity spectra of snow, copper-coloured birch bark and
white birch bark measured in Abisko in 2010 (upper panel).
Lower panel shows a typical sky irradiance spectrum for Abisko
replicated from Ovhed and Holmgren (1995), and the irradiance
expected from snow, copper bark and white bark, where each is
calculated by multiplying the sky irradiance by the relevant
reflectivity.
Ovhed 1995. On applying these reflectivities to a typical sky
spectrum for the Abisko region (replicated from Ovhed 1995),
white bark produces a flatter spectrum in the visible region (400
to 700 nm) than the sky, which gives considerably higher ir-
radiance at the blue end. This supports the argument for using
the blue pane (
400 to 500 nm). However, the irradiance re-
flected from white bark is often higher than from blue sky and
even some clouds. In this Note the shapes of the graphs are
exploited, taking the red pane into account. This involves sub-
tracting a proportion of the red pixel values (with digital
number (DN) between 0 and 255) from the blue

R
B pixel
values to produce a new grayscale image
g
I, defined as:
=
g
IBfr
R (1)
where r
f
is a number between 0 and 1. If
g
I is then nor-
malized back to the range 0:255, the difference between sky
and tree is magnified. Finally, a suitable threshold (=0 to
b
255) should be chosen and applied to
t
g
I, producing a binary
image.
Figure 2 shows an example hemiphoto in which the sunlit
white bark produces many bright patches on trees. In the blue-
pane-only image (upper row) these were too bright to remove
through simple thresholding without starting to turn many sky
pixels black as well. On applying the new method with a value
of r, a subjectively better binary image was produced.
The trunks come out black at a lower threshold and the vegeta-
tion as a whole is filled out to more accurately represent
=0.5f
Figure 2.
Blue colour pane from an Abisko birch hemiphoto (top left) and the
subjectively best binary image (top right) acquired by manually
increasing the threshold until just before some sky pixels began
turning black. By subtracting a fraction of the red pane (see Equation
(1)) and renormalising (centre left), it is easier to correctly classify
bright areas of bark without affecting the sky. This produces a
different value of sky view fraction (vs).
Copyright © 2013 SciRes. 71
T. D. REID, R. L. H. ESSERY
the real canopy. This causes a reduction in
s
v from 0.628 to
0.569.
Incorporating this colour pane manipulation in a graphical
user interface, using sliding bars to choose the best com-
binations of r
f
and b, has proven useful for analyzing
several hundred hemiphotos from the Abisko region. Such
functionality is not included in packages such as GLA, which
allows users to select an individual colour pane but doesn’t
allow panes to be combined in more complex ways. The im-
plementation of Equation (1) or variations on it would be very
straightforward and possibly beneficial for other forest situa-
tions.
t
This Note presents no automated method for removing snow-
covered topography from hemiphotos. Snow could be removed
by calculating horizon angles from a DEM Dozier 1990. Alter-
natively one could use a camera sensitive to near-infrared radi-
ation (NIR) Chapman 2007, because as Figure 1 shows there is
a difference between visible and NIR reflectance for snow. For
this Note, snowy horizons were shaded by hand, which can be
done quite quickly and accurately.
Branch Joining Algorithm
It is reasonable to assume that there should be no flying
twigs in a hemiphoto, and all dark areas of tree should be con-
nected to the ground forming one continuous dark object in the
image. When applying a brightness threshold to a gray colour
pane, the priority is to remove all the sky pixels, but even with
additional measures such as those described in Section 3, some
brighter tree pixels will be saturated to the extent that they will
be wrongly classified as sky. This means that the thresholded
image has many twigs and branches appearing fragmented.
This problem is enhanced for canopies with bright, reflective
bark such as mountain birch.
For these reasons a branch joining algorithm (BJA) was de-
veloped to connect the disjointed parts of thresholded images.
Figure 3 illustrates this process. A connected-component algo-
rithm was used to identify individual blobs in the image. For
each blob, the BJA identifies the two furthest-apart pixels (pix-
els 1 and 2 in Figure 3), representing the extreme ends of a
branch section, and performs searches in increasing radii
around both those points to find out which of the two is closest
to part of a different blob (pixel 3). The pixel is then joined to
the other blob by a line; the width of this line is determined by
the width of the original blob, defined as the number of pixels
in the blob divided by its length in pixels, rounded to the near-
est whole number. If it turns out that the blob is further than a
specified maximum distance from the nearest other blob, the
blob is deleted. For Abisko hemiphotos 30 pixels was chosen as
the maximum separation, but this could be adjusted for differ-
ent circumstances.
Figure 4 shows the BJA applied to an Abisko hemiphoto. On
applying the BJA, the binary image is filled out considerably.
For comparison, a small cropped part of the photo was proc-
essed by manually (laboriously!) shading in black all the
branches seen in the original image. This manual shading is
also subjective to some extent, but at least gives individual at-
tention to every part of the image, so it is arguably the closest
attainable representation of the truth. Counting the sky pixels in
the three cropped images showed that the threshold-only ap-
proach provides a sky view fraction 3.5% higher than the hand-
drawn method. On applying the BJA, the difference is only
Figure 3.
Schematic illustration of the branch-joining algorithm. For each
isolated blob (e.g. bottom left), the two furthest-apart pixels are
identified (marked 1 and 2). The algorithm then finds out which of 1
or 2 are closest to a pixel of another blob-in this case it is pixel 2,
which is joined to pixel 3 with a line (pixels marked with x) whose
width is determined by the width of the original blob.
Figure 4.
Blue pane of an example hemiphoto (top row) processed by thres-
holding (second row) then by the BJA (third row). The close-up
region in the right column was also subjected to a separate manual
approach with twigs hand-drawn in black on the original image before
thresholding (bottom right).
0.5%.
Figure 5 shows
s
v and calculated for 50 Abisko
hemiphotos with just a threshold applied, and the same images
after applying the BJA. The branch-joining algorithm decreases
P
s
v and increases , with maximum changes of 0.03 in P
s
v
Copyright © 2013 SciRes.
72
T. D. REID, R. L. H. ESSERY
Figure 5.
s
v and p calculated for 50 hemiphotos from Abisko, numbered
according to their sky view fraction. Solid lines show the results
with image thresholding only (TO), and dotted lines are results after
applying the branch-joining algorithm (BJA).
and +0.05 in . As might be expected, the differences are
larger when
P
s
v is smaller, because such hemiphotos contain
more vegetation to be processed by the BJA. A change in
s
v
of 3% may seem a small amount, but could have a significant
effect on calculations of surface radiation balance, especially if
conclusions were applied across large areas of forest.
It should be acknowledged that the BJA is rather ad-hoc in
the way it accounts for canopy geometry. A more sophisticated
BJA was trialled by preferentially attaching individual blobs to
others that lay in the direction of the major axis of the original
blob. This resulted in some twigs being drawn in strange places,
because there are many small blobs of 5 pixels or less in which
the major axis is not obvious. Overall, the simpler approach is
usefully non-specific; further levels of sophistication might make
the algorithm less generalizable to other canopies.
Conclusion
The techniques described in this Note address some of the
difficulties encountered on analyzing hemiphotos of leafless
Arctic canopies, and could provide useful additions to existing
hemiphoto software. The manipulation of individual RGB col-
our panes has proven very useful for correctly classifying bright
areas of canopy that would otherwise be missed out, and the
BJA fills many of the gaps left by the thresholding process.
Both techniques remain to be tested for different leafless forest
types, but for at least the case of boreal birch, they help to
standardize analysis across thresholded images.
The process of producing a binary image from a hemiphoto
will always be subjective to some degree, but these new meth-
ods go some way to reducing that subjectivity. The colour pane
manipulation provides a quick and easy way to correctly clas-
sify sunlit areas of bark, encouraging all users to pay special
attention to those areas; meanwhile the repairs made by the
BJA work to produce similar final results from binary images
that may have had different levels of branch fragmentation
resulting from thresholding by different users. The subjectivity
could be further reduced in future by adapting existing auto-
matic thresholding methods Nobis 2005, Ishida 2004 to work
for leafless canopies.
Most usefully, these methods make it possible to use hemi-
photos taken quickly under automatic camera settings and non-
ideal light conditions—often the only possible way to conduct
fieldwork in the Arctic winter. To improve on this study with-
out increased effort in the field, images should be saved in raw
format to avoid loss of detail on conversion to JPEG and allow
more sophisticated processing Lang 2010. Future work could
compare hemiphoto-based radiation transfer modelling to sub-
canopy radiometer measurements Link 2004 or terrestrial laser
scanning Antonarakis 2009, Cote 2009 to assess the benefits of
different image processing methods. Such independent verifica-
tion would be highly beneficial for forest studies on the whole,
because hemiphotos remain the fastest, cheapest and easiest
way to record forest structure in the field.
Acknowledgements
This work was funded by the UK’s Natural Environment Re-
search Council grant number NE/H008187/1. The authors
would like to thank all the support staff at Abisko Scientific
Research Station (ANS), as well as Nick Rutter, Maya King,
Cécile Ménard, Steve Hancock, Rob Holden, Michael Spencer
and Megan Reid for assistance in the field, Mark Richardson
and Melody Sandells for measuring and providing Abisko re-
flectivity spectra, and Casey Ryan for advice on processing
hemiphotos.
REFERENCES
Antonarakis, A. S., Richards, K. S., Brasington, J., & Muller, E. (2010).
Determining leaf area index and leafy tree roughness using terrestrial
laser scanning. Water Reso u rc es Research, 46, W06510.
Chapman, L. (2007). Potential applications of near infrared hemisphe-
rical imagery in forest environments. Agricultural and Forest Mete-
orology, 143, 151-156. doi:10.1016/j.agrformet.2006.12.006
Côté, J.-F., Widlowski, J.-L., Fournier, R. A., & Verstraete, M. M.
(2009). The structural and radiative consistency of three-dimensional
tree reconstructions from terrestrial lidar. Remote Sensing of Envi-
ronment, 113, 1067-1081. doi:10.1016/j.rse.2009.01.017
Dozier, J., & Frew, J. (1990). Rapid calculation of terrain parameters
for radiation modeling from digital elevation data. IEEE Transactions
on Geoscience and Remote Sensing, 28, 963-969.
Essery, R., Pomeroy, J., Ellis, C., & Link, T. (2008). Modelling long-
wave radiation to snow beneath forest canopies using hemispherical
photography or linear regression. Hydrological Processes, 22, 2788-
2800. doi:10.1002/hyp.6930
Frazer, G., Canham, C., & Lertzman, K. (1999). Gap light analyzer
(GLA): Imaging software to extract canopy structure and gap light
transmission indices from true-colour sheye photographs, users man-
ual and program documentation. Burnaby: Simon Fraser University.
Ishida, M. (2004). Automatic thresholding for digital hemispherical
photography. Canadian Journal of Forest Reseach, 34, 2208-2216.
Kriegler, F., Malila, W., Nalepka, R., & Richardson, W. (1969). Pre-
processing transformations and their effect on multispectral recogni-
tion. Proceedings of the 6th International Symposium on Remote
Sensing of Environment. Ann Arbor, MI: University of Michigan,
97-131.
Lang, M., Kuusk, A., Mttus, M., Rautiainen, M., & Nilson, T. (2010).
Canopy gap fraction estimation from digital hemispherical images
using sky radiance models and a linear conversion method. Agricul-
tural and Forest Meteorol og y, 150, 20-29.
doi:10.1016/j.agrformet.2009.08.001
Link, T., Marks, D., & Hardy, J. (2004). A deterministic method to
characterize canopy radiative transfer properties. Hydrological Proc-
esses, 18, 3583-3594.
Nobis, M., & Hunziker, U. (2005). Automatic thresholding for hemi-
sphericalcanopy photographs based on edge detection. Agricultural
and Forest Meteorology, 128, 243-250.
Copyright © 2013 SciRes. 73
T. D. REID, R. L. H. ESSERY
Copyright © 2013 SciRes.
74
doi:10.1016/j.agrformet.2004.10.002
Ovhed, M., & Holmgren, B. (1995). Spectral quality and absorption of
solar radiation in a mountain birch forest, Abisko, Sweden. Arctic
and Alpine Research, 27, 380-388. doi:10.2307/1552031
Pueschel, P., Buddenbaum, H., & Hill, J. (2012). An efficient approach
to standardizing the processing of hemispherical images for the esti-
mation of forest structural attributes. Agricultural and Forest Mete-
orology, 160, 1-13.
van Gardingen, P., Jackson, G., 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.
Zhang, Y., Chen, J. M., & Miller, J. R. (2005). Determining digital
hemispherical photograph exposure for leaf area index estimation.
Agricultural and Forest Meteorology, 133, 166-181.
doi:10.1016/j.agrformet.2005.09.009