Tree Physiology 24, 1323–1331
© 2004 Heron Publishing—Victoria, Canada
Detection of tree roots and determination of root diameters by ground
penetrating radar under optimal conditions†
CRAIG V. M. BARTON1,2 and KELVIN D. MONTAGU1
1
Forest Research and Development Division, State Forests of NSW, P.O. Box 100, Beecroft, NSW 2119, Australia
2
Corresponding author ([email protected])
Received December 10, 2003; accepted February 20, 2004; published online October 1, 2004
Summary A tree’s root system accounts for between 10 and
65% of its total biomass, yet our understanding of the factors
that cause this proportion to vary is limited because of the difficulty encountered when studying tree root systems. There is a
need to develop new sampling and measuring techniques for
tree root systems. Ground penetrating radar (GPR) offers the
potential for direct nondestructive measurements of tree root
biomass and root distributions to be made. We tested the ability
of GPR, with 500 MHz, 800 MHz and 1 GHz antennas, to detect tree roots and determine root size by burying roots in a
32 m 3 pit containing damp sand. Within this test bed, tree roots
were buried in two configurations: (1) roots of various diameters (1–10 cm) were buried at a single depth (50 cm); and (2)
roots of similar diameter (about 5 cm) were buried at various
depths (15–155 cm). Radar antennas were drawn along
transects perpendicular to the buried roots. Radar profile normalization, filtration and migration were undertaken based on
standard algorithms. All antennas produced characteristic reflection hyperbolas on the radar profiles allowing visual identification of most root locations. The 800 MHz antenna resulted
in the clearest radar profiles. An unsupervised, maximum-convexity migration algorithm was used to focus information contained in the hyperbolas back to a point. This resulted in a
significant gain in clarity with roots appearing as discrete
shapes, thereby reducing confusion due to overlapping of hyperbolas when many roots are detected. More importantly, parameters extracted from the resultant waveform through the
center of a root correlated well with root diameter for the
500 MHz antenna, but not for the other two antennas. A multiple regression model based on the extracted parameters was
calibrated on half of the data (R 2 = 0.89) and produced good
predictions when tested on the remaining data. Root diameters
were predicted with a root mean squared error of 0.6 cm, allowing detection and quantification of roots as small as 1 cm in diameter. An advantage of this processing technique is that it
produces results independently of signal strength. These waveform parameters represent a major advance in the processing of
GPR profiles for estimating root diameters. We conclude that
enhanced data analysis routines combined with improvements
in GPR hardware design could make GPR a valuable tool for
studying tree root systems.
Keywords: GPR, nondestructive root measurement, radar antenna, root biomass.
Introduction
Roots account for between 10 and 65% of a tree’s total biomass, depending on factors such as age, species, water and nutrient availability, and competition (see reviews by Vogt et al.
1996, Cairns et al. 1997, Keith et al. 2000, Snowdon et al.
2000). However, such reviews also highlight the limited understanding we have of tree root systems. Worldwide interest
in carbon sequestration by reforestation and afforestation has
led to increased research on tree root biomass and an increased
demand for accurate methods of estimating belowground biomass stocks (Watson et al. 2000). Furthermore, to refine forest
process-models, a greater understanding of the fate of carbon
allocated to tree root systems is required; however, sampling
tree roots, particularly of larger trees, presents logistical difficulties. Common techniques are limited to combinations of
excavation and coring (e.g., Burrows et al. 2000, Resh et al.
2003, Ritson and Sochacki 2003), both of which are destructive, costly, time consuming and laborious. A nondestructive
sampling technique would, therefore, be of value.
For years, geophysicists have detected features in sub-surface rock with ground penetrating radar (GPR). More recently,
GPR has been used to detect buried pipes, abandoned storage
tanks, artifacts, disturbed soil, caves and land mines. Ground
penetrating radar has also been used to detect tree roots and
map root systems (Hruska et al. 1999, Sustek et al. 1999,
Èermák et al. 2000, Wielopolski et al. 2000) and to determine
root biomass (Butnor et al. 2001, 2003). However, both the
measurement and data processing techniques need considerable refinement before GPR can be widely employed to measure tree root biomass.
Principles of ground penetrating radar
Ground penetrating radar generates a series of broadband electromagnetic pulses that are directed into the ground from a
specially oriented antenna. As the pulse travels through the
ground, it is reflected, deflected and absorbed to varying degrees by the materials (soil, water, rocks, roots) through which
† This paper was among those presented at the third meeting of the International Union of Forest Research Organizations concerning “Dynamics of physiological processes in woody roots,” convened by the School of Plant Biology, University of Western Australia, Crawley, WA.
1324
BARTON AND MONTAGU
it travels. It is the contrast in the dielectric permittivity between the bulk medium and buried objects that causes reflections: the greater the difference in dielectric permittivity, the
greater the coefficient of reflectivity (Conyers and Goodman
1997). A receiver in the antenna picks up the return signal and
the radar control unit records the two-way travel time, amplitude and polarity of the return signal. This information is typically plotted on a vertical scale with time on the ordinate and
signal strength and polarity on the abscissa, and is referred to
as an A-scan or trace. When operating, the antenna is dragged
across the ground along a transect and it fires a pulse at regular
intervals of either time or position. Multiple traces are then
combined to form a typical GPR profile scan (B-scan), with
distance on the abscissa and two-way travel time of signal on
the ordinate, with a grey or false color scale used to represent
the intensity and polarity of the return signal.
The pulse emitted from the antenna has an elliptical cone of
divergence and scans a footprint area beneath it, with the long
axis of the elliptical cone in the direction of travel (Conyers
and Goodman 1997). As the antenna is dragged across the surface, the travel-time for a return signal from a buried object decreases to a minimum when the antenna is directly above the
object then increases again as the antenna moves away from
the object, this results in a characteristic hyperbola in the radar
profile. The clearest hyperbolas are formed by point objects or
by linear objects with the major axis horizontal and normal to
the direction of travel of the antenna. Linear objects with the
major axis in the direction of travel of the antenna produce linear features on the radar profile but do not produce hyperbolas,
whereas other interception angles produce distorted hyperbolas.
The central frequency of the radar unit influences the size of
the footprint, with higher frequencies resulting in a smaller,
more focused footprint (Conyers and Goodman 1997). The
central frequency also influences the depth of penetration and
the resolution of the system. Higher frequencies are attenuated
more rapidly, such that a 100 MHz system will provide information to around 30-m depth, whereas a 2-GHz system is unlikely to provide useful information beyond about 0.2 m.
Typical maximum depths of penetration rarely exceed 20
wavelengths, except in a low attenuation, dielectric medium
(Daniels 1996). The attenuation depends on the magnetic permeability and electrical conductivity of the medium. In soil,
water content and mineralogy have a large impact, with high
water, clay and salt contents attenuating the signal most
strongly. Under unfavorable conditions of wet, calcareous or
clay rich soils, the maximum depth of GPR penetration can be
< 1 m, regardless of the frequency of antenna used (Conyers
and Goodman 1997). Higher frequency systems with their
shorter wavelengths have higher resolution (i.e., the ability to
detect small objects and to discriminate between closely
spaced objects). Thus a compromise between penetration and
resolution must be made and it is an important consideration in
the selection of system frequency.
The aim of our study was to determine if GPR could detect
tree roots and measure root diameter under the optimal condi-
tions of a sand test-bed. We tested the ability of three antennas
of different frequencies to detect roots of various sizes and at
various depths. We also describe a data analysis method that
gives more accurate estimates of root diameter than previously
published methods.
Materials and methods
Site
To create optimal testing conditions, we chose damp sand as
the test medium because it has a permittivity similar to soil but
is homogeneous with uniform bulk density and no structure or
voids that will interact with the radar pulse. The sandpit, 4 ×
4 m square and 2 m deep, was cut into the side of a steep slope.
Wooden walls were lined with polyethylene to prevent lateral
flow of water into the pit from the neighboring soil. The front
wall of the pit had a large opening to facilitate emptying.
The pit was filled with 50 Mg of washed, coarse river sand
(< 3 mm). To avoid edge effects, only the central 2 × 2 × 1.5 m
volume was used as the working area for root burial.
Roots
Roots were obtained from a forestry felling operation in the
Hunter region (NSW). Native hardwood (Eucalyptus sp.) tree
stumps were excavated with a mechanical harvester and a selection of roots was taken to the laboratory for grading. Sections of root, circular in cross section, roughly straight with
minimal taper and 50 cm long, were chosen. The cut ends were
sealed with molten wax to minimize water loss. Subsamples of
each root were taken to determine wood density (about 0.63 g
cm –3) and water content (about 65% gravimetric). Two orthogonal diameters were measured at 10-cm intervals along
the length of each root and the mean diameter calculated.
The working area of the sandpit was divided in half and two
trials were simultaneously undertaken. In the first half, nine
roots of various diameters (1–10 cm) were buried at a depth of
50 cm, and in the second half, eight roots of similar diameter
(about 5 cm) were buried at depths of 15 to 155 cm. All roots
were buried parallel to one another so that the GPR transects
would pass normal to the root axis (see Figure 1) thus avoiding
the complication of distorted hyperbolas that would arise if the
antenna tracked across the root at shallow angles. The horizontal distance between roots was 25 cm for roots buried at one
depth and 31 cm for roots buried at different depths. A further
five roots were buried at 155-cm depth beneath the trials to assess the effect of stacking on the ability of GPR to detect and
size roots.
Data collection
The sandpit was allowed to settle for 4 weeks after filling before the radar was tested. A field-portable GPR system
(RAMAC/GPR, Malå Geosciences, Malå, Sweden) with three
antennas (500 MHz, 800 MHz and 1 GHz) was used. Thirteen
transects 15 cm apart, perpendicular to the long axis of the
roots were established (see Figure 1). Radar profiles were collected along these transects with each of the three antennas.
TREE PHYSIOLOGY VOLUME 24, 2004
MEASURING TREE ROOTS USING GPR
The RAMAC GPR antennas each have a small measuring
wheel that enables their position along a transect to be determined accurately. The measuring wheel of the 1 GHz antenna
was very small, so to prevent it from digging into the sand and
dragging rather than rotating, a sheet of polyethylene was laid
across the sand pit. Each profile consisted of between 294 and
315 traces (horizontal resolution = 0.95 cm) and each trace had
511 samples (vertical time step = 0.08–0.09 ns). Four transects
passed directly over the roots buried at 50 cm (Transects 2–5)
and four transects passed directly over the roots buried at various depths (Transects 9–12). The other five transects passed
either to the side of or between the two strips of buried roots
and enabled determination of the information contained in the
radar side-look.
Data processing
Radar profile normalization, filtration and migration routines
were performed with MathCad 2000 (MathSoft, Cambridge,
MA). The main features of interest in the radar profiles were
the hyperbolas, which indicated a root or other point feature
buried in the sandpit. Several data pre-processing algorithms
were applied in a stepwise fashion to remove background
noise and improve identification of the hyperbolas. First, application of a background removal filter removed the parallel
bands observed in the scans, which result from plane reflectors
such as the ground surface, soil horizons and bands of low frequency noise. Second, removal of the DC offset and DC drift
compensated for any slight drift in the zero in each trace that
make up the full profile (Daniels 1996). Third, application of a
5 × 5 median gradient filter smoothed out any random noise
(Pratt 1991, Olhoeft 2000). This filter acted as a low-pass filter
both along the transect and through time, suppressing trace
and time-dependent noise and spike events such as mobile
phone interference.
1325
Following the pre-processing, the velocity of propagation of
the signal through the sand was determined. The shape of a hyperbola depends on the velocity of propagation of the signal
through the sand and the depth of the object. The deeper the
object and the faster the propagation velocity, the flatter the
hyperbola (Daniels 1996). A modeled hyperbola was overlayed on the profile and the velocity parameter adjusted until
the simulated hyperbola matched the observed hyperbolas.
Based on this approach, the propagation velocity was estimated to be 13.5 cm ns –1. An advantage of using damp sand is
that the signal velocity is uniform throughout the sandpit. Because the strength of the return signal deteriorates exponentially with travel time, a combination of linear and exponential
gain was applied to the profiles to enhance the visibility of hyperbolas and aid in the determination of propagation velocity.
This gain was not used during the migration and parameter extraction described below.
With the signal propagation velocity obtained previously, an
unsupervised maximum-convexity migration algorithm was
applied to each profile, focusing the information contained in
the hyperbolas back to a point. This method assumes a
semi-hyperbolic maximum-convexity function and sums the
value of each separate trace at the point at which it intersects
the semi-hyperbolic focus over the ensemble data set. All
in-phase energy is additive, whereas non-coherent energy is
usually out of phase and sums to zero (Daniels 1996). The algorithm is applied to every point in the GPR scan regardless of
whether or not it is at the peak of a hyperbola. If the point is at
the peak of a hyperbola, then all of the information contained
in the hyperbola is focused to that point; if not, then it focuses
random noise to that point that tends to sum to zero. This unsupervised approach does not require an operator to visually locate hyperbolas on the raw radar scan prior to processing. It
automatically highlights point and line reflectors normal to the
plane of the profile and removes the clutter associated with the
constructive and destructive interference between hyperbolas.
This results in discrete zones at the focal point of each hyperbola in the GPR profile.
After application of the filters and migration algorithm, the
location of the buried roots became evident and visual inspection of the migrated profiles indicated that there might be a relationship between the size of roots and the “wavelength” of
the migrated trace through the root. To investigate this relationship further, the traces passing through the center of each
root were selected for analysis and the times between successive zero crossings of the reflected waveform extracted (see
Figure 5). The first four of these time intervals (referred to as
parameters a–d) were correlated against root size for each of
the antennas.
Model fitting and testing
Figure 1. The positions of roots buried in the sand pit (solid horizontal
lines) and the transects along which the radar antennas were pulled
(broken vertical lines).
A multiple linear regression was performed to determine the
best empirical model relating the extracted parameters (a–d) to
root size. The first three parameters (a–c) explained most of
the variation and were therefore used subsequently (see Figure 6). To test the data dependence of the model, half of the
TREE PHYSIOLOGY ONLINE at http://heronpublishing.com
1326
BARTON AND MONTAGU
data were used to calibrate the model, and the remaining data
were used to test the model; i.e., data from two of the four
transects scanning each of the two halves of the sandpit were
used to calibrate the model (Transects 2, 4, 9 and 11), and the
other four transects that passed directly over roots were used to
test the model (Transects 3, 5, 10 and 12).
Results
After filtering the profiles, hyperbolas were observed for each
of the roots of various diameters buried at 50-cm depth in
scans of all three antennas (Figure 2). Hyperbolas were visible
for all nine roots, although the hyperbola for the smallest root,
the one furthest to the right, was faint. The hyperbolas for the
500 MHz system were the least sharply defined. For each antenna, at least one of the roots buried at 155-cm depth beneath
the row of roots buried at 50-cm depth was visible. For the 500
and 800 MHz antennas, a second root at 50-cm depth was visible, although indistinct. Similar results were observed for the
other three transects that ran directly above the roots buried at
50 cm and for roots buried at different depths (data not shown).
The GPR profiles for roots of the same size buried at different depths showed a similar pattern, with more clearly defined
hyperbolas obtained with the 800 MHz and 1 GHz antennas
compared with the 500 MHz antenna (Figure 3). A common
feature of all profiles was the interference between hyperbolas, making it difficult to determine their exact positions. This
interference increased with decreasing antenna frequency. The
GPR profile of transects that did not pass directly over roots
contained weak hyperbolas indicating some degree of side-
look; it was difficult to extract useful information from these
transects and they were therefore excluded from further
analysis.
The effect of the migration algorithm is illustrated in Figure 4. Figures 4A and 4C show the filtered radar profiles
(500 MHz) of transects directly over the roots, buried at various depths (same root diameter) or buried at one depth (varying root diameter) respectively. Figures 4B and 4D show the
corresponding profiles after migration. The positions of most
of the roots were clearer after applying the migration algorithm. The positions of the roots at various depths can be seen
in Figure 4B. The migrated zones become fainter with depth as
a result of signal attenuation, but the size of the zones appears
relatively constant. In Figure 4D, the positions of the roots buried at 50 cm were clear and the size of the migrated zones appeared to vary with root size. Two of the roots buried at 155 cm
were clearly visible, but the third root was undetectable, probably because it was too close to the side of the sand pit and as a
result, only one leg of the hyperbola contained information.
This also explains why the migrated “blob” for the rightmost
root buried at 50 cm is less well defined than the others.
The apparent size of the migrated zone is a function of the
wavelength of the return wavelet (see Figure 5). Figure 5
shows a typical migrated zone using the 500 MHz antenna and
the corresponding wavelet extracted from the trace through the
center of the zone. Parameters a–d are the time intervals (in ns)
between zero crossings of the waveform. It was fairly straightforward to extract the parameters for the 500 MHz profiles but
more difficult for the 800 MHz profiles and very difficult for
the 1 GHz profiles because of confused waveforms that some-
Figure 2. Radar profiles from three antennas (500 MHz, 800 MHz, 1 GHz) along a transect across the center of nine roots of various sizes (1–10 cm
diameter) buried 50 cm deep in the sand pit. Two of the three roots buried at 155 cm deep can be seen on the 500 MHz image and one root on the
800 MHz and 1 GHz images. Horizontal direction represents distance along the transect and vertical direction represents the travel time of the signal.
TREE PHYSIOLOGY VOLUME 24, 2004
MEASURING TREE ROOTS USING GPR
1327
Figure 3. Radar profiles from three antennas (500 MHz, 800 MHz, 1 GHz) along a transect across the center of eight roots of similar size (5 cm diameter) buried in the sand pit at depths of 15–155 cm.
Figure 4. (A) Radar profile
from the 500 MHz antenna
along a transect across the center of eight roots of the similar
diameter (5 cm) buried at various depths (15–155 cm). The
characteristic hyperbolas are
overlapping and interfering
with each other, making interpretation impossible. (B) Eight
roots clearly visible after the
radar profile in (A) had been
processed with the migration
algorithm. (C) Radar profile
along a transect across the center of nine roots of various
sizes buried 50 cm deep. (D)
Nine roots are clearly visible
after the radar profile in (B)
had been processed with the
migration algorithm. Again the
position of the roots is
obvious. The arrows in (D)
indicate the two roots buried at
155 cm.
TREE PHYSIOLOGY ONLINE at http://heronpublishing.com
1328
BARTON AND MONTAGU
Figure 5. (A) Close up of a
processed 500 MHz profile
showing two roots. (B) Trace
through the center of the first
root as indicated by the vertical line in (A). Parameters a–e
are determined from the trace
as the time (ns) between zero
crossings of the waveform.
times failed to cross the zero line.
There was a positive correlation between root size and parameters a–d for the 500 MHz antenna (Figure 6), which was
especially strong for parameters b and c. Similar plots for the
800 MHz and 1 GHz antennas failed to show such correlations
(data not shown) and so no further analysis using data from
these antennae was performed.
The model fitted to half of the data for the 500 MHz antenna
is shown in Figure 7. The coefficient of determination was
0.89 and the root mean squared error (RMSE) was 0.51 cm.
The fitted model was used to predict root diameters for the
other half of the data set (Figure 8). The residuals indicate that
the model generally performed well, with a slight bias toward
over estimating root size. The RMSE of 0.56 cm indicates that
the size of individual roots could be estimated fairly accurately.
Discussion
Investigation of tree root systems with GPR has been largely
confined to qualitative mapping (Hruska et al. 1999, Èermák
et al. 2000). Attempts to quantify tree root systems have met
with mixed success (Butnor et al. 2001, 2003). Butnor et al.
(2003) found reasonable correlation between root biomass in
the top 30 cm of forest soil in soil cores and information obtained from GPR. Under ideal conditions, we were able to locate buried roots and extract information from a processed
waveform that correlated well with root diameter (Figure 7).
When several closely spaced targets exist, such as tree root
systems, GPR profiles usually produce a confusing profile because of interactions between the hyperbolas from each target.
This is readily illustrated by the profiles in Figure 2, even after
normalization and filtering. However, visual identification of
hyperbolas associated with the tree roots was possible for all
Figure 6. Relationships between waveform parameters
a–d (ns) and root diameter
(cm) for all eight transects using the 500 MHz antenna. The
circles show roots of varying
sizes (1–8 cm diameter) buried
at one depth (50 cm), whereas
the triangles show roots of
similar size (5 cm diameter)
buried at various depths
(15–155 cm).
TREE PHYSIOLOGY VOLUME 24, 2004
MEASURING TREE ROOTS USING GPR
Figure 7. Fitted model relating mean root diameter to parameters extracted from processed GPR data. The model was fitted to one half of
the complete data set. The circles denote roots of various sizes buried
at 50 cm and the triangles denote roots of similar size buried at different depths (15–155 cm). The broken line is the 1:1 relationship. The
fitted model is: diameter = –17.1527 + 0.3066a + 1.31b + 0.70c –
0.032b2 – 0.0014abc, r 2 = 0.89, RMSE = 0.5. See Figure 5 for definitions of parameters a, b and c.
three GPR antennas for most roots down to a depth of 155 cm
and between 1–10 cm in diameter. Even when roots were buried below other roots, the three systems tested were able to detect the deeper buried roots. The depth of penetration of the
1 GHz antenna was better than expected, given that the higher
the central frequency and the shorter the wavelength, the more
quickly the signal is attenuated. Low frequency antennas
(10–120 MHz) generate long wavelength radar energy that
can penetrate up to 50 m under certain conditions, but the maximum depth penetration of a 1 GHz antenna is about 1 m or
less in typical soils (Conyers and Goodman 1997). The depth
of penetration of the 1 GHz antenna in our study reflects the
low attenuation properties of the damp sand used in the test
bed. Under field conditions, with typically higher soil water
1329
and clay content, such depth of penetration would be unlikely.
In terms of simple visual identification of hyperbolas, the
800 MHz antenna gave the clearest radar profiles and appears
to represent a good balance between resolution and depth of
penetration for field surveys.
To overcome confusing GPR profiles and hence improve the
visual identification of root location, we applied an unsupervised maximum-convexity migration algorithm (Daniels
1996). This resulted in a significant gain in clarity with the
roots appearing as discrete shapes (Figure 4). The use of the
migration algorithm was simplified by the uniform propagation velocity throughout the sand medium. It would be expected that, for tree roots, velocity would vary because of
variations in soil water content and composition down the profile. However, more complex migration algorithms exist that
allow for variations in propagation velocity within the measurement volume (Daniels 1996). Thus, it appears that a
migrating algorithm could be applied to data from roots in situ.
We have discussed improved methods for processing GPR
data to assist in the visual identification of root location and
root mapping as used by Hruska et al. (1999). However, if
GPR is to find wide application in tree root studies, techniques
that allow the size of roots to be determined and their biomass
to be estimated are required. Few attempts have been made to
extract quantitative information on tree root systems from
GPR data. Butnor et al. (2001) used signal intensity to obtain
information on root diameter and biomass under field conditions. Roots of known size were buried at 15 and 30 cm and a
correlation was observed between return signal strength and
root diameter that deteriorated rapidly with depth (r = 0.81 to
0.55 at 15- and 30-cm depths, respectively). Butnor et al.
(2001) also scanned several forested areas with GPR and then
excavated roots and looked at the correlation between total
root biomass and a measure of return signal energy; however,
the correlations were weak (r = 0.34 to 0.57), because of the
many confounding influences on signal strength (e.g., depth,
angle of crossing and dielectric properties of the medium).
Butnor et al. (2003) employed a slightly different approach
with better results under field conditions. They applied a hori-
Figure 8. Model testing: the
predicted values of root diameter, using the half of the data
not used to fit the model, plotted against the actual mean
root diameters. The residual
plots show that residuals are
normally distributed with a
slight positive bias (0.2 cm)
and RMSE = 0.56 cm.
TREE PHYSIOLOGY ONLINE at http://heronpublishing.com
1330
BARTON AND MONTAGU
zontal distance normalization filter, a background removal filter, followed by a Hilbert transformation to the raw GPR scans
along transects and then estimated the number of pixels above
a threshold. After adjusting for different relationships between
fertilized and unfertilized plots, the resulting pixel count was
correlated (r = 0.86) with the root biomass in soil cores taken
at intervals along the transects. We also examined the use of
signal strength and found it to be poorly correlated with root
size, primarily because of attenuation of the signal with depth.
Consequently, we conclude that it is unlikely that quantitative
information on root size can be obtained from signal strength
because of the large number of confounding influences that
need to be taken into account.
The major finding of our study was that root diameter could
be estimated from the waveform parameters of migrated hyperbola from the 500 MHz antenna (Figures 5–7). Our model
provided good predictions of actual root diameter ranging
from 1 to 8 cm (Figure 8). A major advantage of the processing
technique is that it produces results that are independent of signal strength. Thus, the diameters of roots buried at different
depths could be estimated with no corrections for depth, because the waveform parameters used in the model (measured
time between zero crossings; Figure 5b) were unaffected by
depth for roots buried at depths of 15 to 155 cm deep. In contrast, strength of the signal varied considerably as shown in
Figure 4B where the visibility of the same diameter roots decreases with depth. The use of waveform parameters represents a major advance in the processing of GPR profiles for the
study of tree root systems.
The roots used in our study were not perfect cylinders and
the diameters reported are actually the mean diameter of orthogonal measurements at 10-cm intervals along the root.
There was some variation in diameter and non-circularity in
cross section with the mean coefficient of variation of diameter along a root being 10%. Therefore some of the variation in
the model fit will be a result of this variation in diameter along
the root.
From theory, we expected the higher frequency antennas,
with their shorter wavelength and higher resolution, to provide
the most information about root diameter. However, it was extremely difficult to extract information that correlated with
root size from the migrated traces from either the 1 GHz or
800 MHz antenna. There may be useful information in these
data, in which case the approach we adopted may be inappropriate for those frequencies. The relationship we derived between the waveform parameters from the 500 MHz system and
root diameter is purely empirical and the underlying physical
principles are not fully understood. Improved algorithms designed for complex forest ecosystems could possibly be developed in partnership between biologists and those with more
expertise in the relevant field of physics. However, even without a full understanding of the physics, our approach provides
a useable method of nondestructively determining root diameter with a reasonable degree of precision and accuracy, particularly given that the 500 MHz antenna has good depth
penetration under a wide range of field conditions.
Variation in soil water content over time or from site to site
will influence the signal attenuation and propagation velocity
and consequently the strength of the return signal. This would
have implications for algorithms that rely on signal strength to
relate biomass to GPR data (e.g., Butnor et al. 2001, 2003);
however, it is uncertain how such variation in water content
would influence the parameters extracted from the return
signal used in this study.
More work is required to determine if the waveform parameter approach can be adapted for use under field conditions.
This study was conducted under optimum conditions with a
uniform medium, roots widely spaced and not overlapping,
and the antenna tracked at 90 degrees to the roots. Under field
conditions, roots grow at varying angles, varying depths, often
in clumps, and in a heterogeneous medium, resulting in more
complex radar images. We did not assess the impact of the antenna crossing the root at angles other than 90° as would occur
in the real world, but it is likely that the relationship between
the parameters we have extracted and root size would break
down under such conditions. Improvements in antenna design
by using stepped frequency (Jensen and Gregersen 2000),
cross polarization and steerable beams (Utsi 2000) combined
with optimized survey grids and data analysis routines could
advance the use of GPR for tree root studies. By running
transects close to one another and at a range of angles across a
plot a vast amount of information would be obtained that could
be analyzed with the type of algorithms currently used in medical scanners to yield information on root architecture and biomass. Given the absence of other techniques, apart from
excavation, for measuring tree root systems, further study of
GPR for studying tree root systems seems worthwhile.
Acknowledgments
We thank Tim Pippett of Alpha Geoscience Pty. for his assistance collecting the GPR data. This study was funded under program A2 of the
CRC for Greenhouse Accounting.
References
Burrows, W.H., M.B. Hoffmann, J.F. Compton, P.V. Back and L.J.
Tait. 2000. Allometric relationships and community biomass estimates for some dominant eucalypts in Central Queensland woodlands. Aust. J. Bot. 48:707–714.
Butnor, J.R., J.A. Doolittle, L. Kress, S. Cohen and K.H. Johnsen.
2001. Use of ground-penetrating radar to study tree roots in the
southeastern United States. Tree Physiol. 21:1269–2001.
Butnor, J.R, J.A. Doolittle, K.H. Johnsen, L. Samuelson, T. Stokes
and L. Kress. 2003. Utility of ground penetrating radar as a root
biomass survey tool in forest systems. Soil Sci. Soc. Am. J. 67:
1607–1615.
Cairns, M.A., S. Brown, E.H. Helmer and G.A. Baumgardner. 1997.
Root biomass allocation in the world’s upland forests. Oecologia
111:1–11.
Èermák, J., J. Hruska, M. Martinkova and A. Prax. 2000. Urban tree
root systems and their survival near houses analyzed using ground
penetrating radar and sap flow techniques. Plant Soil 219:1–2.
Conyers, L.B. and D. Goodman. 1997. Ground-penetrating radar: an
introduction for archaeologists. AltaMira Press, Walnut Creek,
CA, 232 p.
TREE PHYSIOLOGY VOLUME 24, 2004
MEASURING TREE ROOTS USING GPR
Daniels, D.J. 1996. Surface-penetrating radar. The Institute of Electrical Engineers, London, U.K., 300 p.
Hruska, J., J. Èermák and S. Sustek. 1999. Mapping tree root systems
with ground-penetrating radar. Tree Physiol. 19:125–130.
Jensen, O.K. and O.G. Gregersen. 2000. Stepped-frequency GPR for
line detection using polarization-dependent scattering. In Proc.
Eighth Int. Conf. on Ground Penetrating Radar. Eds. D.A. Noon,
G.F. Stickley and D.Longstaff. SPIE Vol 4084, Washington, DC,
pp 724–730.
Keith, H., D. Barrett and R. Keenan. 2000. Review of allometric relationships for estimating woody biomass for New South Wales, the
Australian Capital Territory, Victoria, Tasmania and South Australia. National Carbon Accounting Tech. Report No. 5b, Australian
Greenhouse Office, Canberra, Australia, 123 p.
Olhoeft, G.R. 2000. Maximizing the information return from ground
penetrating radar. J. Appl. Geophys. 43:175–187.
Pratt, W.K. 1991. Digital image processing. 2nd Edn. Wiley, New
York, 698 p.
Resh, S.C., M. Battaglia, D. Worledge and S. Ladiges. 2003. Coarse
root biomass for eucalypt plantations in Tasmania, Australia:
sources of variation and methods for assessment. Trees 17:
389–399.
Ritson, P. and S. Sochacki. (2003). Measurement and prediction of
biomass and carbon content of Pinus pinaster trees in farm forestry
plantations, south-western Australia. For. Ecol. Manage. 175:
103–117.
1331
Snowdon, P., D. Eamus, P. Gibbons, K. Khanna, H. Keith, J. Raison
and M. Kirschbaum. 2000. Synthesis of allometrics, review of root
biomass and design of future woody biomass sampling strategies.
National Carbon Accounting System Tech. Report No. 17, Australian Greenhouse Office, Canberra, Australia, 137 p.
Sustek, S., J. Hruska, M. Druckmuller and T. Michalek. 1999. Root
surfaces in the large oak tree estimated by image analysis of the
map obtained by the ground penetrating radar. J. For. Sci. 45:
139–143.
Utsi, V. 2000. Development and different modes of operation of a
multireceiver ground penetrating radar system. In Proc. Eighth Int.
Conf. on Ground Penetrating Radar. Eds. D.A. Noon, G.F. Stickley
and D.Longstaff. SPIE Vol. 4084, Washington, pp 351–355.
Vogt, K.A., D.J. Vogt, P.A. Palmiotto, P. Boon, J. O’Hara and H. Asbjornsen. 1996. Review of root dynamics in forest ecosystems
grouped by climate, climatic forest type and species. Plant Soil
187:159–219.
Watson, R.T., I.R. Noble, B. Bolin, N.H. Ravindranath, D.J. Verardo
and D.J. Dokken. 2000. Land use, land-use change and forestry.
Special report of the IPCC, Cambridge University Press, Cambridge, 377 p.
Wielopolski, L., G. Hendry, J. Daniels and M. McGuigan. 2000. Imaging tree root systems in situ. In Proc. Eighth Int. Conf. on Ground
Penetrating Radar. Eds. D.A. Noon, G.F. Stickley and D. Longstaff. SPIE Vol. 4084, Washington, DC, pp 642–646.
TREE PHYSIOLOGY ONLINE at http://heronpublishing.com