StemVolume - and Above-Ground - Biomass - Estimation of - Pine - Trees - From - LiDAR

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1

Stem Volume and Above-Ground Biomass Estimation


of Individual Pine Trees From LiDAR Data:
Contribution of Full-Waveform Signals
Tristan Allouis, Student Member, IEEE, Sylvie Durrieu, Cédric Véga, and Pierre Couteron

Abstract—The diameter at breast height (DBH) is the most to define effective plans for sustainable forest management
extensively measured parameter in the field for estimating stem and climate change mitigation. Field-measured tree diameter
volume and aboveground biomass of individual trees. However, at breast height (DBH) and total height are generally used as
DBH can not be measured from airborne or spaceborne light
detection and ranging (LiDAR) data. Consequently, volume
the main input variables for estimating aboveground biomass
and biomass must be estimated from LiDAR data using other of individual trees through allometric relationships. Forest in-
tree metrics. The objective of this paper is to examine whether ventories are generally based on field sampling approaches, but
full-waveform (FW) LiDAR data can improve volume and biomass spatially explicit field measurements at the forest stand scale
estimation of individual pine trees, when compared to usual dis- are required in order to implement sustainable management
crete-return LiDAR data. Sets of metrics are derived from canopy practices and to perform detailed ecological studies. Therefore,
height model (CHM-only metrics), from the vertical distribution
of discrete-returns (CHM+DR metrics), and from full-waveform improved assessment of how anthropogenic and environmental
LiDAR data (CHM+FW metrics). In each set, the most relevant factors affect the distribution of aboveground biomass [1] has
and non-collinear metrics were selected using a combination of become crucial.
methods using best subset and variance inflation factor, in order Remote sensing data with field control measurements pro-
to produce predictive models of volume and biomass. CHM-only vide effective solutions to assess spatially explicit inventories
metrics (tree height and tree bounding volume [tree height x
crown area] provided volume and biomass estimates of individual
from individual tree to global scale and in inaccessible areas.
trees with an error (mean error standard deviation) of 2% Among the existing remote sensing techniques, LiDAR (light
26% and 15% 49%, which is equivalent to previous studies. detection and ranging) delivers useful information to describe
CHM+FW metrics did not improve stem volume estimates (5% the vertical structure of vegetation [2]–[4] with a lower sensi-
31%), but they increased the accuracy of aboveground biomass es- tivity to signal saturation compared with radar instruments [5].
timates ( 4% 31%). The approach is limited by the delineation Because airborne and spaceborne LiDAR sense the forest from
of individual trees. However, the results highlight the potential
of full-waveform LiDAR data to improve aboveground biomass above, they do not provide direct access to the tree DBH. Con-
estimates through a better integration of branch and leaf biomass sequently, other tree metrics must be derived from LiDAR data
than with discrete-return LiDAR data. so that they can be used in allometric relationships for the esti-
Index Terms—Algorithm design and analysis, forestry, geophys- mation of volume and biomass.
ical signal processing, laser radar, predictive models, remote mon- Stand volume and aboveground biomass estimations at stand
itoring, vegetation mapping. level are mainly based on metrics derived from the signal am-
plitude [6] when large footprint profiler sensors (laser beam
diameter of 10 to 70 m on the ground) are used. The signal
I. INTRODUCTION backscattered by the ground and the canopy can also be inte-
grated. These integrals and their ratio may be used in addition
S TEM volume and aboveground biomass are two com-
monly used parameters for the assessment of forest pro-
ductivity and carbon sequestration rates, and they are needed
to the signal amplitude [7], [8]. Inventories are performed at a
higher spatial resolution using small-footprint LiDAR data (in
tens of centimeters, with a measurement density varying from
0.5 to several per m [9], [10]) acquired by airborne scanner
sensors. At the plot level, volume and biomass are correlated
Manuscript received September 29, 2011; revised January 24, 2012 and June
07, 2012; accepted July 17, 2012. This work is part of the ExFOLIO project with metrics derived from the vertical distribution of echoes
and was realized thanks to the financial support of the CNES (Centre National (height percentiles), maximum height, mean height, and coef-
d’Études Spatiales) and the French Languedoc-Roussillon region. ficient of variation of heights [11], [12]. Studies have also been
T. Allouis is with Irstea, UMR TETIS, 34196 Montpellier Cedex 5, France.
He is also with the L’avion Jaune, 34980 Montferrier-sur-Lez, France (e-mail: performed at the individual tree level, allowing to directly link
[email protected]). field measurements with LiDAR metrics. They generally rely on
S. Durrieu is with Irstea, UMR TETIS, 34196 Montpellier Cedex 5, France. a single-tree detection algorithm for extracting LiDAR metrics,
C. Véga is with the Institut Français de Pondichéry, UMIFRE 21 CNRS-
MAEE, 11, Pondicherry 605 001, India. He is also with Irstea, UMR TETIS,
followed by a selection procedure and a prediction module [13].
34196 Montpellier Cedex 5, France. Some studies have estimated tree biomass using height per-
P. Couteron is with the Institut de Recherche pour le Développement, UMR centiles (e.g., [14], [15]). Popescu [16] assessed biomass using
AMAP, 34398 Montpellier cedex 5, France.
Color versions of one or more of the figures in this paper are available online
total height and crown diameter derived from a canopy height
at http://ieeexplore.ieee.org. model (CHM), and Chen et al. [17] introduced the use of a
Digital Object Identifier 10.1109/JSTARS.2012.2211863 “canopy geometric volume” metric derived from CHM for stem

1939-1404/$31.00 © 2012 IEEE


This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 1. Study site and plots (white dot circles) location. Hemispherical photos show plots condition.

volume estimation. All these research efforts attest to the utility mountainous area were afforested in the late nineteenth century
of measuring the vertical structure of trees including the crown with black pine (Pinus nigra ssp. nigra [Arn.]). Today, stand
dimensions and the total height [18]. densities varies from 100 stems/ha to over 2,000 stems/ha.
The recent development of small-footprint full-waveform The soil on these black marl badlands is affected by intense
LiDAR sensors has provided new opportunities for vegetation erosion, but the afforestation was found to substantially reduce
studies [19]. Unlike usual small-footprint discrete-return Li- erosion by trapping eroded material [23]. Today, the stands
DARs that record only the most significant echoes, this new are over-mature, and management plans must be developed to
generation of sensors records the entire backscattered energy. renew the forest. This study site is particularly suited to the
New developments in signal processing lead to an increased assessment of LiDAR efficiency in various terrain conditions
number of extracted echoes compared with discrete-return because the mean slope is approximately 53% and can locally
systems [20], thus providing additional information such as reach up to 100%.
echo shape or area under the waveform that might be used to
refine vegetation structure measurements [21], [22]. B. Reference Data
In this paper, we examine whether full-waveform LiDAR Field inventory data were collected during December 2007
metrics combined with commonly used LiDAR variables (i.e., on six circular plots with a 15 m radius. The plots are repre-
total tree height and tree crown diameter) can improve volume sentative of several stand development stages. The mean plot
and biomass estimates at the tree level. These estimates were heights ranged from 9 m to 22 m. Thanks to the homogeneous
compared with those performed using the usual variables alone composition of the stand, we could perform our study using 61
or in combination with percentile metrics derived from discrete- reference trees [16], [24].
return LiDAR data. Each plot center was geo-located using a differential GPS
and a total station. The following characteristics were measured
II. MATERIALS for all trees having a diameter at breast height (DBH) greater
than 7 cm: tree position (i.e., X and Y coordinates calculated
using distance and angle measurements from the plot center),
A. Study Site
tree DBH (measured using a tape), tree height and crown
The study area is located in the Haute-Bléone forest base height (measured using a Vertex III ultrasonic clinometer
(44.16 N, 6.32 W) in the southern French Alps (see Fig. 1), [Haglof Inc]). The plot characteristics are summarized in
and close to the city of Digne-les-Bains. These 108 ha of Table I. On two plots, the maximum extent of tree crowns was
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ALLOUIS et al.: STEM VOLUME AND ABOVE-GROUND BIOMASS ESTIMATION OF INDIVIDUAL PINE TREES FROM LIDAR DATA 3

TABLE I
PLOT CHARACTERISTICS. MEAN VALUES OF INDIVIDUAL TREES STANDARD DEVIATION ARE GIVEN

measured in two directions (i.e., North-South and East-West) The dataset was composed of the raw full-waveform signals
using a crown mirror and a tape. (1D intensity profiles digitized at 1 GHz [15 cm]) and the dis-
The reference stem volumes were calculated for each tree crete returns (3D point cloud) extracted from the raw signals by
from ground-measured DBH and total height using the fol- fitting sums of Gaussian functions [20].
lowing equation developed for Pinus nigra within France
by the French Institute for Agricultural Research [25]. The III. METHODS
equation was built using a destructive sampling of 1755 trees An overview of the method is provided in Fig. 2 and a sum-
mainly located in the southeast of France, with DBH ranging mary of retrieved metrics is provided in Table II.
from 4 cm to 30 cm and heights ranging from 10 m to 55 m.
A. Deriving Metrics From the CHM
A 0.5 m CHM resolution was obtained by subtracting the
Digital Terrain Model (DTM) from the Digital Surface Model
(1) (DSM), both resulting from LiDAR point cloud processing.
DSM cells (0.5 m resolution) were assigned the elevation of
where V is the stem volume in cm , H is the total height in cm, the highest echo within it, and an Inverse Distance Weighting
and C is the trunk circumference at breast height in cm. (IDW) interpolation was performed. Due to a lower density of
No allometric equation was available for aboveground ground points, the DTM was computed on a 1 m grid using
biomass (AGB) estimation of black pine in France. Because the ordinary kriging interpolation of last echoes that actually
our study site is within the Mediterranean climate zone, we reached the ground, and identified after an echo classification
used an equation developed for black pine in Spain, which scheme described in Véga et al. [28].
is principally located in the Mediterranean climate zone. The Segmentation was performed on the CHM to extract and char-
model was established by Montero [26] on 253 Pinus nigra acterize individual tree crowns. The tree crown extraction was
trees located in various Spanish regions, with DBH ranging based on the identification of local maxima on a smoothed ver-
from 7 cm to 70 cm. sion of the CHM (using a Gaussian low-pass filter) and the sub-
sequent modeling of elliptical crowns issued from crown dimen-
(2) sion measurements in various directions (see Véga and Durrieu
[29] for technical details) (Fig. 3).
where AGB-level is in kg, and DBH is the trunk diameter at For each crown, the tree total height was es-
breast height in cm. timated by selecting the value of the highest pixel of the
Reference volume and biomass were computed for each in- CHM inside the modeled crown, and the crown projected
dividual tree using the previous equations. Averaged values are area computed from the previously deter-
given in Table I along with other mean tree characteristics. mined elliptical crowns. From these two metrics, we also
derived the tree bounding volume , which is
C. LiDAR Data similar to the “canopy geometric volume” [17]. is
The data acquisition was performed in April 2007 using a the elliptic cylinder of height and base
RIEGL LMS-Q560 system. This sensor is a small-footprint full- . The accuracy of
waveform airborne laser scanner (see Wagner et al. [27] for the metrics derived from the CHM was estimated using
system design and specifications) which was operated at a pulse ground-truth measurements.
rate of 111 kHz with a scan angle of 22.5 . The flight alti-
tude was approximately 600 m above ground level, leading to a B. Deriving Metrics From the Discrete-Return LiDAR Data
footprint diameter of approximately 0.25 m. The average mea- (DR Metrics)
surement density was approximately 5 shots/m (i.e., 5 wave- Height percentiles are related to the vertical distribution of
forms/m ). LiDAR echoes reflected on the vegetation. The elevations of
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 2. Method overview.

Fig. 3. Identification of tree crowns (shade) and crown apex (star) on the 15 m
radius plot 2 (dashed).

discrete-returns located inside each modeled tree crown were


subtracted from the corresponding terrain elevation retrieved
from the DTM. To remove the influence of undergrowth veg- Fig. 4. Sum of full-waveform signals falling within crown boundaries (repre-
etation, only echoes higher than 3 m above the ground were sented by the cylinder). (a) scheme of a tree, (b) signal selection and (c) resulting
cumulative waveform.
considered. According to previous studies (e.g., [14], [15]),
25th , 50th , 75th , 90th and the LiDAR scan angle. Waveforms that were fully located
height percentiles (also called height quantiles), mean height within modeled tree crowns were then summed to obtain a cu-
and coefficient of variation were re- mulative waveform per tree (Fig. 4). The five footprints (30 cm
trieved from the distribution of echoes and were retained as in diameter) per square meter were considered to provide rep-
discrete-return (DR) candidate variables. resentative measurements of the forest structure.
We also derived a vegetation profile from the cumulative
C. Deriving Metrics From Full-Waveform Signals and waveform. The amplitude of the cumulative waveform was
Vegetation Profiles (FW Metrics) first corrected to account for the reflectance difference between
The raw small-footprint waveforms were georeferenced in a ground and vegetation. According to an analysis of the signal
global coordinate system using the GPS position of the plane dynamics (transmitted and received amplitude ratio) performed
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ALLOUIS et al.: STEM VOLUME AND ABOVE-GROUND BIOMASS ESTIMATION OF INDIVIDUAL PINE TREES FROM LIDAR DATA 5

TABLE II
SUMMARY OF THE METRICS DERIVED AT THE TREE SCALE USED IN THIS STUDY

on the ground and vegetation single echoes, the average ground where is the vegetation profile amplitude at range from
reflectance was 0.52 and the vegetation reflectance was 0.33 on the plane with , and is the recorded ampli-
our study site. Then, the cumulative waveform was corrected tude as a function of range. The correction is computed for each
from the laser beam attenuation along its travel through the interval between two successive ranges from the atmosphere to
vegetation. To achieve this light attenuation correction we used signal extinction .
a model previously adapted [33], [34] from the MacArthur Cumulative waveforms and vegetation profiles provide com-
and Horn equation [32]. This model transforms the cumulative plementary information to describe the structure of tree crowns.
waveform into a vegetation profile to provide an accurate Therefore, vegetation profiles do not account for absolute
description of the vertical structure of a tree: signal amplitude and large tree crowns (in area projected on
the ground) cannot be distinguished from small ones. However,
vegetation profiles provide an improved description of the
(3) vertical structure of a tree. Cumulative waveforms allow us to
distinguish crown width because the signal amplitude tends to
increase with the crown projected area. However, cumulative
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

CHM+FW metrics, predictive linear models were computed


using an all-possible-subset regression. The objective is to
test all possible subsets of the set of metrics to find the linear
combination of metrics that provides the best estimation of
volume (1) or biomass (2) according to the Bayesian informa-
tion criterion (BIC) [35]:

(4)

where is the number of parameters in the model (including


the intercept), is the number of observations, and is the
maximized value of the likelihood function.
By measuring the model goodness-of-fit while imposing a
penalty for increasing the number of explanatory variables, the
BIC addresses the problem of an artificially increased likelihood
when a new variable is added in the model. The all-possible-
subset regression was automatically performed thanks to the R
Fig. 5. The cumulative waveform (left), the corresponding vegetation profile
after correction of laser attenuation (right), and the derived metrics.
software function called ‘bestglm’ [36] using the whole set of
observations (61 trees). The set of variables in the model having
the lowest BIC was retained as the “best subset”.
waveforms do not provide a normalized vertical distribution of
While indicators related to the same tree inevitably have some
the vegetation structure. inter-dependence, the collinearity between metrics can lead to
Several metrics were consequently derived from both cu- model over-fitting, and this is liable to lead to weak predictions
mulative waveforms and vegetation profiles. These metrics when using a new dataset. Consequently, collinearity in a model
are summarized in Table II and illustrated in Fig. 5. was verified for each of the all-possible-subset models by com-
was determined as the maximum amplitude of each cumulative puting the Variance Inflation Factor (VIF). To achieve this, each
signal (except the ground component, i.e. higher than 3 m above metric was expressed as a linear combination of the other met-
the last LiDAR echo), which was consequently related to the rics in the subset, and the VIF of each metric was computed as
number of signals located within tree boundaries.
was the height of the maximum profile amplitude (except
the ground component) and is assumed to correspond to the (5)
height where the crown is either the largest or the densest.
The crown base height corresponds to the where is the coefficient of determination of the linear model
greatest decrease in the profile, detected as the position of the developed for each metric of a given subset.
minimum value of the profile derivative [34]. Signal integrals, The set of variables having the lowest BIC and with all vari-
corresponding to areas under the curve, were computed for ables having a VIF lower than 10 [37] was retained as the “best
both the cumulative waveform and profile, from a few meters independent subset”.
above the tree to the ground ( and ). Similarly, 2) Building Final Models and Assessing the Estimates Accu-
and were computed from a few meters above the racy: A cross-validation was performed to build and assess the
tree to three meters above ground to remove the influence of accuracy of the final predictive models of volume and biomass
undergrowth. and were then computed as ratios using metrics selected in the “best subset” and in the “best in-
between and the energy reflected on the ground (area dependent subset”. In compliance with the rules of the leave-
under the curve, computed from the position of the maximum one-out cross-validation method (LOOCV) [38], a single ob-
amplitude to 15 cm above the latter position) and the servation was retained from the data set as the validation data,
and profile ground component, respectively. The result is a and the remaining observations were used as training data. The
crown porosity index. process was repeated to use each observation once as the vali-
dation data, thus assessing the ability of the model to fit inde-
pendent measurements.
D. Building and Comparing Models for Volume and Biomass The mean percentage error (MPE) was used to express in a
Assessment generic percentage term the error of volume and biomass esti-
Regression models were computed to estimate volume mations:
and biomass using metrics derived from the CHM alone
(CHM-only) or combined either with discrete-return metrics (6)
(CHM+DR) or with full-waveform metrics (CHM+FW). The
estimates were then compared with ground-truth values ((1)
and (2)). In accordance with some previous research [11], [12], where is the actual value and is the estimated value for
we developed linear regression models. each step of the LOOCV.
1) Metric Selection: Thanks to the relatively low number For each step of the LOOCV a single value of the adjusted co-
of candidate variables among the CHM-only, CHM+DR or efficient of determination was also calculated, in order
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ALLOUIS et al.: STEM VOLUME AND ABOVE-GROUND BIOMASS ESTIMATION OF INDIVIDUAL PINE TREES FROM LIDAR DATA 7

Fig. 6. Power regression models relating field-measured DBH and (a) crown area retrieved from the CHM and (b) tree height retrieved from the CHM. The gray
curves were fitted using (a) the 4/3 exponent of West et al. (2009) and (b) the 2/3 exponent of the same study.

to evaluate the model’s goodness-of-fit while taking into ac- LiDAR-derived metrics to predict tree biomass from airborne
count the differences in the number of explanatory variables in or spaceborne platforms.
the model:
C. Volume Estimation Models
The models’ characteristics are summarized in Table III.
1) CHM-Only Metrics: The best-subset model was of the
(7) following form:

(8)
where is the observed value, is the modeled value, is the
number of metrics in the model and is the number of This model, whose coefficients were averaged from the 61
pairs. The final value was obtained as the mean of each models developed using the LOOCV method, had an average
value. of 0.93, leading to an error of 3% 21%, but it
Model coefficients were also averaged in order to produce presented a collinearity between and
final predictive models. (VIF of 19.7 and 31.4, respectively). The best independent
subset, which was the second best-subset, retained and
IV. RESULTS :

A. Accuracy of CHM Metrics


(9)
Tree heights were retrieved from the CHM with a bias of
0.09 m and a standard deviation of 0.57 m. The corresponding This latter model had the same (0.93), but it provided
RMSE was also 0.57 m. The mean crown diameter was esti- slightly scattered estimates (2% 26%).
mated with a 1.48 m RMSE on two reference plots. 2) CHM-DR Metrics: The best-subset model had the fol-
lowing form:
B. LiDAR Metrics in Allometric Models
Theoretically, the crown area was found by West et al. [30]
to equate to , while empirical data in a highly diverse
neo-tropical forest yielded exponent values within the range (10)
of 1.33–1.39 [31], which is close to the theoretical value of
4/3. In our case, a 1.77 exponent was found instead of 1.33 This model had an of 0.94 and an estimation error of
(4/3) (Fig. 6) by fitting a power model on actual data. However, . In addition to and , a
the strong relationship between crown dimension and the DBH strong collinearity was found between and
was confirmed . The relationship between tree (102.5 and 110.2, respectively). The best independent subset
height and DBH was also verified using a power model (Fig. 6) was the 19th ranked subset. This subset had a quite similar
, but the exponent value of 0.92 was not very close (0.93) but provided slightly more accurate estimates
to the theory of West et al. [30] (0.67) or to the empirical data (1% 22%), and it had the following form, where
of Muller-Landeau et al. [31] (0.59–0.60). Such relationships was replaced by compared with the CHM-only model
are a consistent property of forest trees, even if species- and ((11)):
context-specific features can be expected to modulate the expo-
nent values. Consequently, it makes sense to consider tree height
and crown area as analogous to height and DBH, and to use (11)
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

3) CHM-FW Metrics: The best-subset model had the fol- 3) CHM-FW Metrics: The best-subset model had the fol-
lowing form: lowing form:

(17)
(12)
This model had an of 0.91, and it provided estimates
This model had an of 0.95 and provided estimates with with an error of 12% 54%, but it contained collinear vari-
an error of 1% 27%, but it contained collinear variables ables ( and ). The third ranked subset
( and ) for the first 520 ranked subsets. was selected as the best independent subset, with the following
The 521st ranked subset was selected as the best independent form, providing a quite similar (0.90) but a lower error
subset, with the following form, and provided a slightly lower ( 4% 31%):
(0.93) as well as less accurate estimates (5% 31%):

(18)
(13)
4) Comparison of the Different Models: We first noticed
4) Comparison of the Different Models: We first noticed that was selected in every model and was
that the tree bounding volume metric was selected not selected in only one model. The is clearly higher
in every estimation model. Despite the similar in the for the CHM+FW models (0.90–0.91) than for the CHM-only
different models (ranging from 0.93 and 0.95), the best esti- (0.87) and CHM+DR models (0.87–0.88). The DR metrics did
mates were made using either CHM-only metrics, if we con- not seem to increase the accuracy of biomass estimations. The
sider the best subset of variables, or with CHM+DR metrics, most accurate estimates were performed with CHM+FW met-
if we consider the best independent subset. However, the ad- rics using the best independent subset, mainly due to the use of
ditional DR metrics did not seem to increase the accuracy of metric. Both bias and dispersion are lower when using
the stem volume estimations. Additional full-waveform LiDAR in the best independent subset (bias of 4% vs. 15%
data led to slightly worse estimates compared with the CHM- without and standard deviation of 31% vs. 49%). How-
only models. ever, the same tendency was not observed using the best subset.
The decrease in MAPE was 9% compared with both CHM-
D. Biomass Estimation Models only and CHM+DR models when using the best independent
subset. The decrease was 4% and 6% compared with the CHM-
The models’ characteristics are summarized in Table III. only and CHM+DR models, respectively, when using the best
1) CHM-Only Metrics: The best-subset model had the fol- subset.
lowing form:

V. DISCUSSION
(14)
This model, whose coefficients were averaged from the 61 A. Usefulness of Full-Waveform LiDAR Data for Volume and
models developed using the LOOCV method, had an average Biomass Assessment at the Tree Level
of 0.87, leading to an error of 15% 49%. This model
The accuracy of our estimates using CHM-only or CHM+DR
did not produce any collinearity between variables and this best
metrics are in accordance with previous studies, having gener-
subset was also the best independent subset.
ated a 15% RMSE for volume estimation [15] and a 33% RMSE
2) CHM-DR Metrics: The best-subset model had the fol-
for biomass estimation [16].
lowing form:
In this study, we tested whether the use of additional metrics
related to the vertical structure of trees could improve volume
and biomass estimates performed using CHM-only metrics. The
(15) tree bounding volume was selected in all devel-
oped models, and therefore seems to be even more useful for
This model had an of 0.88 and an estimation error of predictions of volume and biomass, which tends to confirm the
12% 47%. A strong collinearity was observed between findings of Chen et al. [17]. Additional metrics derived from
and (VIF of 96.8 and 109.4, respectively). discrete-return data did not improve the results. Additional full-
The best independent subset had quite similar (0.87) waveform metrics did not have a positive influence on volume
and estimation error ( 15% 49%); it had the following form, estimates, but they improved the accuracy of biomass estima-
where the collinear variables were replaced by : tions. Thus, we can consider that stem volume is well described
by crown dimension (related to DBH, see Section IV-B), tree
height metrics ( or ) or both . But
(16) the biomass estimates are more accurate using , which is
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ALLOUIS et al.: STEM VOLUME AND ABOVE-GROUND BIOMASS ESTIMATION OF INDIVIDUAL PINE TREES FROM LIDAR DATA 9

TABLE III
RESULTS OF METRICS SELECTION AND MODEL ACCURACY ASSESSMENT USING METRICS FROM THE “BEST-SUBSET” ( , I.E. WITH LOWEST
BAYESIAN INFORMATION CRITERION) AND FROM THE “BEST INDEPENDENT SUBSET” (BEST-SUBSET WITH ALL VARIABLES HAVING A VARIANCE INFLATION
FACTOR (VIF) 10). ADJUSTED AND MEAN PERCENTAGE ERROR (MPE) ARE AVERAGED RESULTS OF THE CROSS-VALIDATION METHOD (61
VALIDATIONS DATA PER MODEL)

related to the vertical structure of a tree and consider branch and bust estimator due to excessive variability among trees. An error
leaf biomass more effectively. in porosity ratios ( and ) may have occurred due
However, was the only selected full-waveform metric. to the presence or absence of undergrowth, which modified the
The crown height was not considered as a ro- ground signal amplitude. Profile integrals ( and )
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

were not selected, mainly because they were only related to the Draix through the ORE program, and special thanks go to Lau-
vertical distribution of the vegetation and did not account for rent Albrech, from UMR TETIS-Cemagref, for his help with the
changes in crown dimensions (projected area and crown height). collection of field data.
Signal integrals , which gave information on both the
vertical and horizontal (amplitude) distribution of the vegeta- REFERENCES
[1] T. Urquiza-Haas, P. M. Dolman, and C. A. Peres, “Regional scale vari-
tion density proved to be a more valuable parameter for biomass
ation in forest structure and biomass in the Yucatan Peninsula, Mexico:
estimation. Effects of forest disturbance,” Forest Ecol. Manag., vol. 247, no. 1–3,
pp. 80–90, 2007.
B. Limits of the Method Regarding Stand Characteristics [2] M. A. Wulder, C. W. Bater, N. C. Coops, T. Hilker, and J. C. White,
“The role of LiDAR in sustainable forest management,” Forest.
For this study, an automatic tree-to-tree comparison between Chron., vol. 84, no. 6, pp. 807–826, 2008.
reference and estimated data was performed. Consequently, [3] M. van Leeuwen and M. Nieuwenhuis, “Retrieval of forest structural
parameters using LiDAR remote sensing,” Eur. J. Forest Res., vol. 129,
we needed a perfect match between automatically segmented pp. 749–770, 2010.
crowns and field-measured tree positions. For that reason, [4] J.-M. Monnet, J. Chanussot, and F. Berger, “Support Vector Regression
we only retained plots on which the algorithm presented in for the Estimation of Forest Stand Parameters Using Airborne Laser
Scanning,” IEEE Geosci. Remote Sens. Lett., vol. 8, no. 3, pp. 580–584,
Véga and Durrieu [29] offered a 100% crown identification 2011.
success rate, in plots of up to 500 stems/ha (the reference [5] M. L. Imhoff, “Radar backscatter and biomass saturation: Ramifica-
plots used in this paper had an average of 140 trees/hectare). tions for global biomass inventory,” IEEE Trans. Geosci. Remote Sens.,
vol. 33, no. 2, pp. 511–518, 1995.
Satisfactory success rate can occur in higher density plots, but [6] L. I. Duncanson, K. O. Niemann, and M. A. Wulder, “Integration of
crown omission increased because suppressed trees were not GLAS and Landsat TM data for aboveground biomass estimation,”
detected. It would have been even more difficult to identify Can. J. Remote Sens., vol. 36, no. 2, pp. 129–141, 2010.
[7] J. E. Means, S. A. Acker, D. J. Harding, J. B. Blair, M. A. Lefsky, W.
understory trees in multilayered forests, even if some studies B. Cohen, M. E. Harmon, and W. A. McKee, “Use of large-footprint
were performed using LiDAR point distributions when very scanning airborne LiDAR to estimate forest stand characteristics in the
dense point clouds are available (e.g., Jaskierniak et al. [39]). western cascades of Oregon,” Remote Sens. Environ., vol. 67, no. 3,
pp. 298–308, 1999.
In complex or dense stands, where the identification of in- [8] Q. Chen, “Retrieving vegetation height of forests and woodlands over
dividual tree crowns is not possible, the estimation of biomass mountainous areas in the Pacific Coast region using satellite laser al-
directly at the plot level could provide an alternative solution. timetry,” Remote Sens. Environ., vol. 114, no. 7, pp. 1610–1627, 2010.
[9] M. Maltamo, J. Hyyppä, and J. Malinen, “A comparative study of the
In this case, biomass and volume estimations can be performed use of laser scanner data and field measurements in the prediction of
based on the vertical distribution of LiDAR discrete returns (see crown height in boreal forests,” Scand. J. Forest Res., vol. 21, no. 3,
[12]), or from full-waveform signals summed at the plot level pp. 231–238, 2006.
[10] S. Magnussen, E. Næsset, and T. Gobakken, “Reliability of LiDAR
for simulating a large-footprint signal (which has been success- derived predictors of forest inventory attributes: A case study with
fully used to predict biomass; see [40]). However, the impact Norway spruce,” Remote Sens. Environ., vol. 114, no. 4, pp. 700–712,
of topography on the signal profile is an issue that must be ad- 2010.
[11] J. E. Means, S. A. Acker, B. J. Fitt, M. Renslow, L. Emerson, and C.
dressed before deriving metrics from vegetation profiles to as- J. Hendrix, “Predicting forest stand characteristics with airborne scan-
sess volume and biomass. To that aim, an analysis of a DTM ning LiDAR,” Photogramm. Eng. Remote Sens., vol. 66, no. 11, pp.
derived from last echoes to assess the local slope could be con- 1367–1371, 2000.
[12] Y. Li, H. Andersen, and R. McGaughey, “A Comparison of statistical
sidered prior to summing full-waveform signals [41]. methods for estimating forest biomass from light detection and ranging
data,” Western J. App. Forest., vol. 23, no. 4, pp. 223–231, 2008.
[13] M. Dalponte, L. Bruzzone, and D. Gianelle, “A system for the es-
VI. CONCLUSION timation of single-tree stem diameter and volume using multireturn
Our study confirmed the usefulness of tree metrics derived LiDAR data,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 7, pp.
2479–2490, 2011.
from a CHM for stem volume and aboveground biomass predic- [14] Z. J. Bortolot and R. H. Wynne, “Estimating forest biomass using small
tions based on airborne LiDAR data. After building estimation footprint LiDAR data: An individual tree-based approach that incorpo-
models from CHM-only metrics, CHM combined with discrete rates training data,” ISPRS J. Photogramm., vol. 59, no. 6, pp. 342–360,
2005.
return metrics (CHM+DR) and CHM combined with full-wave- [15] S. Tonolli, M. Dalponte, L. Vescovo, M. Rodeghiero, L. Bruzzone, and
form metrics (CHM+FW), we concluded that the tree bounding D. Gianelle, “Mapping and modeling forest tree volume using forest
volume, obtained by multiplying the crown area by tree height, inventory and airborne laser scanning,” Eur. J. Forest Res., vol. 130,
no. 4, pp. 569–577, 2010.
is one of the most useful metrics for volume and biomass estima- [16] S. C. Popescu, “Estimating biomass of individual pine trees using air-
tions. We also observed that the use of additional FW metrics did borne LiDAR,” Biomass Bioenerg., vol. 31, no. 9, pp. 646–655, 2007.
not seem to improve volume estimation but led to an improve- [17] Q. Chen, P. Gong, D. Baldocchi, and Y. Q. Tian, “Estimating basal area
and stem volume for individual trees from LiDAR data,” Photogramm.
ment of biomass estimation compared with models developed Eng. Remote Sens., vol. 73, no. 12, pp. 1355–1365, 2007.
using CHM-only or CHM+DR metrics. The use of FW met- [18] M. Dalponte, N. C. Coops, L. Bruzzone, and D. Gianelle, “Analysis on
rics allowed for a better accounting of the aboveground biomass the use of multiple returns LiDAR data for the estimation of tree stems
volume,” IEEE J. Sel. Topics Appl. Earth Observ., vol. 2, no. 4, pp.
made up of branches and leaves. 310–318, 2009.
[19] C. Mallet and F. Bretar, “Full-waveform topographic LiDAR: State-of-
the-art,” ISPRS J. Photogramm., vol. 64, no. 1, pp. 1–16, 2009.
ACKNOWLEDGMENT [20] A. Chauve, C. Vega, S. Durrieu, F. Bretar, T. Allouis, M. Pierrot-De-
The authors would like to thank GIS Draix for providing the seilligny, and W. Puech, “Advanced full waveform LiDAR data echo
detection: Assessing quality of derived terrain and tree height models
full-waveform LiDAR data and for helping with the ground in an alpine coniferous forest,” Int. J. Remote Sens., vol. 30, no. 19, pp.
truth surveys. They are grateful to INSU for its support to GIS 5211–5228, 2009.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ALLOUIS et al.: STEM VOLUME AND ABOVE-GROUND BIOMASS ESTIMATION OF INDIVIDUAL PINE TREES FROM LIDAR DATA 11

[21] J. Heinzel and B. Koch, “Exploring full-waveform LiDAR parameters [41] T. Allouis, S. Durrieu, and P. Couteron, “A new method for incorpo-
for tree species classification,” Int. J. Appl. Earth Observ. Geoinform., rating hillslope effects to improve canopy-height estimates from large-
vol. 13, no. 1, pp. 152–160, 2011. footprint LiDAR waveforms,” IEEE Geosci. Remote Sens. Lett., vol.
[22] J. D. Muss, D. J. Mladenoff, and P. A. Townsend, “A pseudo-wave- 9, no. 4, pp. 730–734, 2012.
form technique to assess forest structure using discrete LiDAR data,”
Remote Sens. Environ., vol. 115, no. 3, pp. 824–835, 2011.
[23] F. Rey, “Influence of vegetation distribution on sediment yield in
forested marly gullies,” Catena, vol. 50, no. 2–4, pp. 549–562, 2003. Tristan Allouis received an Engineer’s degree
[24] D. Kwak, W. Lee, H. Cho, S. Lee, Y. Son, M. Kafatos, and S. Kim, “Es- (Master degree) in information technology and
timating stem volume and biomass of Pinus koraiensis using LiDAR electronics from the esiea (École Supérieure d’In-
data,” J. Plant Res., vol. 123, no. 4, pp. 421–432, 2010. formatique Électronique Automatique) in 2008,
[25] J. Toth, J. Turrel, and M. Turrel, “La productivité du Pin noir d’Autriche and a Ph.D. degree in Environmental Science from
dans le Sud-Est de la France,” Revue forestière française, vol. 2, pp. AgroParisTech in 2011. He was a finalist of the stu-
111–121, 1983. dent paper contest at the 2010 IEEE Geoscience and
[26] G. Montero, “Cuantificacion de la biomasa forestal aerea y radical de Remote Sensing Symposium. His research activities
distintas especies arboreas,” Montes y energyıas renovables. Ponencias focuses on developing new methods to process both
y Comunicaciones Santiago de Compostela, pp. 115–131, 2004. bathymetric and topographic LiDAR data. He is
[27] W. Wagner, A. Ullrich, V. Ducic, T. Melzer, and N. Studnicka, currently with a private company (L’Avion Jaune
“Gaussian decomposition and calibration of a novel small-footprint SARL) where he is developing an ultra-light laser scanner system for drones
full-waveform digitising airborne laser scanner,” ISPRS J. Pho- and other ultra-light aircrafts.
togramm., vol. 60, no. 2, pp. 100–112, 2006.
[28] C. Véga, S. Durrieu, and T. Allouis, “A LiDAR filtering algorithm for
terrain modeling in complex forested environments,” Comput. Geosci., Sylvie Durrieu received her graduate degree (1989)
vol. 44, pp. 31–41, 2012. and her Ph.D. (1994) from the ENGREF (National
[29] C. Véga and S. Durrieu, “Multi-level filtering segmentation to mea- School for Agricultural Engineering, Water and
sure individual tree parameters based on LiDAR data: Application to Forests), a French Post-graduate Institute in Science
a mountainous forest with heterogeneous stands,” Int. J. Appl. Earth and Engineering. From 1994 to 1999 she worked
Observ. Geoinform., vol. 13, no. 4, pp. 646–656, 2011. at the French National Forest Inventory (IFN) and
[30] G. B. West, B. J. Enquist, and J. H. Brown, “A general quantitative was in charge of the remote sensing R&D activities.
theory of forest structure and dynamics,” Proc. Nat. Acad. Sci. USA, Since 1999 she has been working at Irstea (National
vol. 106, no. 17, pp. 7040–7045, 2009. Research Institute of Science and Technology for
[31] H. C. Muller-Landau, R. S. Condit, J. Chave, S. C. Thomas, S. A. Environment and Agriculture) as a researcher in the
Bohlman, S. Bunyavejchewin, S. Davies, R. Foster, S. Gunatilleke, TETIS lab. Her research interest is the use of remote
N. Gunatilleke, K. E. Harms, T. Hart, S. P. Hubbell, A. Itoh, A. R. sensing technologies, mainly optical imagery and both terrestrial and airborne
Kassim, J. V. LaFrankie, H. S. Lee, E. Losos, J. Makana, T. Ohkubo, R. LiDAR, to improve forest ecosystem monitoring and management. She is also
Sukumar, I. Sun, M. N. Nur Supardi, S. Tan, J. Thompson, R. Valencia, involved in projects aiming at developing a spaceborne vegetation LiDAR
G. V. Muñoz, C. Wills, T. Yamakura, G. Chuyong, H. S. Dattaraja, S. system.
Esufali, P. Hall, C. Hernandez, D. Kenfack, S. Kiratiprayoon, H. S.
Suresh, D. Thomas, M. I. Vallejo, and P. Ashton, “Testing metabolic
ecology theory for allometric scaling of tree size, growth and mortality Cédric Véga received a master degree in Ecology
in tropical forests,” Ecol. Lett., vol. 9, no. 5, pp. 575–588, 2006. from Paul Sabatier University (Toulouse, France)
[32] R. H. MacArthur and H. S. Horn, “Foliage profile by vertical measure- and a Ph.D. degree in Environmental Sciences
ments,” Ecology, vol. 50, no. 5, pp. 802–804, 1969. from UQAM, Montréal, Canada. He is currently
[33] M. A. Lefsky, D. Harding, W. B. Cohen, G. Parker, and H. H. Shugart, heading the Laboratory of Applied Informatics and
“Surface LiDAR remote sensing of basal area and biomass in decid- Geomatics at the “Institut Français de Pondichéry”,
uous forests of eastern Maryland, USA,” Remote Sens. Environ., vol. India. His research interests lie in the area of remote
67, no. 1, pp. 83–98, 1999. sensing of forest ecosystems, particularly using
[34] T. Allouis, S. Durrieu, J.-S. Bailly, P. Chazette, J. Cuesta, C. Véga, P. LiDAR, photogrammetry and very high resolution
Flamant, and P. Couteron, “Potential of an ultraviolet, medium-foot- optical imagery. He is particularly focusing on
print lidar prototype for retrieving forest structure,” ISPRS J. Pho- developing methods for both monitoring forest
togramm., vol. 66, no. 6, pp. S92–S102, 2011. structure and dynamics, and estimating forest biophysical parameters.
[35] Neter, Wasserman, and Kutner , Applied Linear Statistical Models, 2nd
ed. New York: Irwin, 1985.
[36] A. McLeod and C. Xu, Package ‘bestglm’ ver. 0.31. [Online]. Avail- Pierre Couteron received his Ph.D. in Tropical
able: http://cran.r-project.org/web/packages/bestglm/, 2010 Ecology from the Paul Sabatier University of
[37] R. M. O’Brien, “A caution regarding rules of thumb for variance infla- Toulouse III (France) in 1998. In 1988–2002 he was
tion factors,” Qual. Quant., vol. 41, no. 5, pp. 673–690, 2007. Assistant Professor in the tropical forestry teaching
[38] R. R. Picard and R. D. Cook, “Cross-validation of regression models,” department of ENGREF, in Montpellier (France),
J. Amer. Statist. Assoc., vol. 79, no. 387, pp. 575–583, 1984. and was Head of this department in 2002–2005. In
[39] D. Jaskierniak, P. N. Lane, A. Robinson, and A. Lucieer, “Extracting 2005–2008 he was Head of the Ecology Department
LiDAR indices to characterise multilayered forest structure using mix- of the French Institute of Pondicherry (India). In
ture distribution functions,” Remote Sens. Environ., vol. 115, no. 2, pp. 2008, he became Directeur de Recherche of IRD at
573–585, 2011. the AMAP lab, which is associated to Montpellier II
[40] M. A. Lefsky, W. B. Cohen, D. J. Harding, G. G. Parker, S. A. Acker, University and devoted to Botany and Plant/vegeta-
and S. T. Gower, “LiDAR remote sensing of above-ground biomass tion modeling. In 2011 he became Head of AMAP (50 permanent scientists).
in three biomes,” Global Ecol. Biogeogr., vol. 11, no. 5, pp. 393–399, His main scientific topics are modeling for community ecology and vegetation
2002. dynamics, image texture analysis, spatial statistics, tropical vegetation.

You might also like