Generating Pit-free Canopy Height

Models from Airborne Lidar

Anahita Khosravipour, Andrew K. Skidmore, Martin Isenburg, Tiejun Wang, and Yousif A. Hussin

Abstract
Canopy height models (CHMs) derived from lidar data have calculate a raster CHM from the first return lidar points and
been applied to extract forest inventory parameters. How- then extract local maxima from that raster CHM (Hyyppä et al.,
ever, variations in modeled height cause data pits, which 2008; Lim et al., 2003b).
form a challenging problem as they disrupt CHM smooth- The main challenges faced in treetop detection are com-
ness, negatively affecting tree detection and subsequent mission errors (falsely detected trees) and omission errors
biophysical measurements. These pits appear where laser (undetected trees) (Hosoi et al., 2012; Pouliot et al., 2005).
beams penetrate deeply into a tree crown, hitting a lower These errors are mainly attributed to natural variation in tree
branch or the ground before producing the first return. In crown size (Pitkänen et al., 2004) as well as to height irregu-
this study, we develop a new algorithm that generates a larities within individual tree crowns in the input CHM (Sol-
pit-free CHM raster, by using subsets of the lidar points to berg et al., 2006). To address natural variation in crown size,
close pits. The algorithm operates robustly on high-density researchers have developed processing methods that adapt to
lidar data as well as on a thinned lidar dataset. The evalu- the crown (object) size. Pitkänen et al. (2004) developed and
ation involves detecting individual trees using the pit-free tested three different adaptive methods for individual tree
CHM and comparing the findings to those achieved by us- detection based on canopy differences. Wulder et al. (2000)
ing a Gaussian smoothed CHM. The results show that our proposed the use of a local maxima filter with variable win-
pit-free CHMs derived from high- and low-density lidar data dow sizes. However, if the selected window size is smaller
significantly improve the accuracy of tree detection. or larger than the crown size, then the commission or omis-
sion error, respectively, will increase. In order to select the
correct window size, Popescu and Wynne (2004) introduced
Introduction an adaptively varying window technique, based on the idea
The use of airborne Light Detection and Ranging (lidar) has that a moving local maxima filter should be adjustable to an
been increasing in forestry. Lidar is capable of providing appropriate width to account for different crown sizes.
accurate three-dimensional information on forest structure To address irregularities in crown height, a number of
(Lim et al., 2003a), contributing significantly to the improved researchers have suggested pre-processing CHMs to reduce
accuracy of forest inventories (Magnussen et al., 2010; Yu commission and omission errors (Bortolot and Wynne, 2005;
et al., 2011) and subsequent biophysical parameters such as Brandtberg et al., 2003; Chen et al., 2006; Solberg et al., 2006).
biomass (Nelson et al., 1988; Popescu, 2007). Irregularities in canopy surface elevation, also called “data
Typically, a lidar-derived Canopy Height Model (CHM) or a pits,” form a challenging problem due to their disruptive
normalized Digital Surface Model (nDSM) is used for extract- influence on a CHM, reducing accuracy in tree detection and
ing relevant forest inventory information, such as detecting subsequent biophysical measurements (Ben-Arie et al., 2009;
single trees for subsequent height estimation and crown delin- Gaveau and Hill, 2003; Zhao et al., 2009). For example, Sham-
eation (Bortolot and Wynne, 2005; Forzieri et al., 2009). The soddini et al. (2013) indicated that data pits may significantly
CHM represents absolute canopy height above ground, and it affect the estimation of structural forest parameters, especially
is typically calculated by interpolating the first return lidar basal area and stand volume. Since the processing of raw
points and determining their height above a digital terrain lidar point clouds into a meaningful raster is a composition
model. (Hyyppä et al., 2008; Van Leeuwen et al., 2010). Tree of many different procedures, there is no unified agreement
height measurement and crown delineation mainly rely on on the cause of data pits. Axelsson (1999) found that some in-
the identification of local maxima, with each local maximum formation from raw point clouds with similar x-y coordinates
corresponding to the location of an individual treetop and the and different z values is lost when the points are interpolated
surrounding segments forming the tree crown (Véga and Dur- into a raster. Such lost data become significant when multiple
rieu, 2011). Therefore, to be able to extract relevant structural echoes are registered in a forested area. Ben-Arie et al. (2009)
parameters of trees (e.g., tree height) the correct location of and Véga and Durrieu (2011) stated that the problem of data
single trees in the CHM is of fundamental importance (Chen et pits was due to laser scanning processing and/or postprocess-
al., 2006; Persson et al., 2002; Yao et al., 2012). While some ing of lidar point clouds. Data pits may also occur during
researchers have tried to find local maxima directly in the li- classification of lidar point clouds into ground and non-
dar points (Li et al., 2012), most operational users of lidar first ground points when creating a Digital Surface Model (DSM) or
a Digital Terrain Model (DTM), depending on the classification
Anahita Khosravipour, Andrew K. Skidmore, Tiejun Wang,
Yousif A. Hussin are with the Department of Natural Resourc- Photogrammetric Engineering & Remote Sensing
es, Faculty of Geo-Information Science and Earth Observa- Vol. 80, No. 9, September 2014, pp. 863–872.
tion, University of Twente, P.O. Box 217, 7500 AA Enschede, 0099-1112/14/8009–863
The Netherlands ([email protected]). © 2014 American Society for Photogrammetry
Martin Isenburg is with Rapidlasso GmbH, Friedrichshafener and Remote Sensing
Straße 1, 82205 Gilching, Germany. doi: 10.14358/PERS.80.9.863

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING S e pte m ber 2014 863

technique and lidar point density (Kraus and Pfeifer, 1998). or less). Although less accurate, these low density data are
Leckie et al. (2003) called such data pits “holes” and found used at canopy, as well as at individual tree level (Ke et al.,
them to be caused by merging different lidar flight lines and 2010; Popescu, 2007; Yu et al., 2011). It is, therefore, useful to
by laser beams penetrating through canopy branches and develop a pit removal algorithm that generates accurate CHMs
foliage to the ground below. Persson et al. (2002) stated that without data pits for both high and low lidar point densities.
lidar penetration caused “large height variations” within In this study, we present a new “pit-free” algorithm able
single tree crowns, creating difficulty in recognizing separate to create a pit-free CHM raster. The specific objectives of this
crowns. Gaveau and Hill (2003) reported the penetration of study are (a) to evaluate the applicability of the algorithm at
laser beams into a tree crown to be the cause of canopy height different lidar point densities, (b) to evaluate pit-free CHMs
underestimation. Zhao et al. (2013) described data pits as by visual comparison with smoothed CHMs derived with a
“invalid values” or “abnormal elevation changes”, which standard Gaussian smoothing technique (Dralle and Rudemo,
formed unnatural gray or black holes in CHM images. Many 1996), and (c) to assess and compare the accuracy of indi-
studies have claimed that lidar-derived tree parameters (e.g., vidual tree detection using both the smoothed and the pit-free
height and crown diameter) may be misinterpreted due to the CHMs. To our knowledge, this is the first study that rigorously
pits present in the CHM image (Gaveau and Hill, 2003; Persson investigates how the removal of data pits from CHMs affects
et al., 2002). treetop detection.
A number of studies have recommended image smooth-
ing, using methods such as a mean, median, or Gaussian filter
to reduce the data pits (Brandtberg et al., 2003; Hosoi et al., Materials
2012; Persson et al., 2002; Yu et al., 2011). Typically, im-
Study Area
age smoothing consists of a two-dimensional (2D) isotropic
The Bois Noir (black wood) forest forms part of the Barcelon-
kernel function, as widely used on 2D remote sensing im-
nette basin, which is located in the southern French Alps
ages (Dralle and Rudemo, 1996) in order to remove “impulse
(44°23'N, 06°45'E). The Barcelonnette basin is characterized
noise” including salt-and-pepper noise and random-valued
by the Mediterranean climatic and geomorphological condi-
noise (Chan et al., 2005). However, the use of this technique
tions observed in the southern French Alps (Flageollet et al.,
is not appropriate when processing CHMs derived from highly
1999). The size of the study area is about 1.3 km2 and mainly
accurate three-dimensional (3D) lidar data, because all pixels
covered by coniferous plantation forests and grasslands, all
are altered when smoothing their values according to their
within an elevation range of 1,400 to 2,380 m above sea level.
nearest neighbors (Ben-Arie et al., 2009; Zhao et al., 2013).
The forest predominantly consists of mountain pine (Pinus
Smoothing methods have underestimated true tree height
uncinata) and scots pine (Pinus sylvestris) plus a few Larix
(especially of small trees) (Hyyppä et al., 2000), due to the
decidua and occasional deciduous trees (Populus tremula
omission of treetops, as well as underestimated crown radius,
and Fraxinus excelsior).
due to the reduction of crown shoulders (Solberg et al., 2006).
Consequently, such errors reduce the accuracy of forest bio- Field Measurements
mass and carbon estimations. Field inventory data were collected during two weeks in both
Other researchers have proposed different methods for September 2011 (seven plots) and September 2012 (48 plots).
removing data pits. Leckie et al. (2003) and Popescu and Stratified random sampling determined the position of the cir-
Wynne (2004) recommended that only the highest first return cular plots (r = 12.6 m), using a vegetation type map obtained
in each cell (pixel) be used, instead of using all first lidar from the French forest service (Office National des Forêts,
returns that penetrate the crown. Chen et al. (2006) suggested 2000). The vegetation map was superimposed on an aerial
using a large cell size to minimize height variation within orthophotograph and used in the field in order to improve
crowns. However, large cell sizes (e.g. >0.5 m) reduce the po- visual interpretation. The measurements collected included
tential accuracy of the crown boundary. Ben-Arie et al. (2009) tree location, tree height, tree crown diameter (CD), tree stem
introduced a semi-automated pit-filling algorithm to fill the diameter at breast height (DBH), and tree species determina-
data pits. In this algorithm, a Laplacian filter is applied to the tion (Table 1). The total number of trees sampled was 694.
original CHM to find pits with a user-defined threshold (visual A Nikon hand-held laser rangefinder (electronically mea-
decision). Evaluation of different threshold ranges indicated suring distance and angle) was used for tree height measure-
that choosing an incorrect threshold would cause an omis- ment where the treetops could be seen. The DBH of all trees
sion (under-filling) or commission error (over-filling). Zhao et (larger than 7 cm stem diameter at 1.3 m above ground)
al. (2013) improved the method proposed by Ben-Arie et al. within a plot was measured using a 60 cm caliper. The aver-
(2009) by adding a morphological crown control threshold in age crown diameter of the trees was measured in two perpen-
order to minimize the overfilling problem. They assumed the dicular directions, as direct measurement proved difficult
crown shape of both coniferous and deciduous forest trees to (Song et al., 2010).
be near circular. However, the algorithm could not remove In September 2011, the position of individual trees as
all pits, especially not when a lower branch or leaf inside the well as the central point of each plot was recorded using the
canopy was encountered, rather than the ground. Shamsod- Leica 1200 Differential GPS System. A total station was used
dini et al. (2013) developed an adaptive mean filter method to measure tree positions where the differential GPS signal
using variable window sizes in order to fill the pits. However, was too weak due to canopy density. Setting up the geodetic
it is difficult to select an optimal window size that results in base stations took a 24-hour period of static observation for
high accuracy for all tree structural parameters. each geodetic station (see Razak et al. (2011) for more detail).
Due to advances in lidar technology, survey data with very Popescu (2007) reported that treetop positions may be de-
high point densities (i.e., ~50 points/m2 or more) have become termined with higher accuracy using a CHM image than with
available. Researchers have found that such high-resolution error-prone measurements derived from differential GPS in the
point clouds allow them to extract detailed data on forest field. Therefore, a printed CHM image (at a scale of 1:100) with
structure (Holmgren et al., 2008; Hyyppä et al., 2001; Li et 0.15 m pixel size was used in September 2012 to determine
al., 2013). However, as it is very costly to obtain such high- tree location where the differential GPS signal was too weak.
density lidar data for large forested areas, the data sets used Using the CHM technique for the 2012 fieldwork allowed us to
in practice often are of much lower density (e.g., 4 points/m2 survey many more plots in two weeks than in 2011.

864 Septem b er 2 014 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 Table 1. Descriptive Statistics of the Tree Field Measurements
All (n = 694) deciduous (n = 33) Larix (n = 29) Pinus uncinata (n = 307) Pinus sylvestris (n = 325)
DBH Height CD DBH Height CD DBH Height CD DBH Height CD DBH Height CD
(cm) (m) (m) (cm) (m) (m) (cm) (m) (m) (cm) (m) (m) (cm) (m) (m)
Minimum 7 6 0.5 13 7 1.5 13 6.6 3.7 8 6 0.9 7 6 0.5
Maximum 61 25 9.1 43 24 6.3 61 25 8.9 59 23 9.1 41 25 6.7
Median 23 13.5 2.6 25 15.5 3.8 36 17.6 6 25 13 3.4 20 13.5 1.7
Mean 24.1 13.5 3 24.8 16.8 3.6 34.9 17.8 6.2 26 12.9 3.6 20.9 13.4 1.9
SD 8.2 3.2 1.6 9.4 6 1.3 10.5 4 1.4 8.5 3 1.4 5.4 2.4 0.9

Lidar Data Table 2. The Airborne Laser Scanning Characteristics (Razak et al., 2011)
The lidar data and aerial photographs (0.15 m pixel size) were
simultaneously acquired during the leaf-on season in July Acquisition (month/year) July–2009
2009 using a helicopter flying about 300 m above ground level Laser scanner Riegl VQ480i
(Table 2). A full-waveform airborne laser scanning system
IMU system iMAR FSAS - record 500Hz
(RIEGL VQ-480), which was developed for surveying moun-
tainous forested areas, was utilized by Helimap (Vallet and GPS system Topcon legacy - record 5Hz
Skaloud, 2004). The system provides high-speed data acquisi- Laser pulse repetition rate 300 kHz
tion using a fast line scanning mechanism and a narrow in-
frared laser beam. Moreover, the system performs on-line full Measurement rate Up to 150 000 s–1
waveform analysis (in hardware) to extract discrete returns Laser wavelength Near infrared
from the waveforms. For our data set, the VQ-480 system used
a laser pulse repetition rate of up to 300 kHz and recorded up Beam divergence 0.3 mrad
to five returns for each pulse. The initial aim of this survey Laser beam footprint 75 mm at 250 m
was to accurately map complex landslides in forested terrain
(Razak et al., 2011). Therefore, to increase the point cloud Field of view 60°
density, the area was covered by seven flights. The mean Scanning method Rotating multi-facet mirror
point density was 160 point/m2 with an average distance
between laser points of 0.08 m.
There is a time difference between the collected data by Methods
lidar acquisition in July 2009 and the field measurements in
September 2011 and 2012. However, our study area consists Preprocessing of Lidar Data
of mature forest and is characterized by relatively low tem- Our original lidar dataset was stored separately in adjacent
peratures during most of the year, which results in a relatively irregular tiles and classified into ground-points and non-
low tree growth rate. Therefore, we assume that the difference ground-points. To avoid edge effects along tile boundaries
in tree height due to natural growth between the time of the (Brandtberg et al., 2003), the lidar points were retiled with a 25
lidar survey and the acquisition of ground data is negligible. meter buffer around each tile. This essentially moves potential
artifacts into the tile buffer. Unlike other methods that create
a raster CHM from the difference between two rasters (i.e., by

Figure 1. Diagram of pit-free algorithm methodology.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING S e pte m ber 2014 865

subtracting the DTM raster from the DSM raster), our method first exceeding the respective height thresholds and by rasterizing
height-normalized the raw lidar points and then generated a only those triangles where all three edge lengths fall within
CHM raster. The lidar points were height-normalized by replac- a particular cutoff point. The latter tries to assure that only
ing the elevation of each point (i.e., the z coordinate) with its triangles connecting first returns from the same tree crown are
height above the ground. Consequently, the height of all returns rasterized. Intuitively speaking, this rasterization threshold
classified as ground is zero. Afterwards, the ground returns needs to be larger than the average point spacing, but smaller
were interpolated with a triangular irregular network (TIN) than the space that separates individual trees. We experimen-
constructed through Delaunay triangulating (Isenburg et al., tally found 0.45-meter to be an optimal rasterization thresh-
2006) the x and y coordinates in 2D. Then, the height of non- old for our study area and our lidar point density. Figure 2
ground returns was computed as the vertical distance to the TIN illustrates the TIN generation of normalized raw lidar point
(i.e., the distance in the z direction from the TIN to the points). clouds as well as the partial CHMXX raster created at each stage.
These normalized lidar points were used as input for the pit- Moreover, it shows that the long triangles are removed from
free algorithm, described in the following section. The prepro- each partial CHM raster by the particular threshold.
cessing was implemented using batch-scripting the lastile and The cell size of the raster is a key parameter when creating
lasheight modules of LAStools (rapidlasso GmbH, 2013). a CHM, especially for accurate derivation of individual tree
attributes (Chen et al., 2006). Pouliot et al. (2002) have sug-
Description of Pit-free Algorithm gested a ratio of crown diameter to grid size appropriate for
In this section, we describe how to efficiently generate a pit- defining crown shape. Based on these results, we determined
free raster CHM from lidar point clouds. The pit-free algorithm a grid size of 0.15 m in the CHM to be sufficient for recognizing
comprises two stages. The first stage is to construct a standard a minimum crown diameter of 0.50 m from our field survey.
CHM from all first returns and, most importantly, a number of
partial CHMs from only those first returns that correspond to Composing a Pit-free CHM
higher-up vegetation hits. The second stage is to combine all In a final step, a pit-free CHM raster was created by combining
these CHMs into one CHM based on the highest value across all the partial CHMXX (CHM00 to CHM15) rasters. This process can be
CHMs for each x and y raster position (Figure 1). This algo- envisaged by stacking the CHMXX rasters on top of each other
rithm can easily be implemented in the command line with a in order of height, with CHM00 at the bottom and CHM15 at the
batch script that uses an efficient combination of the las2dem top. This simple cell-based calculation creates a final output
and lasgrid modules of LAStools (rapidlasso GmbH, 2013). raster (i.e., the pit-free CHM) with the output value at each
location being the maximum value of all input rasters, CHM00
Partial CHMs to CHM15, without the need to involve any neighboring cells
Pits in the canopy appear whenever the laser beam is able to (Figure 3).
penetrate deeply through canopy branches before producing
a first return. The depth and the distribution of pits in a stan- Evaluation of the Pit-free Algorithm
dard CHM depend on the crown structure and the diameter of In order to demonstrate the robustness of our algorithm for
the laser beam as well as the sensitivity of the system process- generating a pit-free CHM, a lower-density version of lidar data
ing the returning waveform (Gaveau and Hill, 2003). Instead was created artificially from the original lidar survey. The
of hitting the highest point of the canopy, the laser pulses pit-free CHMs derived from both lidar datasets were evaluated
may produce their first return when they hit a lower branch by visually comparing them with the standard CHMs after ap-
or even after they penetrate all the way through the crown to plying a Gaussian smoothing filter.
the ground. Hence, the depth of different canopy pits var-
ies greatly, making it impossible to use a fixed threshold to Thinning Lidar Data
define (and potentially remove) them (Ben-Arie et al., 2009). The low-density version of lidar data was created using a
To address this issue, a set of partial CHMs was computed by simple point thinning algorithm, keeping only every nth (for
excluding all first returns below a certain height so that each example, the 15th) first return along the time line from the
CHM represents only some higher part of the canopy. The fun-
original dataset. By applying the keep-every-nth-return filter to
damental idea is to compute the shape of the canopy at differ- the first returns of each pulse, filtering takes place per-pulse
ent levels. The American Society for Photogrammetry and Re- rather than per-return. This step is important for two reasons:
mote Sensing (ASPRS) classifies lidar points into three layers: first, applying the filter to all returns of a high-return pulse
low vegetation (0.5 m <height ≤2.0 m), medium vegetation positively affects the odds regarding the survival of its first re-
(2.0 m <height ≤5.0 m), and high vegetation (5.0 m <height) turn; and secondly, subsequent CHMXX production only utilizes
(ASPRS, 2008). We used a similar layering of the vegetation the first return of each pulse. The algorithm was implemented
to construct our partial CHMs. All first return points were used in LAStools with a batch script that used every 15th first return
to construct the first CHM (CHM00). This CHM is the standard from the entire point dataset and created a poor low-density
CHM that other researchers have typically generated from the
dataset with mean point density of 7 points/m2.
first return lidar points (Hyyppä et al., 2008). The second CHM We should point out that keeping every nth first return does
(CHM02) was constructed by including first returns at a height not properly simulate a survey undertaken at higher altitude
of 2 and above. The first returns of ground level, low vegeta- with a slower scan rate. The diameter of the laser beam on the
tion, and medium vegetation were removed in the third CHM canopy increases with the altitude of the aircraft’s flight path
(CHM05). The fourth CHM (CHM10) and fifth CHM (CHM15) were due to beam divergence of the laser (Lim et al., 2008). In an
constructed by excluding first returns from heights less than actual survey of lower density lidar, fewer laser pulses would
10 and 15 m, respectively (Figure 1). Distance intervals of 5 sample the ground, each with a wider footprint and thus more
m, as used in this study, are called height thresholds and aim likely to interact with the canopy higher-up (Hall et al., 2009).
to preserve the original morphological structure of the tree However, in our artificially created lower density survey, the
crowns. We applied CHM00 to CHM15, as only 5 percent of the beam width remains unchanged and the likelihood of the thin
trees in the study area are taller than 20 m. laser beams penetrating deeply into the canopy remains high.
This creates the desired low-density data set of “especially
Generating Partial CHM Rasters poor quality.”
The partial CHMXX rasters were generated by triangulating only The pit-free algorithm created a pit-free CHM with a pixel
those first returns with a height (i.e., normalized z coordinate) size of 0.50 m, derived from the low-density lidar data.

866 Septem b er 2 014 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 Figure 3. Subset of pit-free chm.

size (Dralle and Rudemo, 1996). In all our comparisons (on

CHMs with 0.15 m and 0.5 m pixel size, based on high and
low density data, respectively), a Gaussian filter with a 5 ×
5 kernel was used, as the next kernel size down of 3 × 3 had
no noticeable effect regarding removing data pits from the
standard CHMs.
Individual Tree Detection
For individual tree detection we used the variable window
technique, developed by Popescu and Wynne (2004), which
automatically detects treetops. This approach uses a local
maxima method with a circular variable window size to lo-
cate treetops with a well-defined crown in coniferous forests
(Popescu, 2007). The variable window size of local maxima
operates by assuming there is a relationship between tree
height and crown size and uses these parameters to detect
treetops (see Popescu and Wynne (2004) for more details).
Accuracy Assessment of Individual Tree Detection
Detection accuracy was assessed by comparing the automati-
cally detected trees with the trees measured in the field.
Errors of omission and commission were assessed by tree
species and by stem diameter. For each tree species, the clos-
est detected treetop within the reference crown boundary
was considered to be a correctly detected tree. If more than
one treetop is detected, the others are considered commission
errors. Omission errors occur when no treetop is detected
within the boundary. The overall accuracy of tree detection
was computed by the accuracy index (AI) as defined by Pou-
liot et al. (2002):

Figure 2. The generated TINs and rasters of partial chms for AI (%) = [(n – O + C)/n] × 100
an individual tree.
where n is the number of reference trees in the study area, O
is the omission error, and C is the commission error. Based
Experimenting with this rasterizing process, we found 1.5 m on the DBH distribution (Table 1) the range in DBH of the trees
to be an optimal rasterization threshold for such a low-densi- was divided into three classes: DBH less than 20 cm, DBH from
ty lidar dataset. 20 to 40 cm, and DBH greater than 40 cm. The proportion of
correctly detected trees was calculated as a percentage of the
Smoothing CHM Using Gaussian Filter total number trees in each class.
The degree of smoothness of a CHM is determined by the
standard deviation (Gaussian scale) and the window kernel

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING S e pte m ber 2014 867

Results
Comparing the Pit-free CHMs with Smoothed CHMs
The efficacy of the pit-free algorithm was visually compared
with the result from a 5 × 5 Gaussian filter. Figure 4 shows
that the data pits (small dark squares) within the tree crowns,
which are clearly visible in the standard CHM, are not present
in the pit-free CHM. The Gaussian filter also removed pits, but
some pits remained compared to the pit-free CHM generated
from the high-density dataset (Figure 4 (left side)). The filter
over-smoothed the CHM image derived from the low-density
dataset, especially around the shoulders of tree crowns (Fig-
ure 4 (right side)). The algorithm is seen in Figure 4 to operate
robustly even with the poor quality, thinned lidar data. In
our experiments, the algorithm effectively removed the pits
as well as preserved the edges of canopy gaps and crowns for
both the high- and the low-density datasets (Figure 4, pit-free
CHMs, left side and right side, respectively).

Assessing Individual Tree Detection

The variable window technique was applied to both the
smoothed and pit-free CHMs. Figure 5 shows examples of correct-
ly detected trees, omission errors, and commission errors for the
pit-free CHM derived from high-density lidar data (white points
and circles show the position of field trees and their crown size,
respectively; gray stars indicate automatically detected treetops).
Table 3 presents the number and percentage of correctly
detected trees, omission, and commission errors, as well as
the overall accuracy index based on high-density lidar data.
As can be seen in the table, the total accuracy index for tree
detection from the pit-free CHM (74.2 percent) is higher than
from the smoothed CHM (70.6 percent). For the 33 measured
deciduous trees (Fraxinus excelsior, Populus tremula), the de-
tection accuracy was 51.5 percent for the pit-free CHM and 45.5
percent for the smoothed CHM. For coniferous trees, the over-
all detection rates were better. In addition, we tested whether
there is a statistically significant difference between the num-
ber of correctly detected trees using the CHMs derived from Figure 4. Visual representations of the pit-free algorithm
the Gaussian method and the pit-free method. The Chi-square compared to the standard chm and Gaussian filter for both
test showed a statistically significant difference between the high-density and low-density lidar data.

Figure 5. Example of detected trees, illustrating correctly identified crowns, omission errors, and commission errors of the pit-free
CHM derived from high-density lidar data.chm

868 Septem b er 2 014 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

Table 3. Tree Detection Results for the Smoothed chm and the Pit-free chm; Both Models have a Pixel Size of 0.15 m and were Derived from High-Density
Lidar Data
Field-measured trees Gaussian Smoothed CHM Pit-free CHM
number Species Correct Omission Commission AI Correct Omission Commission AI
of trees n (%) n (%) n (%) (%) n (%) n (%) n (%) (%)
33 deciduous trees 19 (57.5) 14 (42.5) 4 (12.1) 45.5 21 (63.6) 12 (36.4) 4 (12.1) 51.5
29 Larix decidua 26 (89.6) 3 (10.4) 3 (10.4) 79.3 25 (86.2) 4 (13.8) 2 (6.8) 79.3
325 Pinus sylvestris 261(80.3) 64 (19.7) 13 (4.0) 76.3 273(84.0) 52 (16.0) 23 (7.0) 76.9
307 Pinus uncinata 205(66.8) 102 (33.2) 1 (0.3) 66.4 225(73.3) 82 (26.7) 0 (0.0) 73.3
694 Total 511(73.7) 183(26.3) 21 (3.0) 70.6 544(78.4) 150 (21.6) 29 (4.1) 74.2

Table 4. Tree Detection Results for the Smoothed CHM and the Pit-free chm; Both Models have a Pixel Size of 0.50 m and were Derived from Low-Density
Lidar Data
Field-measured trees Gaussian Smoothed CHM Pit-free CHM
number Species Correct Omission Commission AI Correct Omission Commission AI
of trees n (%) n (%) n (%) (%) n (%) n (%) n (%) (%)
33 deciduous trees 17 (51.5) 16 (48.5) 10 (30.3) 21.2 20 (60.6) 13 (39.4) 8 (24.2) 33.3
29 Larix decidua 26 (89.7) 3 (10.3) 6 (23.0) 69 26 (89.7) 3 (10.3) 1 (3.4) 86.2
325 Pinus sylvestris 257(79.0) 68 (21.0) 139 (42.7) 36.3 268(82.5) 57 (17.5) 35 (10.7) 71.7
307 Pinus uncinata 184(60.0) 123 (40.0) 81 (26.3) 33.6 205(66.8) 102 (33.2) 4 (1.3) 65.5
694 Total 484(69.8) 210 (30.2) 236 (34.0) 35.7 519(74.8) 175 (25.2) 48 (6.9) 67.7

number of correctly detected trees in the smoothed CHM (511) Table 5. Proportion of Correctly Identified Trees for the Three Stem Diam-
and the pit-free CHM (544): x2 = 4.302, and p = 0.038. eter Classes for Each chm and Point Density.
The results of the tree detection assessment in evaluating All (n = 694)
the pit-free algorithm based on poor quality lidar data can be
seen in Table 4. As expected, the decrease of the point density Portion of DBH range <20 20 to 40 >40
from 160 points/m2 to 7 points/m2 resulted in a very low (cm) (n = 258) (n = 397) (n = 39)
accuracy (i.e., more commission error) when using Gaussian High- smoothed CHM 156 (60.4 %) 322 (81.1 %) 33 (84.6 %)
filters for both deciduous and coniferous trees. The pit-free density
CHM improved the total accuracy of the tree detection in Lidar Pit-free CHM 184 (71.3 %) 326 (82.1 %) 34 (87.1 %)
comparison to the smoothed CHM, resulting in Accuracy Index
Low- smoothed CHM 130 (50.3 %) 322 (81.1 %) 32 (82 %)
values of 67.7 percent and 35.7 percent, respectively. Con- density
sidering all 694 reference trees, 74.8 percent of the trees were Lidar Pit-free CHM 160 (62 %) 327 (82.3 %) 32 (82 %)
detected correctly in the pit-free CHM and 69.8 percent in the
smoothed CHM. The total commission error was 34 percent
for the smoothed CHM, while it was only 6.9 percent for the Discussion
pit-free CHM. The total omission error was 25.2 percent for the One of the challenges in creating a canopy height model
pit-free CHM and 30.2 percent for the smoothed CHM. In addi- through interpolation of first returns is that pits appear when-
tion, the Chi-square test indicated a statistically significant ever a laser beam is able to penetrate a tree crown deeply
difference between the pit-free CHM and the smoothed CHM, before it generates the first return. Based on this observa-
with the pit-free CHM resulting in a higher number of correctly tion we propose a new “pit-free” algorithm, which is able to
detected trees (x2 = 4.403, p = 0.036). remove pits efficiently and shows good potential for improv-
The range of DBH values in the study site (Table 5) allowed ing the detection of forest trees. Our algorithm is novel in that
us to assess how the pit-free CHM compared to the smoothed it removes the data pits directly from the lidar point clouds
CHM for both datasets. Trees with a DBH of more than 20 cm during the CHM creation. Other approaches, including auto-
were successfully detected with both different types of CHM mated methods (e.g., a smoothing filter) and semi-automated
and data. However, a large portion of the missed trees had a methods (i.e., a pit-filling algorithm) do not accurately target
DBH of less than 20 cm, with 71.3 percent and 62 percent of only the pits, but alter all pixels of the CHM raster (Ben-Arie
them correctly detected with a pit-free CHM using high- and et al., 2009; Shamsoddini et al., 2013). Moreover, automated
low-density data, respectively. Only 60.4 percent and 50.3 and semi-automated methods do not consider the natural 3D
percent of trees this size were detected with a smoothed CHM structure of individual tree crowns when removing height
derived from high- and low-density data, respectively. The variations. They consider the crown shape of forest trees as a
Chi-square test showed a significant difference between the circle (in nadir view on the raster CHM) (Zhao et al., 2013).
pit-free CHM and the smoothed CHM regarding the number Our algorithm makes use of the knowledge that a tree
of correctly identified trees with DBH less than 20 cm (high- crown tends to create lidar returns at a wide range of heights
density: p = 0.009; low-density: p = 0.008). However, there and generates a set of partial CHMs to remove the pits. There
was no significant difference between the types of CHM in are two important thresholds in generating the partial CHMs:
detecting trees with a DBH of more than 20 cm (high-density: p (a) the height thresholds, which define the set of partial CHMs,
= 0.0714; low-density: p = 0.646) or more than 40 cm (high- and (b) the rasterization threshold, which includes only those
density: p = 0.745; low-density: p = 1.000). triangles with edge lengths below a particular value. There is

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING S e pte m ber 2014 869

a direct relationship between the number of partial CHMs (e.g., CHMs from last lidar returns can also be used for locating
CHM02 to CHM15) and the number of height thresholds (e.g., 2, 5, individual trees. Last returns might show greater variation in
10 and 15 m). When we decrease the spacing between height height as the laser pulses will either penetrate all the way to
thresholds and consequently use a larger number of partial the ground, or reach only the middle of the tree, or already
CHMs, shallower pits (i.e., with less variation in height) are get stuck in the crown. We believe that our pit-free algorithm
removed from the CHM across all vegetation levels. When we benefits CHMs generated from both first and last return data.
increase the rasterization threshold and consequently include The results indicate that the pit-free algorithm is especially
larger triangles in the partial CHMs, larger pits (i.e., covering successful for lidar data with many pits, such as our artificial
greater areas) are removed as well. Both thresholds, but espe- example data where a small footprint laser beam samples the
cially the latter, change the appearance of the morphological canopy with low density. Here the pit-free algorithm results
structure of the tree crown. In this study, we experimentally in an accuracy index of 67.7 percent. This is particularly im-
selected height thresholds every 5 m for the partial CHMs of pressive compared to the Gaussian filtering, which produces
the taller trees (i.e., high vegetation class). The assignment of a low accuracy index of 35.7 percent. Though many studies
height thresholds was successfully applied here to CHMs with have underlined the importance of high density lidar data to
average tree heights of 13.5 m, but some variation in DBH, as improve tree detection, the result of this study suggests that
shown in Table 5; further research will establish performance laser point density has less impact on tree detection than the
of such thresholds on CHMs generated for different types of pit removal method. The study by Shamsoddini et al. (2013)
forest. We discovered that 0.45 m and 1.5 m are suitable also revealed that the accuracy of tree attribute estimation,
rasteriziation thresholds for the high-density and low-density using low point density lidar data, will vary significantly
lidar data, respectively (i.e., the length of three pixels in depending on the particular pit removal algorithm chosen.
each CHM). A larger threshold may increase the omission The results of this study demonstrate with low point density
error in tree detection because it then becomes more likely lidar data, our pit-free algorithm will yield a higher accuracy
for triangles from neighboring crowns to join together during regarding tree detection than Gaussian smoothing will.
rasterization. Although we evaluated the suitability of these
thresholds for detecting individual trees, it would be useful
to study their effect regarding the morphological structure of Conclusions
tree crowns further. In this study, we introduced a novel pit-free algorithm that
Our results show that the pit-free algorithm is visually su- can construct pit-free CHMs directly from lidar data. The
perior to the Gaussian smoothing filter. Data pits are evident impact of the pit-free algorithm on tree detection accuracy
in the CHM00 raster (standard CHM) generated through standard was investigated and compared with the Gaussian smoothing
Delaunay TIN rasterizing of all first returns. Nearby points method. The results reveal that our algorithm can be adapted
with similar x and y coordinates but different heights tend to work with different lidar point densities, and demonstrates
to form small and steep triangles in the Delaunay TIN. When a statistically significant improvement in the accuracy of tree
the x and y raster position of a CHM pixel falls within such a detection. It will be interesting to investigate whether our pit-
triangle, the linearly-interpolated height value of this cell may free CHMs also improves height estimation, crown segmenta-
drastically differ from that of neighboring cells. The Gaussian tion, and subsequent biophysical parameters such as biomass.
filter generates new values for all cells based on the values of
their nearest neighbors, thus affecting the original structure of
the tree crown, especially when using low-density lidar data. Acknowledgments
By contrast, our pit-free algorithm removes pits from the CHM The authors would like to express their gratitude to Menno
without altering the value of any other cells and thus pre- Straatsma for delivering high quality lidar data. We also
serves the original structure of tree crowns, for small trees as would like to thank Fatemeh Hatami, Sunil Thapa, and Col-
well, for both datasets. The pit-free algorithm offers the possi- lins Byobona Kukunda for their assistance with the fieldwork
bility to efficiently acquire detailed information on tree crown in 2011 and 2012.
characteristics that play a critical role in a variety of applica-
tions (e.g., ecological, hydrological, and meteorological) sensi-
tive to vegetation evolution at local and regional scale. References
ASPRS, 2008. LAS file format specifications, URL: http://www.asprs.
We also evaluated the algorithm by comparing the ac-
org/a/society/committees/standards/asprs_las_format_v12.
curacy of tree detection using the pit-free CHM versus the pdf, American Society for Photogrammetry and Remote Sensing
smoothed CHM. The statistical analysis showed that there is (ASPRS), Bethesda, Maryland (last data accessed: 30 June 2014).
a significant difference, with the pit-free CHM producing a Axelsson, P., 1999. Processing of laser scanner data - Algorithms and
higher number of correctly detected trees with high-density applications, ISPRS Journal of Photogrammetry and Remote
(p = 0.038) as well as low-density (p = 0. 036) lidar data. The Sensing, 54(2-3):138–147.
result indicated the robustness of the pit-free algorithm at Ben-Arie, J.R., G.J. Hay, R.P. Powers, G. Castilla, and B. St-Onge, 2009.
individual tree level in removing data pits for all crown sizes. Development of a pit filling algorithm for lidar canopy height
The Gaussian method, however, reduces the effect of the pits models, Computers & Geosciences, 35(9):1940– 1949.
based on a predefined window kernel size that cannot be Bortolot, Z.J., and R.H. Wynne, 2005. Estimating forest biomass using
optimal for all different crown types, especially at tree level small footprint LiDAR data: An individual tree-based approach
(Shamsoddini et al., 2013). Compared to the Gaussian filter, that incorporates training data, ISPRS Journal of Photogramme-
our algorithm is also significantly better at correctly detecting try and Remote Sensing, 59(6):342–360.
small trees (DBH of less than 20 cm) for both lidar datasets. Brandtberg, T., T.A. Warner, R.E. Landenberger, and J.B. McGraw,
Because the pit-free algorithm does not smooth the CHM, it 2003. Detection and analysis of individual leaf-off tree crowns
does not remove true treetops of small trees as happens with in small footprint, high sampling density lidar data from the
the Gaussian filter (Hyyppä et al., 2000; Kaartinen et al., 2012; eastern deciduous forest in North America, Remote Sensing of
Environment, 85(3):290–303.
Säynäjoki et al., 2008; Solberg et al., 2006). Our algorithm
was only evaluated for CHMs generated from first lidar returns Chan, R.H., H. Chung-Wa, and M. Nikolova, 2005. Salt-and-pepper
noise removal by median- type noise detectors and detail-pre-
as such CHMs are commonly used for tree detection. How-
serving regularization, IEEE Transactions on Image Processing,
ever, the study of Hyyppä et al. (2012) showed that creating 14(10):1479–1485.

870 Septem b er 2 014 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

Chen, Q., D. Baldocchi, P. Gong, and M. Kelly, M., 2006. Isolating Li, W.K., Q.H. Guo, M.K. Jakubowski, and M. Kelly, M., 2012. A
individual trees in a savanna woodland using small footprint new method for segmenting individual trees from the lidar
lidar data, Photogrammetric Engineering & Remote Sensing, point cloud, Photogrammetric Engineering & Remote Sensing,
72(8):923–932. 78(1):75–84.
Dralle, K., and M. Rudemo, 1996. Stem number estimation by kernel Lim, K., P. Treitz, K. Baldwin, I. Morrison, and J. Green, 2003a. LiDAR
smoothing of aerial photos, Canadian Journal of Forest Re- remote sensing of biophysical properties of tolerant north-
search, 26(7):1228–1236. ern hardwood forests, Canadian Journal of Remote Sensing,
Flageollet, J.-C., O. Maquaire, B. Martin, and D. Weber, 1999. Land- 29(5):658–678.
slides and climatic conditions in the Barcelonnette and Vars Lim, K., P. Treitz, M. Wulder, B. St-Onge, and M. Flood, 2003b.
basins (Southern French Alps, France), Geomorphology, 30(1- LiDAR remote sensing of forest structure, Progress in Physical
2):65–78. Geography, 27(1):88–106.
Forzieri, G., L. Guarnieri, E.R. Vivoni, F. Castelli, and F. Preti, 2009. Magnussen, S., E. Næsset, and T. Gobakken, 2010. Reliability of
Multiple attribute decision making for individual tree detec- LiDAR derived predictors of forest inventory attributes: A case
tion using high-resolution laser scanning, Forest Ecology and study with Norway spruce, Remote Sensing of Environment,
Management, 258(11):2501–2510. 114(4):700–712.
Gaveau, D.L., and R.A. Hill, 2003. Quantifying canopy height under- Nelson, R., W. Krabill, and J. Tonelli, 1988. Estimating forest biomass
estimation by laser pulse penetration in small-footprint airborne and volume using airborne laser data, Remote Sensing of Envi-
laser scanning data, Canadian Journal of Remote Sensing, ronment, 24(2):247–267.
29(5):650–657. Office National des Forêts, 2000. The National Forests Office (France),
Hall, R., R. Watkins, D. Heggem, K. Jones, P. Kaufmann, S. Moore, and URL: http://www.onf.fr/, Paris, France (last data accessed: 30
S. Gregory, S., 2009. Quantifying structural physical habitat at- June 2014)).
tributes using LIDAR and hyperspectral imagery, Environmental Persson, A., J. Holmgren, and U. Soderman, 2002. Detecting and mea-
Monitoring and Assessment, 159(1):63–83. suring individual trees using an airborne laser scanner, Photo-
Holmgren, J., A. Persson, and U. Soderman, 2008. Species identifica- grammetric Engineering & Remote Sensing, 68(9):925–932.
tion of individual trees by combining high resolution LIDAR Pitkänen, J., M. Maltamo, J. Hyyppä, and X. Yu, 2004. Adaptive meth-
data with multi-spectral images, International Journal of Remote ods for individual tree detection on airborne laser based canopy
Sensing, 29(5):1537–1552. height model, International Archives of Photogrammetry, Re-
Hosoi, F., H. Matsugami, K. Watanuki, Y. Shimizu, and K. Omasa, mote Sensing and Spatial Information Sciences, 36(8):187–191.
2012. Accurate detection of tree apexes in coniferous canopies Popescu, S.C., 2007. Estimating biomass of individual pine trees us-
from airborne scanning light detection and ranging images ing airborne lidar, Biomass and Bioenergy, 31(9):646–655.
based on crown-extraction filtering, Journal of Applied Remote Popescu, S.C., and R.H. Wynne, 2004. Seeing the trees in the forest:
Sensing, 6. Using lidar and multispectral data fusion with local filtering
Hyyppä, J., H. Hyyppä, D. Leckie, F. Gougeon, X. Yu, and M. Mal- and variable window size for estimating tree height, Photogram-
tamo, 2008. Review of methods of small-footprint airborne laser metric Engineering & Remote Sensing, 70(5):589–604.
scanning for extracting forest inventory data in boreal forests, Pouliot, D., D. King, and D., Pitt, 2005. Development and evalua-
International Journal of Remote Sensing, 29(5):1339–1366. tion of an automated tree detection delineation algorithm for
Hyyppä, J., O. Kelle, M. Lehikoinen, and M., Inkinen, 2001. A monitoring regenerating coniferous forests, Canadian Journal of
segmentation-based method to retrieve stem volume estimates Forest Research, 35(10):2332–2345.
from 3-D tree height models produced by laser scanners, IEEE Pouliot, D.A., D.J. King, F.W. Bell, and D.G. Pitt, 2002. Automated tree
Transactions on Geoscience and Remote Sensing, 39(5):969–975. crown detection and delineation in high-resolution digital cam-
Hyyppä, J., U. Pyysalo, H. Hyyppä, and A. Samberg, 2000. Elevation era imagery of coniferous forest regeneration, Remote Sensing of
accuracy of laser scanning-derived digital terrain and target Environment, 82(2–3):322–334.
models in forest environment, Proceedings of the 4th Workshop Rapidlasso GmbH, 2013. LAStools Rapid LiDAR Processing, URL:
on Lidar Remote Sensing of Land and Lea, Dresden, Germany: http://www.rapidlasso.com, Gilching, Germany (last data ac-
EARSeL. cessed: 30 June 2014).
Hyyppä, J., X. Yu, H. Hyyppä, M. Vastaranta, M., Holopainen, A. Razak, K.A., M.W. Straatsma, C.J. van Westen, J.P. Malet, and S.M. de
Kukko, H. Kaartinen, A. Jaakkola, M. Vaaja, J. Koskinen, and P. Jong, 2011. Airborne laser scanning of forested landslides char-
Alho, 2012. Advances in forest inventory using airborne laser acterization: Terrain model quality and visualization, Geomor-
scanning, Remote Sensing, 4(5):1190–1207. phology, 126(1-2):186–200.
Isenburg, M., Y. Liu, J. Shewchuk, and J., Snoeyink, 2006. Streaming Säynäjoki, R., P. Packalén, M. Maltamo, M., Vehmas, and K. Eerikäin-
computation of Delaunay triangulations, ACM Transactions on en, 2008. Detection of aspens using high resolution aerial laser
Graphics (TOG), ACM, pp. 1049–1056. scanning data and digital aerial images, Sensors, 8(8):5037–
Kaartinen, H., J. Hyyppä, X. Yu, M. Vastaranta, H. Hyyppä, A. Kukko, 5054.
M. Holopainen, C. Heipke, M. Hirschmugl, F. Morsdorf, E. Shamsoddini, A., R. Turner, and J.C. Trinder, 2013. Improving lidar-
Næsset, J. Pitkänen, S. Popescu, S., Solberg, B.M. Wolf, and based forest structure mapping with crown-level pit removal,
J.-C. Wu, 2012. An international comparison of individual tree Journal of Spatial Science, 58(1):29–51.
detection and extraction using airborne laser scanning, Remote
Sensing, 4(4):950–974. Solberg, S., E. Naesset, and O.M. Bollandsas, 2006. Single tree seg-
mentation using airborne laser scanner data in a structurally
Ke, Y., L.J. Quackenbush, and J. Im, 2010. Synergistic use of Quick- heterogeneous spruce forest, Photogrammetric Engineering &
Bird multispectral imagery and LIDAR data for object-based Remote Sensing, 72(12):1369–1378.
forest species classification, Remote Sensing of Environment,
114(6):1141–1154. Song, C., M.B. Dickinson, L. Su, S. Zhang, and D. Yaussey, 2010.
Estimating average tree crown size using spatial information
Kraus, K., and N. Pfeifer, 1998. Determination of terrain models in from Ikonos and QuickBird images: Across-sensor and across-
wooded areas with airborne laser scanner data, ISPRS Journal of site comparisons, Remote Sensing of Environment, 114(5):1099–
Photogrammetry and Remote Sensing, 53(4):193– 203. 1107.
Leckie, D., F. Gougeon, D. Hill, R. Quinn, L. Armstrong, and R. Vallet, J., and J. Skaloud, 2004. Development and experiences with a
Shreenan, 2003. Combined high-density lidar and multispectral fully-digital handheld mapping system operated from a helicop-
imagery for individual tree crown analysis, Canadian Journal of ter, Proceedings of the XX ISPRS Congress, Istanbul,Turkey.
Remote Sensing, 29(5):633–649.
Van Leeuwen, M., N.C. Coops, and M.A. Wulder, 2010. Canopy
Li, J., B. Hu, and T.L. Noland, 2013. Classification of tree species surface reconstruction from a LiDAR point cloud using Hough
based on structural features derived from high density LiDAR transform, Remote Sensing Letters, 1(3):125–132.
data, Agricultural and Forest Meteorology, 171- 172(0):104–114.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING S e pte m ber 2014 871

Véga, C., and S. Durrieu, 2011. Multi-level filtering segmentation to Yu, X., J. Hyyppä, M. Vastaranta, M., Holopainen, and R. Viitala,
measure individual tree parameters based on Lidar data: Ap- 2011. Predicting individual tree attributes from airborne laser
plication to a mountainous forest with heterogeneous stands, point clouds based on the random forests technique, ISPRS
International Journal of Applied Earth Observation and Geoin- Journal of Photogrammetry and Remote Sensing, 66(1):28–37.
formation, 13(4):646–656. Zhao, D., Y. Pang, Z. Li, and G. Sun, 2013. Filling invalid values in a
Wulder, M., K.O. Niemann, and D.G. Goodenough, 2000. Local lidar-derived canopy height model with morphological crown
maximum filtering for the extraction of tree locations and basal control, International Journal of Remote Sensing, 34(13):4636–
area from high spatial resolution imagery, Remote Sensing of 4654.
Environment, 73(1):103–114. Zhao, K., S. Popescu, and R. Nelson, 2009. Lidar remote sensing of
Yao, W., P. Krzystek, and M. Heurich, 2012. Tree species classification forest biomass: A scale- invariant estimation approach using air-
and estimation of stem volume and DBH based on single tree borne lasers, Remote Sensing of Environment, 113(1):182–196.
extraction by exploiting airborne full-waveform LiDAR data,
Remote Sensing of Environment, 123(0):368–380. (Received 19 September 2013; accepted 23 December 2013;
final version 03 April 2014)

872 Septem b er 2 014 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING