Sie sind auf Seite 1von 13

G Model

JAG-1117; No. of Pages 13 ARTICLE IN PRESS


International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and


Geoinformation
journal homepage: www.elsevier.com/locate/jag

Forest inventories by LiDAR data: A comparison of single tree


segmentation and metric-based methods for inventories of a
heterogeneous temperate forest
Hooman Lati a, , Fabian E. Fassnacht b , Jrg Mller c , Agalya Tharani d , Stefan Dech a,e ,
Marco Heurich c
a
Department of Remote Sensing in Cooperation with German Aerospace Center, University of Wrzburg, Oswald-Klpe-Weg 86, 97074 Wrzburg, Germany
b
Institute for Geography and Geoecology, Karlsruhe Institute of Technology, Kaiserstr 12, 76131 Karlsruhe, Germany
c
Department of Nature Protection and Research, Bavarian Forest National Park, Freyunger Str. 2, 94481 Grafenau, Germany
d
Institute of Remote Sensing, Anna University, Sardar Patel Rd., Chennai, India
e
German Remote Sensing Data Center (DFD), German Aerospace Center (DLR), 82234 Wessling, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Inventories of temperate forests of Central Europe mainly rely on terrestrial measurements. Rapid alter-
Received 7 April 2015 ations of forests by disturbances and multilayer silvicultural systems increasingly challenge the use of
Accepted 15 June 2015 conventional plot based inventories, particularly in protected areas. Airborne LiDAR offers an alternative
Available online xxx
or supplement to conventional inventories, but despite the possibility of obtaining such remote sensing
data, its operational use for broader areas in Central Europe remains experimental. We evaluated two
Keywords:
methods of forest inventory that use LiDAR data at the landscape level: the single tree segment-based
LiDAR
method and an area-based method. We compared a set of structural forest attributes modeled by these
Forest structure inventory
Single tree segment-based method
methods with a conventional forest inventory of the highly heterogeneous forest of the Bavarian For-
Area-based method est National Park (Germany), which partially includes stands affected by severe natural disturbances.
Spatial model Area-based models were accurate for all structural attributes, with cross-validated average root mean
Landscape level management squared error ranging from 3.4 to 13.4 in the best modeling case. The coefcients of variation for the
mapped area-based estimations were mostly minor. The area-based estimations were varied but highly
correlated (Pearsons correlations between 0.56 and 0.85) with single tree segmentation estimations;
undetected trees in the single tree segmentat-based method were the main sources of inconsistency. The
single tree segment-based method was highly correlated ( 0.54 to 0.90) with data from ground-based
forest inventories. The single tree-based algorithm delivered highly reliable estimates for a set of forest
structural attributes that are of interest in forest inventories at the landscape scale. We recommend LiDAR
forest inventories at the landscape scale in both heterogeneous commercial forests and large protected
areas in the central European temperate sites.
2015 Elsevier B.V. All rights reserved.

1. Introduction et al., 2005 for an in-depth denition of structural attributes).


Owing to the increasing occurrences of biotic and abiotic dis-
Forest structure is described by various attributes associ- turbances (windthrows, bark beetle infestations, and wildres),
ated with spatiotemporal properties of forest stands, including forests have become more heterogeneous (Seidl et al., 2014). The
height, canopy closure, understory cover, diameter at breast height increasingly prescribed close-to-nature forest management has
(DBH), and volume of standing and lying trees (see McElhinny also contributed to the heterogeneity, particularly in protected
areas, such as national parks (Boncina, 2011). Since conventional
methods of forest inventory, which involve eld samplings of var-
ious intensities, succeed in capturing only a limited portion of this
Corresponding author. Fax: +49 931 31896380. heterogeneity, remote sensing data have opened new horizons for
E-mail addresses: hooman.lati@uni-wuerzburg.de (H. Lati), area-wide mapping from tree to landscape scales (e.g., McRoberts
fabian.fassnacht@kit.edu (F.E. Fassnacht), joerg.mueller@npv-bw.bayern.de
and Tomppo, 2007). Moreover, remote sensing methods are more
(J. Mller), agalya.tharani@daad-alumni.de (A. Tharani), stefan.dech@dlr.de
(S. Dech), marco.heurich@npv-bw.bayern.de (M. Heurich).

http://dx.doi.org/10.1016/j.jag.2015.06.008
0303-2434/ 2015 Elsevier B.V. All rights reserved.

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
2 H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx

cost effective and can be easily applied in remote areas with limited In this study, we compare two methods of estimating forest
access, as is typical for protected areas (Heurich, 2008). structural attributes for inventories by LiDAR data using the exam-
Light detection and ranging (LiDAR) systems provide accurate ple of a heterogeneous Central European forest at the landscape
data for describing forest structure because of their ability to mea- level in the Bavarian Forest National Park (Germany). We compared
sure the inherent three-dimensional structure of stands via discrete the single tree segment-based method, an area-based method, and
return or full waveform measurements (reviewed in Lati, 2012). conventional forest inventory on plots in the study area. We extrap-
The assets of LiDAR data have been discussed based on case stud- olated the results of models and their uncertainty to the entire
ies on a wide variety of forest ecosystems worldwide, ranging from national park via wall-to-wall mapping. We tested whether the
boreal (Hyypp et al., 2008) to temperate (Lati et al., 2012; Tsui area-based LiDAR method produces models and area-wide maps
et al., 2013) and tropical (Drake et al., 2002; Treuhaft et al., 2010) that support the management measures used in the heterogeneous
ecosystems (see also Fassnacht et al., 2014 for a systematic review). forest stands and whether the applied single tree segment-based
Most works published to date share a common interest in inte- method delivers results sufcient for landscape level management
grating LiDAR surveys, specically their models, validations, and without the need for complementary eld data.
wall-to-wall maps, into the existing frameworks of forest inven-
tory on national (McRoberts et al., 2010), provincial (state-wide) 2. Materials and methods
(Johnson et al., 2014), and enterprise levels (Straub et al., 2013),
with the ultimate goal of a full forest inventory with LiDAR data. 2.1. Study area
Recent advances in theory and practice facilitate the use of LiDAR
data for supporting forest inventories (White et al., 2013). Wulder The study area includes the entire Bavarian Forest National Park
et al. (2013) have summarized the current pros and cons that sup- (24,220 ha) at the Czech-German border (49 3 19 N, 13 12 49 E).
port or constrain the design of inventory systems consisting only of The southern part of the current park was designated as the rst
LiDAR data; most of the arguments are somewhat related to strate- national park of Germany in 1970 and was extended to the north-
gic, logistic, and nancial planning issues beyond the technique west in 1997. Characterized by a rough topography with elevation
itself. ranging from 600 to 1453 m a.s.l., the national park features a
Nsset (2004), Nsset et al. (2004) and Hyypp et al. (2008) cool and humid climate with heavy and long-lasting snow cover
summarized the status quo of LiDAR-based studies with a specic at higher elevations (Ewald et al., 2014). The study area mainly
focus as baseline for forest inventory in Scandinavia. In the USA, encompasses three major forest communities, including sub-alpine
state-wide applications are increasingly supported due to the avail- stands dominated by Norway spruce (Picea abies L. Karst) and occa-
ability of affordable LiDAR data covering large areas (Skowronski sionally mountain ash (Sorbus aucuparia L.), slopes covered by a
and Lister, 2012). For such a forest inventory in Canada, White mixture of Norway spruce, silver r (Abies alba Mill.), European
et al. (2013) published instructional best practices with optimal beech (Fagus sylvatica L.), and sycamore maple (Acer pseudoplatanus
LiDAR specications for technical instrumentation, data processing, L.), and valley bottoms dominated by Norway spruce, moun-
ground sampling, and modeling, and indicated the ground data col- tain ash, and birches (Betula pendula Roth and Betula pubescens
lection (sampling design, intensity and measured individual trees) Ehrh.) (Heurich and Neufanger 2005). In the mid-1990s, spruce-
to be of major signicance in such surveys. dominated stands in the national park were massively attacked by
An existing ground forest inventory can act as a basis for the spruce bark beetle (Ips typographus L.). By 2013, this resulted in
subsequent LiDAR forest inventory systems by providing not the death of mature spruce stands over an area of 7000 ha. Accord-
only essential reference data for the model building process, but ing to the non-intervention strategy of the park administration,
also an additional source of a priori knowledge when design- snags are not removed from the sites (Lausch et al., 2013).
ing an improved eld survey. The number of eld samples can
then eventually be reduced, making the inventory cost effective 2.2. Ground-based forest inventory data
(Gobakken et al., 2013). However, the performance of LiDAR inven-
tory approaches, in particular when extrapolating the estimations Within the national park, 5841 permanent inventory plots in
to landscape levels, can be inuenced by various ground sampling a systematic grid of 200 m 200 m have been established. We
properties (Lati and Koch, 2012). selected 118 representative circular inventory plots by principal
Despite substantial advances in LiDAR methods, a forest inven- component analysis using the structural target parameters of this
tory based solely on LiDAR at the landscape level is still an study as attributes. Within each 500 m2 plot, we measured the stem
experimental concept in Central European forests because of their position, DBH (7 cm), and height of each living dead tree and snag
stand heterogeneity, effects of natural disturbances, and forest in 2013. These conventional forest inventory ground data were
management regimes. A LiDAR forest inventory relies primarily used for training and validation of models based on LiDAR data.
on the quality of structural attributes calculated from previously The position of each plot was measured at least twice for 30 min
delineated single trees. Heurich (2008) presented a watershed seg- with differential GPS using post-processing routine.
mentation method for tree delineation, which, when applied to
discrete LiDAR data, successfully delineates up to ca. 80% of the 2.3. Processing ground-based forest inventory data
individual trees of the top story of the stand. Reitberger et al. (2009)
then introduced single tree segmentation using full waveform The raw forest inventory ground data were pre-processed for
LiDAR data and normalized cut segmentation. With this method, training and validation of area-based models of LiDAR data. The
the full information content of the 3D point cloud can be utilized standing volume of each tree was calculated from equations inte-
to detect single trees, which results in a notably improved segmen- grating DBH, height, and tree species-specic constant parameters
tation of smaller trees in mixed stands compared to that achieved of merchantable wood in Bavaria (Franz, 1971) derived by Kennel
by watershed segmentation. Based on the method of single tree (1973). Lying dead trees were excluded when tree volumes were
segmentation, Yao et al. (2012a) introduced an approach to derive calculated. Species-specic single tree volumes were aggregated on
diameter at breast height (DBH) and stem volume of single trees. the plot level. In addition, the structural attributes mean DBH, basal
However, this method has not yet been generalized for typical for- area-weighted DBH, mean height, dominant height, basal area-
est inventories, i.e., plot sampling and other structural attributes. weighted height, and the number of trees per ha were calculated for
each plot. The formulas for calculating the structural attributes are

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx 3

Table 1
Formula used to calculate basal area-weighted DBH, dominant height, basal area-weighted height, and standing volume per plot.

Name Formula Parameters Description


nn=1 (DBHi bai )
Basal area-weighted DBH DBHba = DBH = diameter of a tree at breast The sum of products of all tree DBHs and their
nn=1 bai
height corresponding basal area, divided by the sum
bai = basal area of a tree of all basal areas within a given plot

1

n20%

Dominant height Hdom = Hi H = sum of height of the 20% highest The mean height of the 20% highest trees
n20%
trees in the plot within a given plot
n=1
n20% = number of the 20% highest trees
in the plot
nn=1 (Hi bai )
Basal area-weighted height Hba = H = height of a tree The sum of products of all tree heights and
nn=1 bai
bai = basal area of a tree their corresponding basal area, divided by the
sum of all basal areas within a given plot
 2
DBHp
Standing volume V = Hp f f = form factor Standard single tree volume function (Kennel,
2
DBHp = estimated DBH 1973)
Hp = estimated tree height

summarized in Table 1 and described in detail in Heurich (2006). reader is referred to Yao et al. (2012a) and Reitberger et al. (2009)
The calculated attributes on single tree level were aggregated to for details on single tree segmentation and retrieval of their struc-
sample plot level. tural attributes and to Yao et al. (2014) for analysis of the sensitivity
The national park has a number of stands with various num- of the single tree detection algorithm. Briey, following decompo-
bers of snags that were previously infested to a major extent by sition and calibration of the full waveform LiDAR data, the forest
bark beetles. The dataset studied here included approximately 263 stands are converted to numerous voxels, from which single trees
snags distributed within 54 sample plots. Since growth models are are segmented by means of normalized cut segmentation. The input
commonly based on excluding the snags from forest growth projec- parameters of the segmentation process include x, y, and z coordi-
tions, we excluded al the existing snags from the available dataset nates of the laser beams, in which the pulse width and intensity of
prior to modeling. The mean and standard deviations of all forest the waveform decomposition are integrated. The basis of the tree
attributes are summarized in Table 2. segmentation is to solve a cost function that aims at minimizing
the similarity between two given adjacent segments. The segmen-
2.4. Full waveform LiDAR data tation process is eventually followed by a stem detection approach
to delineate single trees in the overstory and understory (Reitberger
Full waveform LiDAR data across the national park was collected et al., 2009).
with a Riegl 680i laser scanner (350 KHz, nominal point density To model the standing volume of each segmented tree, the
3040 points per m2 ; altitude 650 m. The data was recorded at a crown laser points are separated from the remaining points of a tree
0.32 m footprint) by Milan Flug GmbH within three days in June by nding the crown base height. Then, the crown volume (Vcrown )
2012 under leaf-on condition. These data was used for deriving the is estimated by triangulating the previously found crown points
digital terrain model (DTM) and all further LiDAR metrics. with 3-D alpha shapes. This is followed by deriving the 2-D convex
Prior to area-based forest structure modeling, the original hull of the crown points as the tree crown area (Acrown ). The crown
LiDAR data were pre-processed as follows. The coordinates of the height (Hcrown ) is represented by the difference between the lowest
individual returns and intensity values were rst extracted by and highest crown points, and the tree height (Htree ) is derived by
decomposition of the LiDAR waveforms using Gaussian functions. taking the distance between the maximum height within a given
Rsn segment and the digital terrain model.
The intensity values were corrected by the applying Rn
with RS
Ref The volume of each segmented tree is then estimated by a subse-
as the traveling distance of the laser beam from the sensor to the quent multiple linear regression of Vcrown , Acrown , Hcrown , and Htree .
reecting object, RRef as a reference distance, and n = 2.1 as a cali- Using these four input parameters, a similar regression approach is
bration factor (Reitberger et al., 2009). The sum of waveforms was used to estimate the DBH of each segmented tree (Yao et al., 2012a).
normalized to a maximum value of 1. They were then truncated In our study, we also derived a set of additional structural
at the end of the ground reections to avoid the presumable effect attributes, including basal area-weighted DBH, dominant height,
of post-pulse oscillators on ground reections, which can result basal area-weighted height, and number of trees per ha (Table 1).
in amplitudes falsifying the LiDAR metrics. The amplitudes of the We created buffer zones of 500 m2 around the center coordinates
sum waveforms were then accumulated. The ground height was of the forest inventory plots and aggregated the tree attributes
extracted from the DTM. LiDAR metrics were then calculated in (average values of DBH and height) on the sample plot level.
separate 0.1 m height layers along the stand height; pulse distances An important property of the single tree segment-based method
in the waveform were varied, if necessary. The nal metrics were is that it is based entirely on LiDAR information and does
written in a raster format of 10 m 10 m pixel size. not depend on additional ground reference data except for the
above-mentioned general relationships between DBH and standing
2.5. Segmentation of single trees from full waveform LiDAR data volume, which are assumed to be relatively stable over the average
age of trees in the study area.
In an analysis of 18 sample plots of 2 selected stands in the
national park, Yao et al. (2012a) developed a method to automat-
ically derive tree height, DBH, and standing volume from single 2.6. Area-based method using LiDAR metrics
tree-based data extracted by normalized cut segmentation accord-
ing to Reitberger et al. (2009). Here, the algorithm was reran on The area-based models were based on using the entire wave-
the 2012 full waveform LiDAR data of the entire national park to forms (i.e., laser beams) which were aggregated on each inventory
obtain the structural attributes of the single tree segments. The sample plot. That is, all returns within a given plot were used

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
4 H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx

Table 2
Summary of forest stand statistics within the snags-excluded dataset. Standard deviations are given in parentheses.

Species No. of Mean height Dominant Basal DBH [cm] Basal Volume per ha Trees per ha
stratum plots [m] height [m] area-weighted area-weighted [m3 ]
height [m] DBH [cm]

Coniferous 30 15.79 (10.05) 21.23 (12.51) 15.70 (10.46) 21.85 (12.14) 21.54 (12.89) 353.55 (328.52) 727.33 (644.77)
Deciduous 48 20.26 (5.91) 28.41 (6.81) 20.61 (6.45) 23.70 (7.91) 24.65 (9.69) 362.99 (236.66) 593.75 (490.21)
Mixed 38 20.26 (7.84) 29.79 (9.70) 20.72 (8.59) 25.80 (10.51) 26.90 (12.50) 568.88 (374.43) 695.26 (469.120)

for calculating the applied LiDAR metrics. As inputs to area-based tree segmentation, we considered two variants. In the rst case,
models of forest structural attributes, LiDAR metrics were pre- unavailable plot values from the single tree segmentation method
selected based on both our previous experiences across the study caused by undetected tree segments were replaced by 0. In the sec-
site and the recently-published systematic review of Fassnacht et al. ond variant, such plots were removed from the analysis. For each
(2014a). The selected height metrics included the percentage of case, Pearsons correlation coefcient was calculated to compare
laser points between minimum and maximum height, maximum the results.
height, median height, standard deviation of height, and number of Modeling (training and prediction), evaluation, and paralleliza-
laser points. When we derived the height metrics, a height thresh- tion processes were implemented in R (R Development Core Team,
old of 2 m above ground was considered, which was shown to be 2015) by combining caret (Kuhn and Johnson, 2013), kernlab
effective when focusing on extracting LiDAR beams for modeling (Karatzoglou et al., 2004), raster (Hijmans and VanEtten, 2012) and
tree canopy (e.g., Packaln and Maltamo, 2006) and which avoids a snow (Tierney et al., 2013) libraries. A summary of the methodology
presumable intermixing effect that would otherwise exist between as shown in Fig. 1.
the applied LiDAR height and the understory metrics.
In addition, three density metrics (as suggested by Nsset,
3. Results
2004) were used with the aim of representing the pene-
tration rate of laser beams in different stand stories. These
3.1. Area-based models
included laser point density at the top stand height, laser point
density at 50% of the stand height, and laser point density
Our comparisons of the statistical modeling approaches indi-
at 10% of the stand height. Finally, two specic understory
cated that the RF method was superior in performance. Therefore,
metrics were also added that describe the entire understory
we will only show the results of this method; the performance of
[n(height < 0.5 m)/n(height < 2 m) for all returns] as well as the near-
the other tested modeling approaches (GLMBOOST, SVM, GP and
ground vegetation [n(height < 2 m)/n(height < 60 m) for all returns].
LMSTEP) were inferior to RF in both accuracy (i.e., relative bias)
Our methodology for building area-based models roughly fol-
and precision (i.e., relative RMSE) and therefore were not included
lows those described by Fassnacht et al. (2014a) and Lati et al.
here. The performance of the area-based RF model for each of the
(2015). For each given structural attribute, an initial matrix of
seven structural variables (predicted values) was compared with
response and predictor variables was initially formed, in which the
the values obtained in the conventional ground forest inventory
total number of plots was split into three subgroups of (nearly)
(measured values) (Fig. 2). The area-based RF models yielded vari-
equal sample size to ensure a balanced sampling from the entire
ous rates of prediction error, with best performances for dominant
range of values in the subsequent step. For each subgroup, 500 boot-
height (mean relative RMSE% = 3.44). The mean relative RMSE% of
strap resamples (Efron and Tibshirani, 1994) with replacements
all other structural attributes ranged between 5.45 (mean height)
were drawn, which were nally aggregated across the subgroups
and 13.52 (trees per ha).
to form 500 input datasets for modeling.
The area-based predictions showed overall negligible mean rel-
The input datasets were t by ve commonly applied parametric
ative bias values. The highest biases were observed for volume per
and nonparametric methods: random forests (RF), support vector
ha with a minor value of 0.40% (Fig. 2).
machines (SVM), Gaussian processes (GP), stepwise linear regres-
When we constructed wall-to-wall maps of seven cross-
sion (LMSTEP), and boosted generalized linear model (GLMBOOST).
validated, area-based RF models of seven structural attributes
Methodological details on the applied modeling schemes (except
(Fig. 3), the spatial representation of the structurally heterogeneous
GLMBOOST) are given in Fassnacht et al. (2014a) and Lati et al.
forest stands within the national park was realistically represented
(2015). For GLMBOOST, the reader is referred to Buehlmann (2006).
by excluding the snags from the inventory data prior to modeling.
Three-fold cross validations, each with ve repetitions, were used
This can be best seen in the areas located in northern parts of the
for validating the models. Model performances were assessed with
national park, where vast amounts of snags are present as a result
relative root mean squared error (RMSE) and relative bias from
of periodic bark beetle infestations. The realistic representation of
the mean reference values. The grid-based mean values of 500
the structural attributes when snags were excluded from the inven-
model runs were used for wall-to-wall mapping of the structural
tory data is also apparent in generally low values of coefcients of
attributes and their coefcient of variation on the level of the entire
variation as mapped across the study area (Fig. 4).
national park. To accomplish this, the above-mentioned LiDAR met-
rics were re-calculated on a raster grid, with each cell covering
approximately the area of an inventory plot. 3.2. single tree segment-based models

The single tree segment-based estimations of structural


2.7. Comparison of area-based method, single-tree segmentation attributes showed generally high but variable levels of correlation
method, and conventional forest inventory with the conventional ground inventory data (Fig. 5). The unde-
tected tree segments negatively contributed to the correlations.
We compared the average results of each structural attribute on The single tree segment-based estimations and conventional forest
the plot scale estimated using conventional forest inventory, the (ground) inventory values showed correlations ranging between
single-tree segment-based method, and the area-based method. 0.53 (trees per ha) and 0.89 (dominant height) (Fig. 5). Most impor-
To determine how the results were inuenced by the success in tantly, high levels of correlation were observed for the structural

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx 5

Fig. 1. Summary of the methodology of the study.

attribute volume per ha (up to 0.85). The lowest levels of correlation trees (Fig. 6). Predictions of the area-based model agreed fairly well
observed were for the structural attribute trees per ha. with those of the single tree segmentation method; the highest and
When plots with undetected tree segments were ignored in the lowest Pearsons correlations were observed for dominant height
single tree segmentation method, the correlations between the two (0.90) and trees per ha (0.60), respectively. The correlation for trees
datasets were considerably improved (Appendix 1). For an exam- per ha was mainly affected by the generally lower number of delin-
ple of dominant height, Pearsons correlations up to 0.87 were eated trees in single tree segment-based estimation than in that
observed. modeled by area-based data.
Despite similar trends, correlations for all the structural
3.3. Comparison of the area-based and single tree segment-based attributes except volume per ha (0.009 decrease) and trees per ha
models (0.052 decrease) were higher when undetected single trees were
ignored in the single tree segmentation data (Appendix 2). For
The single tree segment-based values of structural attributes example, the two methods were over 87% correlated for volume
showed generally high, yet variable levels of correlation with those per ha as a prominent allometric attribute. Nevertheless, the occa-
estimated by area-based models. The level of correlation depended sional undetected tree segments were responsible for some degree
on how successful the single tree segmentation method segments of lack of correlation between the two methods.

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
6 H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx

Fig. 2. Performance of cross-validated, area-based, random forest models for seven structural attributes. The individual response is plotted against the predicted values from
500 models based on bootstrap resampling data.

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx 7

Fig. 3. Wall-to-wall maps of cross-validated, area-based, random forest models of seven structural attributes. The pixel values (gray scale) represent the mean values from
500 models based on bootstrap resampling data.

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
8 H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx

Fig. 4. Wall-to-wall maps of prediction coefcients of variation for cross-validated, area-based, random forest models of seven structural attributes. The pixel values (gray
scale) represent the mean prediction coefcients of variation from 500 models based on bootstrap resampling data.

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx 9

Fig. 5. Comparative agreement between ground inventory data (gray points) and single tree segment-based estimates (black points) of seven structural attributes. The black
points with values = 0 in the single tree segment-based estimations represent the forest plots containing undetected tree segments.

4. Discussion in Central Europe. Both methods use full waveform LiDAR metrics.
To estimate the attributes, the area-based method uses plot-scale
We evaluated and compared area-based and single tree metrics as predictor inputs, whereas the single tree segment-based
segment-based approaches for estimating a set of fundamental for- method uses allometric tree features extracted from previously
est structural attributes on a structurally complex landscape level delineated tree segments from LiDAR data.

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
10 H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx

Fig. 6. Comparative agreement between area-based random forest estimates (gray points) and single tree segment-based estimates (black points) of seven structural
attributes. The black points with values = 0 in the single tree segment-based estimations represent the forest plots containing undetected tree segments.

The area-based method yielded good estimates of all seven As expected, the RF model outperformed other tested modeling
structural attributes. The applied bootstrapping allowed the popu- approaches by returning overall higher relative RMSE and lower
lation structure to be approximated from the available sample set, relative bias values. Despite the structural diversity within the
and the high number of cross-validated models enabled the pre- unmanaged stands of the national park and the inherent subsam-
diction variance for each given structural attribute to be inferred. pling effects caused by the RF approach, the area-based predictions

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx 11

showed relatively low variances for all attributes. Since the larger 2012b); therefore, the results need to be compared with our area-
range of a response value can possibly lead to an increased devia- based predictions to clarify the applicability of this component in
tion in prediction (Fassnacht et al., 2014a), a reduced range of values segment-based approach.
may decrease that deviation. However, this is subject to a good Despite moderate correlations for the structural attribute trees
model t for a given response variable. In our study, all ts yielded in per ha (Fig. 6), the single tree segment-based estimates corre-
generally good results, with negligible plot-based prediction biases. lated well with the area-based estimations for all other structural
For example, compared to available records from Central European attributes. Overall, in terms of trees per ha, the correlation was
mixed stands (reviewed by Fassnacht et al., 2014a), our area-based affected by the considerably fewer trees estimated by the single
models returned lower error rates for the standing volume per tree segmentation method. This is presumable, owing to the vari-
ha. One explanation to this will be the use of metrics that were able success of the single tree segment-based method in detecting
extracted from full waveform data. Since we included laser point single trees in such richly structured stands (see also Solberg et al.,
density at multiple stand height levels among our predictors as a 2006). Yao et al. (2012a) reported a high rate of correctly segmented
measure of penetration, we assume this largely reects the differ- trees in the upper stand layer (86%) and a lower success rate in
ent structure of coniferous and deciduous stands, particularly when intermediate and lower stand layers when leaf-on laser scanner
LiDAR data is recorded during the leaf-on period. data were used. This is considered to be one of the main draw-
Comparing our results with those of an earlier study from the backs of the method and is partially a function of whether LiDAR
same area (Heurich and Thoma, 2008) provides a more intuitive data are collected during the leafed or leaess period. Yao et al.
mean of evaluation for area-based estimations: Our models led to (2012a) showed that even when normalized cut segmentation is
prediction errors that slightly outperformed those of that study. used, the detection rate in intermediate and lower stand layers did
However, both studies report values within the same range for not exceed 32% and 17%, respectively. This fundamentally differs
all modeled structural attributes. Especially noteworthy are the from an area-based method, which relies on data of trees from a
coefcients of variation of the predicted values (Fig. 4); the high- ground survey of all vertical stand layers with a DBH greater than
est prediction variability was observed for stands in the northern 7 cm. Another factor that affects the agreement between the two
parts of the national park, which have heterogeneous understory tested approaches is under-segmentation, which can arise when
coverage. Even if these stands are not considered, the prediction the single tree segment-based method is used for dense young
coefcients of variation for DBH (both mean and basal area- forest stands; where one single segment could possibly consist of
weighted) and height (mean, basal area-weighted, and dominant multiple young trees.
heights) were often lower than those observed for volume per ha The correlations of the single tree-based estimates and the
and trees per ha. area-based estimates for all other allometric stand attributes were
As described in the Section 2.3, snags were excluded from the notably high. Larger trees, particularly broad-leaved trees, in the
inventory data prior to modeling. This obviously increased model highest stand layer contribute greatly to the available standing vol-
performances which can be attributed to the elimination of the sub- ume and biomass (e.g., Lutz et al., 2012). In a LiDAR inventory, this
stantial effects caused by snags, which have a DBH comparable to contribution is yet only marginally affected by how successfully
that of vital trees but partially or totally lack crowns. This would the individual trunks in the understory are detected. The single
in turn affect the distributional behavior of laser beams to vari- tree segment-based and area-based results showed notably higher
ous degrees. The observed performances for such snags-excluded correlations by ignoring the undetected trees in segment-based
sample plots support the ndings of Heurich and Thoma (2008), approach. However, one should not neglect this as a potential draw-
who discussed the magnitude of effect caused by large snags on back which can affect the practical application of such method for
the accuracy of the estimated height in live stands. However, this e.g., future tree-based growth simulations. A tree-based growth
also led to high prediction variability in stands dominated by snags model does not only focus on growth projection of mature trees
(Fig. 4), because these stands are then largely unconsidered due to from the overstory of stands, but should also incorporate the
the absence of sample plots representing snags. Apart from this, growth of understory elements including seedlings, saplings and
this will not considerably change the estimation of trees per ha, tree regeneration (e.g., Falkowski et al., 2010; Eerikinen et al.,
since the ability of laser metrics to recognize stems is independent 2014). This drawback could be minimized by improving the single
of their state of vitality. tree detection rate in lower stand layers. Nevertheless, the single
We did not divide the area-based models into strata-specic tree segment-based method applied here is able to support applied
models to directly compare area-based estimations and single tree forest inventory by providing broadly required forest attributes
segmentation estimations. In this regard, automatic recognition on single tree level and over wide areas such as BFNP. Having
of the tree type in single tree segmentation methods could occa- this in mind and by comparing area-based predictions to ground
sionally yield results different from those determined in a eld inventory data suggests that the single tree segment-based esti-
inventory (note the various overall accuracies reaching up to 95% mates still provide evidence of its high potential as a practical tool
for species classication of leaf-on LiDAR data as reported by Yao in landscape-level forest inventory. This method relies solely on
et al., 2012a). Furthermore, the improved prediction performance LiDAR data and temporarily stable empirical information (e.g., allo-
of area-based models by pre-stratied sampling units has pre- metric equations) and therefore does not entail the use of ground
viously been shown to be marginal if the sample size effect is inventory data.
minimized (Lati et al., 2015). However, the high potential of segment-based method might
Nevertheless, area-based models of forest structure generally be partially site specic and thus needs to be tested in struc-
lack a built-in tree type recognition method and depend either turally different mixed stands, especially with various degrees
on an external algorithm applied to optical imagery (e.g., Lati of coniferous/deciduous mixtures, provided that tree allometries
et al., 2012; Fassnacht et al., 2014b) or on information from a are comparable. In addition, site-specic differences cannot be
ground inventory. This is inconvenient compared to the applied neglected because these factors (e.g., increasing altitude) can inu-
single tree segment-based method in which a species classication ence the existing allometric relationships amongst tree height,
from LiDAR data was initially implemented. Another, presumably DBH and growing stock. Our results also revealed that improv-
advantageous feature of the applied single tree method is its abil- ing the single tree-based method by, e.g., minimizing the number
ity to discriminate dead and living trees. This was tested in small of undetected trees, can improve its correlation with area-based
area level but not yet implemented on landscape level (Yao et al., estimations (Appendix 2).

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
12 H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx

The different levels of correlation between single tree segment- detection of single trees in mature stands and in the overstory with
based estimates and ground inventory data also suggest similar area-based models of 3D structure across old and young stands,
trends: Except for the results obtained for trees per ha, the Pear- Based on our results, we recommend the use of LiDAR for prac-
sons correlations were within a satisfactory range (r2 = 0.610.89, tical applications in forestry and conservation at the landscape
as shown in Fig. 5), which strongly supports their application scale in temperate stands similar to those of the Bavarian For-
as a surrogate for ground inventories on the tested landscape est National Park. Especially the single tree segmentation method
level. However, trees per ha showed lower level of association. offers a cost-effective alternative to conventional forest invento-
In presence of an enhanced tree detection (which would enable a ries. For this national park, the effort of recording data can be
higher detection rate in lower stand stories), the agreement would reduced from several months to several days, decreasing the total
unquestionably boost, since the major obstacle here seems to be costs by up to 90%. The new methodology offers more informa-
the underestimated rate of detected trees in the segment-based tion than conventional forest inventories. In particular, the single
approach. An alternative would be to estimate undetected trees tree segment-based method opens up new possibilities in digital
using known tree distribution functions (see Maltamo et al., 2004). forestry, e.g., monitoring of forest stands, evaluating silvicultural
The results obtained with LiDAR data might be improved if a higher regimes, integral monitoring of forest biodiversity and, subject to
point density were used, which would help detect small trees and renements, applying single tree-based growth models.
snags, or if data were collected during the leaf-off period, which
could improve detection of deciduous trees.
Acknowledgements
The correlations between single tree segment-based estimates
and ground inventory data could also be slightly decreased by slight
We thank Dr. Peter Biber (Chair of Forest Growth, Technis-
positional inaccuracies of ground inventory data, and in turn area-
che Universitt Mnchen) for providing scripts on species-specic
based predicted data. The sample set used here was collected with
models of standing volume; the eld crew, in particular Olaf Schu-
differential GPS devices and contained positional errors of mul-
bert, for collecting required eld data; and Prof. Peter Krzystek
tiple degrees, ranging from 0.1 to 5 m (with 100 points being
(Hochschule fr angewandte Wissenschaften Mnchen) for pro-
more accurate than 50 cm), depending on the number of solutions
viding the database and results of the single tree segmentation
available at the time of sampling. We acknowledge that this might
analysis. Karen A. Brune undertook a professional English language
have introduced an unknown degree of uncertainty in our mea-
editing.
surements. However, the global navigation satellite system (GNSS)
positional accuracy for a state-wide Bavarian forest inventory is
9.7 m on average (Springer, 2014). Therefore, the positional error Appendix A. Supplementary data
in our study fully lies within the acceptable range for operational
surveys. Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.jag.2015.06.008

5. Conclusions
References
We evaluated two main methods of estimating structural forest Boncina, A., 2011. Conceptual approaches to integrate nature conservation into
attributes single tree segment-based and area-based method forest management: a Central European perspective. Int. Forest. Rev. 13, 1322.
that embrace various degrees of cost effectiveness for landscape- Buehlmann, P., 2006. Boosting for high-dimensional linear models. Ann. Stat. 34
(2), 559583.
level forest inventories. The area-based model yielded outstanding Drake, J.B., Dubayah, R.O., Knox, R.G., Clark, D.B., Blair, J.B., 2002. Sensitivity of
performances for nearly all structural attributes. The wall-to-wall large-footprint LiDAR to canopy structure and biomass in a neotropical
pixel maps fully represented the structural diversity of the forest rainforest. Remote Sens. Environ. 81 (23), 378392.
Eerikinen, K., Valkonen, S., Saska, T., 2014. Ingrowth, survival and height growth
stands and enabled visualization of the modeling uncertainty. Our
of small trees in uneven-aged Picea abies stands in southern Finland. For.
results suggested that both methods are appropriate for opera- Ecosyst. 1, 5, http://dx.doi.org/10.1186/2197-5620-1-5
tional forest inventories, and each method has some advantages Efron, B., Tibshirani, R.J., 1994. An Introduction to the Bootstrap. CRC Press, 465 pp.
ISBN 0-412-04231-2.
over the other. The area-based method lacks a built-in species
Ewald, M., Claudia Dupke, C., Heurich, M., Mller, J., Reineking, B., 2014. The
discrimination component, whereas the single tree segmentation application of LiDAR remote sensing and activity sensors to analyze winter
method yields some inconsistencies caused by problems in detect- habitat selection of European roe deer. Forestry 2014 (5), 13741390.
ing single trees in the understory and in young, dense forest stands. Falkowski, M.J., Hudak, A.T., Crookston, N.L., Gessler, P.E., Uebler, E.H., Smith,
A.M.S., 2010. Landscape-scale parametrization of a tree-level forest growth
This drawback can constrain its applicability for practical inven- mdoel: a k-nearest neighbor imputation approach incorporating LiDAR data.
tory applications such as tree-based growth models. However, this Can. J. For. Res. 40, 184199.
approach still enables to provide a set of essential forest attributes Fassnacht, F.E., Hartig, F., Lati, H., Berger, C., Hernndez, J., Corvaln, P., Koch, B.,
2014a. Importance of sample size, data type and prediction method for remote
on single tree scale for relatively big areas. The single tree segmen- sensing-based estimations of aboveground forest biomass. Remote Sens.
tation method has recently been further developed by integrating Environ. 154, 102114.
an algorithm to detect snags (Polewski et al., 2015a,b). This updated Fassnacht, F.E., Neumann, C., Frster, M., Buddenbaum, H., Ghosh, A., Clasen, A.,
Joshi, P.K., Koch, B., 2014b. Comparison of feature reduction algorithms for
method should in the future be compared to our models here, in classifying tree species with hyperspectral data on three Central European test
which snags were previously excluded from the inventory data. sites. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 7 (6), 25472561.
Nevertheless, this component needs to be further rened by, e.g., Franz, F., 1971. Grundlagen und Verfahren standortbezogener Leistungsschtzung.
Forschungsbericht der FFA Mnchen (2) (In German).
including external spectral information from aerial imagery. In
Gobakken, T., Korhonen, L., Nsset, E., 2013. Laser-assisted selection of eld plots
addition, care must be taken in improving the tree detection algo- for an area-based forest inventory. Silva Fenn. 47 (5), id 943.
rithm for structurally rich stands, e.g., by using LiDAR data from Heurich, M., 2006. Evaluierung und Entwicklung von Methoden zur
automatisierten Erfassung von Waldstrukturen aus Daten ugzeuggetragener
inventories of trees during the leaess period. Tree species detec-
Fernerkundungssensoren. Forstlicher Forschungsbericht Mnchen Nr. 202,
tion in both area-based and single tree segmentation algorithms ISBN 3-933506-33-6, https://mediatum.ub.tum.de/node?id=603731 (accessed
could be enhanced by incorporating terrestrial laser scanning, at 10.02.15.).
least for a number of structurally representative stands. A fusion Heurich, M., Neufanger, M., 2005. Die Wlder des Nationalparks Bayerischer Wald.
Ergebnisse der Waldinventur 2002/2003 im geschichtlichen und
of area-based and single tree segment-based methods would fur- waldkologischen Kontext. Wissenschaftliche Schriftenreihe der
ther pave the way for operational forest inventories by combining Nationalparkverwaltung Bayerischer Wald, Band 16, 175 (In German).

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008
G Model
JAG-1117; No. of Pages 13 ARTICLE IN PRESS
H. Lati et al. / International Journal of Applied Earth Observation and Geoinformation xxx (2015) xxxxxx 13

Heurich, M., 2008. Automatic recognition and measurement of single trees based Polewski, P., Yao, W., Heurich, M., Krzystek, P., Stilla, U., 2015. Detection of single
on data from airborne laser scanning over the richly structured natural forests standing dead trees from aerial color infrared imagery by segmentation with
of the Bavarian Forest National Park. Forest Ecol. Manage. 255, 24162433. shape and intensity priors. ISPRS Annals of the Photogrammetry, Remote
Heurich, M., Thoma, F., 2008. Estimation of forestry stand parameters using laser Sensing and Spatial Information Sciences. 43. W4 181-188 Part: 2.
scanning data in temperate, structurally rich natural European beech (Fagus R Development Core Team, 2015. R: a language and environment for statistical
sylvatica) and Norway spruce (Picea abies) forests. Forestry 81 (5), 645661. computing. R Foundation for Statistical Computing. Vienna, Austria. http://
Hijmans, R., VanEtten, J., 2012. R-package Raster user guide. R Development Core www.R-project.org. (accessed 10.02.15.).
Team. Available at: http://cran.r-project.org/web/packages/raster/raster.pdf Reitberger, J., Schnrr, C., Krzystek, P., Stilla, U., 2009. 3D segmentation of single
(accessed 10.02.15.). trees exploiting full waveform LiDAR data. ISPRS J. Photogramm. Remote Sens.
Hyypp, J., Hyypp, H., Leckie, D., Gougeon, F., Yu, X., Maltamo, M., 2008. Review of 64, 561574.
methods of small-footprint airborne laser scanning for extracting forest Seidl, R., Schelhaas, M.-J., Rammer, W., Verkerk, P.J., 2014. Increasing forest
inventory data in boreal forests. Int. J. Remote Sens. 29 (5), 13391366. disturbances in Europe and their impact on carbon storage. Nat. Clim. Change
Johnson, K.D., Birdsey, R., Finley, A.O., Swantaran, A., Dubayah, R., Wayson, C., 4, 806810.
Riemann, R., 2014. Integrating forest inventory and analysis data into a Skowronski, N.S., Lister, A.J. 2012. The utility of LiDAR for large area forest
LiDAR-based carbon monitoring system. Carbon Balance Manage. 9 (3), 11pp. inventory applications. Proceedings of Moving from stands to trends: Forest
Karatzoglou, A., Smola, A., Hornik, K., Zeileis, A., 2004. Kernlab- An S4 package or inventory and analysis symposium 2012. Gen. Tech. Rep. NRS-P-105.
kernel methods in R. J. Stat. Softw. 11 (9), 20p. Newtown Square, PA: U.S. Department of Agriculture, Forest Service, Northern
Kennel, E. 1973. Bayerische Waldinventur. Forstliche Forschungsberichte Research Station, USA.
Mnchen. Nr. 11 Ludwig Maximilians Universitt, Mnchen (In German). Solberg, S., Nsset, E., Bollandss, O.M., 2006. Single tree segmentation using
Kuhn, M., Johnson, K., 2013. Applied Predictive Modeling. Springer airborne laser scanner data in a structurally heterogeneous spruce forest.
Science + Business Media, 995 New York. Photogramm. Eng. Remote Sens. 72, 13691378.
Lausch, A., Heurich, M., Fahse, L., 2013. Spatio-temporal infestation patterns of Ips Springer, S., 2014. Zur Gte von GNSS-Einmessungen im Wald bei Forstinventuren.
typographus (L.) in the Bavarian Forest National Park, Germany. Ecol. Indic. 31, B.Sc thesis at the faculty of forest sciences and resource management.
7381. Technische Universitt Mnchen, Freising-Germany, 65 S (in German).
Lati, H., 2012. Characterizing forest structure by means of remote sensing: a Straub, C., Tian, J., Seitz, R., Reinartz, P., 2013. Assessment of Cartosat-1 and
review. In: Escalante, B. (Ed.), In Remote Sensing: Advanced Techniques and WorldView-2 stereo imagery in combination with a LiDAR-DTM for timber
Platforms. Intech Open Access Publisher, Zagreb, Croatia, pp. 428. volume estimation in a highlystructured forest in Germany. Forestry 86,
Lati, H., Koch, B., 2012. Evaluation of most similar neighbour and random forest 463473.
methods for imputing forest inventory variables using data from target and Tierney, L., Rossini, A.J., Li, N., Sevcikova, H., 2013. snow: Simple Network of
auxiliary stands. Int. J. Remote Sens. 33 (21), 66686694. Workstations. R Development Core Team. Available at: http://cran.r-project.
Lati, H., Nothdurft, A., Straub, C., Koch, B., 2012. Modelling stratied forest org/web/packages/snow/snow.pdf. (accessed 10.02.15.).
attributes using optical/LiDAR features in a central European landscape. Int. J. Treuhaft, R.N., Goncalves, F.G., Drake, J.B., Chapman, B.D., dos Santos, J.R., Dutra,
Digit. Earth 5 (2), 106132. L.V., Graca, P.M.L.A., Purcell, G.H., 2010. Biomass estimation in a tropical wet
Lati, H., Fassnacht, F.E., Hartig, F., Berger, C., Hernndez, J., Corvaln, P., Koch, B., forest using Fourier transforms of proles from lidar or interferometric SAR.
2015. Stratied aboveground forest biomass estimation by remote sensing Geophys. Res. Lett. 37, L23403.
data. Int. J. Appl. Earth Obs. Geoinf. 38, 229241. Tsui, O.W., Coops, N.C., Wulder, M.A., Marshall, P.L., 2013. Integrating airborne
Lutz, J.A., Larson, A.J., Swanson, M.E., Freund, J.A., 2012. The ecological importance LiDAR and space-borne radar via multivariate kriging to estimate
of large-diameter trees in a temperate mixed-conifer forest. PLoS One 7 (5), above-ground biomass. Remote Sens. Environ. 139, 340352.
e36131, http://dx.doi.org/10.1371/journal.pone.0036131 White, J., Wulder, M.A., Varhola, A., Vastaranta, M., Coops, N.C., Cook, B.D., Pitt, D.,
Maltamo, M., Eerikinen, K., Pitknen, J., Hyypp, J., Vehmas, M., 2004. Estimation Woods, M., 2013. A best practices guide for generating forest inventory
of timber volume and stem density based on scanning laser altimetry and attributes from airborne laser scanning data using an area-based approach.
expected tree size distribution functions. Remote Sens. Environ. 90 (3), Information Report FI-X-010. The Canadian Wood Fibre Centre, Victoria,
319330. British Columbia. ISBN 978-1-100-22385-8.
McElhinny, C., Gibbons, P., Brack, C., Bauhus, J., 2005. Forest and woodland stand Wulder, M.A., Coops, N.C., Hudak, A.T., Morsdorf, F., Nelson, R., Newnham, G.,
structural complexity Its denition and measurement. For. Ecol. Manage. 218, Vastaranta, M., 2013. Status and prospects for LiDAR remote sensing of
124. forested ecosystems. Can. J. Remote Sens. 39 (81), 8185.
McRoberts, R.E., Tomppo, E.O., 2007. Remote sensing support for national forest Yao, W., Krzystek, P., Heurich, M., 2012a. Tree species classication and estimation
inventories. Remote Sens. Environ. 110, 412419. of stem volume and DBH based on single tree extraction by exploiting airborne
McRoberts, R.E., Tomppo, E.O., Nsset, E., 2010. Advances and emerging issues in full-waveform LiDAR data. Remote Sens. Environ. 123, 368380.
national forest inventories. Scand. J. For. Res. 25, 368381. Yao, W., Krzystek, P., Heurich, M., 2012b. Identifying standing dead trees in forest
Nsset, E., 2004. Practical large-scale forest inventory using a small-footprint areas based on 3D single tree detection from full waveform lidar data. ISPRS
airborne scanning laser. Scand. J. For. Res. 19, 164179. Ann. Photogramm. Remote Sens. Spatial Inf. Sci. I-7, 359364.
Nsset, E., Gobakken, T., Holmgren, J., Hyypp, H., Hyypp, J., Maltamo, M., Nilsson, Yao, W., Krull, J., Krzystek, P., Heurich, M., 2014. Sensitivity analysis of 3D
M., Olsson, H., Persson, ., Sderman, U., 2004. Laser scanning of forest individual tree detection from LiDAR point clouds of temperate forests. Forests
resources: the Nordic experience. Scand. J. For. Res. 19, 482499. 5, 11221142.
Packaln, P., Maltamo, M., 2006. Predicting the plot volume by tree species using
airborne laser scanning and aerial photographs. For. Sci. 52 (6), 611622.
Polewski, P., Yao, W., Heurich, M., Krzystek, P., Stilla, U., 2015. Active learning
approach to detecting standing dead trees from ALS point clouds combined
with aerial infrared imagery. IEEE/ISPRS Workshop Looking from above: When
Earth observation meets vision - EARTHVISION 2015 in conjunction with
CVPR2015, Boston, USA.

Please cite this article in press as: Lati, H., et al., Forest inventories by LiDAR data: A comparison of single tree segmenta-
tion and metric-based methods for inventories of a heterogeneous temperate forest. Int. J. Appl. Earth Observ. Geoinf. (2015),
http://dx.doi.org/10.1016/j.jag.2015.06.008

Das könnte Ihnen auch gefallen