3-D mapping of a multi-layered Mediterranean forest using ALS data

Remote Sensing of Environment 121 (2012) 210ndash223

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment

j ourna l homepage wwwe lsev ie r com locate rse

3-D mapping of a multi-layered Mediterranean forest using ALS data

Antoacutenio Ferraz abc Freacutedeacuteric Bretar d Steacutephane Jacquemoud b Gil Gonccedilalves ce Luisa Pereira fMargarida Tomeacute g Paula Soares g

a Universiteacute Paris Est IGN Laboratoire MATIS 73 avenue de Paris 94165 Saint-Mandeacute Cedex Franceb Institut de physique du globe de Paris - Sorbonne Paris Citeacute Universiteacute Paris Diderot UMR CNRS 7154 Case 7011 35 rue Heacutelegravene Brion 75013 Paris Francec Instituto de Engenharia de Sistemas e Computadores de Coimbra rua Antero de Quental no 199 3000-033 Coimbra Portugald CETE Normandie Centre Laboratoire des Ponts et Chausseacutees 10 chemin de la Poudriegravere BP 245 76121 Grand Quevilly Francee Departamento de Matemaacutetica Universidade de Coimbra Apartado 3008 3001-454 Coimbra Portugalf Universidade de Aveiro Escola Superior de Tecnologia e Gestatildeo de Aacutegueda apartado 473 3754-909 Aacutegueda Portugalg Universidade Teacutecnica de Lisboa Instituto Superior de Agronomia Centro de Estudos Florestais Tapada da Ajuda 1349-017 Lisboa Portugal

Corresponding author at Institut Geacuteographique Navenue de Paris 94160 Saint Mandeacute France

E-mail address antonioferrazignfr (A Ferraz)

0034-4257$ ndash see front matter copy 2012 Elsevier Inc Alldoi101016jrse201201020

a b s t r a c t

a r t i c l e i n f o

Article historyReceived 11 July 2011Received in revised form 18 January 2012Accepted 21 January 2012Available online xxxx

KeywordsAirborne laser scanningLiDARMulti-layered forestUnsupervised segmentationMean shift algorithmFuel mappingVertical stratificationTree crown3-D mappingGround vegetationUnderstoryOverstory

This study presents a robust approach for characterization of multi-layered forests using airborne laser scan-ning (ALS) data Fuel mapping or biomass estimation requires knowing the diversity and boundaries of theforest patches as well as their spatial pattern This includes the thickness of the main vegetation layersbut also the spatial arrangement and size of the individual plants that compose each stratum In order todecompose the ALS point cloud into genuine 3-D segments corresponding to individual vegetation featuressuch as shrubs or tree crowns we apply a statistical approach based on the mean shift algorithm The seg-ments are progressively assigned to a forest layer ground vegetation understory or overstory Our methodrelies on a single biophysically meaningful parameter the kernel bandwidth which is related to the local for-est structure It is validated on 44 plots of a Portuguese forest composed mainly of eucalyptus (Eucalyptusglobulus Labill) and maritime pine (Pinus pinaster Ait) trees The number of detected trees varies with thedominance position from 986 for the dominant trees to 128 for the suppressed trees Linear regressionmodels explain up to 70 of the variability associated with ground vegetation and understory height

copy 2012 Elsevier Inc All rights reserved

1 Introduction

Forests woodlands and shrub formations are very important eco-systems because they provide foundations for life on Earth throughtheir ecological functions regulation of climate and water habitatfor animals and supply of food and goods They exhibit various cano-py structures from homogeneous to heterogeneous and from single-to multi-layered (Landsberg amp Gower 1997) Today we know the hori-zontal structure that describes the patchiness in forest stands betterthan the vertical structure which is difficult to quantify and yet isan important characteristic (Hall et al 2011) The canopy layers(overstory understory and ground vegetation) are distinct from eachother in their density thickness and water content A better appraisalof this vertical arrangement at high spatial resolution would bevaluable for many applications in forestry (Ares et al 2010) carbon

ational Laboratoire MATIS 73

rights reserved

cycle studies (Moore et al 2007) and ecology (Brokaw amp Lent 2000Camprodon amp Brotons 2006) As an example foresters use fuelmodelsfor predictingfire behavior (Pyne et al 1996) and fire behaviormodelssuch as FARSITE (Finney 2004) or BehavePlus (Andrews et al 2005)require information about vegetation strata thickness to detect areaswhere fire easily propagates and spreads (Anderson 1982 Sandberget al 2001)

Airborne laser scanning (ALS) is an active remote sensing tech-nique that provides georeferenced distance measurements betweena remote sensing platform and the surface (Mallet amp Bretar 2009Shan amp Toth 2009) In recent years it has been applied over naturallandscapes to extract terrain elevation (Bretar amp Chehata 2010 Krausamp Pfeifer 1998) classify land cover (Antonarakis et al 2008 Asner etal 2008 Breidenbach et al 2010 Hyyppauml et al 2008 Yoon et al2008) evaluate wildlife habitat (Clawges et al 2008 Martinuzzi et al2009) estimate biomass (Asner et al 2010 Garciacutea et al 2010 Zhaoet al 2009) and assess fuel characteristics (Andersen et al 2005Hollaus et al 2006 Mutlu et al 2008 Riantildeo et al 2003) Dependingon the nature of the target a single pulse emission may induce one or

Fig 1 Regular grid superimposed on the land cover map of the study area

Table 2Field inventory of ground vegetation and understory all stands

Mean height (m) Cover

Groundvegetation

Understory Groundvegetation

Understory

Minimum 015 0 2 0Maximum 13 6 100 95Mean 053 241 521 156Standard deviation 03 164 33 202

Fig 2 Age class of the stands The black bars correspond to the eucalyptus and the graybars to the pines

Table 3Field inventory statistics for trees in mature eucalyptus and pine stands

211A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

several backscattered echoes As the laser beam penetrates down intothe forest canopy layers an unstructured 3-D point cloud that is a dis-crete model of the target is obtained There are two main spatial scalesfor tackling the extraction of forest parameters from ALS data at theplot scale the biophysical variables are averaged over an area encom-passing several trees (eg mean canopy height biomass stem densityleaf area index) while at the individual scale they are estimated for asingle tree (eg tree height crown diameter crown base height)

Vertical stratification has been assessed at the plot scale (Maltamoet al 2005 Riantildeo et al 2003 2004 Zimble et al 2003) Morsdorf etal (2010) use the ALS intensity to discriminate different vegetationstrata They apply a supervised cluster analysis assuming that somespecies have a better light reflection ratio than others This methodworks fairly well in forest ecosystems made of mono-species strataThe intensity is somewhat difficult to analyze because it depends onthe sensor as well as on the geometry orientation and optical prop-erties of the target (leaves branches trunks) Some authors delineatevegetation strata by fitting continuous probability distributions likethe Weibull distribution or mixture models to the ALS density pro-files (Coops et al 2007 Dean et al 2009 Jaskierniak et al 2010Maltamo et al 2004) However plot-based methods are not the mostappropriatemeans to describe the vertical stratification of complex eco-systems such as Mediterranean forests that are characterized by anopen dominant canopy and a lush undergrowth made of herbaceousand woody plants (Di Castri 1981) These are often highly fragmentedforests the stratification ofwhich varies locally due to small ownershipsadministered according to different management rules (EEA 2008)

So far single-tree based methods rely on a canopy height model(CHM) which is an oversimplified representation of reality in verticallyheterogeneous canopies (Hyyppauml et al 2004 Morsdorf et al 2004Persson et al 2004 Popescu amp Wynne 2004 Solberg et al 2006) Inorder to investigate the spatial pattern of dominated trees some

Table 1Biophysical characteristics of stand 30

Height class (m) Species Dominance Mean height (m) Cover

0ndash2Ferns 95

12 50Ulex 5

2ndash8Acacia 70

60 8Pinus 30gt8 Eucalyptus 100 212 20

authors developed multi-stage approaches For instance Richardsonand Moskal (2011) first delineate groups of trees in the CHM and thencalculate the number of trees by fitting a statistical relationship to thecorresponding point cloud distribution Reitberger et al (2009) identifythe taller trees within each group determine the stem position andapply a normalization-cut segmentation method to extract the smallerones Despite good performance these approaches are site-dependentbecause they require several empirical parameters Moreover they donot properly address the issue of vertical stratification in multi-layeredforests because even if they delineate the topmost tree crowns manyALS points corresponding to ground or understory vegetation remainunassigned

Therefore it seems that an approach that simultaneously seg-ments vertical and horizontal structures of forest canopies is lackingThis paper validates a segmentation method based on the mean shiftalgorithm This method has been tested on a 3-D point cloud acquiredwith a small-footprint ALS in a multi-layered Mediterranean forestWe first present the experimental data and the algorithm The seg-mentation of the forest into different strata and the derivation ofthe geometry of individual vegetation features are then detailed

2 Experiment

21 Study area

The study area is located near the city of Aacutegueda in northwestPortugal (40deg36prime N 8deg25prime W) It covers 9 km2 and its altitude varies

DBH(cm)

CBH(m)

Totalheight(m)

Crowndepth(m)

Atypicalshape()

Eucalyptus Minimum 15 25 37 04

172Maximum 700 225 354 142Mean 97 94 133 39Standard deviation 53 39 52 24

Pine Minimum 49 71 8 04

77Maximum 414 223 25 90Mean 235 137 175 38Standard deviation 83 48 50 15

Fig 3 Mean crown depth of dominant codominant and dominated eucalyptus (blackbars) and pine (gray bars) trees per stand Standard deviations are plotted aboveeach bar and the dots represent the minimum and maximum values

212 A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

from 70 m to 220 m with slopes ranging from 25 to 342 Thelandscape is predominantly composed of woodlands dominated byeucalyptus (Eucalyptus globulus Labill) with some stands of mari-time pine (Pinus pinaster Ait) Shrublands are also present as wellas agricultural fields The eucalypts grow in pure and mixed standsthe management of which is mainly done by 3ndash4 short rotations ofabout 10ndash12 years to supply raw materials to the Portuguese pulpand paper industry Despite a limited spatial extension the studyarea displays various kinds of stands and trees in terms of age andcanopy structure The lower strata are composed mainly of suppressedtrees (eucalyptus pine acacia and oak) gorse bush (Ulex spp) heath(Erica spp Pterospartum spp) ferns and herbaceous plants

22 Field data collection

The forest inventorywas performed in the framework of a Portugueseresearch project in accordance with a field protocol recommendedby the Portuguese National Forest Inventory (AFN 2009) The super-imposition of a 325 mtimes325 m regularly spaced grid on a land covermap (DGRF 2005) led to the selection of 45 plots covered mainlyby eucalypts and 2 plots covered mainly by pines (plots 100 and200 Fig 1) The coordinates of the plot centers correspond to thegrid cell centers they were staked out in the field using GPS orwhen the signal was too weak traditional terrestrial surveying tech-niques If the plot center was inaccessible due to dense shrubby veg-etation it was shifted to one of the eight points located at a distanceof 50 m in all cardinal and intercardinal directions Three eucalyp-tus plots could not be sampled Each plot actually consists of twoconcentric circles an outer (400 m2) and an inner (200 m2) circlehereafter called plot and subplot They were delimited using a deca-meter and the trees were numbered using a marker pen The fieldoperators defined different forest stands ie uniform plant

Table 4ALS acquisition parameters

ALS sensor RIEGL LMS-Q560

Wavelength 1550 nmScan angle 45degPulse rate 150 kHzEffective measurement rate 75 kHzBeam divergence 05 mradGround speed 4626 msFlying height 600 mSwath width 479 mSwath overlap 70Nominal distance between two lines 150 mFootprint diameter 30 cmSingle run density 33 ptm2

Expected final point density 99 ptm2

communities in terms of species age and spatial arrangement(Stokes et al 1989) We use stand and substand to designate the for-est stands corresponding to plot and subplot If a plot containedmorethan one stand only the stand coincident with the plot center wasdescribed Among the forest biophysical variables measured duringthe field work the vertical structure (at the stand level) and thesize and shape of individual trees (at the substand level) were care-fully investigated (Pereira et al 2009)

The vertical structure of a stand was described by seven heightclasses (0ndash06 m 06ndash1 m 1ndash2 m 2ndash4 m 4ndash8 m 8ndash16 m and gt16 m)that could be aggregated in situ to better represent the vegetation strata(Table 1) The mean height and percent cover of each stratum were

Fig 4 Mean shift segmentation applied to (a) 2-D and (b) 3-D simulated tree crownsThe initial kernel bandwidths with different vertical and horizontal components arerepresented by cylinders The mean shift vectors are represented by arrows that definethe successive positions of the kernel bandwidth (dashed cylinders) All the data pointsthat have converged to the same mode (filled gray sphere) are grouped together Thegray lines in (c) correspond to the trajectory of random points

Fig 5 Mean shift segmentation of a simulated forest scene using (left) h=(132)m (middle) h=(335)m and (right) h=(69)m The cylinders correspond to the differentkernels and the gray spheres represent the calculated modes

213A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

visually estimated by the field operators Note that one treemay belongto several height classes

All trees were assigned a class and a dominance position (domi-nant codominant dominated and suppressed) Calipers gave a directmeasurement of the diameter at breast height (DBH) whereas thetotal height and the crown base height (CBH) were measured usingeither a telescopic tape measure or a Vertex hypsometer We onlyconsidered trees higher than 2 m with a DBH larger than 5 cm Notethat the forest inventory data are usually acquired with a lower geo-metric accuracy than the ALS data To improve the accuracy a localgeodetic network made of 41 pairs of GPS-derived points was builtin the same map projection as the ALS data (Fig 1) to survey thetree positions using total stations (Gonccedilalves amp Pereira in press) All thedata were subsequently integrated into a single three-dimensionalgeometry

Fig 6 Mean shift segmentation algorithm at the plot level (left) and subsequent histogramare the ground vegetation understory and overstory thicknesses

23 Characteristics of the stands

Table 2 sums up the main structural characteristics of ground veg-etation and understory The large range of percent cover indicatesthat the canopy varies from sparse to very dense

The forest is highly variable in terms of tree age architecture andmetrics The eucalyptus stands are between 1 and 13 years old whilethe two pine stands are 30 and 60 years old (Fig 2) In total there are12 plots with juvenile stands (1ndash4 years) and 32 plots with maturestands (gt 4 years)

The plots contain one (59) or more (41) forest stands They maydisplay an intrinsic structural heterogeneity the architecture of thetrees differs depending on whether they grow in the middle of theforest or near roads and clearings In open space areas the treestend to expand horizontally to search for light reducing their apical

(right) htus and htos are the understory and overstory height thresholds Agv Aus and Aos

Fig 7 Horizontal (gs Gaussian profile surface) and vertical (gr Epanechnikov profilecurve) kernel profiles The point and color bar indicates their weight in the calculationof the kernel barycenter (For interpretation of the references to color in this figurelegend the reader is referred to the web version of this article)

214 A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

dominance Heterogeneity also influences ground vegetation andunderstory since clearings let direct sunrays reach the lowest strataAbout 50 of the measured stands are considered to be heterogeneous

The stands can also be sorted according to three regenerationmethods forest planting produces the so-called high forests (euca-lyptus and pine) coppicing a traditional method of woodland man-agement that consists in pruning trees to near the base allows thestumps to regenerate over-vigorous coppiced trees (eucalyptus) andwhen after cutting a stand contains trees that are left to grow to fullheight it belongs to the category coppice-with-standards (pine)Twenty-five stands are allocated to high forest 16 to coppice and 3 tocoppice-with-standards Table 3 summarizes the main structural char-acteristics of mature eucalyptus and pine trees as well as the percent-age of trees with atypical shape crooked leaning and broken treesSpecimens belonging to juvenile stands are not processed as individualsbut as a forest stratum

Fig 3 details the crown depth in terms of minimum maximummean and standard deviation for each stand Suppressed trees thatare poorly represented in the point cloud are omitted

24 Airborne laser scanning data

The data were acquired on July 14 2008 in a full-waveform modeusing a LiteMapper 5600 airborne LiDAR system (Table 4) which dig-itizes the waveform of the echo signal for every emitted laser pulseThe company in charge of the airborne measurements delivered boththe raw and processed laser data The digitized waveforms were con-verted into echo signals each laser pulse giving rise to 1ndash5 ALS points(RiANALIZE RIEGL 2011a) The position and orientation of the plat-form which are given by onboard GPSIMU measurements were cor-rected by analyzing overlapping laser strips from the calibration flightlines (TerraMatch Burman amp Soininen 2004) These parameterstogether with the GPS measurements acquired during the flight usinga reference ground station provided a point cloud in the WGS84UTMzone 29N coordinate system for further processing (RiWORLD RIEGL

Fig 8 (a) Original point cloud measured on plot 17 (b) MS algorithm applied using aradially symmetric kernel and a 3 m bandwidth (c) MS segments corresponding tomore than five ALS points

Fig 9 Workflow of the adaptive mean shift algorithm

215A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

2011b) Systematic height errors were finally removed by using groundcontrol data spread over the study area

The average point density within each plot is of 95 ptm2

(min=47 ptm2 max=155 ptm2 σ=19 ptm2) To calculate theeffective height of the objects in the scene ground and vegetationpoints were separated (TerraScan Soininen 2010) A Delaunay trian-gulation was then generated to produce a 03 mtimes03 m digital terrainmodel which was used to normalize the point cloud Note that thepoints filtered as ground were kept in the dataset

3 Methodology the mean shift algorithm

An ALS point cloud can be regarded as a multimodal distributionwhere each mode here defined as a local maximum both in densityand height corresponds to a crown apex In this study we investigatethe ability of the mean shift (MS) algorithm to extract the modes ofthe point cloud Due to the complexity of the forest stands whichmix shrubs suppressed trees and dominant trees a single kernelbandwidth is unsuitable To improve the segmentation of individualvegetation features we propose to apply a bottom-up iterative meth-od based on an adaptive MS algorithm which sequentially segmentsindividual vegetation features

31 Background

The mean shift has been primarily applied to image segmentation(Comaniciu ampMeer 2002) Here we explore its potential for segment-ing a three-dimensional point cloud The Parzen window (or kerneldensity estimation) technique is a method for estimating the proba-bility density function (PDF) of a random variable X that is distributedin a d-dimensional space Rd Each point Xi contributes to the PDFbased on its distance from the center of the volume where the dataare distributed The estimated PDF is

f hK Xeth THORN frac14 1nhd

Xnifrac141

KXminusXi

h

eth1THORN

where n is the number of samples of the random variable K is thechosen kernel function and h called the bandwidth is the smoothingparameter that determines the contribution of each sample K is anon-linear function of the distance from the data points to X Wedefine a radially symmetric kernel that satisfies K(X)=ckdtimesk(X2)where ckd is a normalization constant which makes K integrate toone and k is called the kernel profile The algorithm tries to determinelocal maxima of the density function f(X) which correspond to thezeros of the gradient nabla f(X)=0 Assuming that g is the derivative ofthe kernel profile g(X)=minusk (X) and G the corresponding kerneldefined by G(X)=cgdtimesg(X2) where cgd is another normalizationconstant Comaniciu and Meer (2002) calculate the density gradientestimator as

nablaf hK Xeth THORN frac14 f hG Xeth THORN 2ckdh2cgd

mhG Xeth THORN eth2THORN

with mhG(X) the mean shift vector

mhG Xeth THORN frac14

Pnifrac141

Xi gXminusXih

2

Pnifrac141

g XminusXih

2 minusX eth3THORN

The mean shift is the difference between the weighted mean(G-distance) using the kernel G for weights and X the center of the

kernel mhG(X) can be inferred from Eq (2)

mhG Xeth THORN frac14 h2cgd2ckd

nablafhK Xeth THORNfhG Xeth THORN

eth4THORN

Eq (4) shows that at location X the mean shift vector computedwith kernel G is proportional to the normalized density gradient esti-mate obtained with kernel K Thus it always points toward thedirection of the maximum slope of the density function The proce-dure does not need to evaluate the density function fhK itself butonly the kernel profile g In a multidimensional space the kernel isusually split into two or more kernels Here we separate the horizon-tal and vertical domains The MS vector is then defined as

mh G Xeth THORN frac14

Pnifrac141

Xi gs XsminusXs

ihs

2

gr XrminusXri

hr

2

Pnifrac141

gs XsminusXsi

hs

2

gr XrminusXri

hr

2 minusX eth5THORN

where the superscripts s and r refer to the horizontal and verticaldomains gs and gr are the two associated kernel profiles hs and hr thetwo bandwidths and Xs and Xr the two components of the vectors Atstep t the iterative process can be written as

Xtthorn1larrXt thornmhG Xt

eth6THORN

2-D and 3-D synthetic tree crowns were simulated to test the per-formance of the MS algorithm Fig 4a and b shows that points con-verge toward the modes This procedure can be easily extended to adistance-based segmentation technique if all the data points that con-verge toward the same mode are grouped together (Fig 4c) All themodes inscribed in a sphere with radius 1 m are considered as a sin-gle mode

216 A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

32 Determination of the bandwidth

The choice of the kernel bandwidth is critical because it stronglyimpacts on the results Setting a small value produces several distinctmodes (local basins of attraction) while setting a large one aggre-gates small structures into larger ones (large basins of attraction)The determination of an optimal value is actually a major challengeThe thickness of the forest strata generally increases with heightie scrubby vegetation is typically thinner than overstory Threesegmentations have been applied to a simulated scene using differ-ent bandwidths (Fig 5) The smaller bandwidth that is optimal forground vegetation tends to fragment the trees into numerous seg-ments (Fig 5a) Increasing the bandwidth definitely improves thesegmentation of the understory without effect on the taller trees(Fig 5b) Finally the optimal bandwidth for the overstory causesunder-segmentation of the scene (Fig 5c) Worse yet dense groundvegetation tends to attract a sparse understory overestimating thethickness of this layer Thus using a single scale over the entire spaceis not suitable for the analysis of forest environments The issue ofbandwidth selection has been studied for the purpose of multiscale

Fig 10 Segmentation of plot 30 with htus=1m and htos=8m The black dots correspondnext iteration (andashb) First iteration w=0m and hgv=(11) (cndashd) Second iteration w=2(a) and (f) respectively correspond to the field-measured and ALS-derived mean height of gcolor in this figure legend the reader is referred to the web version of this article)

segmentation using either multispectral or hyperspectral images(Bo et al 2009 Comaniciu 2003 Huang amp Zhang 2008) VariablebandwidthMS has already been proved to converge and even to sur-pass fixed bandwidth MS (Comaniciu amp Meer 2002)

In order to properly segment individual vegetation features a dif-ferent bandwidth is assigned to each vegetation stratum The thickerthe forest layer the larger the bandwidth Since vegetation volumesare better predicted if the stratum thickness is known the first stageof the algorithm consists in plotting the height histograms of the forestplots in order to identify the strata overstory understory and groundvegetation A first pass of the MS algorithm is applied to the ALS pointcloud to compute their basins of attraction Eq (5) is applied to theALS points using the uniform kernel profile on both components

gs Xs frac14 1 if Xs le10 otherwise

and gr Xr frac14 1 if Xr le10 otherwise

eth7THORN

Thus in such a case the ratio in Eq (5) is simply the mean of theALS points contained within a cylinder of radius hs height hr cen-tered in X To remove the influence of the horizontal coordinates hs

to the ALS points that remain unlabelled after an iteration and that are inputs for them and hus=(2335) (endashf) Third iteration w=95m and hos=(4365) The lines inround vegetation (green) and understory (red) (For interpretation of the references to

217A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

is set to the plot diameter (~22 m) and hr is defined as the value thatforces the ALS points to converge toward twomodes We set hr=1 mas an initial estimate and increment it to obtain these two modes(Fig 6a) The borderline between the basins of attraction of eachmode defines the overstory height threshold htos (Fig 6b) We as-sume that a plot holds a single layer when htosb1 m and two layerswhen htosb5 m otherwise a third layer may exist In this case theunderstory height threshold htus is set to 1 m Afterwards one can eas-ily compute the thickness of the overstory (Aos) understory (Aus) andground vegetation (Agv)

Finally the kernel bandwidth h=(hshr) corresponding to thecrown segmentation is adapted to the vegetation architecture to ac-count for the aspect ratio of tree crowns so the vertical (hr) and hor-izontal (hs) components may be different (Morsdorf et al 2004)Based on the current ALS dataset we find that the tree crown heightis at least two thirds larger than the crown diameter while groundvegetation is spherical (hgvs =hgv

r ) Then equalizing the two verticalbandwidths hos

r and husr to half the thickness of the layers avoids

under-segmentation in bilayered forests (Eqs 8ndash9) Since groundvegetation is always considered as a uniform layer the bandwidthhgv is set to the corresponding thickness in both directions (Eq 10)

hos frac142hros3

Aos

2

eth8THORN

hus frac14

2hrus3

Aus

2

eth9THORN

hgv frac14 AgvAgv

eth10THORN

33 Adjustment of the kernel profile

We design a 3-D kernel profile as the product of two profilesto compute the modes of the point cloud ie the crown apices

Fig 11 Original point cloud for (a) plot 47 only composed of pine trees and (c) plot 16 mheights of ground vegetation (green) and overstory (blue) are represented by the lines in tground vegetation understory and overstory calculated from the individual vegetation featuin both figures (For interpretation of the references to color in this figure legend the reade

Whereas the horizontal profile searches for the local density max-ima the vertical one dealswith the local heightmaxima The horizontalkernel profile gs follows a Gaussian function

gs xeth THORN frac14 exp minusγ xk k2

eth11THORN

with γ=5 Isotropic kernels are standard in image segmentationwhere emphasis is put on bandwidth selection (Comaniciu 2003Singh amp Ahuja 2003) Asymmetric kernels have been used in videotracking to adapt to the structure of moving targets eg an airplaneor a human body (Wang et al 2004 Yi et al 2008 Yilmaz 2007)In this study an asymmetric kernel is applied to the vertical compo-nent in order to assign a higher weight to the highest points withinthe bandwidth (Fig 7) Therefore the MS vector converges towardthe local height maximum Following Yilmaz (2007) and Yi et al (2008)we first create a mask of the foreground object

mask Xieth THORN frac14 1 if Xrminus h4leXr

ileXr thorn h2

0 otherwise

8lt eth12THORN

And the kernel value is the distance between one data point andthe boundary of the mask

dist Xieth THORN frac14 minXrminushr

4

minusXr

i

3hr

8

Xr thorn hr

2

minusXr

i

3hr

8

8gtgtgtltgtgtgt

9gtgtgt=gtgtgt

if mask Xieth THORN frac14 1

0 otherwise

8gtgtgtgtgtltgtgtgtgtgt

9gtgtgtgtgt=gtgtgtgtgt

eth13THORN

where 3hr8 is a normalizing factor equal to half the bandwidth ofthe asymmetric kernel Using an Epanechnikov profile the weight of

ade of two stands Both plots do not display understory layers and the measured meanhe figures (b) MS individual vegetation features from (a) (d) Canopy height model ofres computed in (c) The surveyed tree metrics are also shown (line segments in black)r is referred to the web version of this article)

Table 5Linear regression parameters for ALS-derived versus field-measured vegetation meanheight () The results only concern juvenile stands Negative values mean anunderestimation

Number of stands Outliers R2 RMSE (m) Δh (m)

Ground vegetation 44 3 070 015 0Understory 32 5 068 096 044Overstory () 10 2 092 031 minus012

218 A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

each point is calculated using

gar Xieth THORN frac14 1minus 1minusdist Xieth THORNk k2 if mask Xieth THORN frac14 10 otherwise

eth14THORN

In the case of an asymmetric kernel the MS vector in Eq (5) canbe then rewritten as

mh G Xeth THORN frac14

Pnifrac141

Xi gs XsminusXs

ihs

2

gar Xieth THORNPnifrac141

gs XsminusXsi

hs

2

gar Xieth THORNminusX eth15THORN

Fig 12 Analysis of the R2 (left axis) and the RMSE (right axis) for height estimation as a funplots used to calculate these statistics is inscribed in the bars

Note that the profile is still radially symmetric (Eq 14) The neigh-borhoods accounted for in the calculation of mhG(X) are selected asa function of an asymmetric bandwidth The weighted distance be-tween points is the product of the two kernels which makes themethod more robust (Fig 7) For instance overlapped crowns mayalso correspond to local density maxima Whereas the horizontal pro-file tends to converge to such zones the vertical profile forces the MSvector to converge on the local height maximum ie the crown apexConversely when undergrowth and overgrowth vegetation interpen-etrate the vertical profile tends to converge toward the upper plantsIn such a case the horizontal profile helps the MS vector to stabilizeon the crown apex of the lower plants which is supposed to be dens-er than the crown base of the upper plants

34 Pre-processing of the point cloud

In a forest canopy the laser beams hit leaves branches andtrunks Since the point cloud is very scattered keeping all points sig-nificantly overestimates the number of individual vegetation featuresas well as the estimation of the stratum height In order to identify thecrown elements in the 3-D point cloud the mean shift (Eq 5) has

ction of the percent cover for (a) ground vegetation and (b) understory The number of

Fig 13Modeled vs field-measured CBH for (a) eucalypts (∘ dominant loz codominant Δdominated suppressed) and (b) pine trees

219A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

been applied to each plot using a uniform kernel (Eq 7) and thebandwidth h=(hshr) with hs=(33)m and hr=3m If all seg-ments containing less than 5 points are removed from the data setbecause of their poor topological structure the bandwidth is largeenough to keep the most significant vegetation features (Fig 8) How-ever this technique may remove suppressed trees that are poorlyrepresented in the point cloud due to occlusion that masks someparts of the canopy volume

35 Extraction of individual trees and refinement of the forest strata

The algorithm involves two or three iterations (Fig 9) It first com-putes a set of mean shift vectors using the ALS points (Eq 15) whichare all considered as seeds The vectors search for the local highestdensity direction with the appropriate bandwidth The latter is select-ed by calculating the 5th height percentile of the current point cloudw In the first iteration the bandwidth is set to hgv (Fig 10a) sincew always tends toward 0 m A trajectory links every ALS point witha certain mode A vegetation feature having a mode lower than htusis considered as ground vegetation (Fig 10c green ellipsoids) Atthe end of the first iteration the corresponding ALS points are re-moved from the point cloud The calculation of w in the second itera-tion defines the bandwidth and therefore the number of iterations(two or three) The bandwidth is hus if wbhtos orand htos if wgthtos

The second iteration extracts the understory which correspondsto vegetation features with modes ranging between htus and htos(Fig 10e red ellipsoids) The third iteration identifies the overstoryas vegetation features with modes higher than htos (Fig 10f blue el-lipsoids) Applying a threshold to the mode space allows definitionof fuzzy frontiers between the strata This is physically meaningfulcompared to a simple vertical stratification based on height thresh-olds After each iteration removing points already assigned improvesthe segmentation by reducing the influence of the denser layersThus when two regions of different densities are close together thepoints belonging to sparser regions are likely to be aggregated bythose belonging to the denser ones This effect is obvious in Fig 5bwhere the forest strata are either overestimated or underestimated

4 Results

This section discusses the results of the algorithm over 44 plotsThey are validated in terms of the forest vertical stratification aswell as the identification of individual trees

41 Segmentation of forest strata

The mean height of ground vegetation is calculated as the 90thheight percentile (Riantildeo et al 2007) of the corresponding laser points(green ellipsoids of Figs 10f and 11b) Unlike other approaches wekeep all the points including ground reflections which justify such ahigh value The 50th height percentile is naturally used to calculatethe mean heights of understory (Fig 10f red ellipsoids) and overstory(juvenile stands Fig 11d) (Peterson 2005)

Linear regression analysis allows investigation of the strength ofthe relationship between the ALS-derived and field-measured heightsof each forest stratum (Table 5) The outliers that represent about7 and 16 of the plots in ground vegetation and understory respec-tively are identified after Huber (1981) and removed from the linearregressions A linear model with a satisfactory RMSE explains 70 ofthe variability associated with ground vegetation height Note therefinement accomplished by the algorithm initially set to a 1 mthreshold (Fig 6) the computed height ranges from 015 m to 125 mThe number of retrieved layers is inherent to the forest patternAlthough all mature plots were initially divided into three stratastands 9 29 45 46 and 47 converge toward only two strata(Fig 11andashb) which means that the echoes reflected by the trunks

are successfully identified Due to the lack of understory the con-dition wgthtus is verified earlier in the second iteration and con-sequently the kernel bandwidth is immediately optimal for theoverstory stratum The MS algorithm also works on plots contain-ing several stands the vertical stratification of which varies radi-cally (Fig 11d) The mean height of the understory is overestimatedThe linear model explains 68 of the variance (Table 5) This may bedue to the assignment of suppressed trees to this layer contrary tofield measurements These trees can be considered as understorysince they grow below the canopy and do not receive direct sunlightAs expected the estimates of overstory mean height are more accuratefor the juvenile stands (Table 5)

Fig 12 showshow the percent cover affects the estimation of groundvegetation and understory height Ground vegetation is surprisingly notmuch affected with R2 varying from 070 to 080 and RMSE lower than002 m (Fig 12a) As for the understory the percentage of explainedvariance increases with the percent cover while the RMSE decreases(Fig 12b) A higher percent cover indicates more plant material and ahigher proportion of laser pulses hitting the canopy Therefore thediscrete model of vegetation generates a better estimate of forest pa-rameters The understory height is more accurate when the percent

Fig 14 Flowchart of the reference trees (RT) and ALS segments (S) linkage method

Table 6Tree identification () In total there are 167 suppressed reference trees but 50 thathave been classified as understory are not taken into account

Tree Dominanceposition

Referencetrees

Identified FP

DT DTminusFN

Eucalyptus Dominant 146 145 (993) 144 (986)

60 (92)Codominant 176 163 (926) 150 (852)Dominated 210 138 (657) 129 (614)Suppressed 117 17 (145) 15 (128)

Pine 52 50 (961) 48 (923) 0Total 701 513 (732) 486 (693) 60 (86)

220 A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

cover exceeds 10 thus a post-processing analysis for identifyingsparse canopies may improve the results

We are interested in comparing our results with CBH which playsa greater role in forest stratification Fig 13 compares the field-measured CBHs with those modeled by selecting the lowest pointssorted out as overstory in 03 mtimes03 m areas (Fig 10f and Fig 11bblue and colored ellipsoids) The missing pixels were generated usinga Delaunay triangulation Such a surface explains 76 of the variabilityof the pine CBH but it poorly characterizes the eucalyptus standswhich are more heterogeneous (Fig 13)

42 Identification of individual tree crowns

As in Solberg et al (2006) and Reitberger et al (2009) the 3-Dsegmentation of individual tree crowns is validated by comparingfield measurements with ALS segments (Figs 11b and 14) A segmentis linked with a reference tree provided that i) the distance dS-RT islower than 70 of the mean distance dNT between eight neighboringtrees and ii) the height values of at least 50 of the ALS points of SZS 50 are contained between the CBH and the tree height

If a segment is assigned tomore than one reference tree the farthesttree from the reference tree is considered a false negative (FN) In orderto quantify the remaining omission errors the neighborhood ofunlinked reference trees was analyzed using a cylinder of radius15 m If there is at least one laser point linked with another refer-ence tree within this volume the current one is also called a falsenegative Thus the FN class means that the tree crown was detected bythe ALS but the algorithm failed to see it as a tree This is the case whentwo crowns were clustered in the same segment If no laser point be-longs to this buffer area a reference tree is declared as an undetectedtree (UT) Finally segments linkedwith any reference tree are classifiedas false positive (FP) This classmay contain vegetation features wrong-ly assigned to the overstory eg tall shrubs but also trees located out-side the substand boundary when their crowns fall inside and are notsurveyed Thus the detected trees (DT) quantify the performance ofALS in characterizing the forest (Table 6)

As expected the detection rate decreases with dominance positionThe estimation error of biomass or basal area should vary accordingly

(Persson et al 2002) To report the number of trees missed by themethod we can sum the omission errors introduced by the algorithmie DTminusFN They are actually low compared to those introduced bythe ALS (07 74 43 17 and 38 percentage points for dominant co-dominant dominated suppressed and pine respectively) The percent-age of FP or commission error equals 86 which is in good agreementwith other studies In a forest mainly covered with Norway spruceEuropean beech fir and sycamore maples Reitberger et al (2009)detect 66 of the reference trees (upper layer 88 intermediatelayer 35 lower layer 24) with a commission error of 11 In aNorway spruce forest Solberg et al (2006) announce a global detec-tion rate of 66 (dominant trees 93 codominant trees 63 sub-dominant trees 38 and suppressed trees 19) with a commissionerror of 26 It is unclear whether the omission errors reported byother studies are due to the inability of the ALS to characterizetree crowns or to the algorithm itself Therefore it is tricky to com-pare our results with the literature since the forest architecture andthe ALS configuration both have an important effect on the accuracyof the different methods

Although the present method searches for local density maximain the point cloud it is not affected by the point density variabilitybecause the MS is a kernel gradient estimator ie it does not evalu-ate the density function itself but normalized local gradients Thusprovided that the local density and height gradients point towardthe crown apices the point density at which the crowns are sampled

221A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

has only a slight impact on the mode search ie on the identification ofindividual vegetation features

43 Validation of tree height and CBH

Fig 15 correlates the ALS-derived and field-measured tree height(Fig 15a and 15c) and CBH (Fig 15b and 15d) for the identified treesCharacterization of the CBH greatly improves in eucalyptus standswhen individual trees are first extracted (Figs 13a and 15b) while itis slightly better in pine stands (Figs 13b and 15d) Table 7 showsthat ourmethod globally underestimates the tree height with a limitedinfluence of the dominance position The slopes of the linear regressionsalmost equal 1 the R2 vary between 091 and 095 and the RMSE be-tween 075 m and 090 m These results are comparable with those ofother studies that show that ALS data tend to underestimate tree height(Gaveau amp Hill 2003 Hyyppauml et al 2008)

Our method overestimates the CBH of 129 m for eucalyptus anda positive correlation with the dominance position is obvious Thelinear regressions follow the same trends with an R2 increasing from058 (dominant) to 071 (suppressed) and an RMSE decreasing from280 m (dominant) to 130 m (suppressed) The crown base is not aswell delineated for eucalyptus as for pine Suppressed trees are morecompact than taller trees the shape of which is more complicatedwith small dead branches lying on the stems Moreover the reflectionof the laser beam on a curved branch can be located under the field-measured CBH This variable is actually difficult to survey because ofits approximate definition it can be viewed as the height of the firstbranch along the stem or as the height where the crown bulk densityexceeds a critical threshold of 0011 kgm3 (Scott amp Reinhardt 2001)The pine CBH is underestimated by 066 m mainly because of deadbranches that were not measured in the field Many ALS points corre-sponding to trunks are also clustered together with crowns particularlyin the old stands Compared to eucalypts and young pines trunks of old

Fig 15 ALS-derived vs field-measured tree height (andashc) and CBH (bndashd) for eucalyptus (

pines are well represented in the point cloud Other methods are moresuccessful in removing their reflections (Popescu amp Zhao 2008) but it isunclear whether they would improve the CBH estimation Our resultsagree with other studies in a Scots pine forest Riantildeo et al (2004)claim that ALS overestimates the CBH and obtain R2 values rangingfrom 065 to 068 In Norway spruce and Scots pine forests Holmgrenand Persson (2004) also notice an overestimation by 075 m (R2=084RMSE=282 m) Popescu and Zhao (2008) extract the CBH of pinesand deciduous trees with an RMSE of 208 m and an R2 of 078

5 Conclusion

This study demonstrates the ability of our method to provide gen-uine 3-D segments corresponding to individual vegetation features ofthe main forest layers ground vegetation understory and overstoryUnlike other methods our approach does not rely on a CHM and di-rectly applies to the 3-D point cloud which is an advantage in charac-terizing heterogeneous forests Segmentation occurs in the modespace where vegetation features are more likely to be discriminatedOur maps allow local calculation of specific statistics for each vegeta-tion layer and consequently accurate delineation of forest areas withsimilar horizontal and vertical structures ie forest stands and conse-quently fuel types Moreover our approach introduces a robust dis-crimination between ground vegetation and taller plants

We show that the mean shift algorithm is a reliable technique forfinding the modes in the multi-modal point cloud distribution of amulti-layered Mediterranean forest Due to the complex pattern ofthe forest environment we established a multi-scale approach wheremodes are computed with an adaptive kernel bandwidth optimizedfor each stratum However so far it can only handle forest structureswith a maximum of three layers A more sophisticated method mightbe developed to deal with highly stratified environments

andashb dominant loz codominant Δ dominated suppressed) and pine trees (cndashd)

Table 7Linear regression parameters for data displayed in Fig 15 Negative values mean an un-derestimation while positive values mean an overestimation

Tree Dominanceposition

Δh (m) R2 RMSE (m)

TH CBH TH CBH TH CBH

Eucalyptus Dominant minus023 144 095 058 085 280Codominant minus027 145 095 061 087 270Dominated minus017 103 093 067 090 192Suppressed minus022 073 091 071 075 130All together minus023 129 096 069 086 248

Pine minus028 066 094 079 107 225

222 A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

Our approach relies on only one parameter the three-dimensionalkernel bandwidth Its vertical component is set as a function of thestratum depth and its horizontal component is defined in relation tothe vertical one Therefore the kernel bandwidth has a biophysicalmeaning the width of a crown depends on its length and the depthof a forest stratum on the length of the crowns Note that these corre-lations may vary significantly depending on the tree species and theforest biome Thus it is necessary to determine the validity domainof these kernel bandwidth settings The robustness of the methodwas assessed at four different levels

a) Intra-plot The method is able to depict the real nature of the stra-ta even when the vertical stratification varies within a plot (41of the plots have more than one stand Fig 11d)

b) Intra-stand The bandwidth settings apply well to crowns with dif-ferent volumes from suppressed to dominant trees (Fig 3 andTable 6)

c) Inter-stand The validated stands display structures with differentarrangements from little to lush ground vegetation combined witheither absent or luxurious understory that can co-exist with over-growth vegetation at different growth stages (Fig 2 and Table 2)

d) Inter-plot Our forest is made up of many small properties thatlead to a fragmented landscape The method does a good job ofhandling the point density variability within the study area (Fig 1and Table 4)

Finally the correlation between field measurements and ALS-derived structural characteristics of ground vegetation and understo-ry depends on the forest type and the ALS configuration Such valuesmay be different in forests with more closed canopies or sparser pointclouds

Acknowledgments

This experiment is part of a PTDCAGR-CFL723802006 researchproject A Ferraz holds a fellowship (SFRHBD383902007) fundedby the Portuguese Foundation for Science and Technology Manythanks to Susan L Ustin (UC Davis) for editing the paper IPGP con-tribution no 3257

References

AFN (2009) Instruccedilotildees para o trabalho de campo do Inventaacuterio Florestal Nacional IFN20052009 Direccedilatildeo de Unidade de Gestatildeo Florestal Divisatildeo para a IntervenccedilatildeoFlorestal Lisboa Portugal Autoridade Florestal Nacional 62 pp

Andersen H E McGaughey R J amp Reutebuch S E (2005) Estimating forest canopyfuel parameters using LiDAR data Remote Sensing of Environment 94 441ndash449

Anderson H (1982) Aids to determining fuel models for estimating fire behaviorUSDA forest servicemdashintermountain experiment station 22 pp

Andrews P Bevins C amp Seli R (2005) BehavePlus fire modeling system version 30Users guide revised USDA forest servicemdashrocky mountain research station 132 pp

Antonarakis A S Richards K S amp Brasington J (2008) Object-based land cover clas-sification using airborne LiDAR Remote Sensing of Environment 112 2988ndash2998

Ares A Neill A R amp Puettmann K J (2010) Understory abundance species diversityand functional attribute response to thinning in coniferous stands Forest Ecologyand Management 260 1104ndash1113

Asner G P Hughes R F Vitousek PM Knapp D E Kennedy-Bowdoin T Boardman Jet al (2008) Invasive plants transform the three-dimensional structure of rain for-ests Proceedings of the National Academy of Sciences of the United States of America105 4519ndash4523

Asner G P Powell G V N Mascaro J Knapp D E Clark J K Jacobson J et al(2010) High-resolution forest carbon stocks and emissions in the Amazon Pro-ceedings of the National Academy of Sciences of the United States of America 10716738ndash16742

Bo S Ding L Li H Di F amp Zhu C (2009) Mean shift-based clustering analysis ofmultispectral remote sensing imagery International Journal of Remote Sensing 30817ndash827

Breidenbach J Naeligsset E Lien V Gobakken T amp Solberg S (2010) Prediction ofspecies specific forest inventory attributes using a nonparametric semi-individualtree crown approach based on fused airborne laser scanning and multispectraldata Remote Sensing of Environment 114 911ndash924

Bretar F amp Chehata N (2010) Terrain modelling from lidar range data in naturallandscapes A predictive and Bayesian framework IEEE Transactions on Geoscienceand Remote Sensing 48 1568ndash1578

Brokaw N V amp Lent R A (2000) Vertical structure In M L Hunter (Ed)Maintainingbiodiversity in forest ecosystems (pp 373ndash399) Cambridge University Press

Burman H amp Soininen A (2004) Available online at TerraMatch users guide httpwwwterrasolidfisystemfilestmatchpdf (accessed 6072011)

Camprodon J amp Brotons L (2006) Effects of undergrowth clearing on the bird com-munities of the Northwestern Mediterranean Coppice Holm oak forests ForestEcology and Management 221 72ndash82

Clawges R Vierling K Vierling L amp Rowell E (2008) The use of airborne lidar to as-sess avian species diversity density and occurrence in a pineaspen forest RemoteSensing of Environment 122 2064ndash2073

Comaniciu D amp Meer P (2002) Mean shift A robust approach toward feature spaceanalysis IEEE Transactions on Pattern Analysis and Machine Intelligence 24 603ndash619

Comaniciu D (2003) An algorithm for data-driven bandwidth selection IEEE Transac-tions on Pattern Analysis and Machine Intelligence 25 281ndash288

Coops N C Hilker T Wulder M A St-Onge B Newnham G Siggins A et al(2007) Estimating canopy structure of Douglas-fir forest stands from discrete-return LiDAR TreesmdashStructure and Function 21 295ndash310

Dean T J Cao Q V Roberts S D amp Evans D L (2009) Measuring heights to crownbase and crown median with LiDAR in a mature even-aged loblolly pine standForest Ecology and Management 257 126ndash133

EEA (2008) European forestsmdashecosystem conditions and sustainable use EEA report no32008 Copenhagen (Denmark) European Environment Agency 105 pp

DGRF (2005) 5deg Inventario Florestal Nacional Fotointerpretaccedilao Direcccedilatildeo Geral dosRecursos Florestais Lisboa Portugal 12 pp

Di Castri F (1981) Mediterranean-type shrublands of the world In F Di Castri DGoodall amp R Specht (Eds) Ecosystems of the world Mediterranean-type shrublands(pp 1ndash52) Amsterdam (The Netherlands) Elsevier Scientific Publications

Finney M (2004) FARSITE Fire area simulator-model development and evaluationUSDA forest service research paper RMRS-RP-4 47 pp

Garciacutea M Riantildeo D Chuvieco E amp Danson F M (2010) Estimating biomass carbonstocks for a Mediterranean forest in central Spain using LiDAR height and intensitydata Remote Sensing of Environment 14 816ndash830

Gaveau D amp Hill R (2003) Quantifying canopy height underestimation by laser pulsepenetration in small-footprint airborne laser scanning data Canadian Journal of Re-mote Sensing 29 650ndash657

Gonccedilalves G amp Pereira L (in press) A thorough accuracy estimation of DTM producedfrom airborne full-waveform laser scanning data of unmanaged eucalypt planta-tions IEEE Transactions on Geoscience and Remote Sensing doi101109TGRS20112180911

Hall F G Bergen K Blair J B Dubayah R Houghton R Hurtt G et al (2011) Char-acterizing 3D vegetation structure from space Mission requirements Remote Sens-ing of Environment 115 2753ndash2775

Hollaus M Wagner W Eberhoumlfer C amp Karel W (2006) Accuracy of large-scale canopyheights derived from LiDAR data under operational constraints in a complex alpineenvironment ISPRS Journal of Photogrammetry and Remote Sensing 60 323ndash338

Holmgren J amp Persson A (2004) Identifying species of individual trees using airbornelaser scanner Remote Sensing of Environment 76 283ndash297

Huang X amp Zhang L (2008) An adaptive mean-shift analyses approach for object ex-traction and classification from urban hyperspectral imagery IEEE Transactions onGeoscience and Remote Sensing 46 4173ndash4185

Huber P J (1981) Robust statistics New York Wiley 320 ppHyyppauml J Hyyppauml H Litkey P Yu X Haggreacuten H Ronnholm P et al (2004) Algo-

rithms and methods of airborne laser scanning for forest measurements The Inter-national Archives of the Photogrammetry Remote Sensing and Spatial InformationSciences 36 82ndash89

Hyyppauml J Hyyppauml H Leckie D Gougeon F Yu X amp Maltamo M (2008) Review ofmethods of small-footprint airborne laser scanning for extracting forest inventorydata in boreal forests International Journal of Remote Sensing 29 1339ndash1366

Jaskierniak D Lane P Robinson A amp Lucieer A (2010) Extracting LiDAR indices tocharacterize multi-layered forest structure using mixture distributions functionsRemote Sensing of Environment 115 537ndash585

Kraus K amp Pfeifer N (1998) Determination of terrain models in wooded areas withairborne laser scanner data ISPRS Journal of Photogrammetry and Remote Sensing53 193ndash203

Landsberg J J amp Gower S T (1997) Forest biomes of the world Applications of phys-iological ecology to forest management (pp 19ndash50) San Diego Academic Press

Mallet C amp Bretar F (2009) Full-waveform topographic lidar State-of-the-art ISPRSJournal of Photogrammetry and Remote Sensing 64 1ndash16

223A Ferraz et al Remote Sensing of Environment 121 (2012) 210ndash223

Maltamo M Eerikaumlinen K Pitkaumlnen J Hyyppauml J amp Vehmas M (2004) Estimation oftimber volume and stem density based on scanning laser altimetry and expectedtree size distribution functions Remote Sensing of Environment 90 319ndash330

Maltamo M Packaleacuten P Yu X Eerikainen K Hyyppauml J amp Pitkanen J (2005) Iden-tifying and quantifying structural characteristics of heterogeneous boreal forestusing laser scanner data Forest Ecology and Management 216 41ndash50

Martinuzzi S Vierling L A Gould W A Falkowski M J Evans J S Hudak A T et al(2009) Mapping snags and understory shrubs for a LiDAR-based assessment ofwildlife habitat suitability Remote Sensing of Environment 113 2533ndash2546

Moore P T Van Miegroet H amp Nicholas N S (2007) Relative role of understory andoverstory in carbon and nitrogen cycling in a southern Appalachian spruce-fir for-est Canadian Journal of Forest Research 37 2689ndash2700

Morsdorf F Meier E Koumltz B Itten K I Dobbertin M amp Allgoumlwer B (2004)LIDAR-based geometric reconstruction of boreal type forest stands at single treelevel for forest and wildland fire management Remote Sensing of Environment92 353ndash362

Morsdorf F Maringrell A Koetz B Cassagne N Pimont F Rigolot E et al (2010) Dis-crimination of vegetation strata in a multi-layered Mediterranean forest ecosystemusing height and intensity information derived from airborne laser scanning Re-mote Sensing of Environment 114 1403ndash1415

Mutlu M Popescu S C Stripling C amp Spencer T (2008) Mapping surface fuelmodels using lidar and multispectral data fusion for fire behavior Remote Sensingof Environment 112 274ndash285

Pereira L Gonccedilalves G Soares P Cambra S Carvalho S amp Tomeacute M (2009) Plan-ning and acquisition of control data to validate forest inventory and the estimationof fuel variables derived from LiDAR data and high resolution CIR images Proc 6degCongresso Florestal Nacional Ponta Delgada- Accedilores 6ndash9 Outubro 2009 9 pp

Persson Aring Holmgren J amp Soumlderman U (2002) Detecting and measuring individualtrees using an airborne laser scanner Photogrammetric Engineering and RemoteSensing 68 925ndash932

Persson Aring Holmgren J Soumlderman U amp Olsson H (2004) Tree species classificationof individual trees in Sweden by combining high resolution laser data with highresolution near-infrared digital images International Archives of Photogrammetry36 204ndash207

Peterson B (2005) Canopy fuels inventory and mapping using large-footprint lidar PhDThesis University of Maryland (MD) 218 pp

Popescu S C amp Wynne R H (2004) Seeing the trees in the forest Using LIDAR andmultispectral data fusion with local filtering and variable window size for estimat-ing tree height Photogrammetric Engineering and Remote Sensing 70 589ndash604

Popescu S C amp Zhao K (2008) A voxel-based lidar method for estimating crown baseheight for deciduous and pine trees Remote Sensing of Environment 112 767ndash781

Pyne S J Andrews P L amp Laven R D (1996) Introduction to wildland fire (2ndEdition) New York John Wiley amp Sons 808 pp

Reitberger J Schnoumlrr C Krzystek P amp Stilla U (2009) 3D Segmentation of singletrees exploiting full waveform LiDAR data ISPRS Journal of Photogrammetry and Re-mote Sensing 64 561ndash574

Riantildeo D Meier E Allgoumlwer B Chuvieco E amp Ustin S L (2003) Modeling airbornelaser scanning data for the spatial generation of critical forest parameters in firebehaviour modeling Remote Sensing of Environment 86 177ndash186

Riantildeo D Chuvieco E Condeacutes S Gonzalez-Matesanz J amp Ustin S L (2004) Genera-tion of crown bulk density for Pinus sylvestris L from lidar Remote Sensing of Envi-ronment 92 345ndash352

Riantildeo D Chuvieco E Ustin S L Sala J Rodriguez-Perez J R Ribeiro L M et al(2007) Estimation of shrub height for fuel-type mapping combining airborneLiDAR and simultaneous color infrared ortho imaging International Journal of Wild-land Fire 16 341ndash348

Richardson J J amp Moskal L M (2011) Strengths and limitations of assessing forestdensity and spatial configuration with aerial LiDAR Remote Sensing of Environment115 2640ndash2651

RIEGL (2011) Available online at RiANALYZE httpwwwrieglcomproductssoftware-packagesrianalyze (accessed 21072011)

RIEGL (2011) Available online at RiWORLD httpwwwrieglcomproductssoftware-packagesriworld (accessed 21072011)

Sandberg D V Ottmar R D amp Cushon G H (2001) Characterizing fuels in the 21stcentury International Journal of Wildland Fire 10 381ndash387

Scott J H amp Reinhardt E D (2001) Assessing crown fire potential by linking modelsof surface and crown fire behaviour USDA forest service research paper RMRS-RP-29(pp 9ndash21) Fort Collins CO Rocky mountain research station

Topographic laser ranging and scanning Shan J amp Toth C K (Eds) (2009) Principlesand processing CRC Press 608 pp

Singh M amp Ahuja N (2003) Regression based bandwidth selection for segmentationusing Parzen windows Proc 9th IEEE International Conference on Computer VisionNice (France) 13ndash16 October 2003 (pp 2ndash9)

Soininen A (2010) Available online at TerraScan users guide httpwwwterrasolidfienusers_guideterrascan_users_guide (Accessed 6072011)

Solberg S Naesset E amp Bollandsas O M (2006) Single tree segmentation using air-borne laser scanner data in a structurally heterogeneous spruce forest Photogram-metric Engineering and Remote Sensing 72 1369ndash1378

Stokes B J Ashmore C Rawlins C L amp Sirois D L (1989) Glossary of terms used intimber harvesting and forest engineering General technical report SO-73 USADforest service New Orleans (LA) Southern Forest Experiment Station 33 pp

Wang J Thiesson B Xu Y amp Cohen M (2004) Image and video segmentation by an-isotropic kernel mean shift Proc European Conference on Computer Vision vol 2(pp 238ndash249)

Yi K M Ahn H S amp Choi J Y (2008) Orientation and scale invariant mean shift usingobject mask-based kernel Proc 19th International Conference on Pattern Recogni-tion Tampa (FL) 8ndash11 December 2008 (pp 1ndash4)

Yilmaz A (2007) Object tracking by asymmetric kernel mean shift with automaticscale and orientation selection Proc IEEE Conference on Computer Vision and Pat-tern Recognition Minneapolis (MN) 17ndash22 June 2007 (pp 1ndash6)

Yoon J S Shin J I amp Lee K S (2008) Land cover characteristics of airborne LiDAR in-tensity data A case study IEEE Geoscience and Remote Sensing Letters 9 463ndash466

Zhao K Popescu S amp Nelson R (2009) LiDAR remote sensing of forest biomass Ascale-invariant estimation approach using airborne lasers Remote Sensing of Envi-ronment 113 182ndash196

Zimble D A Evans D L Carlson G C Parker R C Grado S C amp Gerard P D (2003)Characterizing vertical forest structure using small-footprint airborne LiDAR Re-mote Sensing of Environment 87 171ndash182