ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier .com/ locate/ isprs jprs

Fusion of high spatial resolution WorldView-2 imagery and LiDARpseudo-waveform for object-based image analysis

http://dx.doi.org/10.1016/j.isprsjprs.2014.12.0130924-2716/� 2014 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.

⇑ Corresponding author. Tel.: +1 972 883 4134; fax: +1 972 883 6573.E-mail addresses: yxz102020@utdallas.edu (Y. Zhou), ffqiu@utdallas.edu (F. Qiu).

Yuhong Zhou, Fang Qiu ⇑Department of Geospatial Science, University of Texas at Dallas, 800 W Campbell Rd. GR31, Richardson, TX 75080-3021, USA

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

Article history:Received 30 August 2014Received in revised form 6 December 2014Accepted 15 December 2014

Keywords:FusionLiDARImageryLand coverClassificationHigh resolutionMultispectral

High spatial resolution (HSR) imagery and high density LiDAR data provide complementary horizontaland vertical information. Therefore, many studies have focused on fusing the two for mapping geographicfeatures. It has been demonstrated that the synergetic use of LiDAR and HSR imagery greatly improvesclassification accuracy. This is especially true with waveform LiDAR data since they provide more detailedvertical profiles of geographic objects than discrete-return LiDAR data. Fusion of discrete-return LiDARand HSR imagery mostly takes place at the object level due to the superiority of object-based imageanalysis (OBIA) for classifying HSR imagery.

However, the fusion of the waveform LiDAR and HSR imagery at the object level has not been ade-quately studied. To fuse LiDAR waveform and image objects, the waveform for the objects derived fromimage segmentation are needed. However, the footprints of existing waveform are usually of fixed sizeand fixed shape, while those of building are of different size and shape. In order to obtain waveforms withfootprints that match those of image objects, we proposed synthesizing object-based pseudo-waveformsusing discrete-returns LiDAR data by utilizing count or intensity based histogram over the footprints ofthe objects. The pseudo-waveforms were then fused with the object-level spectral histograms from HSRWorldView-2 imagery to classify the image objects using a Kullback–Leibler divergence-based curvematching approach.

The fused dataset achieved an overall classification accuracy of 97.58%, a kappa coefficient of 0.97, andproducer’s accuracies and user’s accuracies all larger than 90%. The use of the fused dataset improved theoverall accuracy by 7.61% over the use of HSR imagery alone, and McNemar’s test indicated that suchimprovement was statistically significant (p < 0.001). This study demonstrates the great potential ofpseudo-waveform in improving object-based image analysis. This is especially true since currently themajority of commercial LiDAR data are of discrete return while waveform data are still not widelyavailable.� 2014 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier

B.V. All rights reserved.

1. Introduction and background

Fine-scale land cover mapping is essential for a variety of appli-cations, especially in urbanized areas. Urban resource manage-ment, maintenance and planning, and pattern analysis all benefitfrom an accurate and detailed land cover classification. To achievethis, two emerging remote sensing techniques, high spatial resolu-tion (HSR) multispectral imagery and high density Light Detectionand Ranging (LiDAR), have been more and more frequently used todevelop fine-scale urban land cover maps (Zhou, 2013).

The recent launch of many commercial HSR sensor systems(such as GeoEye-1, Pléiades-2, and WorldView-3) greatly improvedthe spatial resolution of imagery remotely sensed, with several 1–4 m multispectral bands and a sub-meter panchromatic band. Inconsort with the increasing availability of HSR remote sensors,object-based image analysis (OBIA) techniques have rapidly devel-oped for fine-scale land cover mapping in the last decade(Blaschke, 2010; Berger et al., 2013; Zhou, 2013). OBIA performsimage classification using image objects or segments rather thanpixels as processing units. Image objects are generated throughan image segmentation procedure with each segment composedof spatially adjacent pixels grouped according to some pre-definedhomogeneity criteria (Blaschke, 2010). Many studies have demon-strated that OBIA approaches are superior to the pixel-based image

222 Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232

analysis for HSR imagery (Ke et al., 2010; Arroyo et al., 2010;Sridharan and Qiu, 2013; Zhou, 2013).

Traditionally, OBIA techniques are based on object-level statis-tical summaries, such as the mean and standard deviation of all thepixel values of an image object. However, these object-level statis-tical summaries are representative of the object’s characteristicsonly when the pixel values follow a normal distribution (Pedleyand Curran, 1991; Shackelford and Davis, 2003; Berger et al.,2013). Unfortunately, a non-normal frequency distribution is com-mon for spectral pixels values due to the within-object heteroge-neity of the HSR imagery (Stow et al., 2012; Sridharan and Qiu,2013; Toure et al., 2013). Consequently, these statistical summa-ries may misrepresent the spectral or structural nature of theobjects and mislead the subsequent analysis (Sridharan and Qiu,2013).

To overcome this problem, novel approaches based on object-level spectral frequency distribution (histogram) or cumulative fre-quency distribution have been proposed for object-based imageclassification recently (Stow et al., 2012; Sridharan and Qiu,2013; Toure et al., 2013). By using curve matching approaches theyhave achieved better performances than using the traditional sta-tistical summaries. Compared to statistical summaries, the curveof an object-level frequency distribution provides a more compre-hensive description of the spectral components of an image object,which is sufficient to separate geographic features with distinctspectral characteristics. However, the utilization of this spectralinformation alone in differentiating between spectrally similarbut structurally different features, such as buildings and roads ortrees and grasses, is still challenging due to the limited spectralresolution of most HSR multispectral sensors.

Unlike HSR imagery that provides two-dimensional (2D) hori-zontal spectral information only, the increasingly available LiDARdata offers the third-dimensional (3D) elevation information forgeographic features (Koetz et al., 2007). There are two types ofLiDAR data based on how the signal is recorded: discrete-returnLiDAR and full-waveform LiDAR (Ussyshkin and Theriault, 2011).Discrete-return LiDAR typically records 1 to 6 returns for eachtransmitted laser pulse. The measurements provided by discrete-return LiDAR, such as elevation, intensity, and the elevationderived digital terrain model, have been intensively used for treespecies classification (Zhang and Qiu, 2012), land cover classifica-tion (Sasaki et al., 2012), and 3D building roof construction (Kimand Shan, 2011).

Full waveform, a relatively new product of LiDAR, becomespopular in the last decade. It records the quasi-continuous time-varying strength of the return signal from the illuminated area(i.e. waveform footprint) using small time intervals (e.g., 1nanosec-ond), consequently resulting in thousands of measurements foreach transmitted laser pulse (Alexander et al., 2010; Ussyshkinand Theriault, 2011). Due to this finer vertical resolution, the wave-form offers an enhanced capability to reflect the vertical structuresof geographical objects compared with the traditional discrete-return LiDAR (Zhang et al., 2011; Farid et al., 2008).

To take advantage of waveform LiDAR for land cover classifica-tion, some studies have attempted to extract more returns fromfull waveforms through discretization. The number of resultantreturns can be set to be much more than that of the traditional dis-crete-return LiDAR, and therefore to better represent the verticalstructure of objects (Wagner et al., 2006; Reitberger et al., 2009;Mallet and Bretar, 2009; Yao et al., 2012). The discretized returnsthen can be analyzed by the existing algorithms designed for tradi-tional discrete-return LiDAR. Other studies were based on analyz-ing the waveform-derived metrics, such as the number of echoes,peak amplitude, echo width, and ground peak location, which areextracted to represent the important characteristics of the wave-form shapes (Zaletnyik et al., 2010; Mallet et al., 2011; Wang

et al., 2012; Zhuang and Mountrakis, 2014; Guo et al., 2011). Theresults of these researches demonstrated positive correlationsbetween those waveform-derived metrics and the correspondingbackscattered surface material within the footprints. Nevertheless,waveform LiDAR alone may not sufficiently separate structurallysimilar but spectrally different target features (e.g., road and grass)(Geerling et al., 2007), although it provides much more verticalstructural information than discrete-return LiDAR.

Overall, HSR multispectral imagery and waveform LiDAR datahave their distinct advantages and disadvantages for fine-scaleland cover mapping. HSR multispectral imagery provides accuratespatial information and moderate spectral information but lacksvertical structural information. On the other hand, waveformLiDAR data provide accurate vertical structural information butlimited horizontal spectral information. Fusion of these two datasources is an obvious approach in order to capitalize on theirrespective advantages and compensate for their respective short-comings for fine-scale land cover classification (Lee et al., 2008;Anderson et al., 2008).

According to Zaletnyik et al. (2010), waveform LiDAR data canalso be conceptually considered as a time-varying frequency distri-bution of returning impulses. This leads to the idea that the 3Dwaveform LiDAR data, like the 2D spectral information from theHSR imagery, can also be analyzed as a frequency distribution.The curve of the vertical distribution of waveform contains sub-stantially more information than waveform-derived metrics andthe discretized returns. Therefore, the synergetic use of frequencydistributions for both the object-level spectral data and for thewaveform LiDAR data through data fusion is therefore expectedto deliver better classification results by taking full advantage ofboth the horizontal spectral and vertical structure information.This idea, to the best of our knowledge, has not yet been investi-gated in the literature.

To integrate the waveform LiDAR data into an object-basedimage analysis, a major challenge needs to be overcome beforesubsequent data fusion. Currently the footprints of all full-wave-form LiDAR are of a fixed size and shape, be it large, medium, orsmall, while the footprints of geographic objects vary dramaticallyin size and shape and seldom match those of the waveforms. Thisobstacle makes it difficult to fuse waveform and spectral frequencydistributions of an image object to perform an object based imageanalysis. The solution we provided in this study is to develop anoriginal method to synthesize object-level pseudo-waveforms withvaried footprint sizes and shapes corresponding to that of differentobjects using discrete-return LiDAR data. As a result, we can easilyfuse the synthesized pseudo-waveform curve and spectralhistogram curves at the object level for fine-scale land covermapping.

To assess the object-to-object similarity based on the fused fre-quency distributions, a discrete Kullback–Leibler (KL) divergencebased classifier was proposed. As a non-parametric approach, KLdivergence does not require normality of the probability distribu-tion and has been widely used in speech and image pattern recog-nition (Olszewski, 2012) and in hyperspectral image classification(Ghiyamat et al., 2013). Given that both pseudo-waveform andspectral histogram can be considered as a discrete probability dis-tribution function, the KL divergence based classifier may be usefulfor classification of the fused frequency distributions.

2. Study area and data

2.1. Study area

The study area is situated in the Turtle Creek Corridor in DallasCounty, Texas (Fig. 1), bounded approximately by 96.811� to96.798�W in longitude and 32.8� to 32.82�N in latitude. It covers

Fig. 1. Study area: Turtle Creek area with false color composite (band 7, 5, 3) image of WorldView-2.

Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232 223

2.8 km2, with a mean elevation of 144 m. Gentle slopes character-ize its topography. The site is a typical urban area with a variety ofnatural features, such as individual trees, groups of trees with con-nected foliage, grass and rivers, and man-made features made ofdifferent materials, such as buildings (tile, metal, or concrete)and roads (asphalt, concrete, or composite). This complex mixtureof geographic features poses a challenge to fine-scale land coverclassification using existing remote sensing technologies(Sridharan and Qiu, 2013).

To assess the potential provided by the fusion of LiDAR data andHSR multispectral imagery to improve fine-scale land cover classi-fication, five land cover classes were identified in this study, specif-ically buildings, pavements (such as roads, parking lots, andsidewalks), trees, grass, and water. These classes cover both spec-trally similar but structurally different features (e.g., buildingsand pavements, trees and grasses) and structurally similar butspectrally different features (e.g., pavements and grass).

2.2. HSR multispectral imagery

HSR multispectral WorldView-2 (WV-2) imagery was acquiredfor the study area on April 29th, 2010 (Fig. 1). WV-2 is the firstcommercial HSR satellite that provides eight multispectral bandsalong with a panchromatic band (DigitalGlobe, Inc., 2010a). The

panchromatic band has a spatial resolution of 0.46 m. The multi-spectral bands, each with a spatial resolution of 1.84 m, are coastalblue (400–450 nm), blue (450–510 nm), green (510–581 nm), yel-low (585–625 nm), red (630–690 nm), red edge (705–745 nm),NIR1 (770–895 nm), and NIR2 (860–1040 nm). While the spectralresolution of WV-2 imagery is improved over its predecessors withthe inclusion of four new bands (coastal blue, yellow, red edge, andNIR2), for the fine-scale land cover classification the power of WV-2 imagery lies in its very high spatial resolution. This providesdetailed information regarding the spatial distribution of geo-graphic features, allowing the adoption of object-based imageanalysis. This also permitted us to collect the ground reference datathrough visual interpretation for accuracy assessment (Sridharanand Qiu, 2013).

2.3. Discrete-return LiDAR data

Discrete-return LiDAR data were acquired on September 23rd,2008 using a Terrain Scanning Laser System (Terra Remote Sensing,Inc.) onboard a Piper Navajo fixed-wing aircraft flying at approxi-mately 960 m above ground at a speed of 235 km/hr. The scannerwas operated with a scanning rate of 34 Hz and a laser pulse rate of60 kHz. The wavelength of the laser pulse was 1064 nm. An off-nadir scan angle of 26�, a beam divergence of 0.45 mrad, and an

224 Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232

80% sidelap of the flight lines produced a LiDAR point density ofapproximately 3.5 points per m2 (pts/m2), with an average pointspacing of 0.67 m. A total of 6,281,769 points were collected forthe study area. For each laser pulse, the 3D coordinate (x, y, andelevation) and the associated intensity value for the first and lastreturns was recorded. The US Forest Service suggests that a pointdensity of Ph pts/m2 and the number of returns per laser pulseof >4 are needed for LiDAR applications in forest and structuralmodeling (Laes et al., 2008). The 3.5 pts/m2 density of LiDAR datain this study, along with two returns recorded per laser pulse,may not be ideal for modeling if only the LiDAR data is used. How-ever, when used in combination with HSR imagery, it is sufficient,especially considering it will be used for synthesizing object-levelpseudo-waveforms to approximate the vertical structures of theindividual objects segmented from the HSR imagery.

3. Methodology

The steps for fusing HSR spectral histograms and LiDAR pseudo-waveform at the object level for classification purpose are summa-rized in Fig. 2 and described in detail in the sub-sections which

Fig. 2. General flowcha

follow. They include data preprocessing (e.g., image segmentation),spectral histogram generation, and pseudo-waveform simulationusing discrete-return LiDAR data.

3.1. Preprocessing

The WV-2 data and LiDAR data require several preprocessingsteps before they can be fused for OBIA. For the WV-2 data, someinitial processing was accomplished by the vendor, including cor-rections for internal sensor geometry, optical distortions, scan dis-tortions, line rate-variations, band registration, and radiometricregistration (Sridharan and Qiu, 2013). A transformation from radi-ance to at-surface reflectance was carried out using an algorithmprovided by DigitalGlobe, Inc. (2010b). The corrected WV-2 multi-spectral bands were further pan-sharpened to the resolution of thepanchromatic band using the Gram-Schmdit algorithm (Padwicket al., 2010).

After pan-sharpening, the resulting multispectral bands retainthe original spectral resolution but have a spatial resolution equiv-alent to that of the panchromatic band (0.46 m). The eight WV-2pan-sharpened multispectral bands were then co-registered to

rt of methodology.

Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232 225

the LiDAR data to remove a noticeable geometric discrepancybetween the two which would likely cause a problem for their sub-sequent fusion using spline transformation. To reduce the borderdistortion introduced by the spline transformation, both the refer-ence image derived from LiDAR and the source image covered anarea larger than our study area. A total of 41 manually selectedground control points, evenly distributed across the larger area,were used for this transformation using ArcGIS 10.2. After co-reg-istration, the total Root Mean Square Error (RMSE) was approxi-mately 0.046 m. The WV-2 image for the study area was thenclipped from the larger one and was used for the subsequent anal-ysis. Finally, ENVI Zoom software was used for image segmentationof the preprocessed WV-2 imagery to generate individual homog-enous segments or image objects. Detailed information aboutimage segmentation can be found in Sridharan and Qiu (2013).

After segmentation, image objects belonging to the shadowclass were first separated out using a fuzzy Kolmogorov–Smirnovclassifier (Sridharan and Qiu, 2013). They were excluded fromthe subsequent analysis for the following reason. Shadow is a darkarea where direct sunlight cannot reach due to obstruction by anadjacent higher object(s). Shadows are therefore not true geo-graphic features. They can be over an object of any land cover classor multiple objects of different classes. Since their correspondingpseudo-waveforms are primarily determined by the vertical struc-tures of the objects under the shadow, pseudo-waveforms for sha-dow can vary greatly from each other and can be similar to that ofany land cover class. If included in the subsequent analysis, thepseudo-waveforms of the shadows may cause great confusionsbetween shadow and other land cover classes.

3.2. Object-level spectral histograms

As already noted, object-level spectral histograms have beensuccessfully employed previously to improve object-based image

Fig. 3. Object-level spectral histograms for water, buildin

classification (Stow et al., 2012; Sridharan and Qiu, 2013; Toureet al., 2013). These histograms, by tabulating the frequency ofoccurrence of each brightness value within the object, graphicallyrepresent the within-object spectral frequency distributions, andtheir peaks are usually determined by the dominant materialswithin the object (Jensen, 2004). The spectral histograms for eachsegmented image object for each individual band were generatedin this study with the range of pixel values on the x axis and thefrequency of occurrence of each of these values on the y axis. Thesehistograms were then normalized by the total number of pixelswithin the object to eliminate the influence of the object size, aswell as to obtain the probability distributions to be fused withthe subsequently synthesized pseudo-waveforms.

To examine the object-level spectral histograms, typical objectsof 5 different classes were selected and their histograms weregenerated (Fig. 3). Each object demonstrates distinguishableobject-level spectral frequency distributions. The water objecthas near-normal frequency distributions with a narrow widthand relatively low brightness values in all bands. Its histogramsare very different from those of other four objects in all eightbands, making it easily separable from other objects. Comparedwith the two man-made objects (pavement and building), thebrightness values of the two vegetated objects (tree and grass)are substantially lower for bands 1 (coastal blue), 2 (blue), 3(green), 4 (yellow) and 5 (red), but higher for bands 7 (NIR1) and8 (NIR2). This is primarily because vegetation has a high absorptionfrom chlorophyll in the visible region of the spectrum and highscattering from spongy mesophyll in the NIR region (Jensen,2004) during the leaf-on season of April when the imagery was col-lected. This makes it possible for the two vegetated objects to beeasily separated from the man-made objects, although the differ-ence between them in band 6 is small. The frequency distributionsof the building and the pavement objects only have some minoroverlay in the first six bands and are well separable from each

g, pavement, tree, and grass at all 8 bands of WV-2.

226 Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232

other in bands 7 and 8. This is expected because the two man-made objects selected are made of different materials. However,the spectral frequency distributions of the building and pavementobjects can be very similar to each other in all eight bands if theyare made of the same material, which may make it impossible todistinguish between the two using spectral information alone.For both vegetated objects, their frequency distributions peakedat very low brightness values in bands 1, 2, and 5, again due tothe strong vegetation absorption of blue and red energy by chloro-phyll. Considerable differences in frequency distributions betweenthe tree and the grass can be observed in bands 3, 4, and 6, wherethe tree object tends to have a wide spread in its histograms com-pared to those of the grass object. This is because the grass has amuch more homogeneous surface while the tree object has a roughtexture due to the presence of branches and background soil. Themaximum distinction in frequency distributions between the twovegetated objects occur for bands 7 and band 8, where both exhibithigher peak pixel values, but those of the grass object are higherthan the tree object. This is again because many pixels in the treeobjects may be of branches and background soil, while almost all ofthe pixels in the grass object are of vegetation. It is also noticeablein Fig. 3 that not all of the frequency distributions for the selectedobject are normal; once again justify the using of spectral histo-grams over the summary statistics for object-based imageclassification.

3.3. Object-level pseudo-waveform

Hitherto, the fixed size and shape of full-waveform LiDAR foot-prints has impeded the fusion of full-waveform LiDAR with HSRimagery at the object level because of the very varied size and shapeof the image object footprints. Our solution to overcome this prob-lem is to use discrete-return LiDAR point data to synthesize pseudo-waveforms with a footprint the same as that of an image object.

The idea of synthesizing pseudo-waveforms was introduced byBlair and Hofton (1999), who simulated pseudo-waveforms inorder to mimic the Laser Vegetation Imaging Sensor (LVIS) wave-forms using high resolution elevation data. Popescu et al. (2011)synthesized pseudo-waveforms to simulate the Geoscience LaserAltimeter System (GLAS) waveforms using discrete-return LiDARdata. Both results demonstrated a very high similarity betweenthe recorded waveforms and their corresponding synthesizedpseudo-waveforms. In addition to these two studies that specifi-cally aimed to mimic existing waveforms of LVIS or GLAS, severalothers have also synthesized large-footprint pseudo-waveformsby aggregating small-footprint LiDAR data to model forest verticalstructures (Lovell et al., 2003; Farid et al., 2008; Popescu and Zhao,2008; Muss et al., 2011; Pirotti et al., 2014). In these studies, thepseudo-waveform is synthesized as the vertical frequency distri-bution of either the total count, or the sum of the intensity, ofthe laser returns as a function of height bins within the simulated

Fig. 4. (a) A building (left) in Google Earth, (b) its corresponding 3

footprint. The results demonstrated that the count-based or inten-sity-based pseudo-waveform was able to approximate canopy pro-files based on the metrics derived from pseudo-waveform.

Inspired by the above studies that synthesized pseudo-wave-forms with footprints of fixed size and shape, we also devised anapproach to synthesize object-level pseudo-waveforms, but withfootprints of varied sizes and shapes corresponding to those ofreal-world objects using the discrete-return LiDAR points. The pro-cedure to synthesize such object-level pseudo-waveforms is asfollows.

Step 1: Find the maximum Hmax and minimum Hmin elevation ofall LiDAR points in the study area, which determines the eleva-tion range. Slice this into height bins with a vertical interval of15 cm. The total number of bins (N) for each pseudo-waveformis

N ¼ ðHmax � HminÞ=0:15 ð1Þ

D LiD

Step 2: For each image object in the study area, extract theLiDAR points within the footprint of the object, sort by theirelevation values, and assign to the corresponding bins.Step 3: For each bin, the total count of the laser points producesa count-based frequency distribution and the sum of the inten-sity of the points produces an intensity-based frequency distri-bution. The results are object-level pseudo-waveforms eachsynthesized to the footprint of their corresponding object.Step 4: For data fusion purpose, the object-level pseudo-wave-forms are then normalized using the total number of returnsor the total intensity within the objects to obtain their probabil-ity distributions, so that they can be subsequently fused withthe normalized spectral histograms derived from the HSRimagery.

As an example, Fig. 4a presents an image of a building complexextracted from Google Earth, Fig. 4b shows all the discrete-returnLiDAR points within its footprint, and Fig. 4c illustrates the synthe-sized pseudo-waveform of the building complex. This buildingcomplex is composed of three major parts: the attic, the mainhigh-rise building, and the attached low-rise building. The synthe-sized pseudo-waveform presents three major echoes, with the toppeak corresponding to the flat roof of the attic, the second one cor-responding to that of the high-rise building, and the bottom onecorresponding to that of the low-rise building.

The fusion of the eight normalized spectral histograms and thenormalized pseudo-waveform was implemented by viewing thenormalized pseudo-waveform as a ninth histogram.

3.4. Kullback–Leibler divergence based classification method

KL divergence was originally introduced by Solomon Kullbackand Richard Leibler in 1951 as the dissimilarity measurement

AR points (middle) and (c) the pseudo-waveform (right).

Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232 227

between two probability distributions. For two discrete probabilitydistributions, P and Q, the KL divergence from Q to P is definedusing Eq. (2), with the requirement that both P and Q sum to1(Olszewski, 2012; Mittal et al., 2013). In this study, P refers tothe probability distribution of a spectral histogram or thepseudo-waveform of a reference object, while Q refers to that ofan unknown object, and i refers to the ith brightness value in aspectral histogram or the ith height bin in a pseudo-waveform.

DðPjjQÞ ¼Xn

i

logðPðiÞ=QðiÞÞ � PðiÞ ð2Þ

Since the KL divergence is not symmetric, i.e. divergence from Qto P is not equal to that from P to Q, that is, D(P||Q) – D(Q||P), themutual KL divergence bD between P and Q is calculated as the meanof D(P||Q) and D(Q||P).

bDðPjjQÞ ¼ ðDðPjjQÞ þ DðQ jjPÞÞ=2 ð3ÞbD is the dissimilarity measurement of two objects for a singleband. The dissimilarity between an unknown object and a refer-ence object for all the bands is obtained by computing the arithme-tic mean of the dissimilarities of all bands,

Di ¼XN

j¼1

bDij

!,N ð4Þ

where Di is the mean dissimilarity between an unknown object anda reference object i, N is the number of bands, and bDij is the curvedissimilarity between the unknown object and a reference object iat band j. After the Di between the unknown object and all referenceobjects has been computed, the object is assigned to the class of thereference object that has the smallest dissimilarity (that is, thehighest similarity) with the unknown object.

The KL divergence based classifier was compared with otherobject-level curve matching based classifiers in the literature,including Histogram Matching Root Sum Squared Differential Area(HMRSSDA) (Stow et al., 2012), Fuzzy Kolmogorov–Smirnov basedclassifier (FKS) (Sridharan and Qiu, 2013), and Histogram AngleMapper (HAM) (Toure et al., 2013), in order to evaluate its perfor-mance as a new object-based image classifier.

3.5. Accuracy assessment

For each class, the same representative reference objectsemployed by Sridharan and Qiu (2013) were used for consistencywith previous study, including 15 for buildings, 11 for pavements,6 for water, 13 for trees, and 13 for grass. For each unknown object,the resultant nine histograms were compared with those of a ref-erence object using a KL divergence based curve matchingapproach to assess its closeness to the reference objects.

For accuracy assessment, a set of testing objects were randomlyselected: 199 for buildings, 100 for pavements, 30 for water, 199for trees, and 50 for grass. The accuracy of the classification resultwas assessed using standard metrics derived from confusionmatrices, including the overall accuracy (OA), producer’s accuracy(PA), user’s accuracy (UA), and the kappa coefficient (Congalton,1991). The significance of the differences between two classifica-tions was tested using McNemar’s test (Bradley, 1968; Agresti,1996). This non-parametric test is based on the error metrics ofthe two classifications and is computed as:

v2 ¼ ðf 12 � f 21Þ2=ðf 12 þ f 21Þ ð5Þ

where f12 denotes the number of samples that are correctly classi-fied by classifier one but wrongly classified by classifier two, andf21 denotes that the number of samples that are correctly classifiedby classifier two but misclassified by classifier one. The statistical

significance of the differences between two classifications is deter-mined from the resulting v2 value which has a standard chi-squaredistribution.

4. Results and discussions

This section presents the results from the pseudo-waveformsimulation and the KL based curve matching classification usingthe WV-2 multispectral imagery alone and fused with the LiDARpseudo-waveform. It also compares the KL based curve matchingapproach with three other curve matching approaches, HMRSSDA,HAM, and FKS.

4.1. Pseudo-waveform simulation

To verify the validity of the synthesized pseudo-waveforms,data from the GLAS onboard the Ice, Cloud, and Land ElevationSatellite (ICESat) were used. Six locations where the GLAS wave-form footprints overlapped the study area were identified. GLASuses the near-infrared channel to acquire vertical profiles ofreturned energy to form a waveform. The footprints of GLASwaveforms are elliptical in shape, but their sizes varied substan-tially during the satellites 7-year mission (2003–2009), with amean major axis ranging from about 50 m to 150 m (Attributesfor ICESat Laser Operations Periods, 2012). The areas of the foot-prints for the six selected waveforms were derived from the GLASproduct using the length of their short and long axes and the ori-entation of the axes (Gong et al., 2011). Pseudo-waveforms withfootprints corresponding to those of GLAS were synthesized usingboth the count-based and intensity-based frequency distributionsand then compared with the recorded GLAS waveforms. Fig. 5shows the six GLAS waveforms and their corresponding count-based and intensity-based pseudo-waveforms. A close look atFig. 5 reveals that the synthesized pseudo-waveforms were ableto approximate the general trend of the corresponding recordedwaveforms although they are different in some of the details, pri-marily because the number of LiDAR points used to synthesizethe pseudo-waveforms was not the same as the number ofimpulses used to form the GLAS waveforms. The count-basedpseudo-waveforms and intensity-based pseudo-waveforms werenearly identical to each other. Therefore, only count-basedpseudo-waveforms were used in the subsequent analysis.

Fig. 6 presents individual pseudo-waveforms over a typicalpavement, a flat-roof building, a slant-roof building, a tree, a grass,and a water object. Pseudo-waveforms over the pavement, the flat-roof building, the grass, and the water objects are very similar toeach other, all with a single narrow echo, because a high concen-tration of energy is reflected back at nearly the same time whena laser pulse encounters a flat surface. However, the peak locationof the echo for the flat-roof building differs greatly from that ofother three, due to the height differences between the buildingroof and the other three object surfaces. This makes it easy to sep-arate flat-roof buildings from pavements, grass, and water usingpseudo-waveforms. However, the utilization of pseudo-waveformsalone cannot distinguish between pavements, grass, and waterbecause their pseudo-waveforms are nearly identical. For theslant-roof building, its pseudo-waveform has one wider echo withlocal ups and downs. The wider spread of the echo corresponds tothe slope and the height of the slant roof. The vertical structure ofthe slant-roof building allows its pseudo-waveform to be distinc-tive from those of objects with a flat surface. The pseudo-waveformover a tree has multiple echoes due to multiple returns at the var-ious heights of its crown, making it completely different from thatof all other objects. Therefore, pseudo-waveforms of trees and

GLAS 385035941_9

Height (m)

Ampl

itude

pseudo−waveform

Height (m)

Ampl

itude

GLAS 935105750_23

Height (m)

Ampl

itude

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

pseudo−waveform

Height (m)

Ampl

itude

25 15 5 0 25 15 5 0 15 10 5 0 15 10 5 0

20 15 10 5 0 20 15 10 5 0 14 10 6 2 14 10 6 2

4 3 2 1 0 4 3 2 1 0 15 10 5 0 15 10 5 0

GLAS 935105750_25

Height (m)

Ampl

itude

pseudo−waveform

Height (m)

Ampl

itude

GLAS 935105750_37

Height (m)

Ampl

itude

pseudo−waveform

Height (m)

Ampl

itude

GLAS 984719801_24

Height (m)

Ampl

itude

pseudo−waveform

Height (m)

Ampl

itude

GLAS 984719801_32

Height (m)

Ampl

itude

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

0.0

0.4

0.8

pseudo−waveform

Height (m)

Ampl

itude

Fig. 5. Pairwise GLAS waveform (left) and the corresponding pseudo-waveforms (right): count-based pseudo-waveform in blue and intensity-based in black. (Forinterpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 6. Count-based pseudo-waveforms for a road, a building with flat roof, a building with slant roof, a tree, and grass, respectively.

228 Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232

grass were clearly distinguishable even though they were spec-trally very similar.

4.2. Classification results for WV-2 imagery alone

The KL based classification using the WV-2 data alone providedacceptable results with an overall accuracy of 89.97% and a kappacoefficient of 0.87 (Table 1). For visualization purposes, the classi-fication map is shown in Fig. 7a. Inspection of the classificationmap reveals that the object-based classification using the spectral

imagery alone has considerable confusion between pavements andbuildings, and trees and grasses, due to their spectral similarity.

Class-specific analysis revealed that moderate confusionoccurred between pavements and buildings, with 29 out of 199buildings misclassified as pavements and 15 out 100 pavementsmisclassified as buildings. As shown in Table 1, the PA for thebuilding class is 85.43% and the PA for the pavement class is 85%.All misclassified buildings were incorrectly classified as pave-ments. Similarly, all the misclassified pavements were incorrectlyclassified as buildings. This is a common experience (Lee and

Table 1Confusion matrix of KL-based land cover classification using WV-2 data only.

Building Pavement Water Trees Grass PA (%) UA (%)

Building 170 15 0 0 1 85.43 91.4Pavement 29 85 0 0 0 85 74.56Water 0 0 30 0 0 100 100Trees 0 0 0 193 7 96.98 96.5Grass 0 0 0 6 42 84 87.5OA 89.97%Kappa 0.87

Table 2Confusion matrix of KL-based land cover classification using the fusion of WV-2 dataand pseudo-waveform data.

Building Pavement Water Trees Grass PA (%) UA (%)

Building 195 6 0 0 0 97.99 97.01Pavement 3 92 0 0 1 92 95.83Water 0 0 30 0 0 100 100Trees 1 0 0 199 1 100 99Grass 0 2 0 0 48 96 96OA 97.58%Kappa 0.97

Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232 229

Shan, 2003), since building roofs and pavements are very likely tobe made of the same materials, such as concrete, resulting in sim-ilar spectral characteristics between these two classes. As a conse-quence, buildings and pavements were not well separated fromeach other by using the multispectral imagery alone. The waterclass was well distinguished from the other four classes with aPA of 100% and a UA of 100%. This is ascribed to the distinctivespectral characteristics of water compared with those of the otherclasses. The PA of the tree class is 96.98% and the PA of the grassclass is 84%. Similarly, confusion mainly occurred between grassand trees, with 6 out of 199 tree objects misclassified as grassesand 7 out of 50 misclassified as trees.

4.3. Classification results for fusion of LiDAR pseudo-waveform andWV-2 imagery

The fusion of pseudo-waveform and WV-2 imagery producedconsiderably better results than using WV-2 imagery alone withan overall accuracy of 97.58% compared with 89.97%, and a kappacoefficient of 0.97 compared with 0.87 (Table 2). The PA and the UAfor all classes were over 90%, suggesting that the fusion of LiDAR

Fig. 7. KL curve matching based mapping (a) using WV-2 alone (left)

pseudo-waveform and WV-2 imagery can achieve very high classi-fication accuracy. The classification map based on the fused datasetis shown in Fig. 7b. The most obvious improvement is that somepavements misclassified as buildings using WV-2 imagery alone(Fig. 7a) were now correctly classified using the fused dataset(Fig. 7b).

A detailed comparison with the classification results solelyusing WV-2 multispectral data showed that the omission and com-mission errors for all five classes were reduced when the LiDARpseudo-waveform was incorporated as the ninth histogram. Spe-cifically, the fused dataset improved the overall accuracy by7.61% and the kappa coefficient by 0.11. With respect to the indi-vidual land cover classes, PA improved by 12.56% for buildings,7% for pavement, 3.02% for trees, and 12% for grass class, and theUA improved by 5.6% for buildings, 21.2% for pavement, 2.5% fortrees, and 8.5% for grass.

Most of the improvements in accuracy can be explained by thereduction of the confusions between spectrally similar but struc-turally different objects. Only 1 out of 199 trees were misclassifiedas grass and 2 out of 50 grasses were misclassified as trees. Sincethe pseudo-waveforms of trees differ greatly from grass, combing

and (b) using the fusion of pseudo-waveform and WV-2 (right).

Table 3Percentage point differences between KL and other three curve matching approaches for WV-2 only and fusion.

WV-2 alone Fusion

HMRSSDA HAM KS HMRSSDA HAM KS

PA (%) UA (%) PA (%) UA (%) PA (%) UA (%) PA (%) UA (%) PA (%) UA (%) PA (%) UA (%)

Building 2.5 4.1 6 2.6 �2 4 3.5 6.2 15.6 3.8 8.5 3.8Pavement 7 4.3 5 4.4 9 �3.8 10 8.6 8 22.1 8 14.2Water 3.3 0 0 0 0 0 3.3 0 0 0 0 0Tree �2 4.9 �0.5 1.4 1 1 0 5.1 1.5 2 2.5 2Grass 22 �3.7 2 8.6 2 8.6 24 �4 6 15.6 8 14.5OA (%) 3.46 2.95 1.39 5.19 7.79 5.88Kappa 0.05 0.04 0.02 0.07 0.1 0.08

Table 4p-Values for McNemar’s test for significant differences between KL and other curve matching approaches.

WV-2 alone Fusion

Classifier HMRSSDA HAM KS HMRSSDA HAM KSMcNemar’s test 0.00865 0.010382 0.268382 1.34E�06 3.26E�10 1.81E�07Significance ⁄⁄ ⁄ ⁄⁄⁄ ⁄⁄⁄ ⁄⁄⁄

Note: ⁄ At a significant level of 0.05; ⁄⁄ at a significant level of 0.01; ⁄⁄⁄ at a significant level of 0.001.

Table 5p-Values for McNemar’s test for significant differences before and after fusion for eachcurve matching approach.

Classifier HMRSSDA HAM KS KL

McNemar’s test 1.81E�07 0.00018 6.15E�05 5.42E�10Significance ⁄⁄⁄ ⁄⁄⁄ ⁄⁄⁄ ⁄⁄⁄

230 Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232

LiDAR pseudo-waveform with WV-2 imagery helped to correct themisclassification of trees or grass using WV-2 alone. Similarly, theconfusion between pavements and buildings was also greatlyreduced. Compared with the classification using WV-2 imageryalone, out of 199 buildings, only 9 rather than 29 were misclassi-fied as pavements; out of 100 pavements, only 9 rather than 15were misclassified as buildings. A detailed investigation of theremaining objects with confusion between buildings and pave-ments revealed the following. (1) Some pavements misclassifiedas buildings were elevated roads or bridges, which have similarpseudo-waveform shapes to buildings. (2) Other pavements weremisclassified as buildings due to the gap in acquisition timebetween LiDAR data and WV-2 data. The airborne LiDAR data wereacquired two years earlier than the WV-2 imagery. New buildingshad been constructed by the time the WV-2 imagery was acquired.(3) The rest of the buildings misclassified as pavements or pave-ments misclassified as buildings were those with a very sparseLiDAR point density, causing either incomplete or completely miss-ing pseudo-waveforms, thus the fusion provided no additionalinformation to correct misclassifications.

Comparing the results before and after data fusion, we can seethat the WV-2 multispectral imagery was able to identify mostobjects based on spectral histograms alone. The incorporation ofLiDAR pseudo-waveform as the ninth band histogram was effectiveto further discriminate between the spectrally similar but structur-ally different objects that cannot be correctly identified by usingWV-2 alone. Clearly, the fusion of pseudo-waveform and WV-2data resulted in a significant improvement for object-based imageclassification.

4.4. Comparison of KL with other curve matching approaches

The results discussed so far used a KL based curve matchingapproach. Other possible curve matching methodologies includeHMRSSDA, HAM, and KSD. Table 3 provides percentage point dif-ferences between results from using KL and results from otherthree approaches for both the WV-2 imagery alone and the fuseddata.

When WV-2 imagery was used alone, KL outperformed thethree existing curve matching based classifiers in terms of boththe overall accuracy and the kappa coefficient. McNemar’s test(Table 4) indicated that the performance of KL was statistically

significantly better at the 0.01 level than HMRSSDA and better atthe 0.05 level than HAM, although it was not significantly betterFKS. When each of the classes is examined individually, of the 15comparisons (5 classes by other three curve matching methods),KL achieved a better or an equal PA for 12 of 15 and UA for 13 of 15.

Similarly, KL achieved the highest overall accuracy and thekappa coefficient among the four curve matching based classifierswhen the fused dataset was used. McNemar’s test (Table 4) hasmarkedly smaller p values between KL and other three classifiers,all being statistically significant at better than the 0.001 level, indi-cating a far higher degree of confidence in the ability of KL toachieve superior results for the fused dataset. In other words, whenmore information becomes available through the fusion of the twodatasets, the superiority of the KL approach is more clearly appar-ent. When individual classes are examined, KL achieved a better oran equal PA for all the 15 comparisons and a better or an equal UAfor 14 out of 15.

It can also be observed from Table 3 that HMRSSDA, HAM, andKSD all achieved improved performances (i.e., a higher PA and UA)for all classes when the fused dataset was used instead of the WV-2imagery alone, and McNemar’s test (Table 5) indicated that suchimprovements were statistically significant for all the classifiersat p < 0.001.

5. Conclusions

Waveform LiDAR data provide a better 3D profile of an objectthan widely available discrete-return LiDAR data. To take betteradvantage of discrete-return LiDAR data, in this study we synthe-sized object-level pseudo-waveforms with footprints of variedsizes and shapes from discrete-return LiDAR data. These pseudo-waveforms have addressed the issue of recorded waveforms thathave footprints of a fixed size and shape, which now makes it

Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232 231

possible to use waveform data for object based image classifica-tion. For object-based image classification, previous studies havedemonstrated that object-level spectral histograms provide morecomprehensive information than the traditional object-level statis-tical summaries (e.g., mean and standard deviation) and produceimproved results. Since pseudo-waveforms can also be consideredas histograms, we fused spectral histograms from WV-2 imageryand pseudo-waveform synthesized from discrete-return LiDAR atthe object level in order to make the most of both horizontal spec-tral information and vertical structure information of the objects.The results indicated that the fused data did significantly enhancethe performance of object-based image classification.

To quantitatively measure the curve similarity between anunknown object and a reference object, the discrete KL-basedcurve matching approach was proposed in this study. The KL-basedclassifier was demonstrated to be comparable and in most casessignificantly better than the three existing object-level curvematching based classifiers in the literature (i.e., HMRSSDA, HAM,and FKS). This is especially true when the fused data was used.

There are some issues that need to be addressed in the future inorder to achieve better performances. First, fusing LiDAR pseudo-waveform with HSR multispectral imagery cannot be performedwhere shadows cover an object. Future studies could address thisissue by performing segmentation based on LiDAR data rather thanimage data. As an active remote sensing technique, LiDAR is notaffected by solar illumination and therefore the impact of shadowis greatly reduced (Sugumaran and Voss, 2007). Second, forobjects/segments with a small number of LiDAR points, the derivedpseudo-waveforms were either incomplete or completely missing.Insufficient LiDAR point density limits the ability of pseudo-wave-forms approaches to correctly classify some of the spectrally simi-lar but structurally different objects. Higher-density discrete-return LiDAR points are becoming more available and offer thepotential to improve classification accuracy. Third, waveforms orpseudo-waveforms are sensitive to surface terrain effects (Hilbertand Schmullius, 2012). In our study area, with a level terrain, thiswas not an issue. However, the pseudo-waveform can be morecomplex over hilly areas. Future studies may focus on removingthe terrain influence before synthesizing the object-basedpseudo-waveforms. Fourth, in this study the pseudo-waveformand the eight bands of WV-2 were treated equally during fusion.However, the influences of individual bands on classificationresults may vary greatly. For example, grass and trees are moreseparable in near infrared bands compared to other visible bands.Sensitivity analysis may be conducted to determine the bestweight for each spectral band and the pseudo-waveform to maxi-mize the discrimination ability of the fused data.

Acknowledgements

The authors would like to thank the DigitalGlobe for providingWorldView-2 data and the Dallas Urban Forest Advisory Commit-tee for providing the LiDAR dataset. The research is partially sup-ported by the visiting professor program at King Saud University.

References

Agresti, A., 1996. An Introduction to Categorical Data Analysis. Wiley, New York.Alexander, C., Tansey, K., Kaduk, J., Holland, D., Tate, N.J., 2010. Backscatter

coefficient as an attribute for the classification of full-waveform airborne laserscanning data in urban areas. ISPRS J. Photogramm. Rem. Sens. 65 (5), 423–432.

Anderson, J.E., Plourde, L.C., Martin, M.E., Braswell, B.H., Smith, M.-L., Dubayah, R.O.,et al., 2008. Integrating waveform lidar with hyperspectral imagery forinventory of a northern temperate forest. Rem. Sens. Environ. 112 (4), 1856–1870.

Arroyo, L.A., Johansen, K., Armston, J., Phinn, S., 2010. Integration of LiDAR andQuickBird imagery for mapping riparian biophysical parameters and land covertypes in Australian tropical savannas. For. Ecol. Manage. 259 (3), 598–606.

Attributes for ICESat Laser Operations Periods, 2012. <http://nsidc.org/data/icesat/pdf/glas_laser_ops_attrib.pdf> (last date accessed 28.08.14).

Berger, C., Voltersen, M., Hese, O., Walde, I., Schmullius, C., 2013. Robust extractionof urban land cover information from HSR multi-spectral and LiDAR data. IEEE J.Sel. Top. Appl. Earth Observations Rem. Sens. 6 (6), 1–16.

Blair, J.B., Hofton, M.A., 1999. Modeling laser altimeter return waveforms overcomplex vegetation using high-resolution elevation data. Geophys. Res. Lett. 26(16), 2509–2512.

Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS J.Photogramm. Rem. Sens. 65 (1), 2–16.

Bradley, J.V., 1968. Distribution-free Statistical Tests. Prentice-Hall, EnglewoodCliffs, N.J..

Congalton, R.G., 1991. A review of assessing the accuracy of classifications ofremotely sensed data. Rem. Sens. Environ. 37 (1), 35–46.

DigitalGlobe, Inc., 2010a. The Benefits of the Eight Spectral Bands of WorldView-2.<http://www.digitalglobe.com/sites/default/files/DG-8SPECTRAL-WP-WEB.pdf>(last date accessed 03.06.13).

DigitalGlobe, Inc., 2010b. Radiometric Correction for WorldView-2. <http://www.digitalglobe.com/downloads/Radiometric_Use%20of_WorldView-2_Imagery.pdf>(last date accessed 13.08.13).

Farid, A., Goodrich, D.C., Bryant, R., Sorooshian, S., 2008. Using airborne lidar topredict Leaf Area Index in cottonwood trees and refine riparian water-useestimates. J. Arid Environ. 72 (1), 1–15.

Geerling, G.W., Labrador-Garcia, M., Clevers, J.G.P.W., Ragas, A.M.J., Smits, A.J.M.,2007. Classification of floodplain vegetation by data fusion of spectral (CASI)and LiDAR data. Int. J. Rem. Sens. 28 (19), 4263–4284.

Ghiyamat, A., Shafri, H.Z.M., Mahdiraji, G.A., Shariff, A.R.M., Mansor, S., 2013.Hyperspectral discrimination of tree species with different classifications usingsingle- and multiple-endmember. Int. J. Appl. Earth Obs. Geoinf. 23 (1), 177–191.

Gong, P., Li, Z., Huang, H., Sun, G., Wang, L., 2011. ICESat GLAS data for urbanenvironment monitoring. IEEE Trans. Geosci. Rem. Sens. 49 (3), 1158–1172.

Guo, L., Chehata, N., Mallet, C., Boukir, S., 2011. Relevance of airborne lidar andmultispectral image data for urban scene classification using Random Forests.ISPRS J. Photogramm. Rem. Sens. 66 (1), 56–66.

Hilbert, C., Schmullius, C., 2012. Influence of surface topography on ICESat/GLASforest height estimation and waveform shape. Rem. Sens. 4 (8), 2210–2235.

Jensen, J.R., 2004. Introductory Digital Image Processing: A Remote SensingPerspective, third ed. Prentice Hall.

Ke, Y., Quackenbush, L.J., Im, J., 2010. Synergistic use of QuickBird multispectralimagery and LIDAR data for object-based forest species classification. Rem.Sens. Environ. 114 (6), 1141–1154.

Kim, K., Shan, J., 2011. Building roof modeling from airborne laser scanning databased on level set approach. ISPRS J. Photogramm. Rem. Sens. 66 (4), 484–497.

Koetz, B., Sun, G., Morsdorf, F., Ranson, K.J., Kneubühler, M., Itten, K., et al., 2007.Fusion of imaging spectrometer and LIDAR data over combined radiativetransfer models for forest canopy characterization. Rem. Sens. Environ. 106 (4),449–459.

Laes, D., Reutebuch, S., McGaughey, B., Maus, P., Mellin, T., Wilcox, C., Anhold, J.,Finco, M., Brewer, K., 2008. Practical Lidar Acquisition Considerations forForestry Applications, RSAC-0111-BRIEF1, Salt Lake City, Utah, U.S. Departmentof Agriculture, Forest Service, Remote Sensing Applications Center, 7p.

Lee, D.S., Shan, J., 2003. Combining lidar elevation data and IKONOS multispectralimagery for coastal classification mapping. Mar. Geodesy 26 (1–2), 117–127.

Lee, D.H., Lee, K.M., Lee, S.U., 2008. Fusion of lidar and imagery for reliable buildingextraction. Photogramm. Eng. Rem. Sens. 74 (2), 215–225.

Lovell, J.L., Jupp, D.L.B., Culvenor, D.S., Coops, N.C., 2003. Using airborne and ground-based ranging lidar to measure canopy structure in Australian forests. Can. J.Rem. Sens. 29 (5), 607–622.

Mallet, C., Bretar, F., 2009. Full-waveform topographic lidar: state-of-the-art. ISPRSJ. Photogramm. Rem. Sens. 64 (1), 1–16.

Mallet, C., Bretar, F., Roux, M., Soergel, U., Heipke, C., 2011. Relevance assessment offull-waveform lidar data for urban area classification. ISPRS J. Photogramm.Rem. Sens. 66 (6 SUPPL.), S71–S84.

Mittal, S., Vyas, S., Prasanna, S.R.M., 2013. Analysis of lombard and angry speechusing Gaussian Mixture Models and KL divergence. In: 2013 NationalConference on Communications, NCC 2013, Anonymous 15 February 2013through 17 February 2013, New Delhi ed.

Muss, J.D., Mladenoff, D.J., Townsend, P.A., 2011. A pseudo-waveform technique toassess forest structure using discrete lidar data. Rem. Sens. Environ. 115 (3),824–835.

Olszewski, D., 2012. A probabilistic approach to fraud detection intelecommunications. Knowl.-Based Syst. 26, 246–258.

Padwick, C., Deskevich, M., Pacifici, F., Smallwood, S., 2010. WorldView 2 pan-sharpening, San Diego. In: Proceedings of the ASPRS Annual Conference,Anonymous, California ed.

Pedley, M.I., Curran, P.J., 1991. Per-field classification: an example using SPOT HRVimagery. Int. J. Rem. Sens. 12 (11), 2181–2192.

Pirotti, F., Laurin, G.V., Vettore, A., Masiero, A., Valentini, R., 2014. Small footprintfull-waveform metrics contribution to the prediction of biomass in tropicalforests. Rem. Sens. 6 (10), 9576–9599.

Popescu, S.C., Zhao, K., 2008. A voxel-based lidar method for estimating crown baseheight for deciduous and pine trees. Rem. Sens. Environ. 112 (3), 767–781.

Popescu, S.C., Zhao, K., Neuenschwander, A., Lin, C., 2011. Satellite lidar vs. smallfootprint airborne lidar: comparing the accuracy of aboveground biomassestimates and forest structure metrics at footprint level. Rem. Sens. Environ.115 (11), 2786–2797.

232 Y. Zhou, F. Qiu / ISPRS Journal of Photogrammetry and Remote Sensing 101 (2015) 221–232

Reitberger, J., Schnörr, C., Krzystek, P., Stilla, U., 2009. 3D segmentation of singletrees exploiting full waveform LIDAR data. ISPRS J. Photogramm. Rem. Sens. 64(6), 561–574.

Sasaki, T., Imanishi, J., Ioki, K., Morimoto, Y., Kitada, K., 2012. Object-basedclassification of land cover and tree species by integrating airborne LiDAR andhigh spatial resolution imagery data. Landscape Ecol. Eng. 8 (2), 157–171.

Shackelford, A.K., Davis, C.H., 2003. A combined fuzzy pixel-based and object-basedapproach for classification of high-resolution multispectral data over urbanareas. IEEE Trans. Geosci. Rem. Sens. 41 (10 PART I), 2354–2363.

Sridharan, H., Qiu, F., 2013. Developing an object-based hyperspatial imageclassifier with a case study using worldview-2 data. Photogramm. Eng. Rem.Sens. 79 (11).

Stow, D.A., Toure, S.I., Lippitt, C.D., Lippitt, C.L., Lee, C., 2012. Frequency distributionsignatures and classification of within-object pixels. Int. J. Appl. Earth Obs.Geoinf. 15 (1), 49–56.

Sugumaran, R., Voss, M., 2007. Object-oriented classification of LIDAR-fusedhyperspectral imagery for tree species identification in an Urbanenvironment. In: 2007 Urban Remote Sensing Joint Event, URS, Anonymous11 April 2007 through 13 April 2007, Paris ed.

Toure, S.I., Stow, D.A., Weeks, J.R., Kumar, S., 2013. Histogram curve matchingapproaches for object-based image classification of land cover and land use.Photogramm. Eng. Rem. Sens. 79 (5), 433–440.

Ussyshkin, V., Theriault, L., 2011. Airborne lidar: advances in discrete returntechnology for 3D vegetation mapping. Rem. Sens. 3 (3), 416–434.

Wagner, W., Ullrich, A., Ducic, V., Melzer, T., Studnicka, N., 2006. Gaussiandecomposition and calibration of a novel small-footprint full-waveform

digitising airborne laser scanner. ISPRS J. Photogramm. Rem. Sens. 60 (2),100–112.

Wang, X., Cheng, X., Li, Z., Huang, H., Niu, Z., Li, X., et al., 2012. Lake water footprintidentification from time-series ICESat/GLAS data. IEEE Geosci. Rem. Sens. Lett. 9(3), 333–337.

Yao, W., Krzystek, P., Heurich, M., 2012. Tree species classification and estimation ofstem volume and DBH based on single tree extraction by exploiting airbornefull-waveform LiDAR data. Rem. Sens. Environ. 123, 368–380.

Zaletnyik, P., Laky, S., Toth, C., 2010. LIDAR waveform classification using self-organizing map. In: American Society for Photogrammetry and Remote SensingAnnual Conference 2010: Opportunities for Emerging Geospatial Technologies,Anonymous 26 April 2010 through 30 April 2010, San Diego, CA ed., pp. 1055–1066.

Zhang, C., Qiu, F., 2012. Mapping individual tree species in an urban forest usingairborne lidar data and hyperspectral imagery: AAG remote sensing specialtygroup 2011 award winner. Photogramm. Eng. Rem. Sens. 78 (10), 1079–1087.

Zhang, J., de Gier, A., Xing, Y., Sohn, G., 2011. Full waveform-based analysis for foresttype information derivation from large footprint spaceborne lidar data.Photogramm. Eng. Rem. Sens. 77 (3), 281–290.

Zhou, W., 2013. An object-based approach for urban land cover classification:integrating LiDAR height and intensity data. IEEE Geosci. Rem. Sens. Lett. 10 (4),928–931.

Zhuang, W., Mountrakis, G., 2014. An accurate and computationally efficientalgorithm for ground peak identification in large footprint waveform LiDARdata. ISPRS J. Photogramm. Rem. Sens. 95, 81–92.