Você está na página 1de 25

This article was downloaded by: [University of Santiago de Compostela], [Eduardo

Gonzlez-Ferreiro]
On: 14 May 2013, At: 11:18
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Remote


Sensing
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/tres20

Classification of rural landscapes


from low-density lidar data: is it
theoretically possible?
a

Sandra Bujn , Eduardo Gonzlez-Ferreiro , Laura Barreiroa

Fernndez , Ins Sant , Eduardo Corbelle & David Miranda

Land Laboratory, Department of Agroforestry Engineering ,


University of Santiago de Compostela , Lugo , Spain

To cite this article: Sandra Bujn , Eduardo Gonzlez-Ferreiro , Laura Barreiro-Fernndez , Ins
Sant , Eduardo Corbelle & David Miranda (2013): Classification of rural landscapes from lowdensity lidar data: is it theoretically possible?, International Journal of Remote Sensing, 34:16,
5666-5689
To link to this article: http://dx.doi.org/10.1080/01431161.2013.792230

PLEASE SCROLL DOWN FOR ARTICLE


Full terms and conditions of use: http://www.tandfonline.com/page/terms-andconditions
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation
that the contents will be complete or accurate or up to date. The accuracy of any
instructions, formulae, and drug doses should be independently verified with primary
sources. The publisher shall not be liable for any loss, actions, claims, proceedings,
demand, or costs or damages whatsoever or howsoever caused arising directly or
indirectly in connection with or arising out of the use of this material.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing, 2013


Vol. 34, No. 16, 56665689, http://dx.doi.org/10.1080/01431161.2013.792230

Classification of rural landscapes from low-density lidar data: is it


theoretically possible?
Sandra Bujn*, Eduardo Gonzlez-Ferreiro, Laura Barreiro-Fernndez, Ins Sant,
Eduardo Corbelle, and David Miranda
Land Laboratory, Department of Agroforestry Engineering, University of Santiago de Compostela,
Lugo, Spain
(Received 1 March 2012; accepted 20 November 2012)
Lidar technology has become an important data source in 3D terrain modelling.
In Spain, the National Plan for Aerial Orthophotography will soon release public low
density lidar data (0.51 pulses/m2 ) for most of the country territory. Taking advantage
of this fact, this article experimentally assesses the possibility of classifying a rural landscape into eight classes using multitemporal and multidensity lidar data and analyses the
effect of point density on classification accuracy. Two statistical methods (transformed
divergence and the JeffriesMatusita distance) were used to assess the possibility of
discriminating the eight classes and to determine which data layers were best suited for
classification purposes. The results showed that dirt road cannot be discriminated from
bare earth and that the possibility of discriminating bare earth, pavement, and low
vegetation decreases when using densities below 4 pulses/m2 . Two non-parametric
tests, the KruskalWallis test and the Friedman test, were used to strengthen the results
by assessing their statistical significance. According to the results of the KruskalWallis
test, lidar point density does not significantly affect the classification, whereas the results
of the Friedman test show that bands could be considered as the only parameter affecting
the possibility of discriminating some of the classes, such as high vegetation. Finally,
the J48 algorithm was used to perform cross-validation in order to obtain the most familiar quantitative values in the international literature (e.g. overall accuracy). Mean overall
accuracy was around 85% when the eight classes were considered and increased up to
95% when dirt road was disregarded.

1. Introduction and previous work


Nowadays the need to update existing databases is widely recognized in order to get the
required information and to detect changes in the environment. This updating should allow
for the development of suitable planning tools for different areas. Human observation and
interpretation has been, and in many cases is, one of the main database update methods.
A recent example is SIGPAC (Sistema de Informacin Geogrfica de Parcelas Agrcolas),
the Spanish Land Parcel Information System (LPIS), which was created to comply with EU
regulations. However, human observation and interpretation is a time-consuming method
subject to omission errors and to the interpreters abilities and knowledge (Pacifici et al.
2007).
*Corresponding author. Email: sandra.bujan@usc.es
2013 Taylor & Francis

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5667

New technologies, including remote sensing, have contributed to increasing the speed,
cost-efficiency, and precision associated with data acquisition, and to developing maps with
spatial resolutions and accuracies that were not feasible before (McRoberts and Tomppo
2007). Among these technologies is laser scanning, which has been in use for more than
30 years and can be fully exploited thanks to the development of highly accurate positioning systems such as GPS (Charaniya, Manduchi, and Lodha 2004). In line with these
developments, the Spanish National Plan for Aerial Orthophotography will soon publish
low-density lidar data (0.51 first return/m2 ) for most of the territory of Spain. It is
envisaged that this data will be updated every two years.
Laser scanning data can be used in a wide range of applications, including land-cover
mapping, updating of geographical databases, assessment of natural hazards, environmental
modelling, and detection of changes (Matinfar et al. 2007). The variety of techniques developed for change detection can be classified into two broad categories: (1) those detecting
binary change/no-change information; and (2) those detecting a detailed fromto change
trajectory (Lu, Batistella, and Moran 2004). Yet, these categories can be further divided
into various sub-categories according to the basic unit of analysis (pixel vs. object) or the
nature of the input data (aerial or satellite imagery, laser scanning data, or both).
Most authors have used aerial imagery to discriminate between different land-cover
classes and, with new high resolution commercial satellites, satellite imagery to discriminate between different land-cover classes (Davis and Wang 2002; Pu, Landry, and Yu 2011).
The relatively low spatial resolution of images used to be a limiting factor for the classification of complex urban environments and rural areas, but such a constraint has been
overcome by the availability of images with higher spatial resolution. Actually, these systems allow even for fine-scale detection of physical changes in objects such as buildings,
houses, or roads.
Recently, a group of authors have used high spatial resolution satellite imagery (e.g.
QuickBird and IKONOS), which provides greater potential to extract more detailed information on urban areas (Van der Sande, De Jong, and De Roo 2003; Lu and Weng 2009;
Dengsheng, Hetrick, and Moran 2010). Traditionally, pixel-based detection methods have
been used. Cannavacciuolo et al. (2007) presented a method for unsupervised change detection in RGB-NIR (full spectrum = visible + near infrared) images with 2.8 m resolution.
This method was based on a Bayesian approach with unsupervised parameter estimation
that classified each pixel into two possible classes: change or no-change. However, improvements in segmentation algorithms and the possibility of mitigating georeferencing errors
(Wulder et al. 2008) have encouraged the use of object-based approaches. In this sense,
Im, Jensen, and Tullis (2008) used multitemporal QuickBird imagery to investigate five
change detection methods based on the detection of a fromto change trajectory in order
to analyse how contextual features could improve classification results. In addition, they
investigated whether object-based methods improved change detection as compared to
pixel-based methods. Their research revealed that object-based classifications were more
effective in terms of cost and time than pixel-based classifications and that they reduced
the salt and pepper effect. This effect is caused by the incorrect operation of the sensor or the antenna receiving the signal, which as a result can present some isolated pixels
with increased contrast as compared with neighbouring pixels (Chuvieco 2008). Similarly,
Dengsheng, Hetrick, and Moran (2010) used QuickBird imagery with 0.6 m resolution to
demonstrate that using textures and object-based classification methods instead of per-pixel
spectral-based classification methods significantly improved classification performance in
urban areas and allowed for detection of surface changes. However, their method showed
problems in classifying pasture and grasslands. Despite the high resolution of IKONOS

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5668

S. Bujn et al.

imagery, Van der Sande, De Jong, and De Roo (2003) obtained poor classification results
for residential areas and roads.
Sometimes discrimination problems may stem from radiometric similarities between
classes. For example, the information provided by a single image is not enough for roofs
made of materials whose radiometric behaviour is similar to the behaviour of road materials. However, when differences in radiance values between two dates are taken as an
indicator of change detection, it must be determined whether radiance changes are caused
by a land-use change or by differences in image acquisition, such as atmospheric conditions, illumination, or sensor calibration (Singh 1989). Taking this into consideration,
some authors have claimed that combining aerial or satellite imagery with height data
would improve the identification of the classes of interest (Charaniya, Manduchi, and Lodha
2004; Tymkow and Borkowski 2008) because both types of data are complementary: aerial
imagery provides spectral information and high planimetric accuracy, whereas lidar data
provides high altimetric accuracy and accurate information of elevation, slope, and some
spectral information, such as intensity data; thus, an optimum equilibrium is reached. In this
context, many authors have used pixel-based methods (Gonalves et al. 2007; Champion
et al. 2009). However, the results of pixel-based classifications are not as good as expected
because of the heterogeneity of the spectral information and the spatial complexity of urban
landscapes and rural areas. For this reason, other authors have used object-based methods
(Zhou, Troy, and Grove 2008; Chen et al. 2009; Barreiro-Fernndez et al. 2010; Dinis et al.
2010). Barreiro-Fernndez et al. (2010) compared an object-based classification and an
automatic classification using a support vector machine at a two-level ensemble to identify four classes (conifers, deciduous, scrub, low vegetation, and bare earth) in a
forest area. To this end, they combined lidar height (Digital Canopy Model) and intensity
data with spectral information acquired from an orthophotograph. Using a change detection
approach based on the detection of a fromto change trajectory, Zhou, Troy, and Grove
(2008) combined lidar data and multitemporal colour-infrared digital aerial imagery with
high spatial resolution to analyse changes in an urban area.
Only a few authors have used lidar data alone to perform land-cover classifications;
some of them have assessed the possibility of using intensity images derived from lidar data
instead of multispectral images and have concluded that intensity data could be satisfactorily used (Song et al. 2002). However, these authors have found difficulties in discriminating
between vegetation-covered and built-up areas. Based on their own experiments, Tymkow
and Borkowski (2008) claimed that lidar data alone are not sufficient to perform a landcover classification, particularly in vegetation areas. Contrary to their claim, Brennan and
Webster (2006) identified up to 10 distinct classes in an area in which various land-cover
types were present and obtained a high average accuracy. Other authors used multitemporal
lidar point clouds to detect changes in forest areas (Yu et al. 2004, 2006); they were the first
to demonstrate the applicability of small footprint, high sampling density airborne laser
scanners for the detection of harvested trees, and the estimation of forest growth from two
lidar data sets acquired with a 2 year difference.
Lidar-based classifications of urban or rural areas have become more common over the
years. This article experimentally assesses the possibility of classifying a rural landscape of
complex relief into eight land-cover classes present in the study area: building, high vegetation (mature trees), harvested forest, young trees, bare earth, pavement, dirt road,
and low vegetation (crop and grass field). In addition, it aims to determine the extent to
which lidar point density affects land-cover classification in this area. To this end, multitemporal lidar data acquired in 2004 and 2010 were used. In addition, the following information
layers were calculated for each density data set: intensity image (I image), an image that

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5669

contains information pertaining to the difference between returns (FL image), the normalized digital surface model (nDSM or nM), and the difference between the nDSMs for
2010 and 2004 (DnM image). Based on this data, two statistical methods (transformed
divergence and the JeffriesMatusita (JM) distance (Richards and Jia 2006)) were used to
assess whether it was possible to discriminate the eight pre-defined land-cover classes with
sufficient accuracy to update existing databases. Finally, the J48 algorithm included in the
Rweka package, available from the R free software (R Core Team 2012), was used to perform cross-validation on the training samples. Cross-validation allowed for an approximate
estimation of the reliability of the classification method.

2. Study area and data sets


2.1. Study area
The study area is located in Vilapena (43 24 N, 7 12 W), municipality of Trabada, northwest Spain, and covers about 4 km2 . This area was chosen for its relief, characterized by
a combination of high elevations of terrain and deep, fertile valleys (elevation range of
250530 m). Additionally, this area exhibits a variety of land-use types characteristic of
the region, where cropped fields, vast forests, and small rural settlements converge, providing an ideal location to test the proposed methods, which could be applied more broadly
(Figure 1).

Figure 1.

Location map. Detailed 2004 colour orthophotograph of the Trabada study area.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5670

S. Bujn et al.

2.2. Data sets


Multitemporal airborne laser scanning surveys were conducted with an Optech ALTM
2033 system (Infoterra Ltd., S/N 01D126, Leicester, UK) in 2004 and a Leica ALS50-II
system (Cartogalicia S.L.) in 2010. The point clouds used in this research were recorded
in November 2004 and October 2010. The greatest difference between both systems
was the capability of the Optech ALTM 2033 to record two laser returns for each pulse
regardless of land-cover characteristics. However, the Leica ALS50-II system recorded
multiple returns (up to five) only in places with permeable surfaces. Table 1 summarizes
the technical characteristics of the scanners used to capture the two lidar data sets.
2.3. Preprocessing of lidar data sets
The 2004 data set was used to generate most information layers. The 2004 flight had a
0.2 mrad beam divergence and a flying height of 10701250 m above a mean terrain elevation of 350 m. Consequently, the laser footprint diameter was nominally 0.18 m, whereas
the laser footprint diameter for the 2010 data set (flying height of 1900 m) was approximately 0.31 m. The illuminated area covered by the footprint can be seen as the maximum
achievable spatial resolution in terms of reflectance information (Hfle and Pfeifer 2007)
which, combined with lidar point density, has resulted in lidar-derived images generated
with a spatial resolution of 0.5 m.
From each lidar data set, a DSM and a digital terrain model (DTM) were generated.
The DTM was generated using TerraScan proprietary software for Microstation (Axelsson
2000; Holmgren, Nilsson, and Olsson 2003; Sithole and Vosselman 2004; Donoghue et al.
2007). This software uses the progressive densification algorithm described in Axelsson
(2000). First, a number of local minima were defined. Based on the defined minima, a
sparse triangulated irregular network (TIN) model was generated. As discussed in Axelsson
(2000), the TIN is then progressively densified. In every iteration, points are added to the
TIN if they were below specific thresholds. Such a densification results in representation of
the surface that gradually approaches the true ground surface.
The altimetry of the DTMs was validated by using 100 control points, 20 of which were
located in areas without vegetation. Control points were obtained by using a GPS and a
Trimble 5603 total station (with an approximate accuracy of 12 cm). The root mean square
Table 1.

Technical characteristics of laser scanners.

Date
Sensor
Laser pulse frequency
Scan angle
Scan frequency
Beam divergence
Wavelength
Overlap
No. of passes
Measurement density
Flying height
Planimetric accuracy
Height accuracy

Lidar data set 1

Lidar data set 2

November 2004
Optech ALTM 2033
33 000 Hz
10
31 Hz
0.2 mrad
1064
60%
18
4 pulses/m2
10701250 m
0.60 m
0.15 m

October 2010
Leica ALS50-II
115,700 Hz
12
40 Hz
0.2 mrad
1064
minimum 30%
6
4 pulses/m2
1900 m
0.20 m
0.07 m

International Journal of Remote Sensing


Table 2.

Characteristics of point reduction and DTMs.


Lidar data
sets
4 pulses/m2

No. of Points

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5671

No. of terrain points

DTM accuracy (RMSE) (m)

DTM accuracy (RMSE) (m)


(control points in areas
without vegetation)

2004 lidar
data
2010 lidar
data
2004 lidar
data
2010 lidar
data
2004 lidar
data
2010 lidar
data
2004 lidar
data
2010 lidar
data

2 pulses/m2

1 pulse/m2

0.5 pulse/m2

2,761,543

1,507,422

761,126

380,726

2,602,271

1,392,783

735,308

368,037

410,231

318,058

270,370

174,926

384,820

315,284

235,308

185,052

0.171

0.236

0.239

0.234

0.344

0.417

0.441

0.517

0.105

0.128

0.103

0.110

0.171

0.187

0.199

0.204

error (RMSE) proposed in the NSSDA (National Standard for Spatial Data Accuracy) was
the statistic used to quantitatively assess the accuracy of the DTMs (see Table 2 for terrain
point filtering data and DTM accuracy). DSM and DTM were used to calculate the nDSMs
for each data set. The nDSM represents the absolute height of objects with respect to the
terrain. To calculate the nDSM, the DTM should be subtracted from the DSM.
The difference between the 2010 nDSM and the 2004 nDSM was also calculated by
subtracting each pixel value in the nDSM for the first year (2004) from its corresponding
pixel value in the nDSM for the second year (2010). The resulting difference image (DnM
image) represented the pixel-wise changes that took place in the study area between both
dates. Previously, the difference between DTMs was calculated in order to verify that the
ground surface had not undergone any significant changes between 2004 and 2010 (only
0.1% of the study area showed a difference between DTMs below 1 m and 0.7% of the
area showed a difference over 1.5 m, with a maximum difference between DTMs of 2.69 m
and a minimum difference of 2.34 m, both in the roadside slope area). In addition, lidar
data for 2004 was used to generate an intensity image from the first return alone and an
image that contained information pertaining to the difference between returns (FL image).
Figure 2 shows the information layers obtained from lidar data acquired at a density of
4 pulses m2 .

3. Methods
3.1. Reduction of point density
Original lidar data with different resolutions (obtained from flights that were carried out in
the same area, at the same time and with the same flight parameters, but at different lidar
pulse densities) would be most desirable for assessing the influence of lidar density (note
that this way of proceeding does not allow for the simulation of other potentially influent
parameters such as flight height or scan angle). The lack of such data can be overcome by
artificially reducing the original lidar point cloud (Gonzlez-Ferreiro, Diguez-Aranda, and

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5672

S. Bujn et al.

(a)

(b)

(c)

(d)

Figure 2. Layers obtained from lidar data acquired at a density of 4 pulses m2 . (a) Normalized I
image, (b) nM, (c) DnM, (d) FL image.

Miranda, 2012). In this case, lidar flights had a point density of 4 pulses m2 , which equals
8 returns m2 (one for the first pulse and another one for the last pulse). To analyse the effect
of a systematic reduction in lidar point cloud density on the classification of a complex rural
area, the original lidar point density was reduced and three data sets with densities of 2, 1,
and 0.5 pulses m2 were obtained. The study area was divided into 1 m 1 m cells. Within
each cell, 4, 2, and 1 returns, respectively, were sequentially and randomly chosen. The
random selection was carried out according to the following conditional structure.

The reduction in the number of first returns (type I) and last returns (type II) is
proportional.

For a density of 2 pulses m2 , two type I returns and two type II returns are chosen
within each 1 1 m cell.

For a density of 1 pulse m2 , one type I return and one type II return are chosen
within each 1 1 m cell.

For a density of 0.5 pulses m2 , one type I return is chosen if the number of type I
returns in the cell is higher than the number of type II returns, whereas one type II
return is chosen if the number of type II returns exceeds the number of type I returns.
The original proportion was expected to be maintained because the selection was random and involved a large number of data. Table 2 shows data pertaining to point reduction,
number of filtered points classified as terrain, and accuracy of the generated DTMs for each
of the densities considered.
As shown in Table 2, the decreased accuracy of the DTMs for 2010 with respect to the
accuracy of the DTMs for 2004 was due mainly to the fact that some of the control points

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5673

used to estimate accuracy were located in forest areas. In 2004, such areas were occupied
by young stands and the laser beam could easily penetrate down to the ground. In 2010,
such areas were covered by mature stands with denser foliage, which caused a decrease
in the proportion of laser beams that reached the ground. As a result, the coarseness of
DTM interpolation increased with the decrease in the number of points classified as terrain,
which resulted in a reduction in accuracy. Yet, after having reduced the effects of control
points under tree areas by considering only the points located in grasslands and roads (see
Table 2), we have concluded that the decrease in accuracy was affected also by the specific
characteristics of each flight (angular errors, caused by errors in the inertial measurement
unit (IMU) and angular encoder on the laser scanner mirror).
3.2. Intensity normalization
As defined by Singh, Vogler, and Meentemeyer (2010), intensity is a radiometric constituent of lidar data, which is recorded by the sensor as the amount of energy backscattered
from objects or Earths surface. The intensity values recorded by the sensor remain
unchanged under different illumination conditions (shadows or occlusions) (Donoghue
et al. 2007; Hfle and Pfeifer 2007), but are affected by other factors, among which are the
following.

Topography and properties of the terrain: the roughness, orientation of the surface,
and composition and density of the terrain and of the objects placed on it directly
affect the intensity values recorded by the sensor. These factors are considered intrinsic to objects and generally do not require compensation. In contrast, changes in
terrain elevation affect path length and result in unwanted weakening or strengthening of the intensities (Luzum, Starek, and Slatton 2004; Hfle and Pfeifer 2007;
Mazzarini et al. 2007).
Characteristics of flight and sensor: beam divergence, and variations in flying height
and scan angle alter footprint size and affect path length (Luzum, Starek, and Slatton
2004; Donoghue et al. 2007) and, therefore, intensity values. The footprint of a laser
scanner is usually much smaller than the distance between two laser points. Hence,
the intensities only represent the reflectance properties of a small part of the terrain.
The difference between the footprint size and the distance between two laser points
can be quite large (Vosselman 2002).
Atmospheric conditions: atmospheric conditions may cause an attenuation of the signal that is sent by or reaches the sensor (Luzum, Starek, and Slatton 2004; Donoghue
et al. 2007).
Yet, the effects of these factors on the intensity values vary and have many nuances.
Accordingly, a number of approaches can be used to address the same problem. In this
sense, Hasegawa (2006) claimed that the effect of intensity correction with distance is
sometimes not significant and, consequently, the use of raw intensity values can occasionally be justified. In agreement with Hasegawa (2006), Holmgren and Persson (2004) used
uncorrected intensity values to identify species of individual trees based on low scan angles
(below 20 ), stability of the sensor during data acquisition, and uniform ground topography. Alternatively, Song et al. (2002) used a median filter to reduce the salt and pepper
noise in the image. These authors found that filtering and interpolation can help mitigate
the noise present in the intensity images.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5674

S. Bujn et al.

Other authors acknowledge the need to consider range, incidence angle, and atmospheric attenuation for intensity normalization. For instance, Gross, Jutzi, and Thoennessen
(2008) demonstrated that the same materials can have different values of intensity. They
verified that intensities are strongly correlated with incidence angle, which is given by
the beam propagation direction and the normal vector of an object surface. According to
Gross, Jutzi, and Thoennessen (2008), only 5% of intensity value modifications in urban
areas are caused by the range. Usually, urban landscapes do not have complex terrain, and
the greatest height variations in these areas are caused by buildings, with height values that
are not normally significant to affect intensity values. For this reason, the greatest intensity
disturbance in urban landscapes is caused by the incidence angle.
In this sense, Jutzi and Gross (2010) showed that large intensity variation caused by
surface orientation and beam propagation direction can be reduced, and consequently,
intensity can be improved by using a standard Lambertian reflection model. However, the
improvements obtained in grasslands and forest areas were very small. In addition, Jutzi
and Gross (2010) showed that the incidence angle directly affected intensity values, but the
improvements obtained after normalization for angles below 10 were very small. Likewise,
Hfle and Pfeifer (2007) described two different approaches to intensity data correction,
data-driven correction, and model-driven correction.
Yet, the application of the approaches described in the above paragraphs requires data
that are not always available (e.g. incidence angle and range for each point). Donoghue
et al. (2007) used intensity data to separate spruce and pine species. They separated
intensity values by height because they had ground data of the height of each laser
pulse, which allowed them to eliminate non-vegetation returns. Based on this data, they
normalized intensity values using the method reported by Luzum, Starek, and Slatton
(2004).
Another group of authors (Mazzarini et al. 2007; Tymkow and Borkowski 2008; Wang
and Glenn 2009; Garca et al. 2010; Singh, Vogler, and Meentemeyer 2010) normalized
intensity values to standard range without using the DTM value. In this approach, the
normalized intensity values were obtained by multiplying the raw intensity value by the
quotient of the range of each point (calculated as the difference between average flying height and the height of each point) and the standard range. Intensity as calculated
with this method equals the values of intensity that would have been recorded if the
range was the same for all of the points, such that the effect of path length variations is
eliminated.
In our case, range and scan angle data were not available for every pulse and the terrain
was very steep, with slopes above 40 in some areas and a topographic range of 250530 m
above mean sea level. For this reason, the method suggested by Mazzarini et al. (2007) was
used to normalize the intensity of each lidar point. The range of each point was approximated as the difference between average flying height and the height of each point. Raw
intensity values were not lost during normalization. Rather, the new normalized values
were stored as an additional feature for each pulse. A number of authors (Solberg et al.
2008; Hopkinson and Chasmer 2009) recommended the use of all of the returns to generate the intensity image, but we decided to use only the first return in order to avoid
the effect of mixed pixels in tree areas. The mixed-pixel problem is caused by the mixture of intensity values corresponding to returns that hit tree crowns and intensity values
corresponding to returns that penetrate down to the ground. Finally, a median filter was
applied to the normalized intensity image to mitigate the effect of the noise present in the
image.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5675

3.3. Feature selection methods


The main question we are attempting to address is is it possible to classify a rural landscape
from lidar data alone? To answer this question, we have adopted an experimental approach.
Determining whether a classification can be carried out with the available data before
actually performing it, and identifying the most useful and efficient information greatly
contributes to choosing the most suitable approach and determining whether the classification is feasible. Under this assumption, the need to analyse and select (if necessary) the
most useful data among the available information becomes important.
The methods used to make such an analysis and/or selection are termed feature selection techniques. Many criteria can be used to classify these methods, which can be divided
into two large groups: (1) the methods that use a transformation of input data (e.g. principal components analysis and the KahunenLove transform (Richards and Jia 2006)) and
(2) the methods that assess the features available for each pixel (or object) and choose the
information layers that allow for the best discrimination of classes. According to remotesensing literature, the most popular separability measures in the second approach are the
discriminator indices, namely divergence and transformed divergence (TD), Bhattacharyya
distance, and JM distance (Richards and Jia 2006; Gu et al. 2008; Tolpekin and Stein 2009).
Because transformed measures have been shown to be more powerful than other
approaches (Gambarova et al. 2010), two transformed measures (transformed divergence
and JM distance) have been used in this article to estimate class separability.
Divergence is a measure of the separability of a pair of probability distributions that
has its basis in their degree of overlap. An increase in the value of divergence involves an
increase in separation between classes (based on the bands considered in the calculation).
Because classes in remote-sensing image data are modelled by multidimensional normal
distributions, the specific form of the divergence when the distributions p(x/i ) and p(x/j )
(the values of the ith and jth spectral class probability distributions at the position x) are
normal is (Richards and Jia 2006):
Dij =

1
1
-1
-1
-1
T
Tr [(Ci Cj )(C-1
j Ci )] + Tr [(Ci + Cj )(i j )(i j ) ],
2
2

(1)

where i and j are the two signatures being compared; i and j are means; Ci and Cj are
covariances; Tr is the trace function and T is the transposition function.
The main problem with this measure is that it never becomes saturated, which means
that if we wanted to relate classification accuracy to divergence, accuracy would increase
constantly with an increase in the number of bands used in the classification (Chuvieco
2008). Accordingly, the use of transformed divergence is recommended, insofar as transformed divergence establishes an a priori upper bound, defined as (Richards and Jia 2006;
Schowengerdt 2007):
TD = 2[1 e

Dij
8

],

(2)

where Dij is the divergence and coefficient 2 represents the maximum value of divergence,
although other authors have often used a value of 2000.
The JM distance between a pair of spectral classes under the assumption of Gaussian
distributions was defined by Richards and Jia (2006) as follows:
JM = 2[1 eBij ],

(3)

5676

S. Bujn et al.

where

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

 C +C 

1
i
j
1
1
T Ci + Cj
Bij = (i j )
(i j ) + ln  2
,
8
2
2
Ci Cj

(4)

which is referred to as the Bhattacharyya distance, and where i and j are the two signatures
being compared.
This separability measure estimates the probability of correct classification. As with
TD, the JM distance is used to transform the values of the Bhattacharyya distance to a
pre-specified range. One of the characteristics of the JM distance is that it has negative
exponential growth, i.e. distance growth becomes progressively milder, which allows us
not to overestimate separability for high distance values (Kavzoglu and Mather 2000).
TD and JM distances take values between 0 and 2. A value between 0 and 1.7 (between
two classes) would suggest that identifying both classes is not possible because both have
similar values for the information layers involved in the calculation. Values between 1.7 and
1.9 suggest a poor separability of classes that could be caused by the presence of some
samples with abnormal values or to the presence of similar characteristics for both classes.
Conversely, values above 1.9 would suggest a very good separability (Gibson and Power
2000).
When the number of classes is larger than two, an average TD and an average JM distance are used. Assuming that L classes have been defined, the average TD can be expressed
according to Tolpekin and Stein (2009):
TD =

L1 
L

2
TD .
L(L 1) =1 =+1

(5)

Similarly, the average JM distance can be defined as


JM =

L1 
L

2
JM .
L(L 1) =1 =+1

(6)

Considering these average values, the combination of bands with the maximum TD
(or JM) will be considered as the combination of bands best suited for classification
purposes.
Originally, the statistical methods described in the above paragraphs are used to measure distance between classes, which is equated to separability. The divergence method
assumes that an increase in distance between classes involves an increase in separability,
but actually, this is not true. If divergence is represented as a function of class distance,
we have that the divergence grows indefinitely so the distance between class means also
grows. If this function is compared to the representation of the probability of obtaining a
correct classification (separability) based on distance, given that the function has a vertical
asymptote (i.e. does not grow indefinitely), it can be observed that separability grows more
slowly than distance. On the contrary, the representation of the JM distance as a function
of distance, which shows a horizontal asymptote, matches the representation of separability better. Accordingly, the JM distance is a more reliable estimator of classification
performance (for more details, see Richards and Jia, (2006)).

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5677

These statistical measures of separability have been widely used in remote sensing to
quantitatively select a subset of information layers that provide the greatest level of statistical separability between two spectral classes (Chih-Cheng et al. 1997). Traditionally,
these methods have been used mainly with the spectral bands of satellite or aerial imagery.
Gambarova et al. (2010) used TD and the JM distance to determine the best band combination of four Systme Pour lObservation de la Terre 5 (SPOT5) image data sets for
four types of rare vegetation communities. TD results were more encouraging than JM
distance results, something that was expected because TD overestimates classification performance. The confusion matrix calculated after the classification process confirmed the
results provided by the separability measures used in the analysis. Gu et al. (2008) used
the JM distance to select and evaluate the best textural parameters for the identification of
residential areas on panchromatic IKONOS imagery with 1 m spatial resolution.
However, only a few authors have used these methods with lidar-derived information
layers. Song et al. (2002) assessed the possibility of land-cover classification using lidar
intensity data instead of multispectral data. They used the TD method to assess the separability of intensities on three intensity grids for four classes (asphalt road, grass, house
roof, and tree). They concluded that asphalt roads and some roofs showed similar intensity values. Likewise, they concluded that the intensities for trees and low vegetation were
similar.
In this article, ENVI proprietary software (Exelis Visual Information Solutions,
Boulder, CO, USA) was used to calculate TD and JM distance. Four information layers
generated from lidar data with the four available point densities were used. First, 60 square
samples representative of each of the eight classes present in the study area (bare earth,
building, harvested forest, high vegetation, low vegetation, pavement, dirt road,
and young trees) were randomly identified. Each sample consisted of a cluster of four
0.5 m pixels. These samples allowed us to calculate the mean vector and the covariance
matrix C for every class , which were used as input data for the calculation of the statistical measures of separability. Based on the four information layers for each point density, we
generated images with every possible combination of information layers or bands without
mixing different densities. Later, the statistical measures of separability for the generated
images were automatically calculated. Based on the results, we evaluated the bands that
were best suited to identify the classes of interest, the performance of the two statistical
methods, and the possibility of discriminating all the classes.
Then, we strengthened our results by assessing their statistical significance. First, the
ShapiroWilk test was used to determine whether data came from a normal distribution
(Shapiro and Wilk 1965, 1968). Results suggested that there was strong evidence to reject
H0 . Given that our data did not follow Gaussian distribution, non-parametric tests were
used to assess the effect of density and combination of bands on the determination of TD
and JM in general, and particularly on some specific pairs of classes.
To quantitatively assess the overall (sample size = 28 pairs of classes 11 band
combinations 4 lidar data densities = 1232) effect of the combination of bands and of
lidar point density on the average TD and JM values and, consequently, on the possibility of
discriminating all eight classes of interest during classification, we used the KruskalWallis
test (Kruskal 1952; Hollander and Wolfe 2000). The non-parametric KruskalWallis test
allows us to compare two or more independent samples and to determine whether the
differences between medians are significantly different (H1 -alternative hypotheses) or not
(H0 -null hypotheses).
In addition, we assessed whether the combination of bands and the density of lidar data
significantly affected the discrimination of the pairs of classes obtained (sample size =

5678

S. Bujn et al.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

44). The non-parametric Friedman test (Hollander and Wolfe 2000) was used to compare repeated observations on the same subjects. The Friedman test is an alternative to
the repeated-measures analysis of variance that, like many non-parametric tests, uses the
ranks of the data rather than their raw values to calculate the statistic.
3.4. Classification example
Finally, in order to transfer the obtained results to quantitative concepts that are more
familiar in the literature (e.g. overall accuracy), we performed a cross-validation using the
J48 algorithm included in the R package Rweka (Hornik, Buchta, and Zeileis 2009),
which was recently used by Corbelle-Rico and Crecente-Maseda (2012). Based on the
J48 algorithm and on the 480 training samples used to compute separability indices, a
decision tree was generated and cross-validation was performed.
4. Results and discussion
4.1. Normalized intensity
Figure 3 shows part of the intensity image generated from raw data, the image obtained
after normalization, and the image obtained after the application of a median filter for
data acquired at a density of 4 pulses m2 . The salt and pepper effect is still present after
normalization in tree or built-up areas. However, this effect is clearly reduced after the
application of the median filter.
The maximum, minimum, mean, and standard deviation for each intensity image have
been graphically represented. As shown in Figure 4, maximum values increase far more
than minimum values after normalization because intensity normalization by range has a
greater effect on the highest intensity values. Conversely, in asphalt areas, in which intensity
values are very low, normalization causes only a slight improvement, which is in agreement
with the results reported by Jutzi and Gross (2010).
When a median filter is applied to the normalized image, the maximum value decreases
with respect to the maximum for the normalized image from 264.8 to 133.6 (for the image

(a)

(b)

(c)

Figure 3. Process for intensity normalization. (a) Intensity image, (b) normalized intensity image,
(c) smoothed normalized intensity (SnI) image.

International Journal of Remote Sensing


2 Pulses m2

Min

Max

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

Mean St.Dev
nI

(a)

SnI

1 Pulse m2

Min

Max
I

Mean St.Dev
nI

(b)

SnI

0.5 Pulses m2

120
100
80
60
40
20
0

Intensity

140
120
100
80
60
40
20
0

Intensity

Intensity

Intensity

4 Pulses m2
300
250
200
150
100
50
0

5679

Min

Max
I

Mean St.Dev
nI

(c)

SnI

120
100
80
60
40
20
0
Min

Max
I

Mean St.Dev
nI

SnI

(d)

Figure 4. Graphical representation of the magnitude of a number of metrics for different types of
intensity images.

generated at 4 pulses/m2 ). On the contrary, mean and standard deviation remain practically constant, regardless of the point density at which the intensity image was generated
(aproximately 40 and 20, respectively). This effect becomes less apparent with the decrease
in point density. Mean and standard deviation values remain constant despite variations in
maximum values because the median filter mitigates the noise present in the image, which,
in this case, corresponds to the misleadingly high values.
The intensity normalization increases the mean value of intensities (see Section 3.3).
Such a variation has a direct effect on the calculation of the separability measures that
consider this information layer. However, the benefits of the mitigation of the salt and
pepper effect are expected to be much greater.

4.2. Separability assessment


Separability among the eight classes of interest was statistically assessed using transformed
divergence and the JM distance methods. Figures 56 summarize the results. Both figures
are divided into 11 blocks that correspond to the 11 possible combinations of the four bands
available. Each block is divided into four parts according to point density: the left upper
corner shows the results of TD (see Figure 5) or JM distance (see Figure 6) for a density
of 4 pulses m2 ; the left bottom corner shows the results for a density of 2 pulses m2 ;
the right upper corner shows the results for a density of 1 pulse m2 ; and the right bottom
corner shows the results for a density of 0.5 pulse m2 .
According to the results, the JM distance is more robust against variations in the mean
vector, i.e. the mean matrix used in the calculation of separability measures is very sensitive
to changes. Such a sensitivity has a greater effect on divergence than on the Bhattacharyya
distance (see Equations (1) and (4), and consequently, the effect of sensitivity is greater on
TD than on the JM distance.
The JM distance is a more reliable estimator of classification performance when anomalous values cause variations in the mean vector. For example, using TD and bands I and
DnM (see Figure 5) allows for the discrimination of classes bare earth versus building
and building versus pavement with a density of 0.5 pulses m2 , but not with higher densities, which does not seem a logical result. In contrast, using the JM distance (see Figure 6)
and the same band combination, classes bare earth versus building and building versus
pavement cannot be discriminated at all. These results could be explained as follows: erroneous values affect the mean vector by increasing distance between classes, which causes
considerable increases in divergence (Richards and Jia 2006) and, consequently, on TD.
Such increases result in an overestimation of classification performance.
To quantitatively strengthen these results and to assess both the effect of lidar point density and band combination, we used the KruskalWallis test. Results suggest a significant

nM and DnM
FL and DnM
nM and FL and DnM

nM and FL
I and DnM
nM and I and DnM
nM and I and FL and DnM

Be B Ha H L D Pv Y
Be

Band combination

nM and I
I and FL
nM and I and FL

S. Bujn et al.

I and FL and DnM

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5680

4 pulses m2

1 pulses m2

Ha
H
L
D
Pv

2 pulses m2

0.5 pulses m2

Information layers
nM
I
FL
DnM

Normalized Digital Surface Model


Intensity image
Difference between first and last pulse
Difference Digital Surface Models

(2004)
(2004)
(2004)
(20102004)

Level of separability
Excellent

Figure 5.

Poor

Very poor

Transformed divergence
Class
Be
B
Ha
H

Bare earth
Building
Harvested forest
High vegetation

L
D
Pv
Y

Low vegetation
Dirt road
Pavement
Young forest

Transformed divergence separability for training classes.

effect of band combination on the classification process, for both TD (p-value = 2.15
109 ) and JM distance (p-value = 2.20 1016 ). Conversely, lidar point density does not
significantly affect the classification (p-value TD = 0.92 and p-value JM = 0.80).
Intuitively, one could think that increasing the number of information layers used in the
classification would improve classification results, but this is not always the case. However,
some band combinations provide better results than others. Based on the results for the JM
distance, it can be stated that

nMI versus nMFL versus nMDnM. The comparison of these combinations reveals
that the DnM band is essential in the discrimination of bare earth and high
vegetation from young trees and harvested forest, respectively. In 2004, the pairs
of classes bare earthyoung trees and high vegetationharvested forest had the
same characteristics (same difference between returns, same height over the terrain,
and same intensity). However, in 2010, the heights of the classes young trees and
harvested forest were reversed, i.e. the value of height for young trees increased
from approximately zero in 2004 to more than 4 m in 2010, whereas the value of
height for harvested forest decreased from above 10 m in 2004 to approximately
zero in 2010. This information is summarized in the DnM band, in which negative
values correspond to harvested forest and positive values correspond to young trees.

nM and DnM
FL and DnM
nM and FL and DnM

nM and FL
I and DnM
nM and I and DnM
nM and I and FL and DnM

5681

Be B Ha H L D Pv Y
Be

Band combination

nM and I
I and FL
nM and I and FL
I and FL and DnM

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

4 pulses m2

1 pulses m2

Ha
H
L
D
Pv

2 pulses m2

0.5 pulses m2

Information layers
nM
I
FL
DnM

Normalized Digital Surface Model


Intensity image
Difference between first and last pulse
Difference Digital Surface Models

(2004)
(2004)
(2004)
(20102004)

Level of separability
Excellent

Figure 6.

Poor

Very poor

Transformed divergence
Class
Be
B
Ha
H

Bare earth
Building
Harvested forest
High vegetation

L
D
Pv
Y

Low vegetation
Dirt road
Pavement
Young forest

JeffriesMatusita distance separability for training classes.

IFLDnM. Removing normalized heights (nM) from combination nM-I-FL-DnM


reveals that the nM band is essential to discriminate between tall objects and the
classes that are at ground level, particularly to separate buildings from bare earth,
pavement, and dirt road. If height values are not used, separating these classes
becomes extremely difficult because they all have very similar intensities and their
pulse difference and height variation over time tend to be 0.
nMIDnM. This band combination produces very good results using only three
information layers. For information, layers generated from the 4 pulses m2 data set,
the value of JM is 1.96 (see Figure 7). Contrary to what was expected, intensity was
useless in separating bare earth from pavement and low vegetation with densities
equal to or below 2 pulses m2 . However, these classes could be differentiated when
using a density of 4 pulses m2 (bare earth vs. pavement JM = 1.83 and bare
earth vs. low vegetation JM = 1.88). Low vegetation and pavement could only
be separated using a density of 4 pulses m2 (JM = 1.92). These results are in agreement with the results reported by Song et al. (2002), who concluded that intensity
values would be very useful for classifying grass versus asphalt road (TD = 1.99;
TD = 2.00 in our case). Yet, they did not assess the separability of grass versus
bare earth.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

S. Bujn et al.

1.9

1.9

1.8

1.8

Mean of JM

Mean of TD

5682

1.7
1.6

1.7
1.6

1.5

1.5

1.4

1.4

1.3

1.3
4

0,5

Density (pulses m )
nM-II
I-DnM
nM-FL-DnM

nM-DnM
nM-I-FL
nM-I-FL-DnM

nM-FL
FL-DnM
I-FL-DnM

(a)

Figure 7.

0,5

2
Density (pulses m )

I-FL
nM-I-DnM

nM-I
I-DnM
nM-FL-DnM

nM-FL
FL-DnM
I-FL-DnM

nM-DnM
I-FL
nM-I-FL
nM-I-DnM
nM-I-FL-DnM

(b)

Comparison of bands and densities. (a) Transformed divergence, (b) JeffriesMatusita.

nMFLDnM. The comparison of this combination with nMIDnM reveals that


when intensity is replaced by difference between returns, low vegetation is no longer
separable from dirt road and pavement. In addition, most average values of separability measures decrease down to values below 1.70, as compared to values above
1.80 for combination nM-I-FL (see Figure 7).
nMIFLDnM. As shown in Figure 7, adding the feature difference between
returns to combination nM-I-FL would not contribute to improving classification
performance insofar as the average values of separability measures do not improve
significantly when difference between returns is used. For instance, when the four
information layers generated from the 2 pulses m2 data set were used, the average
JM distance reached a value of 1.92, whereas the average JM distance reached a value
of 1.91 when difference between returns was not used.
Figure 7 allows for the comparison of average separability values according to lidar
point density and combination of bands. As shown in this figure, overall classification performance slightly decreases with the decrease in lidar point density, which is in agreement
with the results of the KruskalWallis test, according to which the decrease in density is
not significant in all cases (p-value JM = 0.80). The largest variations would occur when
using layers nM and DnM with densities of 4 pulses m2 (JM = 1.73) and 1 pulse m2 (JM
= 1.59). In addition, Figure 7 shows that using the difference between returns would not
improve the performance of the classification significantly, as suggested by the analysis of
pairs of band combinations such as IDnM and IFLDnM, nMDnM and nMFLDnM,
and nMIDnM and nMIFLDnM. These pairs are plotted as almost parallel lines that
show a slight difference. Such a difference could be considered as the contribution of the
difference between returns to classification accuracy. Such a difference was slightly greater
for a density of 0.5 pulses m2 .
If we consider only the classes that can be discriminated using the four available bands
(see Figures 5 and 6), we can conclude that using densities equal to or below 2 pulses m2
decreases the possibility of discriminating bare earth, pavement, and low vegetation.
In the other hand, dirt road cannot be discriminated from bare earth using a density
of 4 pulses m2 because both classes share the same characteristics: they are made of the
same material, they are located at the same height over the terrain, they do not generally
vary over time, and they have a zero difference between returns. Yet, within an object-based
approach, these classes could be discriminated using the shape of the objects.

International Journal of Remote Sensing


Table 3.

Friedman test (p-value)a .

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

Combination bands
BeB
BeHa
BeH
BeL
BeD
BePv
BeY
BHa
BH
BL
BD
BPv
BY
HaH
HaL
HaD
HaPv
HaY
HL
HD
HPv
HY
LD
LPv
LY
DPv
DY
PvY

5683

Bands|densityb JM

***
***
***
***
**
***
***
***
***
***
***
***
***
***
***
**
***
0.088

***
***
***
**
***
***
***
***
***
**

Density|bandsc JM

***
0.079
0.455
0.102

**
***
*
***
***
*
**
***
0.083

**
***
0.392

*
0.194
0.392
0.107
0.162
0.215

*
*
***
***
0.951
0.726

Notes: a 0.000 *** ; 0.001 ** ; 0.01 * ; 0.05 ; 0.1 (bold text indicates that the null hypothesis is rejected). b H0 ,
the parameter bands did not affect discrimination between classes. c H0 , the parameter density did not affect
discrimination between classes.

The Friedman test was used to quantitatively assess the effects of the various band combinations and lidar point densities on the discrimination of isolated pairs of classes. Table 3
summarizes the results of the test, which used the JM values for all combinations of data
layers and densities available for the pairs of classes considered. In none of the cases was
density the only variable that affected discrimination between classes. Conversely, bands
could be considered as the only parameter affecting the possibility of discriminating some
of the classes, particularly high vegetation. In addition, both variables affected discrimination in most cases and only for the pair of classes harvested forestyoung trees did none
of the factors contribute to discrimination, which could be explained by the fact that both
classes show completely opposed values in all the available information layers.
4.3. Classification example
Figure 8(a) shows the results of the 25 iterations calculated from cross-validation for overall
accuracy (grey-shaded bars) and cumulative mean accuracy (purple-shaded lines) for the
four lidar point densities available.
Based on the results of the separability indices and considering dirt road as one of the
most problematic categories, we checked the effects of the class dirt road on classification

5684

S. Bujn et al.

(a) 94.0%

Global precision (%)

90.0%
88.0%
86.0%
84.0%
82.0%

11
13
15
No. of iterations

17

19

21

23

25

G. P. 4 pulses (%)

G. P. 2 pulses (%)

G. P. 1 pulse (%)

G. P. 0.5 pulse (%)

Mean 4 pulses

Mean 2 pulses

Mean 1 pulse

Mean 0.5 pulse

(b) 98.0%
97.0%
Global precision (%)

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

92.0%

96.0%
95.0%
94.0%
93.0%
92.0%

11

13

15

17

19

21

23

25

No. of iterations
G. P. 4 pulses (%)

G. P. 2 pulses (%)

G. P. 1 pulse (%)

G. P. 0.5 pulse (%)

Mean 4 pulses

Mean 2 pulses

Mean 1 pulse

Mean 0.5 pulse

Figure 8. Results of the 25 cross-validation iterations using J48. (a) Accuracy of the classification
for every iteration of the algorithm J48, (b) overall accuracy of the classification for every iteration
of the algorithm J48 without considering the class dirt road.

accuracy. To that end, we recalculated overall accuracy and cumulative mean without considering dirt road. Figure 8(b) shows the results of cross-validation without considering
the class dirt road.
The comparison of both cross-validations reveals the influence of the class dirt road
on overall classification accuracy, with a slight increase in overall accuracy when dirt
road is neglected and lidar data density is 4 pulses m2 . Such an increase becomes more
marked for densities below 4 pulses m2 , with more than 10% difference for lidar data
with a density of 0.5 pulses m2 . The cumulative mean accuracy obtained when the eight
classes are discriminated is 84%, a value that increases up to 94.5% when dirt road is
disregarded.
Figures 8(a) and (b) show the role of lidar point density in the discrimination of the
eight classes. As shown in Figure 8(a), lidar point density plays a key role in terms of

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5685

accuracy insofar as mean accuracy values for densities below 4 pulses m2 are about 85%,
whereas accuracy values increase up to 93% for lidar data with a density of 4 pulses m2 .
As shown in Figure 8(b), differences become less pronounced when dirt road is not
considered.
It is very difficult to compare these results with previous studies, because only a reduced
group of authors have carried out classifications using lidar data alone. However, a similar study was presented by Song et al. (2002). They used the TD method to assess the
separability of intensity on three intensity grids to discriminate four classes (asphalt road,
grass, house roof, and tree). They concluded that intensity information was very useful
for discriminating asphalt road versus grass and asphalt road versus tree. Charaniya,
Manduehi, and Lodha (2004) discriminated four classes (trees, low vegetation, roads,
and roofs) using a supervised classification. They claimed that height (normalized height
and height variation) was a key feature in the classification of areas with tree plantations,
whereas using difference between returns sometimes caused a decrease in classification
accuracy. According to our results, this claim cannot be confirmed, but results suggest that
the contribution of difference between returns may not be significant.
Using an object-based approach, Brennan and Webster (2006) identified up to 10 distinct classes in an urban coastal area with four information layers generated from lidar
data, namely intensity, DSM, normalized height, and multiple return image. The overall
accuracy for the classification of the 10 classes was 94%, but accuracy increased up to 98%
when only 7 classes were identified (coniferous, deciduous, intertidal, low vegetation, roads, structures, and water). However, they found difficulties in discriminating
between trees and buldings, because high vegetation was sometimes very dense, which
led to very similar values of difference between returns for both classes. Finally, Yu et al.
(2004) identified up to 75% of harvested trees at single-tree level using multitemporal lidar
data in a forest area, but they could not identify trees where the radius of crown areas was
below 1 m. In addition, they observed that the number of harvested trees obtained from
direct field measurements did not match the number of trees obtained automatically and
attributed this lack of agreement to crown delineation errors.
Yet, a larger number of authors have combined lidar with other data sets (Charaniya,
Manduchi, and Lodha 2004; Tymkow and Borkowski 2008; Chen et al. 2009; Dinis et al.
2010). Auxiliary data such as aerial photographs or satellite imagery can contribute greatly
to the discrimination of vegetation from other land-cover classes (Chen et al. 2009); and
it becomes essential for the identification of different types of crops or forest species
(Vias et al. 2006). Specifically, vegetation indices can be used to discriminate vegetation from other land-cover classes, whereas the temporal resolution of satellite imagery can
be used to discriminate deciduous species from evergreen species. In this sense, for the
eight land-cover classes considered (without identifying crop types or forest species) and
the available data (with homogeneous densities, insofar as there are no areas without data),
including auxiliary data would not greatly affect the results. Yet, the results for discrimination between the classes bare-earth, pavement, low vegetation, and dirt road could
possibly be improved by using some vegetation index.

5. Conclusions
This article has experimentally assessed the possibility of classifying a complex rural landscape using low-density lidar data. Two indicators of separability, transformed divergence
and the JM distance, were used to assess the separability of eight predefined classes.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5686

S. Bujn et al.

Additionally, the effects of lidar point density and combination of bands on separability
was assessed by using the KruskalWallis and the Friedman tests. Finally, cross-validation
was performed to check the accuracy of the classification method.
According to results, it can be concluded that bare earth cannot be fully discriminated from pavement, low vegetation, and dirt road using lidar point densities below
4 pulses m2 . However, based on the JM mean values, the overall performance of the classification is not significantly affected. The same conclusion can be drawn from the analysis
of the mean overall classification accuracy, which amounted to 85% for densities below
4 pulses m2 and 93% for the highest possible density.
The results for average separabilities and the number of classes that could be identified
have revealed that the best band combination is the combination that uses the nDSM, the
intensity image, and the difference between the nDSMs for 2010 and 2004. Nevertheless,
the classes dirt road and bare earth cannot be discriminated because both classes are
made of the same material. Yet, based on the results reported by other authors, both classes
could be discriminated by using an object-based classification that incorporates the shape
feature.
The use of low-density lidar data involves a decrease in the quality of information
layers, which leads to variations in the average values of sampled data. Yet, the Kruskal
Wallis test suggests that density does not significantly affect the classification output. This
claim is supported by the classification accuracy results obtained from cross-validation,
particularly the results obtained when dirt road was disregarded.
When lidar point densities below 4 pulses/m2 were used, the intensity image lost quality. Such a loss of quality decreased the possibility of discriminating among bare earth,
pavement, dirt road, and low vegetation. However, the quality of nDSMs remained
almost unaffected.
Finally, the reduction of lidar point density has only slightly affected the results (particularly when dirt road was not considered). The results obtained from lidar data with a
density of 0.5 pulse/m2 have been more satisfactory than expected, and open the door to the
use of these techniques in Spain, where lidar data with a density of 0.51 pulses m2 will
soon become available for 90% of most Spanish territory. Yet, by increasing lidar data density up to 2 pulses m2 , the accuracy of results would compensate for additional acquisition
costs and would extend the range of possible applications of the acquired data sets.

References
Axelsson, P. 2000. DEM Generation from Laser Scanner Data Using Adaptive TIN Models.
International Archives of Photogrammetry and Remote Sensing 33: 111118.
Barreiro-Fernndez, L., J. Garca-Gutirrez, E. Gonzlez-Ferreiro, S. Bujn, L. Gonalves-Seco, and
D. Miranda. 2010. Land Cover Classification of Forest Areas using LiDAR and Spectral Data.
In Proceedings of the 4th International Conference on the Operational Tools in Forestry using
Remote Sensing Techniques, edited by D. Miranda, J. Suarez, and R. Crecente, September 710,
Lugo, 159162.
Brennan, R., and T. Webster. 2006. Object-Oriented Land Cover Classification of LiDAR-Derived
Surfaces. Canadian Journal of Remote Sensing 32: 162172.
Cannavacciuolo, L., W. Emery, G. Moser, and S. Serpico. 2007. A Contextual Change Detection
Method for High-Resolution Optical Images of Urban Areas. In Proceedings of the Urban
Remote Sensing Joint Event, edited by Institute of Electrical and Electronics Engineers, April
1113, Paris, 17.
Champion, N., Rottensteiner F., L. Matikainen, X. Liang, J. Hyypp, and B. Olsen, 2009. A Test of
Automatic Building Change Detection Approaches. International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences 38: 145150.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5687

Charaniya, A., R. Manduchi, and S. Lodha. 2004. Supervised Parametric Classification of Aerial
LiDAR Data. In Proceedings of the Conference on Computer Vision and Pattern Recognition
Workshop, June 27July 2, Baltimore, 18. Washington, DC: IEEE Computer Society.
Chen, Y., W. Su, J. Li, and Z. Sun. 2009. Hierarchical Object Oriented Classification Using Very
High Resolution Imagery and LIDAR Data over Urban Areas. Advances in Space Research 43:
11011110.
Chih-Cheng, H., A. Fahsi, W. Tadesse, and T. Coleman. 1997. A Comparative Study of
Remotely Sensed Data Classification Using Principal Components Analysis and Divergence.
In Proceedings of the IEEE International Conference on Systems, Man, and Cybernetics, edited
by Institute of Electrical and Electronics Engineers, October 1215, Orlando, FL, 24442449.
Chuvieco, E. 2008. Teledeteccin ambiental. La observacin de la Tierra desde el espacio. 3rd, Ariel
Ciencia. Barcelona: Ariel.
Corbelle-Rico, E., and R. Crecente-Maseda. 2012. An Object-Oriented Approach to Automatic
Classification of Panchromatic Aerial Photographs with GRASS GIS and R. In Geospatial Free
and Open Source Software in the 21st Century, edited by E. Bocher, M. Neteler, W. Cartwright,
G. Gartner, L. Meng, and M. Peterson, 123137. Berlin: Springer.
Davis, C., and X. Wang. 2002. Urban Land Cover Classification from High Resolution MultiSpectral IKONOS Imagery. In Proceedings of the IEEE International Geoscience and Remote
Sensing Symposium, edited by Institute of Electrical and Electronics Engineers, June 2428,
Toronto, ON, 12041206.
Dengsheng, L., S. Hetrick, and E. Moran. 2010. Land Cover Classification in a Complex Urban
Rural Landscape with QuickBird Imagery. Photogrammetric Engineering & Remote Sensing
76: 11591168.
Dinis, J., A. Navarro, F. Soares, T. Santos, S. Freire, A. Fonseca, N. Afonso, and J. Tenedrio. 2010.
Hierarchical Object-based Classification of Dense Urban Areas by Integrating High Spatial
Resolution Satellite Images and LiDAR Elevation Data. In Proceedings of the International
Conference GEOBIA 2010, June 29July 2, Ghent, unpaginated CDROM.
Donoghue, D., P. Watt, N. Cox, and J. Wilson. 2007. Remote Sensing of Species Mixtures in
Conifer Plantations using LiDAR Height and Intensity Data. Remote Sensing of Environment
110, 509522.
Gambarova, Y., A. Gambarov, R. Rustamov, and M. Zeynalova. 2010. Remote Sensing and GIS
as an Advance Space Technologies for Rare Vegetation Monitoring in Gobustan State National
Park, Azerbaijan. Journal of Geographic Information System 2: 9399.
Garca, M., D. Riao, E. Chuvieco, and F. Danson. 2010. Estimating biomass carbon stocks for a
Mediterranean forest in central Spain using LiDAR height and intensity data. Remote Sensing
of Environment 114: 816830.
Gibson, P., and C. H. Power. 2000. Introductory Remote Sensing: Digital Image Processing and
Applications. 1st ed. London: Routledge.
Gonalves, G., L. Seco, F. Reyes, and D. Miranda. 2007. Land Cover Classification of Rural Areas
using LiDAR Data: A Comparative Study in the Context of Fire Risk. In Proceedings of the
SilviLaser 2008: 8th International Conference on LiDAR Applications in Forest Assessment and
Inventory, edited by R. Hil, J. Rosette, and J. Surez, September 1719, Edinburgh, 427436.
Gonzlez-Ferreiro, E., U. Diguez-Aranda, and D. Miranda. 2012. Estimation of Stand Variables
in Pinus Radiata D. Don Plantations Using Different LiDAR Pulse Densities. Forestry 85:
281292.
Gross, H., B. Jutzi, and U. Thoennessen. 2008. Intensity Normalization by Incidence Angle and
Range of Full-Waveform LiDAR Data. In Proceedings of the Commission IV: Internatinal
Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, edited by
T. H. Kolbe, G. Knig, and C. Nagel, July 311, Beijing, 405412.
Gu, J., J. Chen, Q. Zhou, H. Zhang, and L. Ma. 2008. Quantitative Textural Parameter Selection
for Residential Extraction from High-Resolution Remotely Sensed Imagery. In Proceedings of
the Commission IV: Internatinal Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences, edited by T. H. Kolbe, G. Knig, and C. Nagel, July 311, Beijing,
13711376.
Hasegawa, H. 2006. Evaluations of LiDAR Reflectance Amplitude Sensitivity towards Land Cover
Conditions. Bulletin of the Geogrephical Survey Institute 53: 4350.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

5688

S. Bujn et al.

Hfle, B., and N. Pfeifer. 2007. Correction of Laser Scanning Intensity Data: Data and ModelDriven Approaches. International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences 62: 415433.
Hollander, M., and D. A. Wolfe. 2000. Applied Nonparametric Statistical Methods. 3rd ed. Boca
Raton, FL: Chapman and Hall/CRC.
Holmgren, J., M. Nilsson, and H. Olsson. 2003. Estimation of Tree Height and Stem Volume on
Plots Using Airborne Laser Scanning. Forest Science 49: 419428.
Holmgren, J., and . Persson. 2004. Identifying Species of Individual Trees Using Airborne Laser
Scanner. Remote Sensing of Environment 90: 415423.
Hopkinson, C., and L. Chasmer. 2009. Testing LiDAR Models of Fractional Cover Across Multiple
Forest Ecozones. Remote Sensing of Environment 113: 275288.
Hornik, K., C. Buchta, and A. Zeileis. 2009. Open-Source Machine Learning: R Meets Weka.
Computational Statistics 24: 225232.
Im, J., J. Jensen, and J. Tullis. 2008. Object-based Change Detection Using Correlation Image
Analysis and Image Segmentation. International Journal of Remote Sensing 29: 399423.
Jutzi, B., and H. Gross. 2010. Investigations on Surface Reflection Models for Intensity
Normalization in Airborne Laser Scanning (ALS) Data. Photogrammetric Engineering &
Remote Sensing 76: 10511060.
Kavzoglu, T., and P. M. Mather. 2000. The Use of Feature Selection Techniques in the Context
of Artificial Neural Networks. In Proceedings of the 26th Annual Conference of the Remote
Sensing Society, edited by A. Millington and P. Monks, September 1214, Leicester, unpaginated
CDROM.
Kruskal, W. 1952. A Nonparametric Test for the Several Sample Problem. The Annals of
Mathematical Statistics 23: 525540.
Lu, D., M. Batistella, and E. Moran. 2004. Multitemporal Spectral Mixture Analysis for Amazonian
Land-Cover Change Detection. Canadian Journal of Remote Sensing 30: 87100.
Lu, D., and Q. Weng. 2009. Extraction of Urban Impervious Surfaces from an IKONOS Image.
International Journal of Remote Sensing 30: 12971311.
Luzum, B., M. Starek, and K. Slatton. 2004. Normalizing Lidar Intensities. Accessed October 6,
2011. www.aspl.ece.ufl.edu/reports/GEM_Rep_2006_12_001.pdf.
Matinfar, H., F. Sarmadian, S. A. Panah, and R. Heck. 2007. Comparisons of Object-Oriented
and Pixel-Based Classification of Land Use/Land Cover Types based on Lansadsat7, Etm+
Spectral Bands (Case Study: Arid Region of Iran). American-Eurasian Journal of Agricultural
& Environmental Sciences 2: 448456.
Mazzarini, F., M. Pareschi, M. Favalli, I. Isola, S. Tarquini, and E. Boschi. 2007. Lava Flow
Identification and Ageing by means of LiDAR Intensity: Mount Etna Case. Journal of
Geophysical Research-Solid Earth 112: 143.
McRoberts, R., and E. Tomppo. 2007. Remote Sensing Support for National Forest Inventories.
Remote Sensing of Environment 110: 412419.
Pacifici, F., F. Del Frate, C. Solimini, and W. Emery. 2007. An Innovative Neural-Net Method to
Detect Temporal Changes in High-Resolution Optical Satellite Imagery. IEEE Transactions on
Geoscience and Remote Sensing 45: 29402952.
Pu, R., S. Landry, and Q. Yu. 2011. Object-Based Urban Detailed Land Cover Classification
with High Spatial Resolution IKONOS Imagery. International Journal of Remote Sensing 32:
32853308.
R Core Team. 2012. R: A Language and Environment for Statistical Computing. Vienna: R
Foundation for Statistical Computing. ISBN 3-900051-07-02012.
Richards, J., and X. Jia. 2006. Remote Sensing Digital Image Analysis: An Introduction. 4th ed.
Berlin: Springer.
Schowengerdt, R. 2007. Remote Sensing: Models and Methods for Image Processing. 2nd ed. San
Diego, CA: Academic Press.
Shapiro, S. S., and M. B. Wilk. 1965. An Analysis of Variance Test for Normality (Complete
Samples). Biometrika 52: 591611.
Shapiro, S. S., and M. B. Wilk. 1968. Approximations for the Null Distribution of the W Statistic.
Technometrics 10: 861866.
Singh, A. 1989. Review Article Digital Change Detection Techniques using Remotely-Sensed
Data. International Journal of Remote Sensing 10: 9891003.

Downloaded by [University of Santiago de Compostela], [Eduardo Gonzlez-Ferreiro] at 11:18 14 May 2013

International Journal of Remote Sensing

5689

Singh, K. K., J. B. Vogler, and R. K. Meentemeyer. 2010. Estimation of Land-Use in an Urbanized


Landscape using LiDAR Intensity Data: A Regional Scale Approach. In Proceedings of
the Geospatial Data and Geovisualization: Environment, Security, and Society. Special Joint
Symposium of ISPRS Technical Commission IV and AutoCarto 2010, November 1519, unpaginated CDROM. Orlando, FL: American Society for Photogrammetry and Remote Sensing
(ASPRS); International Society for Photogrammetry and Remote Sensing (ISPRS); Cartography
and Geographic Information Society (CaGIS).
Sithole, G., and G. Vosselman. 2004. Experimental Comparison of Filter Algorithms for BareEarth Extraction from Airborne Laser Scanning Point Clouds. International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences 59: 85101.
Solberg, S., R. Hill, J. Rosette, and J. Surez. 2008. Comparing Discrete Echoes Counts and Intensity
Sums from ALS for Estimating Forest LAI and Gap Fraction. In Proceedings of the SilviLaser
2008: 8th International Conference on LiDAR Applications in Forest Assessment and Inventory,
September 1719, 247256. Edinburgh: SilviLaser 2008 Organizing Committee.
Song, J., S. Han, K. Yuand, and Y. Kim. 2002. Assessing the Possibility of Land-Cover Classification
using LiDAR Intensity Data. International Archives of Photogrammetry Remote Sensing and
Spatial Information Sciences 34: 259262.
Tolpekin, V., and A. Stein. 2009. Quantification of the Effects of Land-Cover-Class Spectral
Separability on the Accuracy of Markov-Random-Field-Based Superresolution Mapping. IEEE
Transactions on Geoscience and Remote Sensing 47: 32833297.
Tymkow, P., and A. Borkowski. 2008. Land Cover Classfication using Airborne Laser Scanning Data
and Photographs. International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences 37: 185190.
Van der Sande, C., S. De Jong, and A. De Roo. 2003. A Segmentation and Classification
Approach of IKONOS-2 Imagery for Land Cover Mapping to Assist Flood Risk and Flood
Damage Assessment. International Journal of Applied Earth Observation and Geoinformation
4: 217229.
Vias, O., A. Ruiz, R. Xandri, V. Pala, and R. Arbiol. 2006. Combined Use of Lidar and Quickbird
Data for the Generation of Land Use Maps. In Proceedings of the ISPRS Commission VII
Symposium Remote Sensing: From Pixels to Processes, edited by J. Kerle and A. Skidmore,
38 (Part 7), May 810, Enschede, 812. International Society of Photogrammetry and Remote
Sensing.
Vosselman, G. 2002. On the Estimation of Planimetric Offsets in Laser Altimetry Data.
International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences 34:
375380.
Wang, C., and N. Glenn. 2009. Integrating LiDAR Intensity and Elevation Data for Terrain
Characterization in Aforested Area. IEEE Geoscience and Remote Sensing Letters 6: 463466.
Wulder, M., S. Ortlepp, J. White, and S. Maxwell. 2008. Evaluation of Landsat-7 SLC-off Image
Products for Forest Change Detection. Canadian Journal of Remote Sensing 34: 9399.
Yu, X., J. Hyypp, H. Kaartinen, and M. Maltamo. 2004. Automatic Detection of Harvested
Trees and Determination of Forest Growth Using Airborne Laser Scanning. Remote Sensing
of Environment 90: 451462.
Yu, X., J. Hyypp, A. Kukko, M. Maltamo, and H. Kaartinen. 2006. Change Detection
Techniques for Canopy Height Growth Measurements Using Airborne Laser Scanner Data.
Photogrammetric Engineering & Remote Sensing 72: 13391348.
Zhou, W., A. Troy, and M. Grove. 2008. Object-Based Land Cover Classification and Change
Analysis in the Baltimore Metropolitan Area Using Multitemporal High Resolution Remote
Sensing Data. Sensors 8: 16131636.

Você também pode gostar