Sie sind auf Seite 1von 4

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.


IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 6, NO. 3, JULY 2009

463

Integrating LiDAR Intensity and Elevation Data for


Terrain Characterization in a Forested Area
Cheng Wang and Nancy F. Glenn

AbstractSeparating ground from nonground laser returns


from airborne light detection and ranging (LiDAR) data is a
key step in creating digital terrain models (DTMs). In this letter,
bare-earth and forested surfaces are classified from LiDAR intensity data in a data set from central Idaho, U.S. Next, a Gaussian
fitting (GF) method is applied to determine ground elevations from
LiDAR elevation data according to the land-cover information.
In comparison to ground-based reference data, the GF method
generated an accurate DTM in this study area. Overall, the DTM
underestimated the ground observations by approximately 31 cm.
A combination of LiDAR intensity and elevation data may be
effectively used to develop DTMs in similar terrain of relatively
simple land-cover classes.
Index TermsDigital terrain model (DTM), forested mountain
area, Gaussian fitting (GF) model, light detection and ranging
(LiDAR).

utilization of spectral and spatial information simultaneously


provided by LiDAR may be advantageous over utilizing either
data set alone.
Building upon previous research on LiDAR elevation and
intensity data, this letter develops a new methodology to identify ground-reflected returns in a study area in central Idaho,
U.S. In comparison to the study by Raber et al. [12] who used
LiDAR elevation data to classify land-cover types in a vegetated
area, in this letter, bare-earth and forested surfaces are classified
from LiDAR intensity data. The bare ground is characterized
with a single Gaussian distribution of LiDAR elevations, and
the forested area is characterized by a bimodal ground and
vegetation distribution. The mean value of the ground Gaussian
distribution is assigned to the ground elevation for a specified
local area.

I. I NTRODUCTION

EPARATING ground from nonground laser returns from


airborne light detection and ranging (LiDAR) data is the
first and most critical step in creating a bare-earth digital
terrain model (DTM). Many algorithms have been developed
for identification of ground returns in LiDAR elevation data
[1][8]. Several studies [2][8] iteratively identify the ground
returns originating from the lowest elevations. Selecting the
lowest elevation returns in areas with steep slopes and high
roughness may introduce error, which may be translated into
the derived DTM. Depending on different LiDAR scanning
densities, typical errors in the creation of LiDAR-based DTMs
range from tens of centimeters to several meters (in the case of
outliers) in forested areas with different terrain slopes [2], [4],
[9], [10].
LiDAR systems provide both elevation and intensity records
for each laser return. LiDAR intensity, in theory, is determined
by an objects reflectance, which can be used to identify
land-cover classes when the data are carefully calibrated [11].
Because LiDAR intensity and elevation data are simultaneously
recorded on the same platform, they are coregistered and geometrically represent the same features. Regardless of the difficulties with calibration of the intensity data, the comprehensive
Manuscript received October 14, 2008; revised January 8, 2009 and
February 13, 2009. This work was supported by the National Oceanic and
Atmospheric Administration under Grant NA06OAR4600124.
The authors are with Boise Center Aerospace Laboratory, Idaho State
University, Boise, ID 83702 USA (e-mail: wangchengcn@hotmail.com;
glennanc@isu.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LGRS.2009.2016986

II. D ATA
The study area is a 1-km2 area near Redfish Lake in central
Idaho, U.S. The elevation ranges between 1990 and 2180 m
above sea level, and terrain slopes are as high as 45 . This area
is dominated by three land-cover types: tall evergreen conifer
trees (height larger than 10 m), bare soil, and pavement. There
is also a sparse amount (less than approximately 10% canopy
cover) of low-height shrub (less than 1-m height) beneath the
forest canopy.
The LiDAR data were collected on October 8, 2005, using
an Optech 50-kHz scanning LiDAR system. Both first and
last LiDAR returns were recorded, along with intensity. The
absolute vertical accuracy of the LiDAR data is approximately
10 cm (95% confidence level), as measured by the vendor using
a ground survey of 445 GPS points. The data were collected
from an altitude of approximately 700 m, resulting in a footprint
diameter of approximately 20 cm for each laser pulse and an
average laser scanning density of 8 points/m2 (including both
returns).
Ground reference observations were collected on September
1819, 2007, and include the following: 1) spectroradiometer
reflectance measurements of soil, pavement, and vegetation
(conifer trees and shrub) (ten samples each) and 2) 182 points
of land cover (composing of soil, pavement, shrub, and conifer
trees) and spatial coordinates (xyz) measured by a differentially corrected GPS. The accuracy of the differential GPS was
better than 1 cm in all directions. Although there was a twoyear time delay between LiDAR and field data collection, we
assume negligible changes in terrain elevation and vegetation
heights because the study site is located in a rural environment

1545-598X/$25.00 2009 IEEE

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 30, 2009 at 16:45 from IEEE Xplore. Restrictions apply.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
464

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 6, NO. 3, JULY 2009

Fig. 1. Frequency distributions and GF applications on (a) bare-ground area and (b) vegetated surface using a 4 m 4 m analysis unit.

and the vegetation mainly consists of slow-growing evergreen


trees.
III. M ETHODS
A. LiDAR Intensity Normalization and Segmentation
In theory, LiDAR intensity values depend on surface reflectance, atmospheric transmission, local incidence angle, and
sensor-to-object distance [13], [14]. Kaasalainen et al. [15]
illustrate that LiDAR intensity has a low dependence on the
local incidence angle. Atmospheric absorption plays a minor
role in airborne near-infrared LiDAR intensity data because
of the wavelength ( = 1064 nm) [13]. Adequate use of the
intensity data is also hampered by the calibration that is often neglected during data collection [11], [15]. Owing to the
negligible effects from local incidence angle and atmospheric
conditions and the lack of calibration, the intensity data can be
normalized with respect to the sensor-to-object distance (path
length). The intensity level changes with the inverse square of
the distance [13], [16].

[see (2)]. Shrubs are not considered here (see the discussion
hereinafter)


(zi z g )2
1
(1)
F = exp
2g2
g 2


1
(zi z g )2
F = exp
2g2
g 2


(zi z v )2
1
+ exp
(2)
2v2
v 2
where zi represents a LiDAR-received return and z g and g
denote the mean value and standard deviation of the groundreflected LiDAR elevations within the area, respectively. Since
our objective is to estimate the ground elevation, (1) and (2) are
simplified to
 
2 
(zi bg )
(3)
F = ag exp
cg
 
 
2 
2 
(zi bg )
(zi bv )
F = ag exp
+ av exp
cg
cv
(4)

B. GF of LiDAR Elevation Data


For discrete LiDAR data, due to the high number of laser
pulses penetrating the vegetation canopy, the returns within
a local area will represent both the ground and vegetation
canopy [17]. For a bare-earth area, LiDAR data have a Gaussian
distribution if there are enough samples in a window (unit)
size [Fig. 1(a)]. For a forested area, the frequency distribution of LiDAR elevations is considered to have a bimodal
Gaussian shape with two overlapping single-frequency distributions [Fig. 1(b)]. One frequency is the Gaussian distribution
from the ground-reflected returns, and the other is the nonstandard Gaussian distribution that relates to the obstruction
and occlusion of LiDAR pulses at different vertical levels
within the forest canopy [Fig. 1(b)]. The size of the analysis
unit (area) used for the frequency plot is determined in accordance with both laser sampling density and canopy cover. The
number of laser returns in each sampling unit should be high
enough to satisfy the statistical analysis [or to realize Gaussian
fitting (GF)].
For the method developed here, we assume that F = f (z)
is a frequency distribution of LiDAR elevations within a local
area (unit), and the elevations can be fit by two Gaussian
distributions to represent bare ground [see (1)] and forest

where bg is the ground elevation and ag and cg are related to the


standard deviation of the ground-characteristic Gaussian function and are representative of the ground slope. The parameters
av , bv , and cv refer to the forest canopy. Theoretically, the mean
value bv indicates the average elevation of forest canopy. Since
the aim of this letter was to determine the ground elevation,
the estimation accuracies of the canopy attributes were not
considered here.
A 4 m 4 m analysis unit was used with approximately
130 laser returns in this letter (8 points/m2 ). The GF algorithm
was applied to each unit, and the Gaussian parameters (mean
and standard deviation) were computed from the frequency distribution. Examples of GF on bare-ground and forested surfaces
are shown in Fig. 1. In addition to the GF, two limitations were
applied to the data processing to obtain reliable estimations:
1) Only units with a ground Gaussian distribution with standard
deviation <1 m were retained, which roughly corresponds to
the largest ground slope (45 ), and 2) only the units with
the difference of the two derived mean values (ground and
vegetation canopy) >1 m were retained in order to remove
units with complex vertical structure (e.g., groundshrub or
groundshrubforest canopy).

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 30, 2009 at 16:45 from IEEE Xplore. Restrictions apply.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
WANG AND GLENN: INTEGRATING LiDAR INTENSITY AND ELEVATION DATA FOR TERRAIN CHARACTERIZATION

465

Fig. 2. (a) Normalized LiDAR intensity image (1 km 1 km area). (b) Relationship between field-measured reflectance and normalized LiDAR intensity.
(c) Land-cover classification map of the study area (1 km 1 km area).

Fig. 3. (a) Elevation map of the identified ground returns. (b) Interpolated DTM of the study area. (c) Difference between the field GPS measurements and the
LiDAR derived elevations by the GF method (the dashed line is a 1 : 1 relationship).

IV. R ESULTS
A. LiDAR Intensity Classification Map
The original LiDAR intensities were normalized with a
700-m flight altitude to obtain a normalized intensity image [Fig. 2(a)]. A linear relationship was found between the
normalized LiDAR intensity and the field-measured spectral
reflectance at 1064 nm [Fig. 2(b)]. The mean intensity value
of vegetation and nonvegetation (digital number = 95) was
applied to the normalized intensity data to produce a binary classification map [Fig. 2(c)]. In this map, black pixels
(DN < 95) denote bare ground, while gray pixels (DN >=
95) denote vegetated surfaces. The vegetated pixels contained
either conifer trees or/and shrubs. One-hundred eighty-two field
points (127 vegetation and 55 nonvegetation) were used to
evaluate the classification result, and an 85% overall accuracy
was achieved for this simple-object classification. The classification information was then used to determine the type of
Gaussian model (single modal or bimodal) for ground elevation
identification within each analysis unit.

B. DTM
Fig. 3(a) is the map of identified ground elevations at the
study site. Dark pixels (16%) indicate failed GFs because of the
two limitations used earlier (see the latter part of Section III-B).
Thus, the vegetated pixels were then filtered to only include
those with canopy >1 m. Fig. 3(b) shows the derived DTM with
4-m spatial resolution that is produced by applying the inverse
distance-weighting interpolation algorithm [18] on the identified ground elevations. We calculated the DTM error between
the LiDAR-derived and field-measured ground elevations for

182 sample points. The statistical results illustrate that the


GF method underestimated the field elevations by an average
error of 30.7 cm with a standard deviation of 114.3 cm.
V. D ISCUSSION
Previous studies have developed methods to create DTMs in
forested areas with errors from tens of centimeters to several
meters. For example, Hodgson et al. [9] produced a DTM with
mean errors of 0.21 and 0.46 m on the pine areas with slopes of
0 2 and 2 4 , respectively. Reutebuch et al. [10] found that
a DTM in uncut conifer plantations had a 0.31-m mean error.
Cobby et al. [2] indicated a DTM error of 3.99 m in deciduous
forests on steeper slopes (10 15 ). Clark et al. [4] predicted
a DTM with a mean error of 0.97 m in a tropical landscape
with steep slopes. In comparison to [2] and [4], our results are
an improvement in accuracy. Our results are similar to [9] and
[10], specifically considering that areas with steep slopes are
filtered out in our method. These studies also indicate that the
terrain slope is an important factor to LiDAR-derived DTM
accuracy. However, the DTM error by the GF method is not
related to terrain slope. Terrain slope mainly changes the width
of the Gaussian distribution but does not change the mean value.
Fig. 4 shows the relationship between terrain slope and error
based on 37 field points along a range of slopes (from 0 to 45 ).
Although our results indicate that the GF method can generate
a relatively higher DTM accuracy in forested areas with less
influence from terrain slope, our study used a higher point
density (8 points/m2 ) in relation to other studies (7 points/m2
in [2], 4 points/m2 in [10], and less than 1 pt/m2 in [4] and [9]).
Although complex vertical structures (groundshrubtree)
are sparse at the study site, the two additional limitations
(see Section III-B) of the GF method can effectively remove

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 30, 2009 at 16:45 from IEEE Xplore. Restrictions apply.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
466

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 6, NO. 3, JULY 2009

for that area. While, theoretically, this solution can be realized,


the estimation accuracy will likely decrease for more complicated canopies. In addition, the LiDAR intensity classification
becomes more difficult with an increasing number of land-cover
types.
ACKNOWLEDGMENT
The authors would like to thank S. Ehiniger for the help
in collecting the field measurements and the reviewers of this
letter.
Fig. 4. (a) Contour line (in red) of the study site with the (white points) field
points. (b) Scatter plot between the terrain slope and DTM error using the field
points in (a).

these units. A complex vertical structure may reduce estimation


accuracy for any method to identify ground returns because
of its lower laser penetration rate and because of multiple
returns in the understory. For example, multiple returns from
the understory within trees can cause multimodal Gaussian
distributions. Statistical results illustrate that 16% of the study
area is not suitable for the GF method, but the remaining 84%
of the area is.
We explored the impact of the LiDAR instruments noise
on the GF-derived DTM. Fourteen GPS measurements were
collected on a 4 m 4 m area of flat pavement with less
than 10-cm difference (the differences are likely due to ground
roughness and small GPS errors) among the measurements.
LiDAR elevations of the same area have a range of approximately 40 cm and standard deviation equal to 7 cm. The latter
is the approximate error of the LiDAR instrument (10 cm with
95% confidence level), while ground roughness may attribute
an additional difference.
VI. C ONCLUSION AND F UTURE W ORK
In this letter, we utilized the LiDAR spatial and spectral
information to identify the ground elevations by a GF method
in a forested area. LiDAR intensity data are used to produce the binary classification of ground and vegetation with
85% overall classification accuracy. The GF algorithm is then
applied to the LiDAR elevation data, and single-modal or
bimodal Gaussian functions are fitted according to the landcover type. Validation of the LiDAR-derived DTM was carried
out with field measurements. Results indicate good agreement
(30.7-cm mean difference) between field measurements and
the GF-derived DTM. We also explored the impact of terrain
slope on the identification of ground laser returns.
The GF method is simple to implement in data analysis
tools. In addition, it is flexible and adaptable for different
terrain and canopy conditions. Although we only considered
simple land-cover conditions, GF can be extended to more
complicated land cover by increasing the number of Gaussian
distributions. One distribution would be allocated to each vertical layer reflected by the laser pulse. For example, to identify
the ground elevation in a forested area characterized by three
layers (groundshrubforest), trimodal GF would be applied to
the LiDAR elevation data, and the mean value of the ground
Gaussian distribution would be considered the ground elevation

R EFERENCES
[1] G. Vosselman and G. Sithole, Experimental comparison of filter algorithms for bare-earth extraction from airborne laser scanning point
clouds, ISPRS J. Photogramm. Remote Sens., vol. 59, no. 1/2, pp. 85
101, Aug. 2004.
[2] D. M. Cobby, D. C. Mason, and I. J. Davenport, Image processing
of airborne scanning laser altimetry data for improved river flood modelling, ISPRS J. Photogramm. Remote Sens., vol. 56, no. 2, pp. 121138,
Dec. 2001.
[3] X. Meng, L. Wang, J. L. Silvan-Cardenas, and N. Currit, A multidirectional ground filtering algorithm for airborne LiDAR, ISPRS J.
Photogramm. Remote Sens., vol. 64, no. 1, pp. 117124, Jan. 2009.
[4] M. L. Clark, D. B. Clark, and D. A. Roberts, Small-footprint LiDAR
estimation of sub-canopy elevation and tree height in a tropical rain forest
landscape, Remote Sens. Environ., vol. 91, no. 1, pp. 6889, May 2004.
[5] D. R. Streutker and N. F. Glenn, LiDAR measurement of sagebrush
steppe vegetation heights, Remote Sens. Environ., vol. 102, no. 1/2,
pp. 135145, May 2006.
[6] N. Pfeifer, T. Peiter, C. Briese, and W. Rieger, Interpolation of high
quality ground models from laser scanner data in forested areas, Int.
Arch. Photogramm. Remote Sens., vol. 31, pp. 383388, 1996.
[7] TerraScan for Microstation Users Guide, Terrasolid Limited, Houston,
TX, 1996. TerraScan.
[8] K. Zhang, S. Chen, D. Whitman, M. Shyu, J. Yan, and C. Zhang, A
progressive morphological filter for removing non-ground measurements
from airborne LiDAR data, IEEE Trans. Geosci. Remote Sens., vol. 41,
no. 4, pp. 872882, Apr. 2003.
[9] M. E. Hodgson, J. R. Jensen, L. Schmidt, S. Schill, and B. Davis, An
evaluation of LiDAR- and IFSAR-derived digital elevation models in leafon conditions with USGS Level 1 and Level 2 DEMs, Remote Sens.
Environ., vol. 84, no. 2, pp. 295308, Feb. 2003.
[10] S. E. Reutebuch, R. J. McGaughey, H. E. Anderson, and W. W. Carson,
Accuracy of a high-resolution LiDAR terrain model under a conifer
forest canopy, Can. J. Remote Sens., vol. 29, no. 5, pp. 527535, 2003.
[11] D. N. M. Donoghue, P. J. Watt, N. J. Cox, and J. Wilson, Remote
sensing of species mixtures in conifer plantations using LiDAR height
and intensity data, Remote Sens. Environ., vol. 110, no. 4, pp. 509522,
Oct. 2007.
[12] G. T. Raber, J. R. Jensen, S. R. Schill, and K. S. Schuckman, Creation of
digital terrain models using an adaptive LiDAR vegetation point removal
process, Photogramm. Eng. Remote Sens., vol. 68, no. 12, pp. 1307
1315, 2002.
[13] F. Mazzarini, M. T. Pareschi, M. Favalli, I. Isola, S. Tarquini, and
E. Boschi, Lava flow identification and aging by means of LiDAR intensity: Mount Etna case, J. Geophys. Res., vol. 112, no. B2, p. B02 201,
Feb. 2007.
[14] H. Hasegawa, Evaluations of LiDAR reflectance amplitude sensitivity
towards land cover conditions, Bull. Geogr. Surv. Inst., vol. 53, pp. 43
50, Mar. 2006.
[15] S. Kaasalainen, E. Ahokas, J. Hyyppa, and J. Suomalainen, Study of
surface brightness from backscattered laser intensity: Calibration of laser
data, IEEE Geosci. Remote Sens. Lett., vol. 2, no. 3, pp. 255259,
Jul. 2005.
[16] M. M. Raymond, Laser Remote Sensing: Fundamentals and Applications.
Melbourne, FL: Krieger, Jan. 1, 1992.
[17] K. Kampa and K. C. Slatton, An adaptive multiscale filter for segmenting
vegetation in ALSM data, in Proc. IEEE IGARSS, Sep. 2004, vol. 6,
pp. 38373840.
[18] W. Tobler, A computer movie simulating urban growth in the Detroit
region, Econ. Geogr., vol. 46, no. 2, pp. 234240, 1970.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 30, 2009 at 16:45 from IEEE Xplore. Restrictions apply.

Das könnte Ihnen auch gefallen