Bare earth extraction from airborne lidar data using different filtering methods

Abstract

Over the past years, several filters have been developed to extract bare-Earth points from airborne laser scanner data. we conducted a test to determine the performance of three different filters and applying wavelets for the degradation of blunders in LIDAR data, and to identify directions for future research. Five selected samples have been processed by three participants. In this paper, the test results are presented. The paper describes the characteristics of the used filter approaches. The filter performance is analysed qualitatively. The performance of all filters depends on the complexity of landscapes in which they used. In general, wavelet de-noising is found to perform well and it is hardly recommended to use wavelets for de-noising the LIDAR data. Future research should be directed towards the usage of wavelets in the filtering algorithms.

Full text

BARE EARTH EXTRACTION FROM AIRBORNE LIDAR DATA USING DIFFERENT
FILTERING METHODS
A. Baligh
*
, M. J. Valadan Zoej, A. Mohammadzadeh
Geodesy and Geomatics Engineering Faculty, K.N.Toosi University of Technology, No. 1346,
Vali_Asr St., Tehran, Iran, Postal Code: 1996715433 -
ali_baligh@yahoo.com, valadanzouj@kntu.ac.ir, ali_mohammadzadeh@alborz.kntu.ac.ir
Commission III, WG III/3
KEY WORDS: LIDAR, Point Cloud, Classification, DTM, Landscape, Segmentation
ABSTRACT:
Over the past years, several filters have been developed to extract bare-Earth points from airborne laser scanner data. we conducted a
test to determine the performance of three different filters and applying wavelets for the degradation of blunders in LIDAR data, and
to identify directions for future research. Five selected samples have been processed by three participants. In this paper, the test
results are presented. The paper describes the characteristics of the used filter approaches. The filter performance is analysed
qualitatively. The performance of all filters depends on the complexity of landscapes in which they used. In general, wavelet de-
noising is found to perform well and it is hardly recommended to use wavelets for de-noising the LIDAR data. Future research
should be directed towards the usage of wavelets in the filtering algorithms.
1. INTRODUCTION
The use of airborne Light Detection And Ranging (LIDAR)
technology offers rapid high resolution capture of surface
elevation data suitable for a large range of applications. The
commercial use of LIDAR in the last few years has gained more
importance as more reliable and accurate systems are produced.
While LIDAR has come a long way, the election of appropriate
data processing techniques for the generation of models is still
being researched. In order to produce digital elevation models,
filtering and quality control pose the greatest challenges,
consuming an estimated 60–80% of processing time (Flood,
2001), and so underlining the necessity for research in this field.
Some comparison of known filtering algorithms and problems
have been mentioned in Huising and Gomes Pereira (1998),
Haugerud and Harding (2001), Tao and Hu (2001) and Sithole
and Vosselman (2003). However, an exprimental comparison is
available, although it would be useful to improve the
performance of the different approaches by adding some hints
on how to apply the algorithms. In order to assess the
weaknesses of the different approaches we used the available
reference data.
Therefore, we established a study to compare the performance
of three automatic filters developed to date, with the aim of:
– determining the comparative performance of these filters,
– applying wavelet denoising method for the detection of
blunders in LIDAR data, and
– identifying directions for future research on
filtering point clouds.
To achieve these aims, we used the ISPRS web site at which
datasets were provided for testing. processing all datasets was
not possible, but we used some. Results were received from
three participants. The algorithms are: Hierarchical terrain
recovery algorithm by
utilizing Image pyramid (Hu, 2003),
Filtering of Airborne laser scanner data based on segmented
point clouds (Sithole, 2005), Slope based filtering (Vosselman,
2000). We used these algorithms because of their novel methods
for extracting the bare-Earth from airborne laser scanner point
clouds.
This paper is structured as follows: The laser scanning data used
in this paper is described in Section 2. Section 3 describes three
selected filter algorithms. Section 4 deals with filter concepts.
Section 5 describes the application of wavelet in LIDAR de-
noising and The results of the implementations are described
and discussed in Section 6. In Section 7, the conclusions are
drawn with respect to the objectives set out.
2. TEST DATA
Within the ISPRS web site
(http://www.itc.nl/isprswgIII-
3/filtertest/StartPage.htm)
we found some free datasets which
belongs to the OEEPE project on laser scanning (OEEPE, 2000),
FOTONOR acquired data with an Optech ALTM scanner over
the Vaihingen/Enz test field and the Stuttgart city centre. By the
kind permission of Commission III, Working Group 3, subsets
of this dataset are available for the comparison of filtering
algorithms. Also reference data which was produced by
interactively filtering the datasets is accessible.
2.1. Data provided to the participants
A total of five samples (of urban area) were chosen because
they contained a variety of features that were expected to be
difficult for the filtering. The datasets included terrain with
steep slopes, dense vegetation, densely packed buildings with
vegetation in between, large buildings, multi-level buildings
with courtyards,
ramps, underpasses and etc. The samples of urban areas were
recorded with a point spacing of 1–1.5 m (0.67 points per
square metre). For each of the samples the reference data is
available. This reference data has been compiled by semi-
automatic filtering and manual editing. Each record in the
237

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
reference data files contains the X, Y, and Z coordinates
followed by either a 0 (ground point) or 1 (non-ground point).
Figure.1 shows a part of an urban test site.
Figure 1. a part of an urban test site
3. OVERVIEW OF FILTERING ALGORITHMS
An overview of the algorithms and a characterisation of filters
are given in Table 1. The filtering algorithms are described in
more detail below.
3.1. George Vosselman
The basic idea is based on the observation that a large height
difference between two nearby points is unlikely to be caused
by a steep slope in the terrain. More likely, the higher point is
not a ground point. Clearly, for some height difference, the
probability that the higher point could be a ground point
decreases if the distance between the two points decreases.
They explicitly
define the acceptable height difference between
two points as a function of the distance between the points:
Dh
max(d). In general, this will be a non-decreasing function.
Further details can be found in Vosselman (2000).
3.2. Yong Hu
The algorithm will identify terrain points by finding local
minima and other topographic points, and recover the terrain
surface in a coarse-to-fine manner. First, after screening the
blunders, the scattered 3-D points are transformed into a grid-
based range image by selecting the point of lowest elevation in
each grid. Then, the raw range image is processed to fill void
areas and correct distortions. Then, an image pyramid is
generated.
The top-level image is hypothesized to be a coarse DTM if its
grid size is larger than the largest non-terrain object. Finally, the
coarse DTM is refined hierarchically from the top level to the
bottom level. At each level, denser terrain points are identified,
and the nonterrain points are replaced by interpolated elevations
using surrounding terrain points. The bottom-level image
represents the expected bare Earth surface.
Post-filtering is performed on the resultant DTM to improve its
quality for mapping purposes. This smoothing attempts to
correct the influence of speckles and undesired non-terrain
undulations of the DTM. At last, the void areas in the raw range
image may be duplicated to preserve water regions.
The quality of the derived DTM is subject to thresholding
parameters used in the algorithms. The optimal values of these
parameters may vary with the scene complexity. One can use a
priori knowledge, if available, including the terrain relief range,
the maximum and minimum building sizes, heights and areas,
the maximal tree height, etc to determine the parameters
empirically first, and then adjust them adaptively if multi-return
or intensity data is available. Further details can be found in Hu
(2003).
3.3. George Sithole
The algorithm emphasis is placed on establishing the
topological and geometric relations between bare-Earth and
object surfaces. Filtering is defined as the identification of
surface segments whose perimeter is raised above their
neighbourhood. Meaningful surfaces can be reconstructed for
large objects but not for small objects (too few points).
Therefore, in the algorithm large and small objects are detected
separately. Large objects are treated by segmentation of the
point cloud, while small objects by smoothing of the point cloud
in a later stage.
The segmentation of a point cloud into smooth surfaces is the
first important step of the algorithm. Two points are considered
to be part of the same surface if there is a smooth path of
adjacent points between them. This definition allows for
discontinuities within a surface as long as there is a path around
a discontinuity that connects points on both sides.
In this filter scan line segmentation algorithm is used in a
different manner by definition and segmentation of scan lines
with multiple orientations. A point cloud is sliced into
contiguous profiles, and this slicing is done in several directions.
Once profiles have been obtained, they are segmented to get
line segments that represent continuous planar curves on
surfaces in the landscape. Segmentation of the point cloud is
achieved by overlaying all segmented profiles which lead to the
generation of a graph, and a surface segment is therefore a
connected subgraph. The usage of the profiles in several
directions hence provides an elegant way to combine the profile
segmentation results to a surface segmentation. Two adjacent
parallel profile segments are connected only if there exists a
profile segment with another orientation that contains points of
both these parallel segments.
The algorithm can be described as a procedural stripping away
of objects from the bare-Earth in the following sequence, large
objects, bridges, and small objects. In each step of the sequence
smaller and smaller objects are removed. The explicit detection
of bridges is necessary to ensure the reliability of the detected
bare-Earth. Further details can be found in Sithole (2005).
Participants Filter description
George Vosselman—
Delft University of
Technology
Slope-based filter
Yong Hu—
University of Calgary
Hierarchical terrain
recovery
George Sithole—
Delft University of
Technology
Segmented point clouds
Table 1. Test participants
238

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
4. FILTER CONCEPT
Every filter has an assumption about the structure of bare-Earth
points in a landscape. Here, In the Slope-based algorithm
(Vosselman), the slope or height difference between two points
is important. If the slope exceeds a certain predefined threshold,
then the highest point is assumed to belong to an object. In
hierarchical terrain recovery algorithm (Hu), the top-level of an
image pyramid is assumed to be a coarse bare-Earth. In such
algorithms a smooth condition is established to validate that the
elevation at the current level and the estimated elevation from
the reference level are reasonably close at every point of a local
patch. If any point in the current window violates the smooth
condition, then that point is classified as a non-terrain point. In
Segmentation-based algorithm (Sithole), the assumption is that
every points in a segment must belong to an object if their
segment is above its neighbourhood. It is important to note that
for such a concept to work the segments must delineate objects
and not facets of objects.
5. WAVELET DE-NOISING
Blunder detection of LIDAR data by using wavelets in this
paper is based on the decomposition of the raw LIDAR image
in order to detect salient changes in height. In the context of
Airborne laser scanning data, a filter response represents a
discontinuity caused by a blunder whereas the underlying
landscape do not response. By employing blunder detection
using wavelets, low frequencies such as ground or large objects
are separated from the high frequency components which
represent blunders.
The LIDAR point cloud is first regularly gridded. The resulting
matrix is then decomposed using the Discrete Wavelet
Transform (Goswami and Chan, 1999) into low frequencies and
high frequencies. The energy is evenly distributed among sub-
images and therefore, the amplitudes of sub-images becomes
lower (Bartels, Wei, and Mason, 2005). In this study, level 4
decomposition is applied to the LIDAR data. Discontinuities
give responses in the details depending
on their relative position to the wavelet kernel.
(a) (b)
Figure 2. (a) LIDAR range image, (b) degradation of blunders
by wavelets.
Figure.2(a,b) depicts wavelet de-noising on LIDAR range
image using Symlets(6) wavelet filter (Daubechies, 1992). The
scene in Figure.2(a) represents building blocks of different
height on the terrain with blunders. Applying wavelet de-
noising, blunders in the scene are degraded as shown in
Figure.2(b). As shown in Figure.2(b), two observations can be
made. First, blunders could be successfully degraded. Second, It
can clearly be seen that the filter response depends on the
degree of the salient height, i.e. the larger difference between
adjacent height values the higher is the magnitude of the
responses. However, as expected, it can also clearly be seen that
flat roofs are not degraded as there is no distinctive change in
height.
6. RESULTS
The filter results of these algorithms have been analysed in
various ways. The data of five refrence samples has been used
to assess the performance of the algorithms in several difficult
terrain types.
The quantitative assessment was done by evaluating Type I
(rejection of bare-Earth points) and Type II (acceptance of
object points as bare Earth) errors. It must be stressed that the
results which are presented here do not cover the difficulties in
filtering as observed in the data, yet in general all the filters
worked quite well.
The output of Hu’s filter is gridded, and it is altered in position
from the original. Therefore, DEMs were generated for the
filtered data and the height of the points in the reference data
were compared against these DEMs. The Type I and Type II
errors have to be understood in the context of height comparison
of the reference against the filtered DEMs.
Vosselman
Test
Samples
Type I
errors %
Type II
errors %
Number of
points
Samp11 43.2 5.7 38010
Samp12 23.8 3.5 52119
Samp21 10.4 1.6 12960
Samp31 5.6 1.9 28862
Samp41 51.8 2.6 11231
Hu
Test
Samples
Type I
errors %
Type II
errors %
Number of
points
Samp11 13.6 7.8 38010
Samp12 4.8 2.6 52119
Samp21 1.5 12.7 12960
Samp31 6.1 1.9 28862
Samp41 21.6 1.7 11231
Sithole
Test
Samples
Type I
errors %
Type II
errors %
Number of
points
Samp11 28.3 13.7 38010
Samp12 3.7 2.6 52119
Samp21 3.3 1.7 12960
Samp31 2.3 3.2 28862
Samp41 6.8 4.9 11231
Table 2. percentages of Type I and Type II errors
All the filtering algorithms fails sometimes. This failure is
caused by the fact that filters are blind to the context of
structures in relation to their neighbourhoods. Because of this, a
trade-off is involved between making Type I and Type II errors.
The computed errors (over all the datasets) ranged from 1.5 to
51.8%, 1.6 to 13.7% for Type I, Type II, respectively.
Vosselman’s filter focus on minimizing Type II errors. This
tendency to minimizing Type II errors, partly suggests that they
consider the cost of Type II errors to be much higher than that
of Type I errors. Table 2. shows the percentages of Type I errors
and Type II errors for three tested algorithms.
239

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
7. CONCLUSIONS
In this paper a comparison of three different filtering methods
was presented. The objectives of the study were to: (1)
determine the performance of filter algorithms, (2) applying
wavelet denoising method for the detection of blunders in
LIDAR data and (3) establish directions for future research. In
general, all the filters worked quite well in landscape of low
complexity. As seen in Table 2, the performance of a filter can
differ depending on the feature content of a landscape. The
problems that pose the greatest challenges appear to be complex
cityscapes and discontinuities in the bare-Earth. So, the
combination of different strategies may be a solution. Wavelet
de-noising methods showed their talent in the detection and
degradation of blunders among LIDAR data. Therefore, it is
hardly recommended to use wavelets in other researches and
even in the filtering methods for the sake of object detection.
REFERENCES
Bartels, M. H. Wei, and D. C. Mason. Wavelet packets and
cooccurrence matrices for texture-based image segmentation.
IEEE International Conference on Advanced Video and Signal-
Based Surveillance, 1:428–433, 2005.
Daubechies, I. Ten lectures on wavelets. CBMS-NSF regional
conference series in applied mathematics, 1992.
Flood, M. (2001). Lidar activities and research priorities in the
commercial sector. IAPRS vol. 34 (3W/4, October 22-24,
Annapolis (MD), USA), pp. 678–684.
Goswami, J. C. and A. K. Chan. Fundamentals of wavelets :
theory, algorithms, and applications. New York, Chichester:
Wiley, 1999.
Huising, E. and L. M. G. Pereira (1998). Errors and accuracy
estimates of laser altimetry data acquired by various laser
scanning systems for topographic applications. ISPRS JPRS vol.
53 (5), pp. 245–261.
Hu, Y. (2003). Automated Extraction of Digital Terrain Models,
Roads and Buildings Using Airborne Lidar Data,
http://www.geomatics.ucalgary.ca/links/GradTheses.html
(accessed 30 april. 2008)
OEEPE, 2000. Working Group on laser data acquisition. ISPRS
Congress 2000. http://www.terra.geomatics.kth.se/~fotogram/
OEEPE/ISPRS_Amsterdam_OEEPE_presentation.pdf
(accessed January 22, 2004).
Sithole, G. and G. Vosselman (2003). Automatic structure
detection in a pointcloud of an urban landscape. Proceedings of
2nd Joint Workshop on Remote Sensing and Data Fusion over
Urban Areas (Urban 2003), May 22-23, Berlin, Germany, pp.
67–71.
Sithole, G. (2005). ISPRS WG III/3, III/4, V/3 Workshop
"Laser scanning 2005", Enschede, the Netherlands, September
12-14, 2005.
Tao, C. and Y. Hu (2001). A review of post processing
algorithms for airborne lidar data. Proceedings ASPRS
conference April 23-27, St. Louis Missouri. CDROM, 14 pages.
Vosselman, G. (2000). Slope based filtering of laser altimetry
data. IAPRS, WG III/3, Amsterdam, The Netherlands vol. 33
(B3), pp. 935–942.
240