The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
417
and the surface response (which is unknown at first), the cross
correlation is performed by neglecting the surface response.
(Kirchhof et al., 2008) have designed an iterative process to
refine the matched filter in so as to integrate a modelled planar
response to the matched filter, assuming that laser beam hits a
planar surface with a given slope. The process consists in
iteratively estimating the objects surface response to first
improve range value estimation and discriminate between
partially penetrable objects and impenetrable ones, then to
extract and select surface primitives in 3D points expected to
belong to impenetrable surfaces. Finally the authors estimate the
relative surface slope (RANSAC) and compute the surface
response, used as prior knowledge in a new step.
4. APPLICATIONS
4.1 Forest areas
Many studies have already been carried out for estimating forest
parameters using multiple echo data derived from laser scanning
systems: small footprint LiDAR systems, with high point density,
can be used to extract trees in small areas (Brandtberg et al.,
2003), their height and crown diameter (Persson et al., 2002),
their volume (Naesset and Bjerknes, 2001), to classify them
according to species (Holmgren and Persson, 2004), estimate
their particular characteristics (Andersen et al., 2005) and even
to measure the growth of the forest and detect trees that have
been felled (Yu et al., 2004). Large footprint systems can reach
the ground and the tree tops in dense, tall canopies. Woodland
parameters can be estimated at large scale: density of population,
coverage, biomass, etc. (Means et al., 1999).
Many algorithms have been developed for forest measurements
(Hyyppa et al., 2004). Full-waveform experimental systems
with large footprint developed by the NASA have been
successfully used in forest environments for measuring the
canopy height (Lefsky et al., 1999) or the vertical distribution of
canopy material (Dubayah and Blair, 2000).
Furthermore, the modelling of raw LiDAR signal recorded by
recent small footprint industrial systems has already proved
efficient in increasing the number of detected objects in
comparison with data provided by multi-echo LiDAR systems
for which real-time point extraction method is unknown to the
end user (Persson et al., 2005; Chauve et al., 2007).
Section(m)
Figure 2: Profil of LiDAR points from multiple pulses data (red)
and from full waveform data (green).
For deciduous vegetation, measurements in leaf-off conditions
give a direct impression of the internal structure of the canopy
(the main levels of vegetation) and are very useful to estimate
the real ’’shape” of the trees (Reitberger et al., 2006).
In forested areas, waveforms can be composed of weak returns
from both the top of the canopy and the ground, and also of
distributed backscatters from the different layers of the
vegetation, which leads to groups of overlapping echoes.
Hardware detection algorithms, most of the time using
thresholds, can hardly detect or separate such very low peaks or
groups of echoes (Figure 1). Therefore, processing the
waveforms allows to control the point extraction method and to
make it more efficient: an improved peak detection was shown
to be very successful to extract additional objects in the received
waveforms and to detect from 30% to 130% additional points
depending on vegetation density (Chauve et al, 2007b). These
points are located mainly within the canopy and in highly dense
understory (Figure 2). Detecting such weak echoes allows a
better describtion of the 3D vegetation structure and the ground.
As a consequence, Digital Terrain Model (DTM), Digital
Surface Model (DSM), and the derived Canopy Height Model
(CHM) are expected to be significantly improved, but is still to
be validated with reference data.
4.2 Urban areas
When processing full waveform data on urban areas, one can
observe that the point density is barely improved. Thanks to
waveform analysis, more echoes are found in tree canopies,
whereas LiDAR pulses cannot penetrate rigid, opaque structures
such as buildings and streets. Consequently, a single peak is still
present within the waveform and can be detected by signal
processing algorithms. Multiple echoes can appear on building
edges and superstructures. Indeed, post-processing algorithms
enable to extract weak echoes not found by on-line detection
techniques. In urban areas, it is observed when the laser beam
hits for instance building edges. The resulting waveform is
therefore composed of distributed backscatters of the roof and
the ground, which can often not be separated by hardware
detection algorithm using fixed thresholds.
Modelling LiDAR waveforms permits to extract 3D point
clouds featuring supplementary useful parameters in addition to
the traditional (x, y, z) coordinates and to perform subsequently
point cloud segmentation based on these parameters. The
standard features are range, amplitude and width. Nevertheless,
analysing a 3D point cloud processed from full waveform data
with model parameters does not allow to define particular
behaviour for each urban object. The echo is wider on the
canopy with regard to roads or grass ones but a wide echo with
low amplitude does not necessarily comes from vegetation.
High amplitudes are noticed on grass and asphalt and variable
amplitude on the roofs of buildings, depending on the roof
materials. In fact, roads and building roofs are made from
different types of material and, therefore, have different similar
values to natural objects (Gross et al., 2007). Moreover, it is
visually possible to distinguish between different urban
materials hit by a laser beam but a more in-depth processing is
required to recognise roof materials (Jutzi and Stilla, 2003).
Consequently, simple algorithms aiming at segmenting building
and vegetated areas, e.g., decision tree approach with empirical
thresholds, lead to a certain rate of incorrect classification
(Duong et al., 2006; Ducic et al., 2006). A more reliable
approach to detect vegetated areas is proposed in (Gross et al.,
2007) , based on the eigenvalues of the covariance matrix
computed for each point with the intensity values in a
cylindrical and spherical environment. Another method based on
a supervised classifier (SVM) has been developed (Mallet et al.,
2008) . It provides a suitable segmentation of ground, building
and vegetation areas but requires classical geometric features in
addition to parameters extracted for waveform modelling.
To achieve more advanced point segmentation in urban areas