bul 2004
flown at
" Ottawa
1e of the
features
. bridges
as with
| spread
long the
nparison
ification
spatio-
IS. it As
the two
ons and
ping as
S, more
the test
(GOS:
le.
and the
k points
ersonnel
sed only
d to the
ized.
les over
along a
5 levels
ing the
: Mosaic
mpany's
terrain
>xpected
o 0.23
ters on
) to 0.75
5 to 1.0
| returns
surface
filtering
inate all
ove the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
oround. It was used to produce the digital elevation model
(DEM) called also digital terrain model (DTM). This data set is
named “bald-earth”.
Visual evaluation:
A first general visual evaluation was done by displaying the
LIDAR DSM and DEM as 3D perspective using the ArcGIS
software from ESRI. For a better visualization, the Z dimension
(highs) was exaggerated by factors 3 and 5. Both the DSM and
DEM give a general good image of the terrain and its general
morphology due their high resolution (1m cells). However, the
DSM produced from the "all-returns" data set contains a large
number of ‘spikes’. The DEM produced from the *bald-earth'
data set has no more nosy spikes but is less useful in the
emergency mapping context where it is important to visualize
the terrain surface with all man made objects on it as buildings,
bridges and all infrastructures. Bald earth only is not interesting
for this user group.
A comparison was made between the LIDAR DSM and a
photogrammetric DSM produced from the 1:15 000 scale aerial
photos with a planimetric resolution of a 5 m grid. The
photogrammetric DSM is very smooth and has practically no
spikes and gives a better general image of the nature of the
terrain. However, due to the important difference in the spatial
resolution (5 m cells vs 1 m cells), it is almost impossible to
identify small buildings on the photogrammetric DSM without
further information and the bridges and the water bodies have to
be edited manually. On the LIDAR DSM most of the buildings
are recognizable, except some houses in a residential area
having mature, high trees. Bridges are easy to identify as most
of the railway due to the high resolution of elevations values
allowing to recognize practically all embankments. Due to the
less interest for the 'bald-earth' DEM in the emergency
mapping context, it was decided to use the ‘all-returns’ data for
the other tests.
Some of the spikes generated probably by noise have an
elevation value over 200 m, out of the range of the terrain inside
the test area (between 75m and 120m) even considering the
existing buildings. A threshold was used for cutting off all the
clevations values outside a range of 74 m to 120 m and was
applied to the row data creating a separate data set names
‘Range cut-off’. For noise reduction, the DSM data was filtered
using two simple median filters with a window size of 3 x 3 and
5 x 5. The results were fair but the differences in noise
reduction between the 3 x 3 median filter and the 5 x 5 median
were not significant. However, it seems that same valuable
clevation information might get lost using the 5 x 5 filter and
the recognition of small objects and buildings will be
compromise. Only the 3x3 median filter was used for all further
analysis.
Evaluation of the LIDAR vs GPS checkpoints
a. Direct comparison with GPS check points:
A set of 203 GPS check points were used to evaluate the
precision of the elevations values from the DSM generated from
the ‘all-returns "LIDAR data set after applying the above
described filtering methods: the out of range cut-off and a 3 by
3 median filter. The results are presented in Table 1.
With a value of 1.82 m for the standard deviation of the *Out-
of-range cut-off" data set, the precision seems to be far from the
expectations. However, it should be remembered that the two
data sets have still some noise like we can see from the Min.
and Max. values for the differences. After eliminating the 3
971
highest differences, the standard deviation dropped to 0.64 m, a
value closer to the expectations. Comparing with the *bald-
earth’ data set was not possible as a number of GPS check
points are located on the top bridges and industrial buildings
that were eliminated by the filtering process used to generate the
‘bald-earth’ DEM. It appears that there is no significant bias
with a mean value of 5 and 6 cm.
Min Max Mean Std. dev.
(m) (m) (m) (m)
Out of
range cut- -19.0 10.7 -0.05 1.82
off
Table 1: Differences between GPS and LIDAR DSM
b.- Comparison of terrain profiles
An other set of evaluations was done using the a set of profiles
recorded using the GPS kinematic technique across some
typical features of the test area: across and along a railway
track, along a country road with moderate, constant slope, over
a 5 levels building and the attached flat parking lot. These
profiles were compared with filtered ‘all-returns’ LIDAR data
set and a set of two DSMs created from the digital aerial images
using the Leica Helava DPW. A typical profile evaluation is
shown in Figure 2. It appears that the LIDAR DSM is very
close to the GPS profile. The standard deviation is 0.21 m
computed over all GPS recorded points of the profiles and the
mean value is 0.14 m.
Profile B
90
lier GPS à
—79 —- Photo
— + —LIDAR)
0.0 20.0 40.0 60.0
Figure 2. A profile across a railway embankment
4.3.3 Evaluation of LIDAR intensity data
The LIDAR system used for the test recorded also the intensity
of the signal for each measured point. The intensity value can
be used to produce an image of the terrain that has the
advantage to be recorded in the same time as the LIDAR
elevation data and to have theoretical the same metric
proprieties as an ortho-image produced using the LIDAR DEM.
The main difference with an image recorded by a standard
digital camera is that the LIDAR recorded points do not have a
regular distribution. The quality and the true resolution could be
variable across the recorded area as it is the almost random
distribution of the LIDAR recorded points. For the present test,
the intensity values were used to produce an image with the
resolution of Im by Im. The resolution is significant lower than
the digital image recorded simultaneously but has an important
advantage for the emergency mapping: it can be produced in the
same time as the LIDAR DSM with no more human
intervention.
