The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
procedure in a way that it is possible to control the results of
each single step.
The algorithm is highly customizable, as it is based on a great
number of parameters, even if, in its usage with datasets
morphologically different, it has shown that a specific set of
parameters can suggested as optimal.
Starting from the points classified as single pulse terrain the
DTM can be computed. The used command in this case (always
performed by us and called v.surf.bspline, still free available in
GRASS) interpolates the point data on a regular grid using
bilinear or bicubic splines with Tychonov regularising
parameter (Brovelli and Cannata, 2004).
What we want to show in this paper is not the detailed
functioning of the algorithm, but the quality of the products that
can be obtained with it.
The control was performed comparing our results (that from
here we will call GRASS results) with the results obtained by
TopScan, a German company which performed the same
process on behalf of Sardinia Region by using a preliminary
automatic algorithm followed by a manual control (hereafter
those products are named Sardinia’s products).
Sardinia’s products were checked by visual comparison with
high resolution orthophotos (12.5 cm) and terrain measurements.
For these reasons, we can assume Sardinia’s products as a good
reference.
The control regarded the classification of the raw data (filtering
analysis), the DSM and the DTM. This kind of control can be
considered as a relative control, because the initial data are the
same for both two products, despite the two procedures are
completely different and independent.
To check the absolute precision of GRASS products a
comparison with a new set of points measured with a GPS
(RTK survey) was performed. In this case the data are
completely independent and the larger accuracy of the GPS data
(±0.03 m) than LiDAR (±0.2-K).3 m) ensures a good dataset as
reference.
The last control was to test the performances of the algorithm to
verify the applicability on real cases. The computational cost
depends on the number of the splines used to interpolate the raw
data. A larger number of splines implicates a better resolution
(but not always a better solution) but it increases the
computational time. Thus a compromise between quality and
time becomes necessary.
In the following paragraphs the controls are presented, starting
from the filtering and grid products, up to the algorithm
performances.
2. CLASSIFICATION CONTROL
2.1 Dataset description
The original dataset is compose on 286T0 6 points acquired with
an Optech ALTM 3100. It covers an area of 59.3 km 2 along the
East Coast of Sardinia Region, in the urbanized areas from
Porto Rotondo to San Teodoro. It is composed of 63 strips
acquired in three days, with a sidelap always larger than 50%.
The altitude above the ground is 1000 m and the scan rate is 70
kHz. The areas were mapped with a mean laser spot density of
higher than 1 points/m 2 roughly.
The dataset was filed as first and last pulses in ASCII text files,
reporting the cartographic coordinates (UTM WGS84) and the
intensity. From the ellipsoidal height the orthometric height was
calculated according with Italian quasi-geoid.
The whole dataset was divided in 28 areas, which corresponds
to the municipalities along the considered part of Sardinia’s
coast.
2.2 Reference data description
The raw data were acquired and processed by
TopScan/HANSAER associated with the Italian company
Aerosistemi S.r.l. and the German company Hansa Luftbild
Sensorik und Photogrammetric GmbH. These companies
performed the entire process on behalf of Sardinia Region.
Firstly TopScan performed a classification of the raw data in
three categories: ground, vegetation and buildings points. The
used method is based on a preliminary automatic algorithm
implemented by TopScan itself. This algorithm is able to divide
the ground points from the object points. Then the object points
were divided in buildings and vegetation points. Lastly a
manual correction was performed to correct residual errors due
to misclassifications.
With the points classified as terrain a grid DTM was performed
by using an interpolation. The DSM instead does not require
any preliminary filtering operation and can be obtained with an
interpolation of the data classified as first pulse. TopScan’s
interpolation method was the linear prediction with bell curve as
base function (Kraus and Pfeifer, 2001).
Both DTM and DSM have a resolution of 2 m and were
checked by visual comparison with high-resolution orthophotos
and a spatial DB. The height accuracy instead was checked with
survey measurements (GPS and Total Station).
All these qualities make Sardinia’s products a good reference to
analyze the classification and the DTM/DSM that our method is
able to provide.
2.3 Test of GRASS classification method
To test the results of GRASS filtering algorithm a comparison
with the Sardinia’s classification was performed.
The main problem was to compare sparse points considering
their classification. In fact a simple count of the points which
belong to some determined categories is not sufficient to
compare the classification (e.g. two methods can classify
exactly the same number of points but these have a different
location). That implicates that is necessary to compare each
single point by using its spatial coordinates.
The vast number of points (over 280T0 6 ) implicates more than
10 17 combinations, and it makes impracticable the control itself.
Even if we searched a method to decrease the number of
operation, it could not make the control workable.
For this reason a reduction of the data and a method able to
speed up the control becomes inevitable.
The method used to compare the vector points was based on
their rasterization on a regular grid with square cells. The
resolution was fixed equal to 0.5 m, so that into each cell only a
360