The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
361
point data was included. In some rare cases in a single cell the
presence of several data was noticed (up to 3 or 4 points) and so
also an ambiguity. To solve this problem we used an algorithm
which evaluated the mean height. Another possible solution
could be the reduction of the cell size, but we preferred to avoid
this choice that implicates a larger number of cells.
The results of the rasterization process are 2 maps which
correspond to Sardinia and GRASS classification. The
advantage is the possibility to subtract the map (that is easy
because the procedure is based on raster algebra operations) in
order to obtain a raster map of the height differences. Where the
both maps contained a float value the difference could be
calculated, instead where one or both maps presented a “nodata”
value also the difference map contained this one.
This means that the statistics on the difference map are identical
to the statistic carried out on the original correspondent vector
points (exception made for the very few cells with an
ambiguity).
Starting from the original dataset a representative set was
extracted considering the spatial distribution of the points and
the morphology of the area. It is compose of 5 areas with a
number of points variable from 1.5T0 6 to 8-10 6 . The
morphology of the areas is also variable enough and includes
urban, rural and wooded areas.
Before filtering an outlier rejection was performed. The
function used is v.outliers, which performs an interpolation of
the data on a regular grid and then calculates the differences
between the new surface and the spread points. The residuals
are compared with a fixed threshold: data corresponding to
residual exceeding the threshold are considered as outliers.
The filtering method developed in GRASS needs some input
parameters. This makes the algorithm highly customizable and
capable to elaborate dataset morphologically different. In any
case, a default parameter set is suggested as optimal being the
choice of the parameter often complicated.
The three functions that compose the algorithm were launched
in sequence with the default set of parameters to perform a
classification in 5 representative test areas, obtaining a
subdivision of the points in terrain and object. Sardinia’s
classification instead is based on terrain, vegetation and
building points and considering that the points suitable for the
realization of the DTM are the terrain points only, we compared
the points classified as terrain by us and by TopScan.
The statistics of the difference related to the points classified as
terrain by both procedures are shown in table 1.
AREA
A
B
C
D
E
F
2856668
2250790
4005581
728888
1550308
L
2861842
2253878
4008225
729011
1550749
TS
1885229
1655103
3289329
568432
1214066
GR
2184190
1830501
3664729
636919
1408589
C
1676170
1577317
3233053
525629
1159922
% TS/GR
86,31
90,42
89,76
89,25
86,19
% C / T
88,91
95,30
98,29
92,47
95,54
As it is possible to see the number of points classified as terrain
by GRASS is always larger than TopScan, but the points in
common between the two methods are approximately equal to
TopScan’s terrain points. This means that GRASS finds the
same points of TopScan but in addition finds others points. A
following control demonstrated that the excess GRASS terrain
points are generally located under the vegetation but on the
ground (see. figure 1), and for this reason they were removed by
the manual control of TopScan. That means that the considered
points are not a classification error.
Figure 1. Mismatch between GRASS and TopScan terrain
points classification.
In area B a building vector mask was manually performed
starting from the orthophotos (177 buildings). Then the mask
was rasterized with a resolution of 0.2 m. The basic idea was to
use the mask to control the number of points classified as terrain
that belong to the mask: these points are a classification error.
The number of points classified as terrain but that lie in building
mask is shown in the following table (Table 2).
GRASS
TopScan / Sardinia
mask points
percentage
mask points
percentage
9819
0.53%
5588
0.34%
Table 2. Points classified as terrain that lie into buildings
A followed analysis demonstrated that the points classified as
terrain that lie in the mask are often close to the edges of the
buildings. Figure 2 depicts a generic situation where it is
possible to observe an error along the North edge of the roofs.
For this reason we considered this points still belonging to the
terrain, because they are due to errors committed during the
manual realization of the mask. In fact the orthophotos present a
perspective deformation due to the distance of the considered
buildings from the nadir point and the objects that have a relief
displacements are so distorted.
Table 1. Difference between the classifications
Legend: F, L = first and last pulses; TS, GR = TopScan and
GRASS terrain points, C = corresponding points