B4, 2012
atures
tation and
hillside, data
regularly
road with
unnel, data
uildings with
aps.
ith trains (low
points), data
vegetation,
on river
oads with
a gaps.
, roads with
ia gaps.
Tr
tested by the
2 refers to the
icient value of
Type I, Type Il
nerated by the
st values of the
ly. The standard
ficient in fifteen
rall accuracy i$
aries depending
in the proposed
the best results
ie optimal result
r to analysis the
d on different
it as follows.
International Archives of the Photogrammetry, Remote Sensin
g and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
! Width K | No. of Identification
Sites (m) Del (m) | Levels Tana tolerance
sampl11 2.1 0.6 | 1.6 7 0.4 0.5
[samplik | 2.6 |06|40| 6 | 07 0.5
samp12 0.9 0.6 | 2.0 7 0.7 0.5
samp21 09 | 06} 20 7 0.7 0.5
samp22 0.9 06] 20 7 0.7 0.5
samp22k | 1.5 | 0.6 | 16 7 1.2 05
samp23 09 | 06120 7 12 0.5
samp24 09 1:06 :20 7 1.2 0.5
samp3 1 09 | 06) 20 7 0.7 0.5
samp41 09 | 06 | 20 7 1.8 0.5
samp42 09 1:0.6 |. 2.0 7 0.7 0.5
samp51 09 | 06 } 20 7 0.7 0.5
sampSlk | 2.0 | 0.6 | 2.4 7 0.7 0.1
samp52 09 | 06.20 7 0.7 0.5
samp52k | 40 | 0.5 | 3.0 6 0.7 0.8
samp53 09 .].0.6.1.2.0 7 0.7 0.5
samp53k | 32 1.5 | 2.4 7 3.0 0.5
samp54 27 102] 24 7 0.4 0.5
samp61 09 {06 20 7 0.7 0.5
samp71 09 (06 20 7 0.7 0.5
City vs. Forest
100%
80% -
0%
# City: Sample 11-42
# Forest: Sample 51-71
Table 2. Parameters used in the proposed MTF method
Since the 15 sample sites are divided into three groups, a
comparison is shown in Figure 4. Group 2 sites shows the
lowest errors and highest accuracy rate and kappa coefficient,
which means that the MTF method can handle Group 1 sites
better than the other two groups, this number is also good
enough to compare with filters compared by ISPRS (Sithole and
Vosselman, 2003). The performance of the MTF method on
Group 2 is average. However, the performance on Group 3
shows a very low Kappa coefficient because of the high Type II
errors. Group 3 sites contain features like steep slope and high
percentage of terrain points. Therefore, the MTF method
probably has flaw on process this type of areas.
p
100%
80% -
30, 4— —
Groups
# Group 1: Sample 11,
24, 41, 54
8 Group 2: Sample 12,
21,22 23,31, 42
Group 3: Sample 51,
52,53,61,71
Figure 5. Average values of Type I, Type II errors, Accuracy
and Kappa sorted by City Sites and Forest Sites
The MTF is based on layering, which is a global analysis of the
data height value. Therefore, the terrain points’ portion of all
points can affect the result. Figure 6 demonstrates the MTF
performance based on different terrain point percentage. It
seems along with the growth of the percentage, the errors
especially Type II error become higher, while the accuracy and
kappa become lower. But it needs to be noticed that there is
only one sample for the terrain point percentage smaller than
40%, and the sample which have higher than 80% terrain points
are all in the Group 3. Therefore, the uncertainty of this feature
still requires further discussion.
Terrain point percentage
100% #< 40% : Samp42
80%
60% #40%-60%: Samp
40% La 4154
0 :: 0- o: »amp
poss 21,22,24,51
0% #> 80%: Samp
kg P 52,53,61,71
&° «S $ S 23,61,
A sj m "S
SS «S S
v
Figure 4. Average values of Type I, Type II errors, Accuracy
and Kappa sorted by three groups
The ISPRS dataset is originally sorted as city sites and forest
Sites, the performance of the MTF method on city sites and
forest site are shown in Figure 5. It is obvious that the
performance on city site is better since it has lower Type I, Il
errors and higher accuracy and kappa. The steep slope and the
dense vegetation coverage might be the reason why the MTF
method has an unsatisfactory result on forest sites. The forest
Sites are basically overlapping with the Group 3 sites. The high
Type II error is probably from the buildings on the slope which
Is a difficulty mentioned by Sithole and Vosselman (2003). It is
also the key to improve the value of the Kappa coefficient.
Figure 6. Average values of Type I, Type II errors, Accuracy
and Kappa sorted by percentage of terrain point
In order to know which feature of the data has more influence to
the result, the average standard deviations of the previous three
types of sortation are calculated as shown in Figure 7. The chart
shows that they are all in the same range of each characteristic;
however, the group sortation has the lowest average standard
deviation among the three types of sortation.
Average Standard Deviations
20%
15% # Groups
a | N & City/Forest
0% el * Terrain percent
Figure 7. Average values of Standard Deviations of Type I,
Type II errors, Accuracy and Kappa in three types of sortation
347