1л: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). 1APRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010
5.3 Classification with and without ground points
The limitation of the classification algorithm of the machine
learning including the RF is that a trained classifier strongly
depends on the class with the highest frequency. That is, the
classifier is not suitable for detecting minor classes. The ground
is one of the dominant classes in power-line scene as well as
sub-urban area. The RTF terrain filter described in section 3.1
can remove ground points out of untagged LIDAR data. Once
removing them, the frequency of the remaining classes might be
almost equal. Figure 4 represents the variation of variable
importance considering and not considering ground and 9
features which seem to be critical without ground are only
shown. According to the figure 4, the Height, contributing
feature to terrain detection, appeared to be less important
without ground, but it is still the most critical feature for the
other classes. While, the importance of Anisotropy, Density
Ratio, and HT increased. This means the trained classifier was
changed to be sensitive to the others. Another way to resolve
this limitation is to take out a training set in which all classes
are uniformly distributed (Lodha, 2007). For this test the RF
was run based on feature derived by point-wise approach using
both ground removed TR and TE.
When comparing to TE1G and TEEG in figure 5, the
classification accuracy of vegetation, pylon, and building was
somewhat improved since a fair classifier which is able to care
minor classes. Additionally, the importance of features relevant
to them was increased. However, power-line is nearly same as
previous one in spite of increasing importance of HT. Perhaps
there would be certain features substituting for HT in the sample
with ground.
and entire features (Table 4). Most were classified similarly to
the result of the method using entire features excluding electric
pylon. There were not any features enabling to characterize
pylon among the 9 features. Figure 6 depicts the classification
map for TEEG using only 9 features. Most errors of pylon are
appeared near ground and vegetation because they include
nearby ground and vegetation as their neighbourhoods while
computing the features.
Class
Veget
Wire
Pylon
Bldg
Accuracy(%)
Veget
63,491
48
13
575
99.00
Wire
97
5,831
50
23
97.17
Pylon
103
106
856
2
80.23
Bldg
656
27
0
23,528
97.18
Table 4. Confusion matrix for TREG using 9 important features
Figure 6. Classification map for TEEG using 9 features
Jt
^ <■
У
✓
& </
^ &
s
Figure 4. Variation of variable importance (with TR)
I 1 TR_WiTH_GRND (TRIG) ESSa TR_WrTHOUT_GRND (TREG)
CST33 TE_W(TH_GRND (TEIG) TE_WfTHOUT_GRND (TEEG)
Figure 5. Classification result of TR and TE with and without
ground points
5.5 Sensitivity analysis on segment size
The derived features from two applied volumetric approaches
are very sensitive to the segment size. For determining the best
segment size we perform a sensitivity analysis on it. The 9
features of the TREG are tested for comparing classification
performance with respect to 2m to 7m segment sizes (Figure 7).
As expected, too large and small segment sizes show worse
classifications due to the multi-class points and scarce points in
a segment. Therefore, 3m is chosen as the best segment size.
: *
—"^---74
Л
• Overall
' \
— Power-line
/ \
——*■ Pylon
/ \
— -» — Building
/ \
' \
—
/
1 T г
, !
2m 3m 4m 5m 6m 7m
Segment Size
Figure 7. Comparison of classification performance on segment
size (with TREG)
5.4 Important feature selection
We selected 9 important features for classifying samples
without ground in section 5.3. The TE without ground was
applied for comparing performance when the RF uses 9 features
6. CONCLUSIONS
This research investigates 21 features which are able to
respectively highlight five classes: ground, vegetation, power
line, pylon, and building, which can be typically found in a
