In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3, 2010 
investigated for detecting a vertically formed structure like 
electric pylon. Pylon and high vegetation have high values. 
Continuous off-segment: This is the opposite of continuous 
on-segment. Overhead power-line would have high count 
because of empty space between ground and power-line. 
4. 3D CLASSIFICATION WITH RANDOM FORESTS 
Every voxel and every point respectively posses the 21 features 
derived through the voxel- and point-based feature extraction 
when there are more than a given number of member points. For 
voxel-based feature extraction, the all feature values of each 
voxel are assigned to its member points. After that, the Random 
Forests (RF) is applied to points which a pair of the 21 features 
have been assigned to by the two feature extraction approaches 
in order to tag them into one of following classes: ground, 
building, vegetation, wire, and pylon. The RF is an ensemble 
classifier which is able to generate considerable number of 
decision trees learned by a partial or entire training data set and 
then derive an optimal tree which minimizes the generalization 
error among previously generated trees. For the RF we used a 
customized Weka 3.5.1 which includes some implemented 
functions - the evaluation of variable importance, interactions 
and proximities between individual decision trees, and so on. 
The variable importance of the customized Weka is perfectly 
same as Breiman’s algorithm (Breiman, 2001). For training, the 
RF fixed M = 4 and T = 60, which mean the randomly selected 
feature number at every node split and the number of populated 
trees, in order to lead a learned classifier to be independent on a 
training sample. 
5. EXPERIMENTAL RESULT 
5.1 Training set (TR) and test set (TE) 
The test data set was acquired along to east and west of Folsom, 
California, USA in August for the purpose of power-line 
management in order for violation clearance against vegetation. 
The coverage of the data is approximately 6870(EW)xl263(SN) 
m 2 and it was collected using Riegl Q560 with 30/m 2 of the 
point density on average. Two subsets are taken from regions 
2.5km away in the original data set: a training set (TR) and a 
test set (TE). The both data contain not only terrain, vegetation, 
power-line, power-line tower, and building which are 
interesting classes in this study, but also are manually or semi- 
automatically classified by a worker with a plenty of experience. 
We regard the manual classification output as a ground truth. 
The scene similarity between TR and TE considerably affects 
the classification result of TE. This is because the classifier 
learned by TR depends on the extracted features and they 
computed from TE would be similar to TR’s. Table 1 
summarizes the data characteristics between TR and TE. The 
likeness would be expected to lead a good classification result 
of TE, but our experiments more focus on the comparison of 
two approaches with respect to the feature extraction depending 
on a segment scale (point or voxel) and sensitivity analysis 
regarding the class uniformity. Thus, the main experiments are 
categorized into two tests: voxel-wise versus point-wise feature 
extraction, and classification with and without ground points. 
Another experiment is to determine the best segment size 
through a sensitivity analysis. 
Setting 
TR 
TE 
Points 
194,289 
484,092 
Point density 
38.75/m 2 
32.29/m 2 
Surface slope 
0.5° 
0.05° 
Bldg, roof type 
Gable 
Gable 
Power-line voltage 
115kV 
115kV 
Pylon type #1 
Lattice 
Lattice 
Pylon type #2 
3-level 
3-level 
Tree proximity 
close to building 
and pylon 
close to building 
Table 1. Data similarity of TR and TE 
5.2 Voxel-wise vs. Point-wise Feature Extraction 
General approaches for feature extraction which dealt with 3D 
point cloud average feature values with member points of a 
segment or interpolate them with neighbouring segments. That 
method is called segment-wise feature extraction. In the 
contrary, point-wise method handles the original values of 
points without any interpolation and simplification. We derived 
21 features for each voxel and each point based on two different 
volumetric approaches with respect to segment scale: voxel- 
wise and point-wise. After that, the feature values extracted by 
voxel-based method are assigned to member points of each 
voxel. That is, each point has two values corresponding to a 
feature variable. The RF was applied for comparative analysis 
of the performance of the two extraction approaches. The TR 
was split into 1/10 for training and 9/10 for validating (leave- 
one-out), which is similar to 10-fold cross validation typically 
used in machine learning. 
Class 
Gmd 
Veget 
Wire 
Pylon 
Bldg 
Accuracy(%' 
Gmd 
79,215 
49 
0 
0 
14 
99.92 
Veget 
82 
63,447 
96 
3 
429 
99.05 
Wire 
2 
95 
5,803 
78 
22 
96.72 
Pylon 
2 
55 
105 
912 
4 
84.60 
Bldg 
23 
404 
30 
0 
23,848 
98.12 
Table 2. Confusion matrix for TR using 
(M=4 and T=60) 
voxel-wise method 
Class 
Gmd 
Veget 
Wire 
Pylon 
Bldg 
Accuracy(%' 
Gmd 
79,206 
57 
5 
0 
10 
99.91 
Veget 
111 
63,449 
26 
4 
467 
99.05 
Wire 
3 
93 
5,840 
62 
2 
97.33 
Pylon 
2 
43 
38 
995 
0 
92.30 
Bldg 
26 
679 
46 
0 
23,554 
96.91 
Table 3. Confusion matrix for TR using point-wise method 
(M=4 and T=60) 
The confusion matrices corresponding to the validation result of 
the trained classifiers with respect to the two feature extraction 
approaches are given in table 2 and table 3. The point-wise 
method is more uniform than the voxel-wise method in the 
classification accuracy of each class. This happened because the 
spherical volume for neighbourhoods search is more valid than 
arbitrary segmented voxel in terms of a certain point. 
Nevertheless, the voxel-based approach led for a reasonable 
success rate. In addition to its good accuracy, it could be applied 
for quick classification of a large-scale corridor data thanks to 
its short computing time in feature extraction. However, we 
applied the point-wise approach for all next experiments. In 
both table 2 and table 3 the learned classifiers have a 
considerable uncertainty between vegetation and building 
compared to the others. It is because some of trees stand very 
closely to building and some of the trees’ branches were 
extended right on building roof top in the TR.