In: Paparoditis N., Pierrot-Deseilligny M„ Mallet C., Tournaire O. (Eds). 1APRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
canopy, linear-like features including wires and tree branches, 
and smooth surface such as ground surface, rocks or large 
trunks using range data collected by laser mounted mobile 
mapping system. They demonstrated three object classes can be 
separable by Bayesian classifier which investigates geometric 
salient features captured from laser point distribution in a local 
neighbourhood. The limitation of deterministic classification 
methods mentioned above is the use of user-specified thresholds. 
A new approach for classification of ALS data is to use machine 
learning techniques: Support Vector Machine (SVMs), decision 
tree, boosting, neural networks and so on. The Random forests 
(Breiman, 2001) is recently emerged as a state-of-the-art 
machine learning technique which considers ensembles of 
decision trees, rather than single tree generated by a base 
classifier. The advantage of Random Forest (RF) does not rely 
on use-specific thresholds for its decision. Furthermore, there 
nearly is no limitation with the number of feature variables 
required for decision node splitting. Compared with other 
classifiers, the accuracy of the RF is as good as Adaboost which 
is well known classifier with high success in machine learning. 
Additionally, the RF automatically sorts feature variables on the 
basis of variable importance for best splitting at each node. 
Narayanan et al. (2009) apply ensemble classifiers including the 
RF for generating under-water habitat maps using the Optech’s 
SHOALS system. They choose the best classifier through a 
quantitative comparative analysis of the ensemble classifiers 
and apply the selected classifier for benthic habitat classification. 
Carlberg et al. (2009) labelled the ALS data into a few of object 
classes including water, ground, roof, tree, and others. The 
method was developed based on a cascade of binary classifiers 
specifically trained for the individual class. The unlabelled 
lasers points were progressively classified by a set of the 
proposed binary classifiers using the RF. Chehata et al. (2009) 
produced a RF classifier using several important features among 
designed 17 features for the purpose of urban scene 
classification. The ensemble classification was accomplished 
using 2D rasterized data which might lead to an ambiguity 
between building and ground. 
In this paper we investigate the potential of the RF classification 
for power-line modelling application using the ALS point of 
clouds. An ultimate goal of the research is to apply a 
knowledge-based classifier trained with small training sample to 
large-scale unlabelled power-line corridors. In order to achieve 
this goal, this paper conducts a sensitivity analysis in terms of 
feature extraction scale, feature importance and feature 
distribution over test datasets with or without the separation 
from training data. In addition, the presented research focuses 
on the classification advantages gained by a 3D profile analysis 
which extract classification features computed from vertical 
segments. A vertical distribution of classified features and their 
topological and semantic relations can be used as additional 
reasoning cues for rectifying the RF classification results 
populated in a point-wise processing manner. As address in 
Chehata et al. (2009), an uncertainties where several objects are 
overlapped at one place (e.g., trees standing nearby buildings) 
can be resolved if a 3D analysis for its vertical superposition is 
investigated. This paper is outlined as follows: the next section 
describes a pre-processing procedure to detect terrain features 
for attributing the above-ground points in height and multiple- 
feature extraction for the RF decision. In section 4, we describe 
the Random Forest as a selected ensemble classifier which is 
trained using features computed from 3D volumes. Finally, we 
present experimental results and draw conclusions with respect 
to the sensitivity analysis of the RF in terms of feature 
extraction scale, feature importance and feature distribution 
over test datasets with or without the separation from training 
data. 
3. PREPROCESSING AND FEATURE EXTRACTION 
3.1 Terrain filtering 
The most of power-line networks are built in open spaces where 
ground feature are predominant compared to the occupancy of 
the other features. Under these circumstances, the ground points 
are typically recognized as a dominant feature. If any object (i.e., 
terrain feature for power-line scene) feature over-occupies a 
given scene, its biased feature distribution would cause errors in 
a fair training of the machine learning algorithms. Thus, for 
improving decision tree training, the contribution of the terrain 
feature should be excluded in advance. The efficiency of 
currently existing terrain filtering algorithms which have been 
developed over the past years has been already validated. In this 
study, we use a RTF (Recursive Terrain Filter) developed by 
Sohn and Dowman (2002) to identify the terrain feature. 
However, the classified terrain is taken into consideration while 
computing feature variables. 
3.2 Segmentation 
LIDAR is usually pre-processed in three ways depending on the 
segment scale for the purpose of reducing the complexity and 
simplifying a scene: point-based, 2D grid-based, and voxel- 
based. The first method has an advantage of that each point 
possesses original values on features, but it has a drawback to 
the processing time due to handling individual points. A 
spherical volume is traditionally used to find neighbours of a 
certain point. The second might make a result much more 
quickly, but the height of each point could be neglected due to 
the dimensional reduction and the height interpolation. The last 
way has the benefit of maintaining three dimensional 
information of point and shortening its throughput at the same 
time. 
In the both voxel-based and sphere-based segmentation, the 
most important factor is the segment size. It is generally decided 
by the space which certain patterns can be extracted from the 
member points in. When it is too big, points corresponding to 
two or more different classes could be combined in a huge 
segment. On the other hand, it could be difficult to identify the 
part of a structure due to few points of a voxel if it is too small. 
In this study it is a minimum size to recognize that the member 
points of a segment are a part of a natural or a man-made 
structure, and it is determined by a sensitivity analysis on 
extensive segment sizes as performed in section 5.5. The 
optimal segment size dependent on the unit of 3D point cloud is 
chosen to 3 meter or 10 feet. In spite of the optimal size, 
however, the mixture of points belonging to two or more 
structures surely exists at the place where two structures are 
linked. For instance, all vegetations are always grown on the 
ground and vegetations under the clearance violation reach 
power-lines. The classification errors would frequently occur 
among these segments. The 3m of voxel and 1.5m radius of 
sphere are respectively used for voxel- and point-based feature 
extraction. 
3.3 Feature variables 
The feature variables are computed from the geometry of 
member points within the volumetric space segmented by voxel 
(voxel-based) and sphere (point-based). For point-based feature 
127