In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C, Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France, Septeniber 1-3, 2010 
3D CLASSIFICATION OF POWER-LINE SCENE FROM AIRBORNE LASER 
SCANNING DATA USING RANDOM FORESTS 
H. B. Kim a , G. Sohn a 
a GeoICT Lab, Earth and Space Science and Engineering Department, York University, 4700 Keele St., Toronto, ON M3J 1P3, 
Canada - (hskim, gsohn)@yorku.ca 
Commission III, WG 2 
KEY WORDS: LIDAR, classification, power-line scene, random forests, ensemble classifier, Weka 
ABSTRACT: 
Since the introduction of Airborne Laser Scanning (ALS) know as an alternative aerial-based data acquisition tool, the requirement 
of the 3D model reconstruction in both urban and power-line scenes has dramatically increased. Especially, electric utilities 
including power-line and tower are crucial infrastructures that require considerable resources to be monitored and managed 
effectively. For the establishment of the power-line scene inventory, its geospatial information such as positions and attributes of 
power-line networks should be accurately recorded. This paper presents a 3D classification method to classify power-line scene 
where a few structures including trees, transmission lines and pylons would be vertically overlapped. The research proposes two 
different scales of feature extractions from a volumetric space and its embedded points for taking advantages of full 3D analysis 
against conventional 2D pixel-based analysis. With targeted object instances including ground, vegetation, power-line, pylon and 
building, 21 features to characterize each class are extracted from different segment scale. The Random Forest is investigated as an 
ensemble decision classifier to classify power-line scenes with extracted features. 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 class distribution over test 
datasets with or without the separation from training data. Experiments suggest that an optimized classification performance of 96% 
success rate by Random Forest can be achieved with point-based feature extraction and data sets with relatively equal distribution of 
the training data. 
1. INTRODUCTION 
3D object reconstruction in urban, suburban, and power-line 
scenes has become an interesting issue independently on 
specific data: aerial images, ALS (Airborne Laser Scanning), 
and TLS (Terrestrial Laser Scanning). Currently, the frequency 
of the use of ALS data has dramatically increased compared to 
other sensory data due to its advantageous ability of direct 3D 
measurement with high density, accuracy and foliage 
penetration. Recently, a summary of advanced photogrammetry 
and remote sensing technologies using ALS for scientific and 
engineering applications was addressed in Shan and Toth (2008) 
and Vosselman and Mass (2010). However, most of ALS-based 
researches for 3D object reconstruction mainly focused on few 
numbers of urban (i.e., building, terrain and road) and natural 
(trees, canopy and forest) features. Not many research works 
has been reported in the automation of corridor objects such as 
power-line networks. Power-line network is considered as one 
of the most important infrastructures in North America which 
requires reliable monitoring of its safety. Recently, Jwa et al. 
(2009) introduced an automatic algorithm to reconstruct 3D 
transmission models from ALS point of clouds using non-linear 
least square regression method. Most of the proposed methods 
for reconstructing 3D urban and natural objects require a 
reliable classification of ALS data based which 3D modelling 
algorithm can be applied. The accuracy of 3D modelling is 
primarily subjective to the classification errors. In addition to 
the importance of achieving high quality of classification results, 
its cost-effectiveness should be concerned, in particular as the 
algorithm deals with massive scale of modelling coverage and 
spatial-temporal applications. This is the case for power-line 
risk management using ALS data. This research addresses the 
problem of knowledge-based classification when it is trained 
with relatively small training samples, but tends to be applied to 
large-scale unlabelled data for power-line scene classification. 
2. RELATED RESEARCH 
Classification approaches can be divided into two categories: 
binary classification and multi-class classification. The binary 
classifier aims to classify given scene into pre-dominant two 
features which is often used for solving the fore-to-background 
problem. For instance, Sohn and Dowman (2002) separated 
ground features from non-ground ones which stand above the 
ground by developing RTF (Recursive Terrain Filter). This 
terrain filter progressively finds terrain points through 
evaluating candidate points identified from the iterative TIN 
defragmentation in downward step and upward step. Baillard 
and Maître (1999) represented an objective function of the 
binary classification to separate ground features from above 
ground ones based on the MRF (Markov Random Field) using 
3D information. Rutzinger et al. (2008) developed an object- 
based point cloud analysis for classifying ALS data into 
vegetation and non-vegetation features using the ALS Full- 
waveform-driven information. Multi-class classification aims to 
classify ALS data into multiple object classes rather than two 
instances at one simultaneous process or by a hierarchical multi 
segmentation. In this category, Axelsson (1999) proposed an 
algorithm which enable to segment ALS data per scan line into 
ground, building, or power-line by applying the MDL 
(Minimum Description Length). Not only using ALS data, 
Lalonde et al. (2006) classified the objects on the terrain into 
three classes: rough surface objects such as grass and tree