In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3, 2010 
4. METHODOLOGY 
The combination process was implemented in several stages as 
follow: 
4.1 Filtering of lidar point clouds 
First the original lidar point clouds were filtered to separate on- 
terrain points from points falling onto natural and human made 
objects. A filtering technique based on a linear first-order 
equation which describes a tilted plane surface has been used 
(Salah et al.. 2009). Data from both the first and the last pulse 
echoes w’ere used in order to obtain denser terrain data and 
hence a more accurate filtering process. After that, the filtered 
lidar points were converted into an image DTM. and the DSM 
was generated from the original lidar point clouds. Then, the 
nDSM was generated by subtracting the DTM from the DSM. 
Finally, a height threshold of 3m was applied to the nDSM to 
eliminating other objects such as cars to ensure that they are not 
included in the final classified image. 
4.2 Generation of Attributes 
Our experiments were carried out characterizing each pixel by a 
32-element feature vector which comprises: 25 generated 
attributes, 3 image bands (R. G and B), intensity image, DTM, 
DSM and nDSM. The 25 attributes include those derived from 
the Grey-Level Co-occurrence Matrix (GLCM), Normalized 
Difference Vegetation Indices (NDV1), slope and the 
polymorphic texture strength based on the Forstner operator 
(Fbrstner and Gulch, 1987). The NDV1 values for the UNSW, 
Bathurst and Fairfield test areas were derived from the red 
image and the lidar reflectance values, since the radiation 
emitted by the lidars is in the 1R wavelengths. The resolutions 
of the lidar reflectance data for these study areas are lower than 
that for the images, and this may impact on the ability to detect 
vegetation. Since the images derived for the Memmingen 
dataset include an IR channel, the NDV1 was derived from the 
image data only. The attributes were calculated for pixels as 
input data for the three classifiers. Table 3 shows the attributes 
and the images for which they have been derived. These 
attributes have been selected to be uncorrelated based on the 
problem of correlation between feature attributes. All the 
presented attributes were used for every test area. A detailed 
description of the filtering and generation of attributes process 
can be found in Salah et al. (2009). 
attribute 
Red 
Band 
Green 
Band 
Blue 
Band 
Intensity 
DSM 
nDSM 
PTS 
V 
V 
V 
V 
V 
V 
HMGT 
V 
V 
V 
V 
V 
V 
Mean 
V 
V 
V 
V 
V 
V 
entropy 
V 
V 
V 
V 
V 
Slope 
x 
X 
X 
X 
X 
V 
Table 3. The full set of the possible attributes from aerial 
images and lidar data. V and x indicate whether or not 
the attribute has been generated for the image. PTS 
refers to polymorphic texture strength; HMGT refers 
to GLCM/homogeneity; Mean refers to GLCM/ 
Mean; entropy refers to GLCM/ entropy. 
4.3 Land Cover Classification 
In this work, we have used the Self-Organizing Map (SOM), 
Classification Trees (CTs), and Support Vector Machines 
(SVMs) classifiers to estimate the class memberships required 
for the combination process. 
Support Vector Machines (SVMs) 
SVMs are based on the principles of statistical learning theory 
(Vapnik. 1979). SVMs delineate two classes by fitting an 
optimal separating hyperplane (OSH) to those training samples 
that describe the edges of the class distribution. As a 
consequence they generalize w^ell and often outperform other 
algorithms in terms of classification accuracies. Furthermore, 
the misclassification errors are minimized by maximizing the 
margin between the data points and the decision boundary. 
Since the One-Against-One (1A1) technique usually results in a 
larger number of binary SVMs and then in subsequently 
intensive computations, the One-Against-All (1AA) technique 
was used to solve for the binary classification problem that 
exists with the SVMs and to handle the multi-class problems in 
aerial and lidar data. The Gaussian radial basis function (RBF) 
kernel has been used, since it has proved to be effective with 
reasonable processing times in remote sensing applications. 
Two parameters should be specified while using RBF kernels: 
• C, the penalty parameter that controls the trade-off 
between the maximization of the margin between the 
training data vectors and the decision boundary plus the 
penalization of training errors 
• y, the width of the kernel function. 
In order to estimate these values and to avoid making 
exhaustive parameter searches by approximations or heuristics, 
we used a grid-search on C and y using a 10-fold cross- 
validation. The original output of a SVM represents the 
distances of each pixel to the optimal separating hyperplane, 
referred to as rule images. All positive (+1) and negative (-1) 
votes for a specific class were summed and the final class 
membership of a certain pixel was derived by a simple majority 
voting. 
Self-Organizing Map Classifier (SOM) 
The SOM undertakes both unsupervised and supervised 
classification of imagery using Kohonen’s SOM neural network 
(Kohonen, 2001). SOM requires no assumption regarding the 
statistical distribution of the input pattern classes and has two 
important properties: the ability to learn from input data; and to 
generalize and predict unseen patterns based on the data source, 
rather than on any particular a priori model. In this work (Salah 
et al., 2009), the SOM has 32 input neurons which are: 25 
generated attributes, 3 image bands (R, G and B), intensity 
image, DTM, DSM and nDSM. The output layer of an SOM 
was organized as a 15 x 15 array of neurons as an output for the 
SOM (255 neurons). This number was selected because, as 
recommended by Hugo et al. (2006), small networks result in 
some unrepresented classes in the final labelled network, while 
large networks lead to an improvement in the overall 
classification accuracy. Initial synaptic weights between the 
output and input neurons were randomly assigned (0-1). In the 
output of the SOM, each pixel is associated with a degree of 
membership for a certain class. 
Classification Trees (CTs) 
The theory of Classification trees (CTs) (also called decision 
trees) was developed by Breiman et al. (1984). A CT is a non- 
parametric univariate technique built through a process known 
as binary recursive partitioning. This is an iterative procedure in 
which a heterogeneous set of training data consisting of