In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C... Tournaire O. (Eds). 1APRS. Vol. XXXV1I1. Part ЗА - Saint-Mandé, France. Septentber 1-3. 2010 
INTEGRATING MULTIPLE CLASSIFIERS WITH FUZZY MAJORITY VOTING FOR 
IMPROVED LAND COVER CLASSIFICATION 
M. Salah 2 3 '*, J.C.Trinder a , A.Shaker b , M.Hamed b , A.Elsagheer b 
a School of Surveying and Spatial Information Systems, The University of New South Wales, 
UNSW SYDNEY NSW 2052, Australia - (m.gomah, j.trinder)@unsw.edu.au 
b Dept, of Surveying, Faculty of Engineering Shoubra, Benha University, 108 Shoubra Street, Cairo, Egypt - 
ahmshaker@link.net, prof.mahmoudhamed@ yahoo.com, alielsagheer@.yahoo.com 
ISPRS WG 111/2 "Point Cloud Processing" 
KEY WORDS: Aerial Images, Lidar, GLCM, Attributes, Hybrid Classifier, Fuzzy Majority Voting. 
ABSTRACT: 
In this paper the idea is to combine classifiers with different error types based on Fuzzy Majority Voting (FMV). Four study areas 
with different sensors and scene characteristics were used to assess the performance of the model. First, the lidar point clouds were 
filtered to generate a Digital Terrain Model (DTM), and then a Digital Surface Model (DSM) and the Normalized Digital Surface 
Model (nDSM) were generated. A total of 25 uncorrelated feature attributes have been generated from the aerial images, the lidar 
intensity image, DSM and nDSM. Three different classification algorithms were used to classify buildings, trees, roads and ground 
from aerial images, lidar data and the generated attributes. The used classifiers include: Self-Organizing Map (SOM); Classification 
Trees (CTs); and Support Vector Machines (SVMs) with average classification accuracies of 96.8%, 95.9% and 93.7% obtained for 
SVMs, SOM, and CTs respectively. FMV was then applied for combining the class memberships from the three classifiers. The 
main aim is to reduce overlapping regions of different classes for minimizing misclassification errors. The outcomes demonstrate 
that the overall accuracy as w'ell as commission and omission eivors have been improved compared to the best single classifier. 
1. INTRODUCTION 
Researchers are continually seeking to improve the performance 
of classifiers in remote sensing. Taking advantage of the 
complementary information about image data provided by 
classifiers based on different mathematical concepts, the next 
natural frontier is the integration of multiple approaches into a 
unified framework. The efficient combination of classifiers, 
should achieve better classification results than any single 
classifier. Kanellopoulos et al. (1997) have demonstrated the 
complementary behaviours of neural and statistical algorithms 
in terms of classification errors. Therefore these classifiers 
result in uncorrelated classification errors and hence higher 
accuracies can then be reached by combining them. 
2. RELATED WORK 
For remote sensing applications, Benediktsson et al. (2007) 
have presented a brief summary of recent developments of 
multiple classifier systems (MCS) in w'hich the optimal set of 
classifiers is first selected and then they are combined by a 
specific fusion method. The aim is to effectively merge the 
results of the classifiers taking advantage of the benefits of each 
while reducing their weaknesses. 
More recently, researchers have investigated classifier selection 
for MCS design. Giacinto and Roli (2001) clustered the 
candidate classifiers according to interdependency and selected 
one classifier from each cluster. Hao et al. (2003) also used a 
heuristic search for classifier selection. Mountrakis et al. (2009) 
presented a hierarchical, multi-stage adaptive strategy for image 
classification. They iteratively applied various classification 
methods, e.g., decision trees, neural networks, identified regions 
of parametric and geographic space where accuracy is low, and 
tested the application of alternate methods, repeating the 
process until the entire image was classified. 
Applications of majority voting (MV) for pattern recognition 
have already been studied in detail in Lam and Suen (1997). A 
trainable variant of majority voting is weighted majority voting, 
which applies a weight to each vote. The weight applied to each 
classifier can be obtained for example by estimating the 
accuracies of the classifiers on a validation set. 
Yu-Chang and Kun-Shan (2009) introduced a multiple classifier 
system for land cover classification. The Bagging and Boosting 
algorithms were investigated as a weighting policy and then an 
adaptive thresholding criterion was defined to account for the 
ambiguities between classes. 
Recent work has focused on deriving the uncertainty map of the 
land-cover prediction, which based on the uncertainty of land- 
cover classification for each pixel. Alimohammadi et al. (2004) 
used maximum likelihood classification algorithm to perform 
the classification and generated uncertainty estimation. 
Another technique which is widely studied in classical classifier 
fusion but less addressed in remote sensing is Fuzzy Majority 
Voting (FMV). FMV theory has already been investigated in 
automatic disambiguation of word senses (Le et al., 2007), but 
this is probably the first attempt to use it for combining 
information derived from different classifiers for improvement 
of land cover mapping. FMV has been proposed in this research 
to further improve the classification performances and 
overcome the shortcomings of the previous approaches of 
combining classifiers, such as sensitivity to noise, 
computational load and the need for parametric statistical 
modeling of each data source. The major motivation of our 
work is to establish a framework to combine classifiers with 
Corresponding author.