1
DECISION FUSION OF MULTITEMPORAL SAR AND
MULTISPECTRAL IMAGERY FOR IMPROVED LAND COVER CLASSIFICATION
B. Waske a , J. A. Benediktsson b ’*
a Center for Remote Sensing of Land Surfaces (ZFL), University of Bonn,
53113 Bonn, Germany - bwaske@uni-bonn.de
b ’* Department of Electrical and Computer Engineering, University of Iceland,
107 Reykjavik, Iceland - benedikt@hi.is
KEY WORDS: multisensor imagery, data fusion, classification, SAR, multispectral
ABSTRACT
A strategy for classifying multisensor imagery, consisting of multispectral and SAR data is presented. Each image source is
individually classified by a support vector machine (SVM). In decision fusion the outputs of the pre-classification are combined to
derive the final class memberships. This fusion is performed by another SVM. The results are compared with well-known parametric
and nonparametric classifier methods. The proposed SVM-based fusion approach outperforms all other concepts and improves the
results of a single SVM that is trained on the whole multisensor data set. Moreover the results clearly show that the individual image
sources provide different information and a multisensor approach generally outperforms single-source classifications.
1. INTRODUCTION
Regions with a high degree of agricultural land-use are
investigated in numerous remote sensing based land cover
studies. These areas are characterised by a great temporal
variability and typical spatial patterns of high-frequent land
cover changes between individual agricultural field plots. Thus
single-date approaches are often inefficient, due to great
temporal differences in crop phenology. Multitemporal
applications are more appropriate in this context and improve
the accuracies of existing classifications (e.g., Blaes et al.,
2005).
However, the efficiency of optical imagery is often limited by
weather conditions such as solar illumination and cloud cover.
Hence a reliable generation of image time series within one
growing season is difficult, especially in regions like Central
Europe. This is a drawback, particularly for operational
monitoring systems. SAR data are normally independent from
these factors and in several studies the accuracies were
increased by using multitemporal SAR imagery (e.g., Chust et
al., 2004, Blaes et al., 2005).
Beside the multitemporal data, in other studies, the positive
impact of multisensor data on classification accuracy was
illustrated, e.g., data sets consisting of SAR and multispectral
imagery (Benediktsson and Kanellopoulos, 1999; Chust et al.,
2004, Blaes et al., 2005, Huang, et al., 2007).
The main reason for the success of these studies is the use of the
two different sensor systems that operate in different
wavelengths, ranging from visible to microwave. Consequently
dissimilar land cover information is acquired and the
classification accuracy can be increased by multisensor imagery.
Even if some applications are based on conventional statistical
methods such as the well known maximum likelihood classifier
(Chust et al., 2004, Blaes et al., 2005, Huang, et al., 2007),
those methods are often not ideal for multisensor imagery,
because in the very most cases the class distributions cannot be
modelled by adequate multivariate statistical models
(Benediktsson et al., 1990). Consequently, more sophisticated
strategies are more applicable in this context and Richards
(2005) pointed out that the development of adequate strategies
to combine inherent information content in complementary data
sets is perhaps the greatest ongoing challenge in the field of
remote sensing.
Multiple classifier systems give an interesting approach, which
is applied in several studies to multitemporal and mulitsource
imagery. The general concept of classifier ensemble is based on
training a classifier on resampled input data. Afterwards the
outputs of the independent classifiers are combined to create the
final result.
Brown de Colstoun (2003) applied a decision tree on
multitemporal images from the Enhanced Thematic Mapper-
Plus (ETM+) to differentiate between 11 land cover types.
Ensemble techniques such as boosting were successfully used
to improve the final classification results. In Waske et al. (2006)
a multiple classifier system that is based on a random selection
of input features, was used successfully for classifying
multitemporal Envisat ASAR and ERS-2 imagery. The overall
accuracy was significantly increased compared to the results
achieved by a single decision tree. In Briem et al. (2002)
different single and multiple classifiers were considered for
classifying multisource data. The classifier systems always
outperformed the single algorithms in terms of accuracy.
Gislason et al. (2006) applied the concept of Random Forests to
Landsat MSS image and topographical information and
achieved relatively high accuracies.
Other multiple classifier systems are based on a combination of
different algorithms: In Benediktsson and Kanellopoulos (1999)
a multisensor data set was classified by combining the outputs
of a neural network and a statistical classifier. The two
classifiers were trained individually on the SAR and
multispectral imagery. Afterwards these pre-classification
outputs were combined by decision fusion.
* Corresponding author.