A Distributed Non-Parametric Unsupervised
Classification Algorithm for Remotely Sensed
Optical Imagery
B. Liu, T. Bretschneider
School of Computer Engineering, Nanyang Technological University
N4-02a-32 Nanyang Avenue, Singapore 639798
Tel: +65 – 6790 6045, Fax: +65 – 6792 6559
SINGAPORE
Email: astimo@ntu.edu.sg
ABSTRACT: The previously proposed mean-shift algorithm provides a simple and non-parametric
technique for estimating density gradients, which was generalised to image
classification and segmentation. This paper proposes two extensions to the algorithm, namely an
approach to improve the classification accuracy and the implementation for parallel / distributed
computer architectures. The first issue is addressed by handling the overlapping of ground cover
classes in the spectral space by incorporating the spatial information, while a parallelisation was
developed for the latter issue.
1. INTRODUCTION
The previously proposed mean-shift (MS) algorithm provides a simple and non-parametric tech-nique
for estimating density gradients, which was generalised to image classification and
segmentation (Comaniciu, 2002). It is a partitioning clustering technique augmented by analys-ing
a complex multimodal feature space to delineate arbitrarily shaped clusters in it. Unlike
other approaches this method does not assume any probability density function in advance and
thus is well suited for the application in unsupervised classification with a wide range of possi-ble
input scenes. This is of foremost importance for the targeted platform, i.e. the on-board
execution on the mini-satellite X-Sat (Bretschneider, 2004). Moreover, the algorithm exhibits an
almost linear time complexity and thus is advantageous for the fulfilment of the realtime
processing requirements of the mission. This paper proposes two extensions to the MS algo-rithm,
namely an approach to improve the classification accuracy and the implementation for the
parallel computer architectures of the X-Sat. The emphasis is on a fast and stable classification
algorithm rather than on high precision for a selected number of test scenes since the result is
used as a selection criterion for scenes to be processed further. The analysis of the developed
technique presents the obtained speedup factors by the parallelisation and investigates the in-creased
classification accuracy.
