CLASSIFICATION OF PINGPONG COSMO-SKYMED IMAGERY USING SUPERVISED AND UNSUPERVISED NEURAL NETWORK ALGORITHMS M. Penalver, C. Pratola, I. Fabrini, F. Del Frate, G. Schiavon, D. Solimini University of Rome “Tor Vergata”, DISP. Rome, Italy (delfrate@disp.uniroma2.it ) ABSTRACT The novel instruments of the COSMO-SkyMed (CSK) Earth Observation programme, offer an opportunity to explore at various resolutions the information content of X-band signal backscattered with different polarizations. In spite of their potential to render additional information about an area of interest, speckle noise and artifacts make X-band acquisitions difficult to interpret. This is a motivating scenario to explore what (semi-)automatic procedures might be able to offer. This paper is first attempt to process CSK Stripmap PingPong data using two well-known artificial neural network techniques: the supervised backpropagation multilayer perceptron and the unsupervised self-organizing map. Index Terms— Synthetic aperture radar, Artificial neural networks, Multilayer perceptrons, Self organizing feature maps, Image classification 1. INTRODUCTION In November 2010 it was completed the launch of the Italian Space Agency (ASI) COSMO-SkyMed (CSK) Earth Observation programme [1]. CSK provides higher spatial resolution and accuracy data with a shorter revisit time through a constellation of four low-orbit mid-size satellites, each one equipped with a multi- mode high-resolution synthetic aperture radar (SAR) operating at X-band. The second generation system, planned for 2016 as a continuation of, and overlapping with, the current mission, will be able to simultaneously receive and process dual-polarization data of yet higher spatial resolution. CSK’s novel instruments, therefore, offer an opportunity to explore at various resolutions the information content of X-band signal backscattered with different polarizations. However, where the use of centimetric wavelength in multi-polarization configurations might improve the knowledge about the area of interest, importantly contributing to the production of land cover maps, the strong speckle noise and the presence of artifacts affecting X-band acquisitions make classification a less-than-easy task. Such scenario strongly motivates development of efficient (semi-) automatic processing tools that facilitate taking advantage of the growing amount of available data. Several studies [2-14] have revealed the potential of artificial neural network (ANN) empirical computational models [15,16] in processing remotely sensed imagery. In particular, their parallel data processing capability make ANNs well suited for pattern recognition in multi-source data, with no prior knowledge required about the statistical distribution of the classification classes. The work presented in this paper, carried out in the context of the ASI-funded CAMLAND project, is a first attempt to process CSK Stripmap PingPong data with ANNs seeking (semi-)automatic pixel-based classification of images of a sub-urban area. Specifically, two neural models have been studied: a multilayer fully-interconnected feedforward ANN (Multilayer Perceptron or MLP) trained with the Backpropagation supervised learning algorithm [17], and a topology-preserving self- organizing ANN (Self-Organizing Map or SOM) [18]. After a suitable network architecture has been derived, the latter enables a fully automated process, whereas the former requires human intervention only to provide an adequate set of training patterns. Both mechanisms, when properly trained, allow automatic classification of different images characterized by the same geometry of acquisition. 2. METHODOLOGY A land cover identification (manmade structures vs natural surfaces) has been derived from a 15-meter resolution CSK PingPong image of the Tor Vergata University campus on the outskirts of Rome, Italy (Fig.1). The image was acquired in both HV and HH polarizations in right-looking mode and with an incidence angle of about 25°. The PingPong mode provides medium-resolution, wide swath imagery through a strip acquisition implemented by alternating a pair of Tx/Rx polarizations across bursts (HH, HV, VV and VH). Even if the area of interest is currently affected by a strong urbanization process, large buildings, high-density urban sites and small isolated houses, still alternate with wide green areas and cultivated fields. In such scenario, with fine-scale land cover features not immediately identifiable at the resolution of the 5888 978-1-4673-1159-5/12/$31.00 ©2012 IEEE IGARSS 2012