HYPERTEMPORAL SAR SEQUENCES FOR MONITORING LAND COVER DYNAMICS Paolo Gamba, Fabio Dell’Acqua and Giovanna Trianni Department of Electronics, University of Pavia Via Ferrata, 1, 26100, Pavia, Italy phone: + (39) 0382 985781, fax: + (39) 0382 422583, email: {name.surname}@.unipv.it web: tlclab.unipv.it Keywords: multitemporal analysis, image mining, land cover monitoring. ABSTRACT The availability of long sequences of SAR data virtually at any location on the Earth allows an extensive evaluation of land cover dynamics, which is a well-known extension to image pair multitemporal analysis. In this work we show the first results of an extensive evaluation of one among these “hypertemporal” SAR sequences, with a stress on urban area monitoring. 1. INTRODUCTION The availability of very long time series of SAR data is a very interesting effect of the long and successful missions carried out mainly by the European Space Agency. ERS-1 and ERS-2 have been providing a consistent data set of ob- servations of the Earth surface since the first of these two satellites was launched. While many of current missions (ENVISAT, TerraSAR-X) are not meant to collect continuously data all over the world, other ones are (e.g. ALOS/PALSAR). Moreover, the data already in the archives represents a huge collection of in- formation, which has been underutilized so far. Indeed, the most important application of these long series of data is Differential SAR Interferometry [1], mainly for subsidence monitoring [2,3] and, generally speaking, for effects with a rather long temporal scale involving the third dimension. Of course, multi-temporal SAR data sets has been used in many applications, but the point which we want to stress here is that series of hundreds of observations are available in the archives and little has been done to ex- ploit the potential coming from their temporal extent. To the knowledge of the authors, a few, important, exceptions to this lack of interest are the analysis of long-term behavior of urban areas by means of interferometric coherence analysis [4]. Moreover, application to forest mapping as in [5] and rice monitoring as in [6] has proved feasible, but no general application to many diverse land covers using long mul- titemporal SAR sequences has been considered. As a matter of fact, even in the cited papers no full exploitation of very long temporal series has been attempted. This work is devoted to the analysis of these series, which, in accordance with the usual notion of “hyper-spectral” re- motely sensed images, where the number of spectral bands is very large, will be named throughout this work as “hyper- temporal'” data series. Moreover, the focus of this work will be on amplitude ground range data. This is done on purpose, to show how even this very basic information, an elaborated subset of the rich backscattered values recorded by any SAR sensor in its original slant-range projection may provide interesting results by itself. 2. HYPERTEMPORAL SAR SEQUENCE ANALYSIS Let us assume that a hyper-temporal data series X n , n ∈ 1, ..., N { } is available, with N >> 1 . The simplest approach that could be taken with these data is the same as with multi-spectral images, where X n i , j ( ) has no longer the meaning of a “spectral” response, but it is now the “tempo- ral” response of the imaged ground area at the (i,j)-th pixel location. However, unlike multi-spectral images, where the bands are acquired simultaneously, here the temporal align- ment of the pixels belonging to each temporal band to the previous and the following bands is not guaranteed. Luckily, this problem has already been studied, as for SAR data is concerned, for differential interferometric applications. Therefore, commercial software is already available now to achieve subpixel accuracy in multi-temporal SAR image coregistration. In any case, some deegree of inaccuracy in these procedures could not be avoided. Therefore, in the following the discussion will be split between per-pixel and per-region approaches. The overall processing tree for hy- pertemporal SAR sequence processing in both approaches is shown in fig.1. 2.1 Per-pixel analysis Even when sub-pixel co-registration accuracy might be con- sidered as achieved, amplitude SAR data is still affected by speckle noise. Therefore, unlike multi- or hyper-spectral data, no classification procedure should be directly applied to the multi-temporal pixel X n i , j ( ) . A wiser approach would require a reduction of the noise by means of a mul- titemporal speckle filter like in [7]. While the optimal ap- proach would require to compute the complex correlation between any pair of images to achieve the largest possible improvement in the Equivalent Number of Looks (ENL), the