A preliminary statistical analysis of textural features in RADARSAT-1 images of an urban area Paolo Gamba, Fabio Dell'Acqua Department of Electronics, University of Pavia via Ferrata, 1 – I-27100 Pavia (ITALY) tel.: +39 0382 98 59 23, fax: +39 0382 42 25 83 Email: {paolo.gamba, fabio.dellacqua}@unipv.it Abstract Although fine resolution satellite SAR images are increasingly available, extraction of information from coarse resolution images of urban areas still deserves attention, as much information can be contained in features visible at lower resolutions. Recent papers, indeed, have proven that texture measures carry a significant amount of information at the resolutions typical for ERS, RADARSAT-1 and ENVISAT satellites as well as the Shuttle Imaging Radar missions (SIR), allowing acceptably accurate classification with a limited set of classes, generally with a single “urban” class. We have investigated the separability of more, different urban classes in a small multitemporal, multiangle RADARSAT dataset over the town of Pavia in northern Italy, by computing and comparing statistics of textural features over the different classes. In this paper we present and summarise the results of such investigation, highlighting some characteristic behaviours emerged during our experiments. Seasonal variations seem to have a more significant effect on statistical features than the different incidence angles; standard deviation of textures are more stable than the corresponding means; for a given texture, standard deviations are more sensitive to the particular considered class than the mean texture value. Keywords: SAR images, statistics, textural features, urban areas. 1 INTRODUCTION The characterization of complex urban environments using satellite data is becoming more and more feasible and interesting thanks to the finer and finer ground spatial resolution of the new sensors increasingly available. Not always, however, a very high resolution is needed for extracting information in such areas. As a matter of fact, the scales of urban features may be very different, and even for classification purposes the relevant scale depends on the level of land cover/land use one is seeking to discriminate [1]. Indeed, recently some literature has appeared describing attempts to exploit the information contained in SAR images at the current resolution, as it has been and is provided by ERS, RADARSAT-1 and ENVISAT satellites as well as the Shuttle Imaging Radar missions (SIR). In urban areas, geometric patterns in which the buildings are organised are expected to result in texture patterns in the SAR images, notwithstanding low resolution. Texture measures have been successfully used to classify [2] different, characteristic areas in the urban environment using SAR images whose resolution is coarser than the basic elements of the urban environment [3]. So far, however, this has been done for a very limited set of classes within the urban environment, and this calls for a deeper investigation of the discriminability of different areas within the same town. To this aim we have collected a set of four RADARSAT-1 images over the town of Pavia in northern Italy, acquired at different times of the year and in different modes, which resulted in different incidence angles. We have also collected a very detailed ground truth [4] by direct inspection of the same urban area, resulted in a dense thematic map containing over ten different urban classes, from “urban meadows” to “mixed residential” to “apartments, with a garden”, and so on. Eight different textural features [5] have been computed for each SAR image. The basic statistical moments –average and standard deviation- of the texture measures over pixels belonging to each class have been determined and stored. The choice of the moments was made with a Max Likelihood classifier in mind, whose operation is based on such moments.