Street Tracking Based on SAR Data from Urban Areas Sigurjon Orn Sigurjonsson Jon Atli Benediktsson and Johannes R. Sveinsson Department of Electrical and Computer Engineering University of Iceland Hjardarhaga 2-6, 107 Reykjavik, Iceland sigurjs@ieee.org, {benedikt, sveinsso}@hi.is Gianni Lisini and Paolo Gamba Department of Electronics University of Pavia Via Ferrata 1, 27100, Pavia, Italy {gianni.lisini, paolo.gamba}@unipv.it Jocelyn Chanussot Signal & Images Laboratory - LIS/INGP BP 46 - 38402 St. Martin d’Heres, France jocelyn.chanussot@lis.inpg.fr Abstract—A method for street tracking is proposed. The method consists of two steps. First, a “blob image” of possible street candidates is created. Then, the street segments from that blob image are extracted. Two feature extraction approaches based on mathematical morphology are applied as preprocessing for the street tracking. One method is based on using differential morphological profiles but the other uses morphological opening and closing operators with a rotating structuring element (SE). The method is tested on an AIRSAR image from Los Angeles with and without noise filtering. The obtained results are measured using two indexes; correctness and completeness. Of the two methods used in the feature extraction, the SE rotation appears to give better results. Noise filtering does not have a major effect in street tracking for the AIRSAR image. I. I NTRODUCTION Classification of high-resolution remote sensing images from urban areas has been addressed in [1] and [2] using two main steps: - feature extraction based on the construction of a differential morphological profile (DMP) which characterizes each pixel both in terms of intensity and in terms of local geometry and - classification based on a neural network (eventually after selection of the most significant features). In this paper, a further evolution of [1] and [2] is presented and applied in neural network classification of an AIRSAR image of Los Angeles, California shown in Fig. 1. For the feature extraction step, two methods are used. One applies openings and closings with linear structuring elements under varying angles and the other creates Differential Morphological Profiles as described in [1]. The effect of using speckle filtering prior to the construction of the morphological profiles is also investigated, as is the effect of Alternating Sequential Filters, a family of morphological filters. The maps obtained from the classification of the different morphological profiles are used for street tracking. The street extraction is done in two steps. The first one is aimed at discarding the blobs that do not possess the usual characteristics of the roads such as elongation. This is made by a routine that tries and associate each blob present in the filtered image to a “street prototype” database [3]. If the shape of the blob under test is too different from the ones found in the database, it will be removed from the image. It is possible to remove even only a part of the entire blob that shows too peculiar features. The second step is a skeletonization step reducing the selected blobs to linear segments. To evaluate the quality of the extracted street network, two quantitative indexes are computed, the correctness and completeness indexes [4]. Both of them require the knowledge of the true network and provide a means to understand to what extent the extracted network is similar to the reference one. In particular, completeness represents the fraction of ground truth length extracted while correctness is the fraction of segments’ length belonging to actual roads. II. FEATURE EXTRACTION This section describes two different methods used for fea- ture extraction which - although they are referred to as such - they are actually classification methods themselves. They both rely on neural networks and differ only in the way the input vectors are constructed. On one hand they are created using Differential Morphological Profiles [1] (DMPs) and on the other using morphological openings and closings with linear structuring elements under various angles. Fig. 1. AIRSAR image of Los Angeles, California. 1273 0-7803-9050-4/05/$20.00 ©2005 IEEE. 1273