Remote Sensed Images Segmentation through Shape Refinement G.Gallo, G.Grasso, S.Nicotra and A.Pulvirenti Dipartimento di Matematica e Informatica Universit` a di Catania Viale A.Doria 6, 95125, Catania, Italy gallo,ggrasso,snicotra,apulvirenti @dmi.unict.it Abstract A novel approach to the automatic classification of re- mote sensed images is proposed. This approach is based on a three-phase procedure: first pixels which belong to the areas of interest with large likelihood are selected as seeds; second the seeds are refined into connected shapes using two well known image processing techniques; third the results of the shape refinement algorithms are merged together. The initial seed extraction is performed using a simple thresholding strategy applied to index. Subsequentlyshape refinement through Seeded Region Grow- ing and Watershed Decomposition is applied, finally a merg- ing procedure is applied to build likelihood maps. Experi- mental results are presented to analyze the correctness and robustness of the method in recognizing vegetation areas around Mount Etna. Keywords: Remote sensing, image processing, classifica- tion. 1 Introduction Remote sensed images have a growing relevance as sour- ces of information about resources and environment mon- itoring and planning. The need for better algorithms for extracting, processing and making easily accessible the in- formation coming from satellite images has been perceived as one of the first priorities by the Pattern Recognition and Computer Vision research community. As a consequence the literature about methods and techniques suitable for pro- cessing remote sensed images is by now very large and in- cludes application of many well known methods for classi- fication and image analysis [5, 7, 17, 18]. An important issue in the analysis of remote sensed data is to segment and contour interest regions from a satellite im- age. For example extracting areas containing a particular type of vegetation or water basins. This is a relatively easy task provided that such regions may be spectrally character- ized in a robust and reliable way [6, 7, 8]. Unfortunately this is rarely the case: single pixels may include heterogeneous areas due to the resolution limits of the remote imaging sys- tem. Moreover even a well defined spectral signature may be corrupted by sensor noise, atmospheric disturbances and seasonal variations [6, 14, 15, 17]. In this paper we propose a novel approach to the problem of automatic segmentation of remote sensed satellite images. The method presented here includes three main phases. The first phase consists of a preliminary classification based purely on the grade of affinity of individual pixels to the se- lected spectral signature of interest. Only pixels that belong to the desired class with a high degree of probability are se- lected. Because of this conservative selection criterion the selected areas are generally very fragmented and cover only a minimal fraction of the region of interest - pixels selected in this phase are called ”seeds”. In the second phase the ”seeds” are used as starting points to select larger areas ac- cording to topological closeness, morphological properties and spectral similarity. In particular we propose to use two different well tested algorithms to achieve growing around the seeds: seeded region growing [1, 2, 10, 16] and water- shed decomposition [3, 4, 12, 13, 19]. Details and parame- ter choices to optimise these algorithms for the problem at hand are discussed in the following sections. The choice of these two algorithms has been taken because their outputs are, to a certain degree, complementary. At the end of the second phase all the pixels in a satellite image belong to one of the four subsets of the original im- age: seed pixels, pixels grown around seeds using seeded region growing, pixels grown around seeds using watershed decomposition, and pixels that are left out of the region of interest. This four sets are combined to delineate regions of interest in the third phase using a simple merging proce- dure. We have tested the novel methodology over several exam- ples of images and for different regions of interest. In par- ticular we investigated the effectiveness of the algorithm in