ABSTRACT We present an automated scheme for segmentation of high mountain glaciers using Fast Adaptive Medoid Shift (FAMS) algorithm and Digital Elevation Model (DEM). FAMS is a non-parametric clustering technique that has been optimized and made data driven from its original Medoid Shift algorithm. 6 Band TM sensor satellite images are fed to FAMS as input along with height, slope and gradient information extracted from a DEM. Clean glacier and debris covered glacier are treated separately. Each glacier having its own regional minima and debris is delineated individually. A unique slope-gradient model is used to separate the debris covered portion from its surrounding and extension rocks as well as to exclude the lateral moraine. The proposed model is independent of the DN values of satellite image bands and therefore is able to perform well even in areas where debris covered glaciers exactly resemble the surrounding rocks. Experiments have been carried out on KaraKoram and Hindukush mountain ranges of Asia and validated against supervised manual segmentation results as well as Google Earth TM imagery. Results have shown our fully automated method to be time efficient, robust and accurate. Key Words: Remote sensing, Fast Adaptive Medoid Shift, Glacier Segmentation, Debris covered glaciers. 1. Introduction The best natural indicator of ongoing climate change is a temporal analysis of changes of mountain glaciers [1,2,3]. Glacier segmentation is a prerequisite for monitoring global climate change, global warming, water resources, glacial hazards and many other important life saving tasks [4]. Segmentation of Glaciers is a very trivial task in remote sensing. Glaciers consist of two main parts; the clean white portion and the debris covered portion which mostly resembles the surrounding rocks. Different methods have been adopted for this purpose using satellite imagery. Multi spectral Satellite images from Thematic Mapper (TM) sensor of the Landsat Satellite are often used for the purpose of delineating glaciers [5]. General categories of glacier delineation are described in [6]. The methods used for Glacier classification can be broadly grouped into 5 categories namely False Colour Composites (FCCs) from Landsat Multispectral Scanner (MSS) and Thematic Mapper (TM) digital data, glacier mapping with TM by manual delineation of the glacier outline, segmentation of ratio images with reflectance thresholds, different supervised classification techniques and calculation of glacier reflection with TM and the comparison with ground measurements. Automated supervised classification based on the maximum-likelihood algorithm was done by [9]. The authors maintain that geomorphologic areas like debris may not be correctly mapped using the multispectral classification technique. A supervised maximum likelihood classification method was used by [10] performing trials using a combination of different input bands and image processing techniques like PCA, the ratio TM-4/ TM-5, and the Normalized Difference Snow Index (NDSI). PCI image processing package was used to attain the results. However, to obtain accurate results the debris covered area had to be delineated manually. Lately a semi-automated method was proposed in [11] which employs a three step approach of band ratioing TM Bands 4/5, classification using an intensity hue saturation (IHS) colour space for discarding the vegetation and fixed slope thresh holding (<24°). Although the results are very encouraging yet they do not rule out the requirement of an expert to remove the errors in misclassification. An equally good method was devised by [12] using Advanced Space-borne Thermal Emission and Reflection Radiometer (ASTER) data. The slope thresh hold was set to <12° making the method non-adaptive. However this approach also suffered from the drawbacks of its predecessor. This paper proposes a unique automated method of glacier segmentation using recently proposed [13] Fast Adaptive Medoid Shift Algorithm (FAMS) as a pre- processing step to cluster the satellite image data. The contributions are two folds; 1) segmentation of clean glacier individually and 2) automatic delineation of debris covered glacier using data driven thresholds. Experiments show that our automated proposed method is more than 90% accurate and fully robust. This paper is organized as follows; Section 2 gives an overview of FAMS, section 3 gives details of our proposed glacier segmentation method while results of experiments are analysed in section 4. A CLUSTERING BASED AUTOMATED GLACIER SEGMENTATION SCHEME USING DIGITAL ELEVATION MODEL Syed Zulqarnain Gilani, Naveed Iqbal Rao Electrical Engineering Dept, College of Telecommunication Engineering, National University of Sciences and Technology, Pakistan {zulqarnain.gilani, naveedi}@mcs.edu.pk