Land-cover Classification in SAR Images using Dictionary Learning Gizem AKTAŞ* a , Çağdaş BAK a , Fatih NAR a , Nigar ŞEN a a SDT Space and Defence Technologies, Galyum Blok ODTU Teknokent Ankara, Turkey ABSTRACT Land-cover classification in Synthetic Aperture Radar (SAR) images has significance in both civil and military remote sensing applications. Accurate classification is a challenging problem due to variety of natural and man-made objects, seasonal changes at acquisition time, and diversity of image reconstruction algorithms.. In this study, Feature Preserving Despeckling (FPD), which is an edge preserving total variation based speckle reduction method, is applied as a preprocessing step. To handle the mentioned challenges, a novel feature extraction schema combined with a super-pixel segmentation and dictionary learning based classification is proposed. Computational complexity is another issue to handle in processing of high dimensional SAR images. Computational complexity of the proposed method is linearly proportional to the size of the image since it does not require a sliding window that accesses the pixels multiple times. Accuracy of the proposed method is validated on the dataset composed of TerraSAR-X high resolutions spot mode SAR images. Keywords: SAR, land-cover classification, dictionary learning, sparse representation, superpixel 1. INTRODUCTION In remote sensing applications, classification of the land-cover (like forest, urban, water/ice, farm-land i.e. crop, oil-slick etc) is important for controlling deterioration of the environment and destruction of wetland, for urban region planning, natural resources monitoring, and for collecting information on possible future disasters [1]. Land-cover classification can be implemented using images acquired from various types of sensors. Synthetic Aperture Radar (SAR) is a radar that collects information from the earth surface and generates high resolution images of wide areas. Since SAR is an active sensor it can generate proper images independent of time of day and weather conditions such as fogs, clouds, and rains. SAR is mostly used for surveillance of wide areas, by employing change detection [5], target detection, target recognition [3], [4], [7], automatic extraction of curvilinear features [6], sea / ice monitoring, mining [8], oil spill pollution detection [9], and land-cover classification [10] algorithms. On the contrary, SAR images suffer from speckle noise which may occur due to sensor and/or platform parameters, reconstruction algorithm, weather conditions and atmospheric effects... Therefore, classification algorithms should include routines for solving problem of diversity of noise and features of the images which causes degradation in performance of the algorithms In general, classification algorithms can be analyzed in two steps for land-cover classification in SAR images. The first step is feature extraction from input data set and classifier training by extracted features. The second step is analyzing the performance of trained classifier using seen images (training data) and unseen images (test data). Performance result on seen images shows the success of the selected features and ability of classifier for separating target classes. Performance result on unseen images shows the generalization capability and robustness of the trained classifier. Features are extracted by using two different approaches, the first method uses a sliding window with a fixed size and features are extracted from the pixels inside the sliding window. This sliding window traverses the entire image sequentially and this leads to inefficient use of processing time. In addition, deciding on the size of the window is critical. While the small sized window is computationally efficient; it may not be sufficient to obtain statistically powerful and accurate results. On the other hand, if the window size is increased, computational load increases, and also there is no guarantee to increase the classification accuracy, since there may exist multiple class information in defined window and it causes to decrease the statistical power. The second one extracts the features by segmenting the entire image into small regions. This feature extraction approach uses segmented regions for determining statistical and textural attributes in salient locations. For both method, the features are extracted according to intensity values of the pixels and there are various approaches [11-17] to extract the features like statistical computations, textural descriptions, and filter based analysis. The first one is also called as general features and they are generally computed from the known statistical