Synthetic Aperture Radar Image Clustering with Curvelet Subband Gauss Distribution Parameters

Remote Sens. 2014, 6, 5497-5519; doi:10.3390/rs6065497 OPEN ACCESS remote sensing ISSN 2072-4292 www.mdpi.com/journal/remotesensing Article Synthetic Aperture Radar Image Clustering with Curvelet Subband Gauss Distribution Parameters Erkan Uslu * and Songul Albayrak Department of Computer Engineering, Yildiz Technical University, 34220 Istanbul, Turkey; E-Mail: * Author to whom correspondence should be addressed; E-Mail: ; Tel.: +90-212-383-5764; Fax: +90-212-383-5732. Received: 27 February 2014; in revised form: 29 May 2014 / Accepted: 30 May 2014 / Published: 16 June 2014 Abstract: Curvelet transform is a multidirectional multiscale transform that enables sparse representations for signals. Curvelet-based feature extraction for Synthetic Aperture Radar (SAR) naturally enables utilizing spatial locality; the use of curvelet-based feature extraction is a novel method for SAR clustering. The implemented method is based on curvelet subband Gaussian distribution parameter estimation and cascading these estimated values. The implemented method is compared against original data, polarimetric decomposition features and speckle noise reduced data with use of k-means, fuzzy c-means, spatial fuzzy c-means and self-organizing maps clustering methods. Experimental results show that the curvelet subband Gaussian distribution parameter estimation method with use of self-organizing maps has the best results among other feature extraction-clustering performances, with up to 94.94% overall clustering accuracies. The results also suggest that the implemented method is robust against speckle noise. Keywords: clustering; curvelet transform; synthetic aperture radar; self-organizing maps 1. Introduction Several remote sensing and observation systems are developed for earth surface monitoring, which can be grouped into three main categories: laser-based light detection and ranging (LIDAR), optical sensor-based multi- or hyper-spectral imaging, and microwave-based synthetic aperture radar (SAR). Among these methods, SAR is the most prominent as it has the best atmosphere permeability, better Remote Sens. 2014, 6 5498 resolution and different modes of operation, such as polarimetry and interferometry. SAR imaging is an active imaging system with a microwave transmitter emitting pulsed radio waves and a receiver getting backscattered radio waves. Synthetic aperture utilizes the Doppler effect on microwave-illuminated regions to increase the azimuth direction resolution. The use of the Doppler effect results in increased azimuth resolution with reduced antenna length up to a physically allowed size. Commercially, SAR sensors are carried either by air or satellite platforms. The wavelength used in SAR imaging varies by usage requirements from 65 cm to 0.5 cm. SAR images are contaminated by a form of noise called speckle noise which can be modelled multiplicatively. SAR images are used in areas such as target detection, structure detection, road extraction, ship detection, land use classification, oil spill detection, ice field tracking, disaster aftermath evaluation, etc. These fields of use require a great deal of continuous observation and manual analysis. At this point, the use of automatic analysis tools is inevitable. In SAR literature, pixel-based, region-based and contour-based clustering and segmentation algorithms are applied alone or in a cascaded structure. In [1], iterative region growing with the semantics method based on a Markov random field, edge strength model and region growing is applied for SAR image clustering. In [2], a Markov random field approach for SAR clustering is enriched by introducing a third random variable. Ensemble learning of spectral clustering results based on gray level co-occurancy matrix (GLCM) and wavelet transform is introduced in [3] for SAR imagery. Spectral clustering is carried out by k-means clustering in a projection space, where the transformation matrix is calculated by eigenvectors of the Gaussian similarity matrix of samples. In [4], cascaded implementation of Voronoi tessellation, Bayesian inference and reversible jump Markov chain Monte Carlo (RJMCMC) methods are used for SAR clustering. Voronoi tessellations are used to decompose homogeneous polygonal regions and Bayesian inference and RJMCMC is used for labeling. In [5], the integrated active contour method is introduced. Compared to the active contour method, where image segmentation is defined as an energy minimization problem for a closed curve, the integrated active contour approach defines energy based on the maximum likelihood estimation of parted regions’ gamma distributions. In [6], complex Wishart distribution features are used with Chernoff distance for agglomerative hierarchical clustering. In [7], level set segmentation is used together with the SAR Wishart distribution model. In [8], GLCM calculated on the Gabor filter results in the brushlet space used for SAR clustering. The article is structured as follows: Section 2 gives information about the proposed feature extraction method (curvelet subband µ, σ features), together with benchmark feature sets. In Section 3, the test site, data format and clustering methods implemented are introduced. In Section 4, experimental results are presented with several measures: first, the experimental setup is introduced, followed by a presentation of the accuracies, and finally, clustering maps are given as a means of visual comparison. Section 5 concludes the work emphasizing the important findings. 2. Proposed Method The proposed feature extraction method (curvelet subband µ, σ features) is introduced together with the benchmark methods (original data, speckle reduced data, polarimetric decomposition features) in this section. Remote Sens. 2014, 6 5499 2.1. Benchmark Feature Sets 2.1.1. Original Data The original data is used as a base benchmark feature set for comparison. The original data features are constructed as taking the absolute values of the upper triangular matrix of the coherency matrix. Original data has six features per sample. 2.1.2. H/A/α Polarimetric Decomposition Eigenvalue decomposition of the coherency matrix results in occurrence probabilities of three different scattering processes. The occurrence probabilities Pj (j = 1, …, 3) of these scattering processes are the ratios of relevant eigenvalue λj by the sum of all eigenvalues and can be given in Equation (1) [9]. 𝑃𝑗 = λ𝑗 λ1 + λ2 + λ3 (1) The measure of randomness in the whole scattering process entropy H can be given in Equation (2) based on scattering process probabilities where 0 ≤ H ≤ 1. The lower value of H indicates one dominant scattering process, whereas higher value shows that there is volume scattering and the overall scattering is more random. 3 𝐻=− 𝑃𝑗 log 3 𝑃𝑗 (2) 𝑗 =1 The anisotropy A is the measure of difference in secondary scattering mechanisms and can be given in Equation (3). Anisotropy provides complementary information to ent (...truncated)