Remote Sensing Image Retrieval via Symmetric Normal Inverse Gaussian Modeling of Nonsubsampled Shearlet Transform Coefficients

This paper presents a nonsubsampled shearlet transform (NSST) based remote sensing (RS) image retrieval technique where the NSST subband coefficients are modeled with symmetric normal inverse Gaussian (SNIG) distribution. The NSST is popular for its very good directional selectivity and shift invariance. The SNIG is a four parameter distribution which can describe a wide range of heavy tailed distributions. Through Kolmogorov-Smirnov (KS) goodness of fit test, it is shown that the SNIG best approximates the shape of distribution of detail NSST coefficients. For Maximum Likelihood (ML) estimation of SNIG parameters, an Expectation-Maximization (EM) type of algorithm is used. The RS images are first decomposed with NSST and the feature vector is formed with the estimated SNIG parameters from detail NSST coefficients. Here, the UC Merced land use (UCM) and the High-resolution satellite scene (HRS) datasets are used for the experiments conducted. The results of the proposed method reveal superior performance, when compared with some well known techniques.