ANALYSES OF DESPECKLING ALGORITHMS FOR POLARIMETRIC SAR DATA

Abstract

Polarimetric SAR (PolSAR) systems can characterize different land features based on decomposition parameters. The estimation of these parameters is often biased due to the presence of speckle noise. Speckle noise in PolSAR returns depicts a peculiar signal-dependent phenomenon and is better characterized by a multiplicative model rather than an additive one. Also, as the noise statistics are far from Gaussian, specialized tools for PolSAR speckle filtering different from the standard image denoising tools are required. PolSAR data is necessarily characterized by the polarimetric/ scattering information and its spatial resolution. So, in the process of speckle filtering, these features of the PolSAR data should be significantly preserved. Indeed, this is the main motivation behind the work carried out in this thesis. Accordingly, this thesis is divided into two main areas of the study. The first part concerns about development and analyses of speckle filters for full-pol SAR data. And the second part deals with the development and analysis of speckle filter for hybrid-pol SAR data. In this process, four novel PolSAR speckle filters are developed and presented that apart from reducing noise, preserve the polarimetric information as well as the spatial resolution of the data. First, the l1-NLM filter that performs better than conventional non-local patch-based PolSAR filter in excessive noise is presented. l1- NLM filter is implemented by the famousWeiszfeld’s algorithm to find the weighted l1-norm distance minimization estimate. Second, the CFAR-PolSAR filter that integrates Wishart based pre-classification to PolSAR speckle filtering is demonstrated. CFAR-PolSAR filter achieves extended noise reduction and edge preservation with lesser computations. Third, a texture classification based filter (TCBF) is presented that exploits the differences between texture variations and speckle heterogeneity in the PolSAR data. The speckle noise generates a heterogeneity pattern in PolSAR data that is distinct from textural variations due to heterogeneous media. Also, the K-distribution similarity of covariance matrices is derived. Fourth, a speckle filtering approach named as Stokes based sigma filter (SBSF) based on probability density function of Stokes parameters is presented. Also, a new sigma range calculation algorithm depending on degree of polarization and mean intensities is presented. In order to illustrate the relevance of above PolSAR speckle filters, the experiix Abstract ments are conducted over variety of PolSAR datasets. In this work, two full-pol single-look RADARSAT-2 datasets acquired over Mumbai (India) coastal area and San Fransisco (USA) bay area are used. A four-look full-pol AIRSAR dataset acquired over San Fransisco (USA) bay area is also used. Analysis of the efficacy of SBSF on real hybrid-pol SAR data is demonstrated on single-look hybrid-pol RISAT-1 data acquired over Mumbai city (India). Apart from these real PolSAR datasets, simulated datasets are generated through Monte Carlo simulation approach and analyzed for evaluation of the filtering performance. In summary, this thesis contributes in the development of PolSAR speckle filters that: 1) preserve scattering information, textural information, and data statistics, 2) enhance averaging in homogeneous regions, 3) filter heterogeneous regions with preservation of sharp details and edges, 4) un-filter strong targets, and 5) reduce computational complexity.