Despeckling PolSAR Images with a Structure Tensor Filter


1 CTIM, Universidad de Las Palmas de Gran Canaria, Spain
2 LaCCAN, Universidade Federal de Alagoas, Brazil

Index


Abstract

In this work we propose a new despeckling filter for Fully PolSAR (Polarimetric Synthetic Aperture Radar) images defined by 3x3 complex Wishart distributions. We first generalize the well-known structure tensor to deal with PolSAR data which allows to efficiently measure the dominant direction and contrast of edges. The generalization includes stochastic distances defined in the space of Wishart matrices. Then, we embed the formulation into an anisotropic diffusion-like schema to build a filter able to reduce speckle and preserve edges. We evaluate its performance through an innovative experimental setup that also includes Monte Carlo analysis. We compare the results with a state-of-the-art polarimetric filter.

For a more detailed information about this technique see reference [1] quoted below. Please, cite such reference in your own research works if you use any of the results, data or software included here.


Experimental dataset and results

We have performed experiments with synthetic and real data. We compare our results with a state-of-the-art polarimetric non-local means filter (SDNLM: stochastic distance non-local means) [2].

Related to synthetic data, in the original paper [1], 2000 replicas of 240x240 pixels size with L=3 were generated in order to perform a Monte Carlo analysis. Regarding data from Operational Sensor, a well-known 4 looks intensity PolSAR image belonging to the region of Flevoland in the Netherlands was used.

In both cases, ground truth images are provided in order to asses the filters' performance. Unlike the traditional evaluations carried out on manually selected regions, we propose to use 5 different classes in the case of the synthetic data, and 14 classes for the actual PolSAR data case. In this way, a more precise evaluation is performed, including an assessment of edge preservation naturally. Moreover, using the proposed evaluation, regions with different polarimetric signatures can be evaluated. Furthermore, we also include the measure of the mean preservation index (MPI) in order to ensure the preservation of the mean in the whole image after filtering.

The first row of Table 1 presents the results (Pauli codification) obtained with one the replicas of the synthetic data, while the second one includes the results regarding actual PolSAR data. From left to right we have the images corresponding to the original noisy PolSAR data, the result of the SDNLM filter and of the structure tensor one. You can enlarge the images by clicking on any of them. The implementations described in the software section include both PolSAR data.

Table 1: results of the proposed technique compared with a state-of-the-art technique for synthetic and actual PolSAR data (see [1] for further explanations). From left to right: the original noisy data, SDNLM filter, and structure tensor filter
Synthetic Data
Observed data
Observed data
SDNLM
SDNLM
Structure Tensor
Structure Tensor
Actual PolSAR Data
Observed data
Observed data
SDNLM
SDNLM
Structure Tensor
Structure Tensor


Software

An implementation of the proposed technique (for Microsoft Windows platform) and of the stochastic distance non-local means (SDNLM - in Matlab) filter are available. Using them, you can reproduce the results presented on the paper. Both implementations include:

  • The software/implementation files.
  • The synthetic and actual PolSAR images represented as 9 .txt files with the different bands.
  • A BMP image with the ground truth for each one of the PolSAR data.
  • A readme file with a description of the contains and instructions of use.

Related to the outcomes, the SDNLM filter provides:

  • A BMP image with the Pauli codification of the input noisy image.
  • A BMP image with the Pauli codification of the filtered image.
  • A .mat file named "Observed_Filtered.mat", which contains the original observed data, S, and the filtered one, F.

On the other hand, the outcomes of the structure tensor filter are:

  • A BMP image with the Pauli codification of the input noisy image.
  • A BMP image with the Pauli codification of the filtered image.
  • A text file in CSV format with the statistics for the mean and standard deviation inside the different ROIs and in the whole image, including the equivalent number of looks (ENL) as well, before and after filtering. (Note that the variations of the mean and standard deviation of the filtered image are indicated as percentages of the initial values)

Releases


Acknowledgments

This research has been partially supported by the MINECO project reference MTM2016-75339-P (AEI/FEDER, UE) (Ministerio de Economía y Competitividad, Spain), and by CNPq and Fapeal (Brazil). Synthetic images have been generated by using Matlab, and based on the Wishart distributions described in [2]. Actual PolSAR data, belonging to the region of Flevoland in the Netherlands, is freely available at the ESA sample datasets.


References

  • [1] Santana-Cedrés, D., Gomez, L., Alvarez, L., and Frery, A. C.: Despeckling PolSAR Images with a Structure Tensor Filter. (submitted to the special issue on Advanced Statistical Techniques in SAR Image Processing and Analysis of the Geoscience and Remote Sensing Letters (GRSL) journal) - Under Revision
  • [2] Torres, L., Sant'Anna, S.J., da Costa Freitas, C., and Frery, A.C.: Speckle reduction in polarimetric SAR imagery with stochastic distances and nonlocal means. Pattern Recognition, vol. 47, no. 1, pp. 141-157, 2014.

Copyright

The software, results and the PolSAR data included in this page are distributed under the license: CC Creative Commons "Attribution-NonCommercial-ShareAlike" see http://creativecommons.org/licenses/by-nc-sa/3.0/es/deed.en