INTEGRATION OF INFORMATION-THEORETIC TOOLS FOR POLSAR IMAGEPROCESSING AND ANALYSIS

Sidnei Sant’Anna, Corina da C. Freitas, Leonardo Torres

Divisao de Processamento de ImagensInstituto Nacional de Pesquisas Espaciais

Av. dos Astronautas, 175812227-010 Sao Jose dos Campos, SP – Brasil

Alejandro C. Frery∗

Instituto de ComputacaoUniversidade Federal de Alagoas

Av. Lourival Melo Mota, s/n57072-900 Maceio, AL – Brasil

ABSTRACTPolarimetric Synthetic Aperture Radar (PolSAR) is having anincreasingly positive impact in the Remote Sensing commu-nity since the launch of the first practical fully polarimetricsensor, AIRSAR, in 1985. The availability and relevance ofthese data makes it important to offer users state-of-the-arttechniques for PolSAR image processing and analysis. Oneof the most used platforms is PolSARpro, a freely availablesoftware which includes tools for data processing, extraction,analysis and visualization. As PolSARpro has not yet incor-porated some of the advances stemming from the InformationTheory framework, in this work we present the current state ofthe development of a system which integrates such tools usingTerraLib.

1. INTRODUCTION

Remote Sensing data from Polarimetric Synthetic ApertureRadar (PolSAR) are becoming easily available since the launchof the first practical fully polarimetric sensor, AIRSAR, in1985 [1]. Since then, many sensors have been deployed and therelevance of these images for Remote Sensing is increasinglyrecognized.

Such availability and relevance, along with the peculiarnature of the data [2], makes it important to offer users state-of-the-art techniques for PolSAR image processing and anal-ysis. One of the most used platforms is PolSARpro [3, avail-able at https://earth.esa.int/web/polsarpro/home]. This software includes a number of tools for dataprocessing, extraction, analysis and visualization.

Albeit a complete toolbox, PolSARpro has not yet incor-porated some of the latest advances in PolSAR image pro-cessing and analysis, in particular those stemming from theInformation Theory framework. In this work we presentthe current state of the development of a system which in-tegrates such tools using TerraLib [4, available at http://www.terralib.org/] as platform.

TerraLib’s Web site states that∗Thanks to CNPq, Capes, Fapeal and Fapesp for funding.

TerraLib is a GIS classes and functions library,available from the Internet as open source, allow-ing a collaborative environment and its use for thedevelopment of multiple GIS tools. Its main aimis to enable the development of a new generationof GIS applications, based on the technologicaladvances on spatial databases. TerraLib is a freesoftware.

TerraLib is developed in C++.

2. INFORMATION-THEORETIC TOOLS FORPOLSAR

New tools emerged from Information Theory after the adventof important connections between stochastic divergences, en-tropies, and statistical tests. Pardo et al [5, 6] showed that,under mild conditions,

a) a large class of divergences between distributions can beturned into distances which, once properly scaled, haveknown asymptotic properties, and that

b) a large class of entropies can be compared with testsstatistics which have known asymptotic properties.

Frery et al [7, 8] obtained such statistics under the as-sumption of the scaled complex multivariate complex Wishartdistribution, a widely accepted model for full PolSAR datafrom textureless targets. The authors present examples ofclassification procedures based on those statistics.

These analytic tools were recently used for edge detec-tion [9], and for PolSAR image filtering using NonlocalMeans [10]. Using a numerical approach, Silva et al. [11]devised a supervised classification scheme for bivariateGamma models, useful for dealing with pairs of intensi-ties as those provided by Radarsat-2, Envisat and CosmoSkymed. These tools were developed in several program-ming languages as, for instance, R [12, freely availableat http://www.r-project.org/], OX [13, freelyavailable at http://www.doornik.com/ox/], and

239978-1-4799-7929-5/15/$31.00 ©2015 IEEE IGARSS 2015

IDL/ENVI, as proofs-of-concepts rather than intended forfinal users.

The system here described incorporates those and othertools. The system focus is threefold, namely:

• Ease-of-use: it inherits well-tested user-interactiontools.

• Numerical accuracy: as it employs routines from R,which is well known for being more dependable thanother computational platforms in this respect [14].

• Completeness, since it incorporates not only techniques,but state-of-the-art assessment procedures [15].

3. SYSTEM ARCHITECTURE

The system is conceptually organized in three main operationalparts; cf. Figure 1:

Input: A set of routines for reading and organizing severaltypes of input formats, with emphasis on those currentlyemployed by PolSARpro.

Data Processing: The core of the system, where the proce-dures for estimation, hypothesis testing, classification,filtering and segmentation, as well as the extraction ofquality measures are implemented. It also includes toolsfor PolSAR image stochastic simulation.

Output: A set of routines for saving output data (images,maps, plots and reports) in freely available formats.

figureThree main types of information can be used as input:

cartoon models, training samples or parameters, and images.Cartoon models are specified to serve as classes; cf. Fig 2(a).Training samples are used to obtain the parameters that char-acterize the models used to describe PolSAR data, namely thenumber of looks and the complex covariance matrix. Finally,real or simulated images can be the input, either in covarianceor correlation matrix format.

The data processing part is the system core. It providestools for simulating images, currently under the scaled com-plex Wishart model, but easily extensible to other models [16]as, for instance, the polarimetric K, G0 and G laws, the Kum-merU distribution etc. This simulation requires a cartoonmodel for the classes, and a parameter (or a sample fromwhich the parameter will be estimated) for each class; Fig. 2(b)shows an image formed by data simulated with parametersfrom real samples, assigned to the classes stipulated by thecartoon model shown in . This is described by the dotted greenarrows, and the output, indicated by the solid green arrow, is asimulated image as, for example, the one shown in Fig. 2(c).

PolSAR images can be filtered using the Stochastic Dis-tances Nonlocal Means presented in Ref. [10]. The systemonly needs an image as input (as Fig. 3(a)), and it also outputs

an image (as Fig. 3(b)); cf. dashed light brown and red linesin Fig. 1.

Another segmentation is presented in Fig. 4(b), whichis the result of applying the image segmentation proceduredescribed in Ref. [17] to the original data shown in Fig. 4(a).Using the training samples shown in Fig. 4(c), a classificationcan be obtained using the region classifier system described inRef. [11]; the result is shown in Fig. 4(d). Additionally, as ameasure of the quality of the classification, the same procedurereturns the p-value of each assignment, and its visualization isshown in Fig. 4(e).

The region classifier allows specifying a number of stochas-tic distances, as well as measures of dissimilarity based onentropies. These distances and entropies are from the (h-φ)family of stochastic measures [7, 8], so users can experimentand compare results. The same holds true for the measuresof similarity employed to build the convolution matrices onwhich the Nonlocal Means technique relies.

The pre-processing module is responsible for unifying allpossible input types (PolSARpro, ENVI etc. formats) into aunique raster representation of Hermitian matrices. The systemwill evolve to include techniques for PolSAR decompositionas, for instance, the one proposed by Bhattacharya et al. [18]that is based on stochastic distances.

4. CONCLUDING REMARKS

Using a C++ infrastructure, namely TerraLib, a system forPolSAR image processing and analysis is being developed bygathering a collection of procedures available in a number ofdifferent platforms. Such procedures have in common the useof techniques stemming from Information Theory, applied tothe processing and analysis of this kind of data. The user isnot aware of such diversity; rather than that, he/she is able toexperiment with the many options this approach provides, inan amenable way.

5. REFERENCES

[1] J.-S. Lee and E. Pottier, Polarimetric Radar Imaging:From Basics to Applications, CRC, Boca Raton, 2009.

[2] R. Touzi, W. M. Boerner, J. S. Lee, and E. Lueneburg, “Areview of polarimetry in the context of synthetic apertureradar: concepts and information extraction,” CanadianJournal of Remote Sensing, vol. 30, no. 3, pp. 380–407,2004.

[3] E. Pottier and L. Ferro-Famil, “PolSARPro v5.0: AnESA educational toolbox used for self-education in thefield of POLSAR and POL-INSAR data analysis,” in2012 IEEE International Geoscience and Remote Sens-ing Symposium (IGARSS), 2012, pp. 7377–7380.

240

Fig. 1. System architecture and main elements of each part.

[4] G. Camara, L. Vinhas, K. R. Ferreira, G. R. Queiroz,R. C. M. Souza, A. M. V. Monteiro, M. T. Carvalho,M. Casanova, and U. M. Freitas, “TerraLib: An opensource GIS library for large-scale environmental andsocio-economic applications,” in Open Source Ap-proaches in Spatial Data Handling, G. B. Hall, Ed.Springer, 2008.

[5] L. Pardo, D. Morales, M. Salicru, and M. L. Menendez,“Generalized divergence measures: Information matrices,amount of information, asymptotic distribution, and itsapplications to test statistical hypotheses,” InformationSciences, vol. 84, no. 8, pp. 181–198, May 1995.

[6] L. Pardo, D. Morales, M. Salicru, and M. L. Menendez,“Large sample behavior of entropy measures when pa-rameters are estimated,” Communications in Statistics –Theory and Methods, vol. 26, no. 2, pp. 483–501, 1997.

[7] A. C. Frery, A. D. C. Nascimento, and R. J. Cintra, “Ana-lytic expressions for stochastic distances between relaxedcomplex Wishart distributions,” IEEE Transactions onGeoscience and Remote Sensing, vol. 52, no. 2, pp. 1213–1226, Feb. 2014.

[8] A. C. Frery, R. J. Cintra, and A. D. C. Nascimento,

“Entropy-based statistical analysis of PolSAR data,”IEEE Transactions on Geoscience and Remote Sensing,vol. 51, no. 6, pp. 3733–3743, June 2013.

[9] A. D. C. Nascimento, M. M. Horta, A. C. Frery, and R. J.Cintra, “Comparing edge detection methods based onstochastic entropies and distances for PolSAR imagery,”IEEE Journal of Selected Topics in Applied Earth Obser-vations and Remote Sensing, vol. 7, no. 2, pp. 648–663,Feb. 2014.

[10] L. Torres, S. J. S. Sant’Anna, C. C. Freitas, and A. C.Frery, “Speckle reduction in polarimetric SAR imagerywith stochastic distances and nonlocal means,” PatternRecognition, vol. 47, pp. 141–157, 2014.

[11] W. B. Silva, C. C. Freitas, S. J. S. Sant’Anna, and A. C.Frery, “Classification of segments in PolSAR imageryby minimum stochastic distances between Wishart dis-tributions,” IEEE Journal of Selected Topics in AppliedEarth Observations and Remote Sensing, vol. 6, no. 3,pp. 1263–1273, June 2013.

[12] R Core Team, R: A Language and Environment for Statis-tical Computing, R Foundation for Statistical Computing,Vienna, Austria, 2014.

241

(a) Cartoon model (b) Simulated image (input)

(c) Segmented image (output)

Fig. 2. Classes, single-look simulated image (false colorR=|Shh|2, G=|Shv|2, and B=|Svv|2), and its segmentation.

[13] J. A. Doornik, Object-Oriented Matrix Programming Us-ing Ox, Timberlake Consultants Press & Oxford, London,3 edition, 2007.

[14] M. Almiron, E. S. Almeida, and M. Miranda, “The relia-bility of statistical functions in four software packagesfreely used in numerical computation,” Brazilian Journalof Probability and Statistics, vol. 23, no. 2, pp. 107–119,2009.

[15] S. Foucher and C. Lopez-Martinez, “Analysis, evaluation,and comparison of Polarimetric SAR speckle filteringtechniques,” IEEE Transactions on Image Processing,vol. 23, no. 4, pp. 1751–1764, Apr 2014.

(a) PolSAR image (input) (b) Filtered image (output)

Fig. 3. Pauli decomposition of the original AIRSAR imageover San Francisco, and its filtered version.

(a) PolSAR image (input) (b) Cartoon model (input)

(c) Training samples (input) (d) Classified image (output)

(e) p-value image (output)

Fig. 4. 4-look C-band image (R=|Shh|2, G=|Shv|2, andB=|Svv|2), segmented image (cartoon model), training sam-ples, and the outputs: classification and associated p-values.

[16] X. Qin, H. Zou, S. Zhou, and K. Ji, “Simulation of spa-tially correlated PolSAR images using inverse transformmethod,” Journal of Applied Remote Sensing, vol. 9, no.1, pp. 095082, Jan 2015.

[17] M. F. S. Saldanha, Um segmentador multinıvel paraimagens SAR polarimetricas baseado na distribuicaoWishart, Tese de Doutorado, Instituto Nacional dePesquisas Espaciais (INPE), Sao Jose dos Campos, 2013.

[18] A. Bhattacharya, A. Muhuri, S. De, S. Manickam, andA. C. Frery, “Modifying the Yamaguchi four-componentdecomposition scattering powers using a stochastic dis-tance,” IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing, in press.

242