TupinTU1T024.ppt

24
Tuesday, 26/07/2011, Vancouver, Canada, IGARSS 2011 INFLUENCE OF SPECKLE FILTERING OF POLARIMETRIC SAR DATA ON DIFFERENT CLASSIFICATION METHODS Fang Cao 1 , Charles-Alban Deledalle 1 , Jean- Marie Nicolas 1 , Florence Tupin 1 , Loïc Denis 2 , Laurent Ferro-Famil 3 , Eric Pottier 3 , Carlos López-Martínez 4 1 Institut Télécom, Télécom ParisTech, France 2 Université de Lyon, France 3 Université de Rennes 1, France 4 Universitat Politècnica de Catalunya, Spain

description

 

Transcript of TupinTU1T024.ppt

Page 1: TupinTU1T024.ppt

Tuesday, 26/07/2011, Vancouver, Canada, IGARSS 2011

INFLUENCE OF SPECKLE FILTERING OF POLARIMETRIC SAR DATA ON DIFFERENTCLASSIFICATION METHODS

Fang Cao1, Charles-Alban Deledalle1, Jean-Marie Nicolas1, Florence Tupin1, Loïc Denis2, Laurent Ferro-Famil3, Eric Pottier3, Carlos López-Martínez4

1 Institut Télécom, Télécom ParisTech, France2 Université de Lyon, France3 Université de Rennes 1, France4 Universitat Politècnica de Catalunya, Spain

Page 2: TupinTU1T024.ppt

Telecom ParisTechpage 2

Index

Index

Introduction

Speckle filtering

Decomposition and classification

Conclusion

Page 3: TupinTU1T024.ppt

Telecom ParisTech

Speckle filtering:• A pre-processing step to reduce the speckle noise

before image segmentation or classification

Tested filters: Refined Lee’s filter, IDAN filter and NL-PolSAR filter

Decomposition and classification:

Evaluation of the performance of speckle filtering methods through Cloude–Pottier decomposition and Wishart H/alpha classification

page 3

IntroductionIntroduction

Page 4: TupinTU1T024.ppt

Telecom ParisTechpage 4

Index

Index

Introduction

Speckle filtering

Decomposition and classification

Conclusion

Page 5: TupinTU1T024.ppt

Telecom ParisTechpage 5

Speckle filtering approaches

Page 6: TupinTU1T024.ppt

Telecom ParisTechpage 6

Speckle filtering approaches

Page 7: TupinTU1T024.ppt

Telecom ParisTechpage 7

Speckle filtering approaches

Page 8: TupinTU1T024.ppt

Telecom ParisTechpage 8

Speckle filtering approaches

Page 9: TupinTU1T024.ppt

Telecom ParisTechpage 9

Speckle filtering approaches

Page 10: TupinTU1T024.ppt

Telecom ParisTech

Speckle filtering approaches

San Francisco (JPL L-Band AIRSAR)

Refined Lee IDAN NL-PolSAR

|SHH- SVV| |SHV| |SHH+ SVV|

Page 11: TupinTU1T024.ppt

Telecom ParisTech

Speckle filtering approaches

Flevoland (JPL L-Band AIRSAR)

Refined Lee IDAN NL-PolSAR

|SHH- SVV| |SHV| |SHH+ SVV|

Page 12: TupinTU1T024.ppt

Telecom ParisTechpage 12

Index

Index

Introduction

Speckle filtering

Decomposition and classification

Conclusion

Page 13: TupinTU1T024.ppt

Telecom ParisTech

Coherency matrix: Hermitian, semi-definite positive matrix → diagonalization

Cloude-Pottier Decomposition

• Eigenvalue/eigenvector calculation of the coherency matrix of fully polarimetric SAR data.

• Covering the whole range of scattering mechanisms

• Automatically basis invariant.

Page 14: TupinTU1T024.ppt

Telecom ParisTech

Probability of each 3 scattering mechanism

Entropy H: the global distribution of scattering mechanism

angle: the type of scattering mechanism

Anisotropy A : the two least important scattering mechanism effects

Cloude-Pottier Decomposition

Page 15: TupinTU1T024.ppt

Telecom ParisTech

b

Refined Lee IDAN NL-PolSARSan Francisco by JPL L–Band AIRSAR

Entropy

Alpha

Anisotropy

1.0

0

1.0

0

90°

0

Page 16: TupinTU1T024.ppt

Telecom ParisTech

Entropy

Alpha

Anisotropy

1.0

0

1.0

0

90°

0

b

Refined Lee IDAN NL-PolSAR

The refined Lee filter and the NLPolSAR filters have similar performance. The IDAN filter usually introduces bias in entropy and anisotropy values, which may result to unreliable classification results.

San Francisco by JPL L–Band AIRSAR

Page 17: TupinTU1T024.ppt

Telecom ParisTech

The Wishart H Classification

Building

Water

Forest

Initialization for 8 classes using H/α

Wishart clustering

Convergent?

Cloude-Pottier decomposition

Speckle reduction

No

POLSAR data

Classification resultsH/ initialization: 8 classes

Page 18: TupinTU1T024.ppt

Telecom ParisTech

Wishart clustering

• Supervised algorithm

• Based on the complex Wishart distribution of coherency matrix

• Use maximum likelihood criterion

: the trace of a matrix

The Wishart H / Classification

Distance measure

V : the cluster center coherency matrix

Maximum likelihood criterion

Page 19: TupinTU1T024.ppt

Telecom ParisTech

Decomposition and classification

page 19

AIRSAR ALOS/PALSAR Radarsat–2

Refined LEE

NL-PolSAR

Page 20: TupinTU1T024.ppt

Telecom ParisTech

Decomposition and classification

page 20

AIRSAR ALOS/PALSAR Radarsat–2

Refined LEE

NL-PolSAR

The results of AIRSAR, ALOS/PALSAR and RadarSat-2 data show that the classification results with different sensors are quite similar, except the water area in the AIRSAR data, which is due to the big variation of the incidence angle of the airborne sensor.

Page 21: TupinTU1T024.ppt

Telecom ParisTech

The NL-PolSAR filter has better performance than the refined Lee filter, for example, thegolf course areas and the lakes in the AIRSAR classification results.

Page 22: TupinTU1T024.ppt

Telecom ParisTechpage 22

Index

Index

Introduction

Speckle filtering

Decomposition and classification

Conclusion

Page 23: TupinTU1T024.ppt

Telecom ParisTech

Conclusion

• Comparison of 3 speckle filters:

Refined Lee’s filter, IDAN filter and the NL-PolSAR filter• Comparison of the influence on decomposition and

classification

Cloude-Pottier decomposition & Wishart H/a classification

• Obtained results with different sensors:

Radarsat-2, ALOS/PALSAR and AIRSAR• The NL-PolSAR filter achieves the best performance

in our experimental tests

Page 24: TupinTU1T024.ppt

Telecom ParisTech

Thank you!

page 24