«Department of Geomatics Engineering New Segmentation Algorithms for Dual and Full Polarimetric SAR Data (URL: ...»
Department of Geomatics Engineering
New Segmentation Algorithms for Dual and Full
Polarimetric SAR Data
UNIVERSITY OF CALGARY
NEW SEGMENTATION ALGORITHMS FOR DUAL AND FULL
POLARIMETRIC SAR DATA
SUBMITTED TO THE FACULTY OF GRADUATE STUDIES
IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE
DEGREE OF DOCTOR OF PHILOSOPHY
DEPARTMENT OF GEOMATICS ENGINEERING
A new agglomerative methodology is presented for full polarimetric SAR data segmentation. A new probabilistic distance is proposed for the agglomerative hierarchical merging of small clusters into an appropriate number of larger clusters. The proposed probabilistic distance measures the distance between two complex Wishart distributions, independently of the number of samples in each distribution.
The proposed methodologies, divisive and agglomerative, are developed so that they overcome drawbacks of existing segmentation approaches. The proposed methodologies are applied to SAR data from three spaceborne SAR systems, ALOS, TerraSAR-X and RADARSAT-2. For validation purposes, the segmentation results of the proposed methodologies are compared with results obtained by a number of recognized segmentation approaches. The comparison of results from the developed methods demonstrates significant improvements over conventional methods and overcome the drawbacks identified.
iii Acknowledgements This research work would not be possible without the invaluable help and guidance of a number of great people. I am indebted to my supervisors Prof. Dr. Michael Collins and Prof. Dr. Alexander Braun. Sincere and special thanks go to Prof. Dr.
Vassilia Karathanassi from the Laboratory of Remote Sensing at the National Technical University of Athens, where important parts of this thesis were done during research visits there. All gave their best in providing me with a stimulating and relaxed environment. I deeply appreciate their support which was much more than scientific!
Many thanks go to Prof. Dr. Demetrius Rokos, Prof. Dr. Demetre Argialas, Dr.
Polychronis Kolokousis, Christos Iossifides and Angelos Tzotzos from the Laboratory of Remote Sensing at the National Technical University of Athens for their useful help and discussions. Thanks to my examiners Dr. Michael Sideris, Dr. John Yackel and Dr.
I wish to thank the Japanese Space Agency (JAXA) for providing the PALSAR ALOS SAR data, the German Space Agency (DLR) for providing the TerraSAR-X SAR data and the Canadian Space Agency (CSA) for providing the RADARSAT-2 SAR data.
Furthermore, I gratefully acknowledge the financial support from the Natural Sciences and Engineering Research Council of Canada (NSERC) and the University of Texas at Dallas.
I do not want to forget to thank all the people in Greece and Canada who simply offered me their friendship and did not make me feel alone during my long academic trip (12 years!) far away from my family and my home country Palestine! Many thanks go to my friend Mohannad for the useful discussions we used to have. All my love goes to my father Darwish (the best mathematician ever!), my mother Nadia, my brothers and my darling Nehal.
Table of Contents
List of Tables
List of Figures and Illustrations
Symbols and Abbreviations
CHAPTER ONE: INTRODUCTION
1.1 Literature review
1.1.1 Approaches Based on Divisive Clustering
1.1.2 Approaches Based on Agglomerative Clustering
1.2 Motivation and Problem Statement
1.3 Research Objectives
1.4 Thesis Structure
CHAPTER TWO: BASIC POLARIMETRIC SAR CONCEPTS
2.1 Radar Polarimetry Historical Background
2.2 Polarimetric Scattering Vector
2.3 Deterministic and Non-deterministic Scatterers
2.4 Polarimetric Scattering Mechanisms
2.5 Polarimetric Decomposition Methods
2.5.1 The Pauli Decomposition method
2.5.2 The Cloude-Pottier Decomposition Method
2.5.3 The Freeman-Durden Decomposition Method
CHAPTER THREE: POLARIMETRIC SAR DATA AND STUDY AREAS.................25
3.1 ALOS Full Polarimetric SAR Data
3.2 RADARSAT-2 Full Polarimetric SAR Data
3.3 TerraSAR-X Dual Polarized SAR Data
CHAPTER FOUR: POLARIMETRIC SAR DIVISIVE CLUSTERING
4.2 Delon’s Histogram Thresholding Algorithm
4.3 Divisive Hierarchical Full Polarimetric SAR Segmentation
184.108.40.206 First Segmentation Level
220.127.116.11 Second Segmentation Level
18.104.22.168 Third Segmentation Level
4.3.3 Comparison with the k-means Algorithm
4.4 Divisive Hierarchical Dual Polarimetric Segmentation
vi 4.4.2 Implementation
22.214.171.124 Segmentation Results
CHAPTER FIVE: POLARIMETRIC SAR AGGLOMERATIVE CLUSTERING.........70
5.2 Wishart Chernoff Probabilistic Distance
5.2.1 The Complex Wishart Distribution
5.2.2 Chernoff Error Bound
5.3 Agglomerative Clustering
5.3.2 Implementation and Discussion
5.4 Application to RADARSAT-2 Full Polarimetric SAR Images
CHAPTER SIX: CONCLUSIONS AND FUTURE PERSPECTIVES
6.2 Future Perspectives
Table 2.1: Notable SAR missions
Table 2.2: Pauli matrices and their interpretation
Table 3.1: System parameters of ALOS polarimetric mode
Table 3.2: System parameters of the quad-polarized mode of RADARSAT-2.
............... 28 Table 3.3: System parameters of TerraSAR-X polarimetric mode
Table 4.1: Number of segments for each segmentation case
Table 5.1: Error in classification in percent for selected land cover regions of the POLSAR image.
Table 5.2: Number of classes, temperature and weather condition for each RADARSAT-2 image and acquisition time.
Data source for temperature and weather condition data is the National Climate Data and Information Archive of Canada
Figure 2.1: a) Scattering from smooth surface, b) Scattering from rough surface, c) Double bounce scattering, d) Volume scattering.
Figure 2.2: Radar signal penetration for different bands.
Figure 2.3: Backscattering decomposition into trihedral and two dihedrals with different tilt angles.
Figure 2.4: The α angle values and the corresponding scattering mechanisms.
............... 20 Figure 2.5: Segmentation of the polarimetric SAR data based on the entropy and alpha angle.
Figure 2.6: Segmentation of the polarimetric SAR data based on the entropy, alpha angle and anisotropy.
Figure 3.1: a) RGB of the ALOS polarimetric SAR data (red = T22, green = T33, blue = T11 ), b) AVNIR-2 RGB optical image of the study area (the study area is 83km N-S and 12km W-E), c) A map of the study area.
Figure 3.2: Full polarimetric RADARSAT-2 data of Churchill acquired in: a) October 31, 2009, b) December 18, 2009, c) January 24, 2010, d) April 3, 2010, e) May 21, 2010, and f) July 22, 2010.
Figure 3.3: a) SPOT image of the Churchill area, b) Map of Churchill.
Figure 3.4: a) The TerraSAR-X HH image before speckle filtering, b) The VV image in the same area, c) Optical image of the study area (Source: Google Earth aerial image 31 May 2006).
Figure 4.1: Three-mode scheme divided into three intervals.
Figure 4.2: Flowchart of the proposed multilevel segmentation methodology based on input data from Pauli or Freeman-Durden analysis.
Figure 4.3: Image segmentation by speckle filtering using the refined Lee filter.
........... 42 Figure 4.4: a) Polarimetric SAR image: Red = HH, Green = HV, Blue = VV, b) Data after removing speckle noise using the refined Lee filter three times, c) Pauli analysis images in RGB (Red = double bounce, Green = 45o tilted double bounce, Blue = surface), d) Freeman-Durden analysis images in RGB (Red = double bounce, Green = volume, Blue = surface).
Figure 4.5: a) First segmentation level of Pauli data, b) First segmentation level of Freeman-Durden data, c) Second segmentation level of Pauli data, d) Second
Figure 4.6: Variations of the amplitude values of the Surface scattering mechanism FD (dominant) for the subarea S SV of the Freeman-Durden images
Figure 4.7: Segmentation based on the k-mean clustering algorithm.
a) Segments produced based on Wishart k-means H/α clustering, b) Segments produced based on Wishart k-means SPAN/H/α/A clustering.
Figure 4.8: Sample region selected from the Southern end of the study area.
a) Pauli case, b) Freeman-Durden case, c) Wishart k-means H/α, d) Wishart k-means span/H/α/A.
Figure 4.9: Scheme shows the concept of the proposed segmentation methodology that creates subspaces based on the thresholded histograms.
Figure 4.10: a) The HH+VV after speckle filtering using a refined Lee filter, b) The HH-VV after speckle filtering using a refined Lee filter.
Figure 4.11: Flowchart of the segmentation approach for dual polarized SAR data.
....... 57 Figure 4.12: The HH+VV and HH-VV thresholded histograms and the sub-spaces produced by combing them in a two dimensional histogram-based space.
.............. 58 Figure 4.13: The extracted borders (red colour) of the resulting segments overlain on a composited image (Green = HH+VV, Blue = HH-VV) for a) First segmentation level, b) Second segmentation level, c) Third segmentation level,
d) Fourth segmentation level, (e) Fourth segmentation level produced by the eCognition
Figure 4.14: a), c) and e) Sample areas from the final (fourth) segmentation level produced by histogram thresholding.
b), d) and f) Sample areas from the final (fourth) segmentation level produced by eCognition.