Segmentation-Based Building Analysis from Polarimetric Synthetic Aperture Radar Images [Elektronische Ressource] / Wenju He. Betreuer: Olaf Hellwich

Sujets

• Informatik

Informations

• Publié par: technische_universitat_berlin
• Publié le: 01 janvier 2011
• Nombre de lectures: 11
• Langue: English
• Poids de l'ouvrage: 41 Mo

Extrait

Segmentation-Based Building Analysis
from Polarimetric Synthetic Aperture
Radar Images
vorgelegt von
Master of Engineering
Wenju He
aus China
von der Fakultat IV - Elektrotechnik und Informatik
der Technische Universitat Berlin
zur Erlangung des akademischen Grades
Doktor der Ingenieurwissenschaften
- Dr.-Ing. -
genehmigte Dissertation
Promotionsausschuss:
Vorsitzender: Prof. Dr. Oliver Brock
Gutachter: Prof. Dr.-Ing. Olaf Hellwich Prof. Dr.-Ing. Stefan Hinz (KIT)
Tag der wissenschaftlichen Aussprache: 15 Juli 2011
Berlin 2011
D 83Acknowledgements
In the past four years I have studied Synthetic Aperture Radar (SAR) image processing at
the computer vision and remote sensing group of Technische Universit at Berlin. Professor Dr.-
Ing Olaf Hellwich guides me through the study. I sincerely thank him for providing me the
opportunity to pursue PhD study in the eld of SAR. He helps me to organize and arrange the
study. His patience and kindness impress me.
I bene t a lot from the discussion meetings in our group. I learn from the way the
colleagues do researches. We share knowledge and have interesting discussions. Constructive
suggestions by colleagues are very helpful. Priv.-Doz. Dr. Andreas Reigber shared his profound
knowledge in SAR data processing to me. He leaded me into this eld by explaining the basic
concepts in details. He guided me to implement the basic and advanced SAR processing tech-
niques. Marc J ager guided me into the amazing world of machine learning. He shared a lot of
interesting papers and new ideas to me. He also helped me in e cient IDL programming and
problems with Linux system.
Dr.-Ing. Hongwei Zheng helped me to settle down during my rst year in Berlin. Lots of
thanks are given to Ronny H ansch. We travelled together to several conferences and had many
interesting discussions. I thank Ms. Marion Dennert for her e orts in arranging my studies and
travels. Dr.-Ing. Volker Rodehorst kindly helped me to revise and print several posters. I would
also like to thank him for solving computer problems for me. Dr. Maxim Neumann answered
me some questions of SAR processing during conferences. Sincere thanks are give to Adam
Stanski, Adhish Prasoon, Andreas Friedrich, Anke Bellmann, Cornelius Wefelscheid, David
Bornemann, Esra Erten, Matthias Heinrichs, Oliver Gloger, Saquib Sarfraz, Stefan Stoinski,
Stephane Guillaso and Ulas Yilmaz.
I am grateful to Technische Universit at Berlin for providing me the scholarship for my
study. I sincerely thank Ms. Roswitha Paul-Walz for arranging my scholarship. I am also grate-
ful to our group and Technische Universit at Berlin for sponsoring me to attend the conferences.
The lectures and posters in the conferences enlarge my view of state-of-the-art developments of
SAR technologies.
Some of my work is inspired by Dr. Derek Hoiem. He also helped me on the implemen-
tation details of the algorithms. The work from Dr. Kevin P. Murphy helped me to implement
Conditional Random Fields and Hidden Markov Model.
Many sincere thanks are given to my friends. The friendship gives me con dence, en-
courage and happiness. Finally, I would like to sincerely thank my parents and brother for
continuous support to my study.
3Abstract
High resolution Synthetic Aperture Radar (SAR) imagery has many applications in urban areas,
e.g. land cover classi cation and building displacement measurement. Buildings are evident in
SAR imagery due to their strong backscattering compared to natural environment. Polarimetric
SAR (PolSAR) imagery combines horizontal and vertical polarizations. It provides polarization
features of scatterers on buildings. Polarimetric decomposition aims to interpret the scattering
process as contributions from several mechanisms, e.g. surface, double bounce and volume
scattering. The scattering characteristics derived from PolSAR imagery can be exploited for
object analysis.
In this thesis we analyze buildings in meter-resolution PolSAR imagery in urban areas.
Segmentation and classi cation of PolSAR imagery are investigated. Polarimetric features,
e.g. amplitude, parameters from polarimetric decomposition and coherence, are extracted. We
adapt state-of-the-art feature extraction, segmentation and classi cation framework for urban
area analysis.
Segmentation provides an initial premise for semantic object analysis. The generated
segments provide spatial support for e cient feature extraction. A good segmentation is critical
for region feature extraction and e cient object detection. We adopt watershed, mean shift,
e cient graph-based segmentation and normalized cuts for PolSAR imagery. These algorithms
produce satisfying segmentation results. The segmentation results are evaluated on ground truth
data.
We propose the integration of probabilistic boundary estimation and segmentation from
PolSAR imagery. One framework is spectral graph segmentation based on probabilistic bound-
ary algorithm. Accurate boundaries are obtained through combining di erent types of gradi-
ents. The segmentation results preserve the weak boundaries. Another framework is occlusion
boundary estimation, in which segmentation and boundary extraction are interleaved. The
segmentation results are of the highest accuracy.
Object extraction is achieved by supervised classi cation of the segments. We extract
polarimetric and e ective low-level features, including texton histogram, histogram of oriented
gradients and sale-invariant feature transform descriptor. Texton histogram is well adapted to
PolSAR imagery. The classi cation aims to group the segments into several semantic classes. We
adopt several strategies for grouping. The rst is Conditional Random Fields, which emphasizes
that neighboring segments are prone to belong to a same class. The second is classi cation based
on multiple segmentations algorithm, which explores the capability of a hierarchy of segmenta-
tions providing spatial support for object evidence extraction. The last strategy is exploiting
building alignment angle and evidence from other objects in a Bayesian detection model. The
appearance of a building in a PolSAR image is in uenced by its alignment angle with respect to
the ight trajectory. We extract e ective features and train classi er to identify building align-
ment angle. Experimental results demonstrate the e ectiveness of these classi cation strategies.
Subaperture analysis is an important tool for SAR data processing. Each subaperture
spans a di erent part of the Doppler spectrum and samples object re ections at di erent az-
imuth look angles. The dependency of object scattering on azimuth look angle is modeled by
Hidden Markov Model (HMM), which describes the behavior variations of buildings across the
subapertures. States in the HMM represent representative centers in the feature space. The
state sequence along the subapertures indicates the scattering dynamics, which is valuable for
the analysis of stationary and non-stationary scatterers. The HMM is also able to classify
buildings from clutter and discriminate between buildings with di erent alignment angles.
4Contents
1. Introduction 9
1.1. Synthetic Aperture Radar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.1.1. Polarimetric SAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.1.2. Interferometric SAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2. Characteristics of SAR Imagery . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.1. SAR Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.2. Advantages of SAR Imagery . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.3. Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3. Buildings in SAR Imagery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.1. SAR Imagery of Urban Areas . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.2. Building E ects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.3. Building Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4. Organization and Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2. Polarimetric SAR Imagery 25
2.1. Scattering Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.1. Scattering Coe cient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.2. Scattering Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.1.3. Second-order Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2. Scattering Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3. Polarimetric Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1. Sphere, Diplane and Helix Decomposition . . . . . . . . . . . . . . . . . . 29
2.3.2. Eigenvalue Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.3. Freeman and Durden Decomposition . . . . . . . . . . . . . . . . . . . . . 32
2.4. Polarization Orientation Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5. Subaperture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.6. Coherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.7. Polari