Francesca Bovolo

1
A Multilevel Parcel-Based Technique for Change
Detection in VHR Images Robust to Registration
Noise
Remote Sensing Laboratory Dept. of Information
Engineering and Computer Science University of
Trento Via Sommarive, 14, 38050 Trento, Italy

• Francesca Bovolo
• Lorenzo Bruzzone
• Silvia Marchesi

E-mail lorenzo.bruzzone_at_ing.unitn.it Web page
http//disi.unitn.it/rslab/
2
Outline
Introduction and Aim of the Work
1
Analysis of Registration Noise Properties in VHR
images
2
Proposed Multiscale Approach for Reducing
Registration Noise
3
Experimental Results
4
Conclusion
5
3
Introduction
Very high geometrical resolution (VHR)
multitemporal images, also after proper
registration and geometrical corrections,
generally show a significant misalignment due to
both the different view angle of the sensor and
acquisition geometry.
Original Quickbird Image
Multitemporal False Color Composite
Problem the comparison between non perfectly
aligned pixels leads to a sharp increase of the
number of false alarms.
4
Aim of the Work

• Analyze the statistical properties of residual
  registration noise (RN) in multitemporal very
  high resolution (VHR) images of urban and rural
  areas.
• Propose an novel method robust to registration
  noise for change detection in multitemporal VHR
  images made up of
• A multiscale technique for reducing the impact of
  registration noise without affecting the high
  geometric content of the scene
• An adaptive context-sensitive procedure for
  exploiting the spatial context information.

5
Change Vector Analysis
X1
X1
x2
Multispectral image t1
X1
XD
Vector Difference
SCVs analysis
X2
XD
X1
X2
Change-Detection Map
X1
Multispectral difference image
x1
Multispectral image t2
Definitions
Ac
1. Magnitude-Direction Domain
2. Circle of unchanged pixels
3. Annulus of changed pixels
4. Annular sector of the k-th kind of change
F. Bovolo and L. Bruzzone, A Theoretical
Framework for Unsupervised Change Detection Based
on Change Vector Analysis in Polar Domain, IEEE
Transactions on Geoscience and Remote Rensing,
Vol. 45, No. 1, pp. 218-236, 2007.
6
Registration Noise Properties
Simulated multitemporal data sets
Reference Image
Misregistration 2 pixels
Misregistration 4 pixels
Multitemporal False color composites
Observed effects
Cause

• Non-perfect alignment between multitemporal
  pixels that belong to the same class.

• Higher spreading of the statistical distribution
  of SCVs in the circle of unchanged pixels.

• SCVs that show a distribution similar to the one
  of changed pixels.

• Comparison of pixels which belong to different
  objects (pixels associated with detail and border
  regions).

7
Registration Noise Multiscale Properties
Multitemporal data set (Misregistration 4
pixels)
Full resolution (Level 0)
8
Proposed Multiscale Approach
CVA at Full Resolution
Estimation of the RN Properties
X1
X1
X1
X2
X2
Vector Difference
Multiscale Decomposition
X1
X1
X1
X1
CD Map
Differential Analysis for RN Estimation
SCVs Analysis
VHR Multispectral image t1
X1
Multiscale Decomposition
X1
X2
X2
Multiscale Difference Images
VHR Multispectral image t2
Multitemporal segmentation
X1
X1
X1
X2
Parcel-based Context Sensitive Analysis
9
Estimation of Registration Noise Distribution
Marginal conditional densities of the direction
of pixels in Ac at level i
Level 0
Cn
Ac
p0
pi
Cn
Level i
Ac
10
Estimation of Registration Noise Distribution
We define the marginal conditional density of
registration noise in the direction domain as
0
11
Proposed Multiscale Approach
CVA at Full Resolution
Estimation of the RN Properties
X1
X1
X1
X2
X2
Vector Difference
Multiscale Decomposition
X1
X1
X1
X1
CD Map
Differential Analysis for RN Estimation
SCVs Analysis
VHR Multispectral image t1
X1
Multiscale Decomposition
X1
X2
X2
Multiscale Difference Images
VHR Multispectral image t2
Multitemporal segmentation
X1
X1
X1
X2
Parcel-based Context Sensitive Analysis
12
Data Set Description
Study area City of Trento (Italy). Multitempora
l data set portion (984984 pixels) of two
images acquired by the Quickbird satellite in
October 2004 and July 2006. Objective assess
the effectiveness of the proposed approach.
October 2004
July 2006
Reference Map
13
Experimental Results
Level 0
Level 4
14
Experimental Results
15
Experimental Results
Reference Map
October 2004
July 2006
Standard CVA
Proposed multiscale technique
16
Conclusion

• An analysis of the statistical behaviour of
  registration noise (RN) in VHR multitemporal
  images has been carried out in the context of a
  polar framework for change vector analysis.
• An adaptive technique for reducing the effects of
  residual RN in unsupervised change detection on
  VHR images has been presented, which
• automatically identifies annular sectors affected
  from RN according to a differential analysis of
  the marginal distributions of the direction at
  different scales
• exploits an adaptive parcel-based method for
  generating the change detection map.
• The effectiveness of the proposed technique in
  reducing the effects of the registration noise
  was confirmed by several experiments carried out
  on different pairs of Quickbird multitemporal
  images.

17
Future Developments

• Refine the estimation of the registration noise
  distribution by quantizing the polar domain in
  relatively small resolution cells.
• Extend the experimental analysis to Very High
  Geometrical resolution images acquired by
  different sensors (Ikonos, Spot-5).