multitemp 2005 – biloxi, mississippi, usa, may 16-18, 2005 remote sensing laboratory dept. of...
TRANSCRIPT
MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Remote Sensing LaboratoryDept. of Information and Communication Technology
University of TrentoVia Sommarive, 15, 38050 Trento, Italy
Francesca BOVOLO and Lorenzo BRUZZONE
E-mail: [email protected] Web pages at: http://www.ing.unitn.it/~bruzzone/RSLab
A Wavelet-Based Change-Detection Technique for Multitemporal SAR Images
2MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Outline
Introduction and background.
Aim of the work.
Proposed wavelet-based change-detection approach:
Multiscale decomposition;
Adaptive scale-selection;
Automatic multiscale fusion.
Experimental results.
Conclusions and future developments.
3MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Introduction
Unsupervised change-detection in multitemporal remote-sensing images plays an important role in several application domains (the ground truth on the analyzed problem is not available in many real applications).
Unlike optical multispectral images, Synthetic Aperture Radar (SAR) data are less used in unsupervised change-detection problems. This is due to:
the complexity of SAR data pre-processing;
the presence of multiplicative speckle noise that renders difficult the separation between changed and unchanged classes.
However, the use of SAR images for change detection is particularly attractive from the operational view point as they are not affected by sun light and atmospheric conditions.
4MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Aim of the Work
Develop a novel wavelet-based multiscale approach to unsupervised
change detection in multitemporal SAR images (suitable for single-
channel single-polarization multitemporal SAR data);
Adaptively exploiting information at different scales in order to obtain
change-detection maps that shows:
high accuracy in both homogeneous and border regions;
high geometric fidelity.
5MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Proposed Approach: Block Scheme
Scale-DrivenFusion
X1
t2 SAR image
t1 SAR image
Adaptive ScaleIdentification
ImageComparison
(log-ratio)
“Log-ratio” Image
MultiscaleDecomposition
Change-detection map
Ω ={ωc , ωu }0LRXnLRX
1-NLRX
X2
XLR M
6MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Image Comparison
Difference operator: the distribution depends on both the relative change and the reference intensity level in the original images. Thus changes are not detected in the same way in high and low intensity regions.
Ratio operator: reduces the multiplicative distortion affects common to the two considered images due to speckle noise. The distribution depends only on the relative changes between images.
jjL
D
L
jjLj
jLD
L
D LX
LXL
LXP
21
2111
0! )1( !
! )1(
22121 exp
!1
1),(
Equivalent numberof look of the SAR data
Intensity of SARimages at t1and t2
Usually log-ratio operator is used instead of the ratio as the log-ratio image has a more symmetrical statistical distribution and transforms the residual multiplicative noise model into an additive noise model.
L
R
LR
L
RX
X
L
LXP
212
1
21
221
!1
!12),(
7MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Multiscale decomposition (1)
Objective: compute a multiscale sequence of log-ratio images characterized by different trade-offs between SNR and geometrical-detail content.
l(.)
h(.)
h(.)
l(.)
h(.)
l(.)
Column wiseRow wise
1
0
1
0
1
1
0
1
0
1
1
0
1
0
1
1
0
1
0
1
)2,2()(
)2,2()(
)2,2()(
)2,2()(
D
p
D
q
LLnLR
nHHLR
D
p
D
q
LLnLR
nHLLR
D
p
D
q
LLnLR
nLHLR
D
p
D
q
LLnLR
nLLLR
qjpiXqhphji,X
qjpiXqlphji,X
qjpiXqhplji,X
qjpiXqlplji,X
nLLLRX
Proposed approach: apply iteratively the two-dimensional wavelet transform to the log-ratio image following the Mallat pyramidal algorithm.
8MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Multiscale decomposition (2)
The desired multiresolution sequence
is computed applying inverse wavelet
transform at each resolution level
independently: LL1 LH1 HL1 HH1
Level 1
LL2 LH2 HL2 HH2
Level 2
LLN-1 LHN-1 HLN-1 HHN-1
Level N-1
to all wavelet coefficient after thresholding detail sub-bands;
only to the approximation sub-bands (neglecting detail sub-bands).
LRLR XX 0
(.)l̂ (.)l̂
(.)l̂ (.)l̂ (.)l̂
1LRX2
LRX1NLRX
(.)l̂
9MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Adaptive Scale Identification
2 2
nnL
nLn CV
ji,μji,σji,LCV
Definition: a resolution level is reliable for a given pixel if the pixel does not belong to a border area at that level.
The set of n reliable scales for the generic pixel (i,j), is made up of all the sequential resolution levels that satisfy the following condition:
Homogeneous regionsSij=N-1
Border regionsSij= 0
“Intermediate” regionsLevel 0 < Sij < N-1
Sij is the optimal scale: the lowest resolution level that satisfies the definition of reliable scale for a given pixel.
CV: is the coefficient of variation (normalized standard deviation) computed over the whole image;LCV: is the local coefficient of variation computed over a moving window;
10MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Scale-Driven FusionThree different scale-driven fusion strategies have been considered for generating the change-detection map:
Optimal scale selection (OSS);Fusion at decision level (FDL);Fusion at feature level (FFL).
Optimal scale selection:
nc,k ,ωji,Mωji,M kk withijS 1NS ijand
Pixel label at optimalresolution level Sij
Fusion at decision level:
jiVjiMk
k
kk ,maxarg,
# of times the pixel (i,j) is
assigned to class k over the set of its reliable scales
Pixel label in the finalchange-detection map
11MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Fusion at feature level:
Compute a new set of N images by averaging all possible sequential combinations of reliable scales:
n
h
hLR
nMS
n 01
1XX , with 1,...,1,0 Nn
For each pixel compute the final label as:
ijij
ijij
SSMS
SSMS
TT
ji, ifω
ji, ifωji,M
c
n
X
X
,
Scale-dependent threshold value
Scale-Driven Fusion
12MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Data Set Description
Study area: a forest area in the central Canada (Saskatchewan).
Multitemporal data set: a portion of 350×350 pixels two images acquired by the SAR sensor of ERS-1 the 1st July and the 14th October 1995.
Objective: identify forest fires that affected the considered area between the two acquisition dates.
July 1995 October 1995 Reference Map
13MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Multiresolution Decompositon
Level 4 Level 5 Level 6
Level 1 Level 2 Level 3Level 0 (Log-ratio image)
14MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Experimental Results
Proposed approach
Strategy False Alarms Missed Alarms Overall Errors
OSS 3227 4678 7905
FDL 2634 3777 6411
FFL 2242 3379 5621
Classical algorithms
Strategy False Alarms Missed Alarms Overall Errors
Wavelet de-noising 2769 4243 7012
Lee de-noising 2875 3965 6840
15MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Experimental Results
Classical approach(de-noising with the Lee filter)
Proposed approach(FFL strategy)
Change-detection mapsReference map
16MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Conclusions
A novel adaptive wavelet-based technique for change-detection in multitemporal SAR images has been proposed.
Two novel methodological contributions characterize the proposed method compared with traditional algorithms:
automatic and adaptive selection of the reliable scales to be used in change detection for each pixel;
scale-driven fusion strategies;
The presented technique shows both high sensitivity to geometrical details and a high robustness to noisy speckle components in homogeneous areas;
Experimental results obtained on real multitemporal SAR data confirm the effectiveness of the proposed approach and in particular of the FFL strategy.
17MULTITEMP 2005 – Biloxi, Mississippi, USA, May 16-18, 2005
Future Developments
Explore different strategies for performing the multiresolution decomposition step in order to proper modeling the change information at different resolution levels (e.g., stationary wavelet transform).
Integrate automatic thresholding procedures in the scale-dependent fusion procedure.
Extend the use of the proposed approach to change detection in very high resolution SAR images and multiband and fully polarimetric SAR data.