Recent Advances in Object-based Change Detection.pdf

26
July 25, 2011 Mitglied der Helmholtz-Gemeinschaft Recent Advances in Object-based Change Detection IGARSS 2011, Vancouver Change Detection and Multitemporal Image Analysis I | Irmgard Niemeyer, Clemens Listner Nuclear Safeguards Group Institute of Energy and Climate Research IEK-6: Nuclear Waste Management and Reactor Safety Forschungszentrum Jülich GmbH, Germany

Transcript of Recent Advances in Object-based Change Detection.pdf

Page 1: Recent Advances in Object-based Change Detection.pdf

July 25, 2011

Mitg

lied

de

r H

elm

ho

ltz-G

em

ein

sch

aft

Recent Advances in Object-based Change Detection

IGARSS 2011, Vancouver

Change Detection and Multitemporal Image Analysis I

| Irmgard Niemeyer, Clemens Listner Nuclear Safeguards Group Institute of Energy and Climate Research IEK-6: Nuclear Waste Management and Reactor Safety Forschungszentrum Jülich GmbH, Germany

Page 2: Recent Advances in Object-based Change Detection.pdf

Slide 2

Acknowledgments

German Support Programme for the

International Atomic Energy Agency (IAEA)

Project on satellite imagery analysis and photo

interpretation support“

EC FP7, Global Monitoring for Environment and

Security (GMES)

Current project G-MOSAIC

General R&D interests

Methodological developments, PhD thesis Listner

Page 3: Recent Advances in Object-based Change Detection.pdf

Slide 3

Recent Advances in Object-based Change Detection

Page 4: Recent Advances in Object-based Change Detection.pdf

Slide 4

Very high spatial resolution optical sensors (<1m): WorldView-2

Page 5: Recent Advances in Object-based Change Detection.pdf

Slide 5

Object-based change detection using

IR-MAD

Iteratively Reweighted Multivariate Alteration Detection (IR-MAD) [Nielsen 2007]

Linear transformation of the feature space aimed to enhance the change information in the difference image

Modeling object’s feature vector as random vectors F

and G of length N

Transformation of vectors to enhance relevant changes

var(m1

= a1

TU - b1

TV) → max under the constraint that var(a

1TU) = var(b

1TV) = 1

Further orthogonal variates mi can be computed

Σmi2 ~ Chi2 indicating change probability P(change)

Iteration by weighting with 1- P(change)

Additional step: Application of PCA to U and V

1. Introduction

Page 6: Recent Advances in Object-based Change Detection.pdf

Slide 6

Statistical pixel-based change detection approaches provide good results, but shows limits due to …

• low number of spectral channels or small spectral range covered,

• image registration problems.

Object-based change detection looks promising, but …

• how to connect corresponding objects?

• how to carry out a reasonable segmentation for this task?

Object-based change detection using

IR-MAD

1. Introduction

Page 7: Recent Advances in Object-based Change Detection.pdf

Slide 7

Existing approaches to segmentation

for object-based change detection

Segment I1 and I

2 as stack

• segmentation not adequate for I1

and I2

• shape features cannot be used

Use segmentation of I1 for I

2

• segmentation not adequate for I2

• shape features cannot be used

Independent segmentation

• leads to false-alarm segment changes

• shape features can be used

Time 1

Time 2

Image data

Segmentation levels

Time 1

Time 2

Image data

Segmentation levels

Time 1

Time 2

Image data

Segmentation levels

2. Segmentation

Page 8: Recent Advances in Object-based Change Detection.pdf

Slide 8

Multiresolution segmentation

Region-based bottom-up approach to segmentation

Each segment is a binary tree (leafs=pixel, root=final segment)

Implemented in eCognitionTM

Starts with chessboard segmentation

Selects iteratively a segment X and merges it to a

neighboring segment Y if

)),((min),(

)(ZXdYXd

XNZ

)),((min),()(

ZYdXYdYNZ

TXYd ),(

2. Segmentation

Page 9: Recent Advances in Object-based Change Detection.pdf

Slide 9

Multiresolution segmentation

2. Segmentation

Page 10: Recent Advances in Object-based Change Detection.pdf

Slide 10

Segmentation of identical images up to Gaussian noise (μ=0,σ=0.1) using multiresolution segmentation

Multiresolution segmentation applied to slightly different images

2. Segmentation

Page 11: Recent Advances in Object-based Change Detection.pdf

Slide 11

Multiresolution segmentation adapted

for object-based change detection 1

1. Segment I1 using multiresolution segmentation

2. Apply this segmentation to I2 and recalculate color

heterogeneity

3. Check each merge for consistency with I2 using a

predefined test

4. Remove inconsistent segments using a predefined removal strategy

5. Re-run the multiresolution segmentation using the so

gained segmentation of I2 as an initial segmentation

2. Segmentation

Page 12: Recent Advances in Object-based Change Detection.pdf

Slide 12

Multiresolution segmentation adapted

for object-based change detection 2

Given segment S3 with children S

1 (seed) and S2

Threshold test

• h(S3) ≤ T

check in I

2 ?

Local best fitting test

• Is S2 the locally best fitting neighbor for S

1 in I2 ?

Local mutual best fitting test

• Are S1 and S2 local mutually best fitting in I

2 ?

Reduce sensitivity of the best fitting tests by using

Tchecktolerance

2. Segmentation

Page 13: Recent Advances in Object-based Change Detection.pdf

Slide 13

Segmentation for object-based change detection Threshold test & universal segment removal strategy

2. Segmentation

Page 14: Recent Advances in Object-based Change Detection.pdf

Slide 14

Segmentation for object-based change detection Local mutual best fitting test & global segment removal strategy

2. Segmentation

Page 15: Recent Advances in Object-based Change Detection.pdf

Slide 15

Segmentation for object-based change detection Local best fitting test & local segment removal strategy

2. Segmentation

Page 16: Recent Advances in Object-based Change Detection.pdf

Slide 16

Segmentation for object-based change detection Threshold test & universal segment removal strategy

2. Segmentation

Page 17: Recent Advances in Object-based Change Detection.pdf

Slide 17

Object correspondence for object- based change detection

Directed Via intersection

1

1

i x i

n

i y k

k=

x = f S ,

y = f Tn

1

2

i x

i y

x = f S ,

y = f S

3. Object correspondence

Page 18: Recent Advances in Object-based Change Detection.pdf

Slide 18

Object-based change detection

4. Experiments

Post-processing Integration to GIS or GDBS

Change classification

Class-based FFN Marpu 2009

Change detection

IR-MAD Nielsen 2007, Listner & Niemeyer

2011b

Segmentation Multiresolution segmentation adapted

to change detection

e.g. Listner & Niemeyer 2010, 2011a,

2011b

Pre-processing Image-to-image registration,

Radiometric normalization

Canty & Nielsen 2009

Page 19: Recent Advances in Object-based Change Detection.pdf

Slide 19

Object-based change detection

4. Experiments

Page 20: Recent Advances in Object-based Change Detection.pdf

Slide 20

Object-based change detection

Segmentation of the bitemporal imagery using threshold test and universal segment removal strategy.

4. Experiments

Page 21: Recent Advances in Object-based Change Detection.pdf

Slide 21

Object-based change detection

Directed change detection. Changes from time 1 to time 2 (left) and from time 2 to time 1 (right).

4. Experiments

Page 22: Recent Advances in Object-based Change Detection.pdf

Slide 22

Object-based change detection

Change detection using intersected objects.

Change detection using MAD

objects.

4. Experiments

Page 23: Recent Advances in Object-based Change Detection.pdf

Slide 23

Object-based change detection Accuracy assessment

Directed change detection: T1T2

Directed change detection: T2T1

Change detection using intersected objects

Change detection using MAD objects

Overall accuracy

0.98 0.98 0.98 0.99

KIA 0.82 0.87 0.77 0.75

4. Experiments

Page 24: Recent Advances in Object-based Change Detection.pdf

Slide 24

Summary

An enhanced procedure for segmentation was introduced and implemented into the change detection workflow.

Moreover, numerically issues in the IR-MAD method were addressed.

The proposed methods showed good results in three experiments using aerial imagery.

Further developments are needed:

• New consistency tests and segment removal strategies;

• methods for enabling the user to easily select the segmentation parameters, e.g. by using training samples;

• implementation as eCognition plugin.

5. Summary

Page 25: Recent Advances in Object-based Change Detection.pdf

Slide 25

Most recent publications

C. Listner and I. Niemeyer (2011a), “Advances in object-based change detection,” Proc. IGARSS 2011, Vancouver, July 2011

C. Listner and I. Niemeyer (2011b), “Object-based change detection,” Photogrammetrie, Fernerkundung, Geoinformation (PFG), vol. 3, 2011 (in print)

Page 26: Recent Advances in Object-based Change Detection.pdf

Slide 26

Thank you for your attention.

Dr. Irmgard Niemeyer Nuclear Safeguards Institute of Energy and Climate Research IEK-6: Nuclear Waste Management and Reactor Safety Forschungszentrum Jülich GmbH in der Helmholtz-Gemeinschaft | 52425 Jülich | Germany Phone / Fax: +49 2461 61-1762 / -2450 Email: [email protected] www.fz-juelich.de/ief/iek-6/