1. General and specific contexts
Method description and Evaluation Criteria
Experiments
Conclusion
2D Change Detection from satellite imagery :
Performance analysis and Impact of the spatial
resolution of input images
Nicolas Champion∗ , Didier Boldo, Marc Pierrot-Deseilligny, Georges Stamon
July 26th , 2011
1/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
2. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
General Context
General Issue
Updating maps → a crucial and topical issue in
most western countries
2/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
3. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
General Context
General Issue
Updating maps → a crucial and topical issue in
most western countries
General trends
to use remote-sensed images to perform the
detection of changes
to start the procedure from satellite images
Quickbird (Bailloeul et al., 2005)
Pl´iades-HR (Poulain et al., 2009)
e
2/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
4. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
General Context
General Issue
Updating maps → a crucial and topical issue in
most western countries
General trends
to use remote-sensed images to perform the
detection of changes
to start the procedure from satellite images
Quickbird (Bailloeul et al., 2005)
Pl´iades-HR (Poulain et al., 2009)
e
Some questions remain unanswered. . .
What is the optimal Ground Sample Distance (GSD) to use for input images ?
2/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
5. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
General Context
General Issue
Updating maps → a crucial and topical issue in
most western countries
General trends
to use remote-sensed images to perform the
detection of changes
to start the procedure from satellite images
Quickbird (Bailloeul et al., 2005)
Pl´iades-HR (Poulain et al., 2009)
e
Some questions remain unanswered. . .
What is the optimal Ground Sample Distance (GSD) to use for input images ?
The main goal of this presentation is provide possible answers to this question
2/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
6. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
General Context
General Issue
Updating maps → a crucial and topical issue in
most western countries
General trends
to use remote-sensed images to perform the
detection of changes
to start the procedure from satellite images
Quickbird (Bailloeul et al., 2005)
Pl´iades-HR (Poulain et al., 2009)
e
Some questions remain unanswered. . .
What is the optimal Ground Sample Distance (GSD) to use for input images ?
The main goal of this presentation is provide possible answers to this question
Methodology used for this paper → Evaluating the method by (Champion et al., 2009) 1
using 3 test areas and 2 different GSD for input images
1. N. Champion, D. Boldo, M. Pierrot-Deseilligny, and G. Stamon. 2D building change detection from high resolution satellite
imagery : A two-step hierarchical method based on 3D invariant primitives. In : Pattern Recognition Letters, vol. 31, pp. 1138–1147, 2010.
2/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
7. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Outline
1 General and specific contexts
2 Method description and Evaluation Criteria
3 Experiments
4 Conclusion
3/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
8. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
9. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
10. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
→ Demolished
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
11. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
→ Demolished / Modified - Errors in the DB
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
12. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
→ Demolished / Modified / New
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
13. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
→ Demolished / Modified / New
Images used to perform change detection
→ Pl´iades-HR ( c CNES)
e
a tri-stereosocpic system
4 channels → RGB + NIR (Near InfraRed)
2 datasets → GSD=50cm and GSD=70cm
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
14. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
→ Demolished / Modified / New
Images used to perform change detection
→ Pl´iades-HR ( c CNES)
e
a tri-stereosocpic system
4 channels → RGB + NIR (Near InfraRed)
2 datasets → GSD=50cm and GSD=70cm
Products derived from input images
DSM (processed with MICMAC 1 )
1 - www.micmac.ign.fr
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
15. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
→ Demolished / Modified / New
Images used to perform change detection
→ Pl´iades-HR ( c CNES)
e
a tri-stereosocpic system
4 channels → RGB + NIR (Near InfraRed)
2 datasets → GSD=50cm and GSD=70cm
Products derived from input images
DSM (processed with MICMAC 1 )
RGB + NIR Orthophotos
1 - www.micmac.ign.fr
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
16. General and specific contexts
Method description and Evaluation Criteria General context
Experiments Specific context
Conclusion
Specific context
Updating the French topographic vector
database (DB)
changes to detect
→ Demolished / Modified / New
Images used to perform change detection
→ Pl´iades-HR ( c CNES)
e
a tri-stereosocpic system
4 channels → RGB + NIR (Near InfraRed)
2 datasets → GSD=50cm and GSD=70cm
Products derived from input images
DSM (processed with MICMAC 1 )
RGB + NIR Orthophotos
vegetation mask (NDVI)
1 - www.micmac.ign.fr
4/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
17. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Outline
2 Method description and Evaluation Criteria
Method Workflow : Step 1/Step 2
Method Description : Step 1
Evaluation criteria
5/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
18. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method Workflow
Input
Data
Database (DB)
to update
satellite
images
DSM
Orthophotos
6/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
19. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method Workflow
Input
Data
Database (DB)
to update
satellite
images nDSM
DSM DTM
Orthophotos
6/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
20. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method Workflow
Input
Step 1
Data Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
DSM DTM
Database
partially
Orthophotos
updated
Identifying
non-changed buildings . . .
. . . and changed buildings
6/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
21. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method Workflow
Input
Step 1
Data Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
6/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
22. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method Workflow
Input
Step 1
Data Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
Step 2 detection
6/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
23. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method Workflow
Input
Step 1
Data Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
Step 2 detection
6/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
24. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Definition
Definition
a Digital Terrain Model (DTM) = a digital representation of the ground surface
DSM
DTM
7/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
25. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method used for processing the DTM
230m
200m
DSM
Method described in (Champion et al., 2009)
N. Champion, D. Boldo, M. Pierrot-Deseilligny and G. Stamon. Automatic estimation of fine terrain models from multiple high
resolution images. In : Proc. of the ICIP Conference, 2009.
8/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
26. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method used for processing the DTM
230m
200m
DSM Building and Vegetation Mask
Method described in (Champion et al., 2009)
N. Champion, D. Boldo, M. Pierrot-Deseilligny and G. Stamon. Automatic estimation of fine terrain models from multiple high
resolution images. In : Proc. of the ICIP Conference, 2009.
8/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
27. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method used for processing the DTM
230m
200m
DSM Building and Vegetation Mask
DTM
Method described in (Champion et al., 2009)
N. Champion, D. Boldo, M. Pierrot-Deseilligny and G. Stamon. Automatic estimation of fine terrain models from multiple high
resolution images. In : Proc. of the ICIP Conference, 2009.
8/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
28. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1)
Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
detection
9/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
29. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1)
Step 1
Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
detection
9/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
30. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1)
Step 1
Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
detection
9/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
31. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1)
Step 1
Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
detection
9/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
32. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1)
Step 1
Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
detection
9/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
33. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Robust Primitives
(1) Information about Overground
based on the nDSM = DSM – DTM
6m
0m
input DSM nDSM
10/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
34. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Robust Primitives
(1) Information about Overground
based on the nDSM = DSM – DTM
(2) Linear Primitives : 2D Contours
extracted from the DSM (Deriche, 1987)
10/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
35. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Robust Primitives
(1) Information about Overground
based on the nDSM = DSM – DTM
(2) Linear Primitives : 2D Contours
extracted from the DSM (Deriche, 1987)
(3) Linear Primitives : 3D Segments
Reconstructed from input multiscopic images (Taillandier and Deriche, 2002)
10/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
36. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1)
Step 1
Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
detection
11/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
37. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1)
Step 1
Matching
Database (DB)
to update
Robust
Primitives
satellite
images nDSM
Decision
updated DSM DTM
DTM
Database
partially
Orthophotos
updated
nDSM Identifying
non-changed buildings . . .
. . . and changed buildings
Overground Mask t Overground Mask t-1
Morphological
comparison
New Building
detection
11/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
38. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
1. A. Desolneux, L. Moisan, and J.-M. Morel, Meaningful alignments. In : International Journal of Computer Vision, 1 :40, 2000.
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
39. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
nDSM
Illustration of the computation of SDissimilarity
1. A. Desolneux, L. Moisan, and J.-M. Morel, Meaningful alignments. In : International Journal of Computer Vision, 1 :40, 2000.
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
40. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
nDSM
Illustration of the computation of SDissimilarity
nDSM
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
41. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
nDSM
Illustration of the computation of SDissimilarity
Building to verify
nDSM
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
42. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
nDSM
Illustration of the computation of SDissimilarity
Area
not covered
by overground pixel
in the nDSM
Building to verify
nDSM
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
43. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
nDSM
Illustration of the computation of SDissimilarity
Area
not covered
by overground pixel
in the nDSM
Another area with
no coverage
Building to verify
nDSM
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
44. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
nDSM
Illustration of the computation of SDissimilarity
Area
not covered
by overground pixel
in the nDSM
Another area with
no coverage
Building to verify
nDSM
nDSM
A (PBuilding (x,y) : nDSM(x,y) < TH )
SDissimilarity =
ABuilding
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
45. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
Decision Rules
Hypothesis : “a few changes in the scene” ⇒ “a building already present in
the database is likely still to be in the scene at the time of investigation.”
1. A. Desolneux, L. Moisan, and J.-M. Morel, Meaningful alignments. In : International Journal of Computer Vision, 1 :40, 2000.
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
46. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
Decision Rules
Hypothesis : “a few changes in the scene” ⇒ “a building already present in
the database is likely still to be in the scene at the time of investigation.”
⇒ A weak indication is enough to validate it during Step 1
1. A. Desolneux, L. Moisan, and J.-M. Morel, Meaningful alignments. In : International Journal of Computer Vision, 1 :40, 2000.
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
47. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
Decision Rules
Hypothesis : “a few changes in the scene” ⇒ “a building already present in
the database is likely still to be in the scene at the time of investigation.”
⇒ A weak indication is enough to validate it during Step 1
Linear Primitives nDSM
⇒ A small value for SDissimilarity or SDissimilarity
1. A. Desolneux, L. Moisan, and J.-M. Morel, Meaningful alignments. In : International Journal of Computer Vision, 1 :40, 2000.
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
48. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
Decision Rules
Hypothesis : “a few changes in the scene” ⇒ “a building already present in
the database is likely still to be in the scene at the time of investigation.”
⇒ A weak indication is enough to validate it during Step 1
Linear Primitives nDSM
⇒ A small value for SDissimilarity or SDissimilarity
{Unchanged Buildings} = Buildingi : SnDSM (Buildingi ) <
Dissimilarity M1
i∈[1,N]
∪
Linear Primitives
Buildingi : SDissimilarity (Buildingi ) < M2
i∈[1,N]
1. A. Desolneux, L. Moisan, and J.-M. Morel, Meaningful alignments. In : International Journal of Computer Vision, 1 :40, 2000.
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
49. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Database Verification (Step 1) : Matching and Decision
based on the a contrario paradigm 1
Dissimilarity Scores
nDSM Linear Primitives
nDSM → SDissimilarity Linear Primitives (2D contours & 3D segments) → SDissimilarity
Decision Rules
Hypothesis : “a few changes in the scene” ⇒ “a building already present in
the database is likely still to be in the scene at the time of investigation.”
⇒ A weak indication is enough to validate it during Step 1
Linear Primitives nDSM
⇒ A small value for SDissimilarity or SDissimilarity
{Unchanged Buildings} = Buildingi : SnDSM (Buildingi ) <
Dissimilarity M1
i∈[1,N]
∪
Linear Primitives
Buildingi : SDissimilarity (Buildingi ) < M2
i∈[1,N]
⇒ Our change detection system is actually a no-change detection system :
changes are inferred from {Unchanged Buildings} by taking the complementary
1. A. Desolneux, L. Moisan, and J.-M. Morel, Meaningful alignments. In : International Journal of Computer Vision, 1 :40, 2000.
12/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
50. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Method output - Illustration on Amiens Pont de Metz (GSD = 50cm)
Change Map Evaluation Map
unchanged modified demolished new TP FN FP TN
13/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
51. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Evaluation Criteria
Success criteria used in this project (Champion et al., 2009) 1
Confusion Matrix
XX
XXX System (S)
Unchanged (S) M&D2 (S) New (S) Bg.3 (S)
Ref.1 (R)
XX
X X
Unchanged (R) # NU(S)→U(R) # NC(S)→U(R) - -
M&D2 (R) # NU(S)→C(R) # NC(S)→C(R) - -
New (R) - - # NN(S)→N(R) # NB(S)→N(R)
Bg.3 (R) - - # NN(S)→B(R) -
1
Ref. = Reference 2 M&D = Modified and Demolished 3 Bg. = Background
1. N. Champion, F. Rottensteiner, L. Matikainen, X. Liang, J. Hyyppa, and B. Olsen, “A test of automatic building change detection
approaches.,” in International Archives of Photogrammetry and Remote Sensing, Vol. 38, Part 3/W4, 2009.
14/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
52. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Evaluation Criteria
Success criteria used in this project (Champion et al., 2009) 1
Confusion Matrix
XX
XXX System (S)
Unchanged (S) M&D2 (S) New (S) Bg.3 (S)
Ref.1 (R)
XXX
X
Unchanged (R) # NU(S)→U(R) # NC(S)→U(R) - -
M&D2 (R) # NU(S)→C(R) # NC(S)→C(R) - -
New (R) - - # NN(S)→N(R) # NB(S)→N(R)
Bg.3 (R) - - # NN(S)→B(R) -
Computing the Recall and Precision for each change class - e.g. :
1
Ref. = Reference 2 M&D = Modified and Demolished 3 Bg. = Background
1. N. Champion, F. Rottensteiner, L. Matikainen, X. Liang, J. Hyyppa, and B. Olsen, “A test of automatic building change detection
approaches.,” in International Archives of Photogrammetry and Remote Sensing, Vol. 38, Part 3/W4, 2009.
14/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
53. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Evaluation Criteria
Success criteria used in this project (Champion et al., 2009) 1
Confusion Matrix
XXX
XXX (S)
System
Unchanged (S) M&D2 (S) New (S) Bg.3 (S)
Ref.1 (R) XX X
Unchanged (R) # NU(S)→U(R) # NC(S)→U(R) - -
M&D2 (R) # NU(S)→C(R) # NC(S)→C(R) - -
New (R) - - # NN(S)→N(R) # NB(S)→N(R)
Bg.3 (R) - - # NN(S)→B(R) -
Computing the Recall and Precision for each change class - e.g. :
#NC(S)→C(R)
RecallM&D =
#NC(S)→C(R) + #NU(S)→C(R)
1
Ref. = Reference 2 M&D = Modified and Demolished 3 Bg. = Background
1. N. Champion, F. Rottensteiner, L. Matikainen, X. Liang, J. Hyyppa, and B. Olsen, “A test of automatic building change detection
approaches.,” in International Archives of Photogrammetry and Remote Sensing, Vol. 38, Part 3/W4, 2009.
14/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
54. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Evaluation Criteria
Success criteria used in this project (Champion et al., 2009) 1
Confusion Matrix
XXX
XXX (S)
System
Unchanged (S) M&D2 (S) New (S) Bg.3 (S)
Ref.1 (R) XX X
Unchanged (R) # NU(S)→U(R) # NC(S)→U(R) - -
M&D2 (R) # NU(S)→C(R) # NC(S)→C(R) - -
New (R) - - # NN(S)→N(R) # NB(S)→N(R)
Bg.3 (R) - - # NN(S)→B(R) -
Computing the Recall and Precision for each change class - e.g. :
#NC(S)→C(R) #NC(S)→C(R)
RecallM&D = / PrecisionM&D =
#NC(S)→C(R) + #NU(S)→C(R) #NC(S)→C(R) + #NC(S)→U(R)
1
Ref. = Reference 2 M&D = Modified and Demolished 3 Bg. = Background
1. N. Champion, F. Rottensteiner, L. Matikainen, X. Liang, J. Hyyppa, and B. Olsen, “A test of automatic building change detection
approaches.,” in International Archives of Photogrammetry and Remote Sensing, Vol. 38, Part 3/W4, 2009.
14/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
55. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Evaluation Criteria
Success criteria used in this project (Champion et al., 2009) 1
Confusion Matrix
XX
XXX System (S)
Unchanged (S) M&D2 (S) New (S) Bg.3 (S)
Ref.1 (R)
XXX
X
Unchanged (R) # NU(S)→U(R) # NC(S)→U(R) - -
M&D2 (R) # NU(S)→C(R) # NC(S)→C(R) - -
New (R) - - # NN(S)→N(R) # NB(S)→N(R)
Bg.3 (R) - - # NN(S)→B(R) -
Computing the Recall and Precision for each change class
1st success criterion (qualitative effectiveness)
⇒ a high recall for all changes = Demolished & Modified and New
1
Ref. = Reference 2 M&D = Modified and Demolished 3 Bg. = Background
1. N. Champion, F. Rottensteiner, L. Matikainen, X. Liang, J. Hyyppa, and B. Olsen, “A test of automatic building change detection
approaches.,” in International Archives of Photogrammetry and Remote Sensing, Vol. 38, Part 3/W4, 2009.
14/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
56. General and specific contexts
Method Workflow : Step 1/Step 2
Method description and Evaluation Criteria
Method Description : Step 1
Experiments
Evaluation criteria
Conclusion
Evaluation Criteria
Success criteria used in this project (Champion et al., 2009) 1
Confusion Matrix
XXX
XXX (S)
System
Unchanged (S) M&D2 (S) New (S) Bg.3 (S)
Ref.1 (R) XX X
Unchanged (R) # NU(S)→U(R) # NC(S)→U(R) - -
M&D2 (R) # NU(S)→C(R) # NC(S)→C(R) - -
New (R) - - # NN(S)→N(R) # NB(S)→N(R)
Bg.3 (R) - - # NN(S)→B(R) -
Computing the Recall and Precision for each change class
1st success criterion (qualitative effectiveness)
⇒ a high recall for all changes = Demolished & Modified and New
2nd success criterion (economical effectiveness)
to limit the number of false alarms ⇒ a high precision for modified and demolished buildings
⇔ a high recall for unchanged buildings (a lower precision for new buildings is not critical)
1
Ref. = Reference 2 M&D = Modified and Demolished 3 Bg. = Background
1. N. Champion, F. Rottensteiner, L. Matikainen, X. Liang, J. Hyyppa, and B. Olsen, “A test of automatic building change detection
approaches.,” in International Archives of Photogrammetry and Remote Sensing, Vol. 38, Part 3/W4, 2009.
14/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
57. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Quantitative Evaluation (50cm)
Unchanged M&D1 New
Amiens Downtown
Recall 95,7 95,45 80,3
Precision 99,8 48,09 28,49
Amiens Pont-de-Metz
Recall 84,98 96,55 92,13
Precision 98,57 69,65 38,86
Toulouse Rocade
Recall 80,52 96,99 76,47
Precision 99,55 37,61 18,44
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
58. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Recall (Qualitative Effectiveness)
Quantitative Evaluation (50cm) very good for modified and
demolished buildings ( 95%)
Unchanged M&D1 New
Amiens Downtown
Recall 95,7 95,45 80,3
Precision 99,8 48,09 28,49
Amiens Pont-de-Metz
Recall 84,98 96,55 92,13
Precision 98,57 69,65 38,86
Toulouse Rocade
Recall 80,52 96,99 76,47
Precision 99,55 37,61 18,44
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
59. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Recall (Qualitative Effectiveness)
Quantitative Evaluation (50cm) very good for modified and
demolished buildings ( 95%)
Unchanged M&D1 New good for new buildings (on
Amiens Downtown average 85%) . . .
Recall 95,7 95,45 80,3
Precision 99,8 48,09 28,49
Amiens Pont-de-Metz
Recall 84,98 96,55 92,13
Precision 98,57 69,65 38,86
Toulouse Rocade
Recall 80,52 96,99 76,47
Precision 99,55 37,61 18,44
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
60. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Recall (Qualitative Effectiveness)
Quantitative Evaluation (50cm) very good for modified and
demolished buildings ( 95%)
Unchanged M&D1 New good for new buildings (on
Amiens Downtown average 85%) . . . but sometimes
Recall 95,7 95,45 80,3 critical (< 80%)
Precision 99,8 48,09 28,49
Amiens Pont-de-Metz
Recall 84,98 96,55 92,13
Precision 98,57 69,65 38,86
Toulouse Rocade
Recall 80,52 96,99 76,47
Precision 99,55 37,61 18,44
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
61. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Recall (Qualitative Effectiveness)
Quantitative Evaluation (50cm) very good for modified and
demolished buildings ( 95%)
Unchanged M&D1 New good for new buildings (on
Amiens Downtown average 85%) . . . but sometimes
Recall 95,7 95,45 80,3 critical (< 80%)
Precision 99,8 48,09 28,49
Precision (Economical Effectiveness)
Amiens Pont-de-Metz
relatively low for new buildings
Recall 84,98 96,55 92,13
Precision 98,57 69,65 38,86
Toulouse Rocade
Recall 80,52 96,99 76,47
Precision 99,55 37,61 18,44
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
62. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Recall (Qualitative Effectiveness)
Quantitative Evaluation (50cm) very good for modified and
demolished buildings ( 95%)
Unchanged M&D1 New good for new buildings (on
Amiens Downtown average 85%) . . . but sometimes
Recall 95,7 95,45 80,3 critical (< 80%)
Precision 99,8 48,09 28,49
Precision (Economical Effectiveness)
Amiens Pont-de-Metz
relatively low for new buildings
Recall 84,98 96,55 92,13
→ < 20% in Toulouse Rocade
Precision 98,57 69,65 38,86
Toulouse Rocade
Recall 80,52 96,99 76,47
Precision 99,55 37,61 18,44
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
63. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Recall (Qualitative Effectiveness)
Quantitative Evaluation (50cm) very good for modified and
demolished buildings ( 95%)
Unchanged M&D1 New good for new buildings (on
Amiens Downtown average 85%) . . . but sometimes
Recall 95,7 95,45 80,3 critical (< 80%)
Precision 99,8 48,09 28,49
Precision (Economical Effectiveness)
Amiens Pont-de-Metz
relatively low for new buildings
Recall 84,98 96,55 92,13
→ < 20% in Toulouse Rocade
Precision 98,57 69,65 38,86
better for demolished and
Toulouse Rocade
modified buildings (70% in
Recall 80,52 96,99 76,47
Amiens Pont-de-Metz)
Precision 99,55 37,61 18,44
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery
64. General and specific contexts
Method description and Evaluation Criteria Evaluation at 50cm
Experiments Comparison 50cm/70cm
Conclusion
Evaluation at 50cm
Analysis
Recall (Qualitative Effectiveness)
Quantitative Evaluation (50cm) very good for modified and
demolished buildings ( 95%)
Unchanged M&D1 New good for new buildings (on
Amiens Downtown average 85%) . . . but sometimes
Recall 95,7 95,45 80,3 critical (< 80%)
Precision 99,8 48,09 28,49
Precision (Economical Effectiveness)
Amiens Pont-de-Metz
relatively low for new buildings
Recall 84,98 96,55 92,13
→ < 20% in Toulouse Rocade
Precision 98,57 69,65 38,86
better for demolished and
Toulouse Rocade
modified buildings (70% in
Recall 80,52 96,99 76,47
Amiens Pont-de-Metz)
Precision 99,55 37,61 18,44
a high recall for unchanged buildings
→ work saved for operators
1 - Modified and Demolished
15/18 Champion et al. IGARSS 2011 2D Change Detection from satellite imagery