SAR information extraction Primitive features for radar data mining

1
SAR information extractionPrimitive features for
radar data mining
M.Quartulli, M.Datcu
2
Approaches to vision

• D. Marr 1982 Vision, Freeman NY
• Feature-based
• Bottom-up
• - Building representations out of signals

• D.Ballard 1991 Animate Vision
• In Artificial intelligence 48, pp. 57-86
• Objective-based
• Well-posed localized problem
• Makes use of prior information

3
KimV primitive feature extraction
4
Approaches for IE
SRTM Baltimore, 25m
XSAR Bern, 30m
RSAT LasVegas, 15m
Intermap Maastricht, 0.5m
5
Bayesian model-based IE

• Existing models or explicit assumptions for p(S)
• Hierarchies of models

6
Bayesian information extraction
GMRF Space Variant G(.) Likelihood
Despeckled backscatter intensity
Backscatter Intensity texture norm
Backscatter intensity
SRTM Baltimore, 25m
7
Bayesian information extraction
GMRF Space Variant Gaussian Likelihood
Clean elevation map
Elevation texture norm
DEM elevation map
SRTM Baltimore, 25m
8
Feature fusion for 3D urban characterization
e.g. as a fusion of features derived from
intensity, height
9
Feature fusion for 3D characterization
dense urbanized
medium urbanized
light urbanized
forest
agricultural
water
Baltimore, USA
10
Building detection from SRTM data
Baltimore, USA
11
Building detection from SRTM data
water/shadow
sparse urbanized
building
forest
agricultural
Baltimore, USA
12
Feature fusion for building reconstruction
by resolving high-resolution detail
Baltimore, USA
13
Feature fusion for building reconstruction
Baltimore, USA
14
Feature fusion for building identification
Intermap Munich East, 0.5m
15
Urban metric SAR information extraction

• Historical buildings, moving vehicles, open
  windows, cars on roofs
• What models? How many parameters?

E-SAR Dresden, 0.5m
16
Urban metric SAR information extraction
17
Marked point processes
S X(s) s in D
X(s) (wi,hi,a,t)
Wx,wy,h,a,t
Wx,wy,h,a,t
18
Marked point processes
S X(s) s in D, X(s) (wi,hi,a,t)
Uniformly distributed
19
Reconstruction of built-up areasthrough marked
point processes
20
Hierarchic geometrical models
A measure of the quality of the
reconstruction p(XD)
21
Hierarchic geometrical models
A measure of the quality of the
reconstruction p(XD) 1/Z exp -U(XD)
22
Likelihood for metric urban SAR
Objects in spatial relationships Slowly varying
small number of scatterers per resolution cell
(NOT speckle)
Intermap Oberpfaffenhofen, 2m
23
Hierarchic geometrical models
The potential U(XD) is decomposed U(XD)
SSij Ul(pkfik) Si Up(xi xj) into interaction
model Up(xixj) intersection
measure(xi,xj) taking into account up to two-term
interactions data model Inverse Gaussian
(slowly varying small number of scatterers)
24
Likelihood Ratio scattering/clutter
Intermap Oberpfaffenhofen, 2m
25
Geometrical modelling of industrial buildings
Intermap Oberpfaffenhofen, 2m
26
Reconstruction of built up areasby stochastic
geometry
Intermap Maastricht, 0.5 m
27
Conclusions

• HR InSAR urban Scene Understanding by
• Pixel-based Information Fusion
• Can fuse/aggregate multiple information layers
• Intensity, Coherence, Phase
• Object-Based Scene Understanding
• explicit Scene Modeling
• better use of Sensor Modeling in
  geometry/radiometry
• results ready for 3d applications