Diapositiva 1

1
University of CassinoSchool of
Engineering University of Firenze
School of Engineering
Image Communication Lab.
Content Extraction from SAR Images via
Segmentation of Textural Features and Region
Based Classification
Maurizio Abbate, Bruno Aiazzi, Luciano
Alparone, Stefano Baronti, Ciro DElia, Gilda
Schirinzi Department DAEIMI, University of
Cassino, Cassino, Italy Institute of Applied
Physics Nello Carrara (IFAC-CNR), Florence,
Italy Department of Electronics
Telecommunications, University of Florence,
Florence, Italy
2

Presentation Outline

• Scenario and motivations
• Proposed method for SAR image segmentation
• Information-theoretic textural features
  extraction
• Segmentation based on Tree-Structured MRF
• Classification achieved by clustering features of
  segments
• Results
• Conclusions

3

Scenario and Motivations

• Image segmentation useful for classification,
  data compression, restoration, etc.
• Image segmentation algorithms exploit image
  models relying on local homogeneity.
• SAR images corrupted by speckle noise.
• Traditional (optical) image segmentation
  algorithms inadequate to SAR images.
• Information-Theoretic Heterogeneity Features
  Aiazzi et al., IEEE TGRS, 2005 suited to
  overcome this problem.

4

Heterogeneity Features
5

Segmentation through TSMRF (1/3)
6

Segmentation through TSMRF (2/3)

• Probabilistic image models allow to state the
  segmentation problem as a maximum a-posteriori
  (MAP) estimation problem.
• MRFs allow the joint a-posteriori probability law
  of a huge number of variables to be represented
  by using local conditional probabilities.
• MAP estimation may be obtained by Iterated
  Conditional Mode (ICM) or Simulated Annealing
  (SA) algorithms.
• Unfortunately the complexity of such algorithms
  depends exponentially on the number of classes.

7

Segmentation through TSMRF (3/3)
8

SAR Data Ground Truth

• NASA/JPL SIR-C C-band polarimetric SAR data,
  acquired on April 16th 1994 over the city of
  Pavia, Italy 787 787 area in the HH channel.
  Pixel spacing 10.5 m.
• Pink high density urban area (2.5) blue
  medium density urban (5.4) green low density
  urban (5.1) red industrial (0.6) black
  vegetated (86.4)

9

Segmentation Results

• Feature Map (Joint Information)

Segmentation Map
10

Overlay with Optical Image (1/5)
11

Overlay with Optical Image (2/5)
12

Overlay with Optical Image (3/5)
13

Overlay with Optical Image (4/5)
14

Overlay with Optical Image (5/5)
15
Classification Procedure

• Calculation of textural features to yield pixel
  vectors.
• Initial centroids based on pixel vectors
  calculated from training set, if ground truth
  data are available, otherwise by clustering
  (either crisp or fuzzy) the set of vectors.
• Iterative reclustering of vectors into
  dynamically upgraded classes based on a
  Mahalanobis-like weighted distance.
• Refinement of centroids and weights through a
  fuzzy nearest-mean reclustering procedure
  enhanced by an entropy minimizing membership
  function (EFNMR) aimed at preserving minor
  classes.
• A crisp classification map is computed at each
  iteration and used as startup map for the next
  step.
• Convergence occurs after 3-4 iterations.

16
Flowchart of EFNMR Classifier
dm(n) distance between pixel vector x(n) and
centroid c(m) K of components of pixel
feature vector ?k feature-dependent weights.
17
Maximum-Entropy Membership Function
Um(n) membership of pixel vector x(n) to
centroid c(m) M of centroids, i.e., of
clusters 0 lt ? lt 1 positive constant optimizing
class separability 1 lt ? lt 2 membership
exponent optimizing convergence.
18
Confusion Matrix (HH only)
Mean score 67.0 average score 47.8 average
score of structured classes (H, M, L, I) 42.0.
19
Classification Map (HH HV)
Mean score 77.2 average score 49.4 average
score of structured classes (H, M, L, I)
42.1.0 vegetation score 83.5.
20

Conclusions and Developments

• We proposed an operational framework in which
  application-dependent processing of remote
  sensing imagery, including SAR, may be cascaded
  to more general methods that are unconstrained
  from specific applications.
• The joint use of information-theoretic
  heterogeneity features and TS-MRF segmentation is
  promising, to overcome limitations deriving from
  the nature of SAR data.
• Preliminary results of classification of
  different types of urban environment by means of
  a feature clustering algorithm suggest possible
  inaccuracies of the ground truth.
• The proposed solutions can be profitably used by
  the processing engine of an image information
  mining system for Earth Observation (EO) data.