An InformationTheoretic Feature for Multitemporal Analysis of SAR Images

1
An Information-Theoretic Feature for
Multitemporal Analysis of SAR Images
Luciano Alparone, Bruno Aiazzi, Stefano
Baronti, Andrea Garzelli Dept. Electronics
Telecommunications, University of Florence,
Florence, Italy Institute of Applied Physics
Nello Carrara (IFAC-CNR), Florence,
Italy Dept. Information Engineering, University
of Siena, Siena, Italy
2
Summary

• Scenario and motivations
• Automated analysis of SAR images ? speckle
• Multitemporal analysis of SAR images ? Log-Ratio
  operator
• Multitemporal changes related to conditional
  information between radar backscatters on two
  pass dates, according to Shannons Theory
• Definition of an information-theoretic change
  feature at pixel level
• Experimental results on two sets of true SAR
  images
• Conclusions

3
Scenario and Motivations

• Synthetic Aperture Radar (SAR) provides
  multitemporal observations of the Earth with all
  weather conditions.
• Both spatial and temporal analysis are hindered
  by the presence of the noise typical of all
  coherent imaging systems called speckle.
• Scene heterogeneity in SAR images has been
  traditionally measured by the local coefficient
  of variation defined as the ratio of local
  standard deviation to local mean.
• Image texture may be characterised by parameters
  derived from the grey-level co-occurrence matrix
  (GLCM), which, however, is sensitive to noise.
• Changes between two pass dates may be highlighted
  by the logarithm of the pixel ratio between two
  co-registered SAR images (Log-Ratio), which
  however, requires heavy pre-processing for
  speckle reduction.

4
Information-Theoretic Problem Statement

• Rationale the negative of logarithm of the
  probability of an amplitude level in one image
  conditional to the level of the same pixel in the
  other image conveys an information on the amount
  of change occurred between the two passes.
• The conditional probability is denormalised in
  such a way that it is one, hence the change
  feature is identically zero, when the joint
  probability attains its maximum on either the row
  or the column.
• Depending on such a normalisation, some changes
  may be emphasised or concealed.
• All joint probabilities are calculated by
  digitising the 2D histogram (scatterplot) of
  locally windowed backscatters, e.g. 256 ? 256,
  and smoothing the outcome by means of a Gaussian
  filter.

5
Multitemporal IT Change Feature

• Take two co-registered SAR images g1(m,n) and
  g2(m,n) of the same scene taken at different
  times and/or by different systems.
• Calculate the joint probability p(i,j), with i
  being the level of g2 and j that of g1 at the
  same pixel position.
• Let us define a normalised conditional probability

• Define the multitemporal information-theoretic
  pixel feature as

• Values of C(m,n) with g1(m,n) lt g2(m,n)
  constitute the map of changes in excess.
• Values of C(m,n) with g1(m,n) gt g2(m,n) are
  complemented to yield the map of changes in
  defect.

6
Multitemporal RadarSat Data Set

• RadarSat image acquired on 06/04/95

• RadarSat image acquired on 02/11/95

• The first date shows an area in Canada.
• The second date shows the same area after a fire.

7
Maps of IT Feature Between 06/04 and 02/11

• Changes in defect

• Changes in excess

• Normalisation of joint PDF by maximum along i

8
Maps of IT Feature Between 06/04 and 02/11

• Changes in defect

• Changes in excess

• Normalisation of joint PDF by maximum along j

9
Map of Log-Ratio Between 06/04 and 02/11

• Changes in defect (lower part of histogram)

• Changes in excess (upper part)

• Pre-processing with Kuans despeckling filter on
  a 9?9 window

10
Multitemporal Data Set of the City of Pavia

• RadarSat image acquired on 20/10/00

• ERS-2 image acquired on 28/10/00

• The first pass date is immediately after a flood
• Eight days after most of flooded lands have been
  drained

11
Maps of IT Feature Between 28/10 and 20/10

• Changes in defect

• Changes in excess

• Normalisation of joint PDF by maximum along j

12
Maps of IT Feature Between 28/10 and 20/10

• Changes in defect

• Changes in excess

• Normalisation of joint PDF by maximum along i

13
Map of Log-Ratio Between 28/10 and 20/10

• Changes in defect (lower part of histogram)

• Changes in excess (upper part)

• Pre-processing with Kuans despeckling filter on
  a 9?9 window

14
Assessment of IT Multitemporal Feature Log-Ratio

• ROI on IT feature (analogously for Log-Ratio)

• Ground truth of changes on ROI

15
ROC Plots of IT Multitemporal Feature Log-Ratio

• A threshold is increased starting from zero
  (rightmost in the plot)
• All pixels whose feature is above the threshold
  are assumed to be changed
• Area below Log-Ratio 0.932
• Area below IT Multitemporal Feature 0.966

16
Conclusions

• A novel pixel feature suitable for change
  analysis was derived from information-theoretic
  concepts.
• The conditional information between radar
  backscatters measured on different times seems to
  bring advantages over conventional Log-Ratio
  analysis for highlighting changes occurred across
  pass dates, thanks to low sensitivity to noise
  and no pre-processing requirements.
• Experiments have demonstrated that the proposed
  information-theoretic change feature is capable
  of providing accurate change maps from couples of
  SAR images.