Image information mining in long temporal SAR sequences over urban areas

1
Image information mining in long temporal SAR
sequences over urban areas
ESA EUSC Workshop 2008
Frascati, March 0406 2008

• Paolo Gamba, Fabio DellAcqua
• Dept. Of Electronics
• University of Pavia
• Giovanna Trianni
• SERCO c/o ESA/ESRIN

Presenter Giovanna Trianni
2
Purpose of the study and state of the art

• Analysis of long temporal SAR series.
• The huge amount of medium resolution SAR data
  made available by ERS 1, ERS 2 and Envisat
  satellites has not been exploited, except for
  agricultural mapping applications (mostly, rice
  and forest)
• Temporal trends in land use changes are
  completely delegated to optical sensors (e.g.
  Landsat).
• This is a first attempt to exploit the
  information contained in long series of temporal
  SAR data.
• Statistical analysis carried out over four
  classes (urban areas, water, urban vegetation and
  extra urban vegetation) with particular
  attention to human settlements.

3
Framework for long temporal SAR sequence analysis
(1)
hyper temporal data series with
Same approach as for multi spectral images
temporal response
Two approaches
PER PIXEL
PER REGION
4
Framework for long temporal SAR sequence analysis
(2)

• Amplitude SAR data are affected by speckle noise.
• No classification procedure should be directly
  applied to the multi temporal pixel
• Reduction of the noise by means of a
  multitemporal speckle filter.
• The optimal approach requires the computation of
  the complex correlation between any pair of
  images.
• Simplified assumption each image is uncorrelated
  to the previous one
• The non correlation assumption doesnt hold for
  urban areas, where coherence is high over long
  time periods.
• Per pixels analysis of urban areas has very
  poor performances
• Only segments of class urban as a whole have
  been considered here, and other statistical
  parameters rather than pixel intensity values .
• Per-segment analysis -gt Segmentation by using
  Corine Land Cover

5
First test set

• October 3rd, 1994 (ERS-1)
• November 9th, 1994 (ERS-1
• July 22nd, 1995 (ERS-1)
• October 29th, 2000 (ERS-2)

• August 13th, 1992 (ERS-1)
• October 22nd, 1992 (ERS-1)
• June 24th, 1993 (ERS-1)
• November 21st, 1993 (ERS-1)

6
Experimental Results
Log-mean
Log-variance
7
Experimental results over Pavia
8
Second test set
73 ERS 1/2 scenes recorded between 1992 - 2000
This data set has been kindly provided by T.R.E.
s.r.l.
9
Experimental results
10
Per-segment approach in land use monitoring (1)
Another interesting application of the per
segment approach is land use monitoring, which
could be implemented both among and within
classes. It is possible to sub-classify different
urban areas with respect to the fact that they
are maintaining, increasing or reducing their
mean scattering value, hint to a stability,
increase or decrease of built-up areas.
11
Per-segment approach in land use monitoring (2)
backscattered values
12
Conclusions

• Long temporal SAR sequences may provide
  interesting information about changes in the
  imaged area.
• As a consequence, very simple analysis tools,
  such as comparison of time trajectories, can be
  used for image mining.
• The methodology is based on extraction of
  significant statistical parameters for each
  segment and their comparison across the whole
  (long) temporal sequence.
• The first application of this procedure is
  naturally change detection, either for sudden
  change extraction or long term change analysis
  (shown in this work).
• Sudden change detection could be instead useful
  to understand where a given transition (e.g. in
  land cover) took place at the same time of a
  known one, by automatically comparing the
  temporal parameter values sequences.
• The framework proposed here is just a first
  attempt to exploit the wealth of information
  somehow hidden in long temporal SAR sequences.
• Future analyses will be devoted to a more precise
  evaluation of the stability of the approach using
  different GIS layers as initial segmentation.