On Mining ERS and METEOSAT multitemporal images

1
On Mining ERS and METEOSAT multitemporal images

• 4th ESA-EUSC Conference on Image Information
  Mining for Security Intelligence, Torrejon Air
  Base, Spain, 27/11/06
• Andreea Julea, Universitatea Bucuresti - LAPI Lab
• Institutul de Stiinte Spatiale Bucuresti,
  Romania.
• Nicolas Méger, University of Savoie - LISTIC Lab,
  France.
• Emmanuel Trouvé, University of Savoie - LISTIC
  Lab, France.

2
Objective

• To find frequent evolutions at the pixel level
  over the same geographical zone.
• To test sequential patterns extraction for
  processing multitemporal images.

3
Dataset
Band 1
Band 2
time
t1
t2
t3
t4
Pixel values are represented by symbols
4
Preprocessing
For each pixel position For each band For
each date of acquisition Report pixels
values using symbols
Band 1
Band 2
t1
t2
t3
t4
5
Base of sequences
6
Frequent Sequential patterns

• Sequential patterns are ordered sets of symbols
• Support measure of a given sequential pattern
  sequences in which it occurs / sequences.
• Sequential patterns whose support is above a
  minimum support threshold s are said to be
  frequent.

7
Example
S ( ) 3/9 S (
) 3/9 S ( ) 4/9
If s 1/3 , above sequential patterns are
frequent.
8
Number of possible patterns
n number of time units
b number of bands
Sj number of symbols for one band j
With only 2 bands, 3 symbols per band and 4 time
units, we have to check the support of 54240
patterns
9
Algorithms

• APrioriAll, SPADE, C-SPADE, GSP, PrefixSpan,
  PrefixGrowth.
• Based on the anti-monotonicity of the support
  measure.
• Example if is not frequent,
  then can not be frequent and
  does not need to be counted.

10
Experiment 1 dataset and preprocessing

• 8 visible band images from METEOSAT satellite
  (256 gray scale format, 2 262 500 pixels)
• Acquired on the 7th, 8th, 9th, 10th, 11th, 13th,
  14th, 15th of April 2006 at 12.00 GMT.
• Re-discretization

pixel values
symbol
meaning
0,50
0
Water, vegetation
50,100
Soil, thin clouds
1
Sand, relatively thick clouds
100,200
2
200,255
Thick clouds, sand, snow
3
11
Experiment 1 quantitative results
Extractions performed using M. J. Zakis
prototype on AMD athlon 64 3000 with 512 MB of
RAM Under Suse Linux 10.
Execution times (s) vs minimum support
Number of frequent patterns vs minimum support
12
Experiment 1 qualitative results
Localisation of 0-gt0-gt3-gt0 (s 17,5)
Localisation of 3-gt3-gt3-gt3-gt3-gt3-gt3-gt3 (s 0,7)
Spatio-temporal loc. of 0-gt0-gt3-gt0 (s 17,5).
In red it occurs on the 1st , 2nd, 6th and 7th
day.
13
Experiment 2 dataset and preprocessing

• 5 interferometric SAR images from ERS satellites,
  acquired betwen July 1995 and April 1996 (1 048
  576 pixels) and covering the Mont Blanc area.
• Two bands were considered amplitude and
  coherence.
• Amplitude and coherence were discretized into 4
  intervals each.

Amplitude in October 1995
Coherence in October 1995
14
Experiment 2 results

• 1 7-gt1 7-gt1 7 (s7,5) is considered of interest
• - symbol 1 denotes low amplitude
• symbol 7 denotes high coherence
• gt it is related to small rocks, grass, snow and
  glacier zones.
• Blues zones on spatio-temporal localisation
  indicate an evolution in October 1995, March and
  April 1996
• gt snow and glacier zones.

Spatio-temporal loc. of pattern 1 7-gt1 7-gt1 7
15
Conclusion

• Results can be obtained using various bands at
  various resolutions.
• Results are coherent with the domain knowledge.
• Influence of discretization and filters has to be
  assessed.
• Extractions at higher levels (objects and
  regions) is a promising way.

16
(No Transcript)