Download Unsupervised Spatio-Temporal Mining of Satellite Image Time Series

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Document related concepts

Cluster analysis wikipedia, lookup

Nonlinear dimensionality reduction wikipedia, lookup

Transcript
ANR-07-MDCO-04
http://www.efidir.fr
Unsupervised Spatio-Temporal Mining of
Satellite Image Time Series
A. Julea, N. Méger, E. Trouvé, Ph. Bolon, C. Rigotti., M-P. Doin and
C. Lasserre
FOSTER kickoff, Lyon, the 21st of January 2011
1
Outline
1. Problem statement
2. SITS and base of sequences
3. Frequent sequential patterns
4. Incorporating spatiality: the FGS-patterns
5. Applications
6. Conclusion
2
1. Problem statement: objectives
To describe a SITS temporally and
spatially.

Unsupervised techniques.

time
3
1. Problem statement: approach


Assumption: pixels relating to a same evolution or
sub-evolution should be numerous and should be
connected.
Constraints


All possible evolutions and sub-evolutions must be
taken into account for all pixels: combinatorial
explosion + millions of pixel evolutions to be mined,
Different occurrence dates of a same sub-evolution
should be considered to be robust to noisy
acquisitions: clustering methods are discarded.
4
2. SITS and base of sequences
Pixel values are described using symbols (e.g., quantization, clustering).
t1
t2
t3
t4
p1
p2
p3
p4
p5
p6
p7
p8
p9
SITS
p1
t1
t2
t3
t4
p2
t1
t2
t3
t4
p3
t1
t2
t3
t4
p4
t1
t2
t3
t4
p5
t1
t2
t3
t4
p6
t1
t2
t3
t4
p7
t1
t2
t3
t4
p8
t1
t2
t3
t4
p9
t1
t2
t3
Base of sequences
t4
5
3. Frequent sequential patterns



Extracted model: sequential patterns
Measure: support, i.e., the number of pixels
supporting an evolution/sub-evolution. It can be
used to extract evolutions/sub-evolutions shared
by at least σ pixels, the frequent sequential
patterns.
σ , termed as the minimum support, is used to set
an active constraint to prune the search space
(anti-monotonicity property).
6
3. Pruning the search space using σ
7
4. Incorporating spatiality: the FGS-patterns

A new measured is proposed: the average connectivity of
evolution α, i.e. the average number of pixels covered by α
in the 8-neighborhood of the pixels covered by α.
Links(α)
support(α)


It is used to extract frequent sequential patterns covering
pixels that are sufficiently connected: a minimum average
connectivity is set, κ. They are termed as Frequent Grouped
Sequential patterns (FGS-patterns).
This constraint is not anti-monotonic but ...
8
4. Pushing the average connectivity constraint

AC(α) <= Links(α)/σ

Thus if Links(α)/σ < k, then AC(α) < k.
•
Asking for patterns having Links(α)/σ >= k is
an anti-monotonic constraint.
•
Strategy: we explore the search space using
Links(α)/σ >= k, and we select patterns having
AC>=k.
•
Speed-up: between 7% and 40%.
•
Complementing to support pruning.
9
5. Applications
Crop monitoring by optical remote sensing
10
Available data
●
ADAM SITS: data assimilation through agro-modeling (http://kalideos.cnes.fr)
●
60 SPOT images between 2000 and 2008.
●
3 bands: B1 in green (0.5-0.59 μm), B2 in red (0.61-0.68 μm), and B3 in near
infrared (NIR 0.78-0.89 μm).
●
Spatial resolution: 20m×20m.
●
Observed scene: rural area in East Bucharest, Romania.
11
Data selection and preprocessing
●
●
●
●
Sub-scene: 1000×1000 pixels.
The ground truth is available for the 2000-2001 period
(fields belonging to the Romanian National Agricultural
Research and Development Institute ,5.9% of the
scene).
20 images between October 2000 and July 2001.
A synthetic band B4 is computed. It gives the
Normalized Difference Vegetation Index (NDVI). B4 =
B3−B2 / B3+B2. Pixel values are quantized into 3
intervals.
12
Quantitative results
Standard PC: Intel Core 2 @ 3GHz, 4 GB RAM,
linux kernel 2.6.3
•
•
•
•
•
σ= 1%
k= 5
Execution times: 750s
Number of FGS-patterns: 474
Maximal FGS-patterns are selected (32 patterns)
.
13
Results (σ=1%, κ=5)
White pixels relate to a single FGS-pattern which characterizes wheat crops.
It covers 61.4% of the wheat crops of the ground truth and
91.3% of the white pixels indeed match wheat crops.
14
5. Applications
Crustal deformation monitoring by differential
SAR interfermotry
15
Available data
●
Lake Mead.
●
About 56 ERS/ENVISAT acquisitions (1996-2008).
●
Spatial resolution: 130m x 130m
●
Interferograms: image of displacements/master image.
+ DEM
Slow and long wavelength ground motion
Problem: the signal is affected by spatialy and temporaly
random atmospheric perturbations …
Ripples, bubbles,
patches, fronts, ....
17
Data selection and preprocessing




20 images
The stratified atmosphere is corrected but the turbulent
atmosphere is still present.
Displacements/master image = f(“real” displacements +
atmosphere of the master image + atmosphere of the
acquisition).
Pixel values(phase differences) are quantized into 3
intervals.
18
Quantitative results
Standard PC: Intel Core 2 @ 3GHz, 4 GB RAM,
linux kernel 2.6.3
•
•
•
•
•
σ= 2%
k= 6
Execution times: 4900 s
Number of FGS-patterns: 10073
Longest pattern are selected (5 patterns)
.
19
Results (σ=2%, κ=6)
A
B
A: the white pixels correspond to the 4 longest evolution (out of 5) stating that those
pixels stick to high positive phase difference values before switching to low negative
ones ( 3>3 … 1>1).
B: these pixels correspond (clear ones) to strong positive regression coefficients
between the phase differences and the water level fluctuations. The atmospheric
patterns are thus not considered and the displacements patterns are focused on.
Results (σ=2%, κ=6)
A
B
A: the white pixels correspond to a single long evolution stating that the pixel values
stick to low negative phase differences (1>1...>1).
B: these pixels (clear ones) correspond to a negative average speed (uplift), in
particular for the Las Vegas area. That can be explained by a decreasing water
pumping from the Las Vegas aquifers.
5. Applications: references
A. Julea, N. Méger, P. Bolon, C. Rigotti, M.-P. Doin, C. Lasserre, E. Trouve and V. Lazarescu., Unsupervised
Spatiotemporal Mining of Satellite Image Time Series using Grouped Frequent Sequential Patterns, IEEE
Transactions on Geoscience and Remote Sensing, to appear, vol. 49, issue 4, 2011, 14 pages.
Julea A., Méger N., Rigotti C., Doin M.-P., Lasserre C., Trouvé E., Bolon P., Lazarescu V., Extraction of
Frequent Grouped Sequential Patterns from Satellite Image Time Series, IEEE Int. Geoscience And Remote
Sensing Symposium (IGARSS 10), CD-ROM , Honolulu, HI, USA, July 2010, 4 pages.
Other datasets and multiband:
Julea A., Méger N., Trouvé E., Bolon P., Rigotti C., Fallourd R., Nicolas J.-M., Vasile G., Gay M., Harrant O. et
al, Spatio-Temporal Mining of PolSAR Satellite Image Time Series, ESA Living Planet Symposium - The 2010
European Space Agency Living Planet Symposium, Bergen, Norway, July, 6 pages.
Julea A., Méger N., Trouvé E., On mining METEOSAT and ERS Multitemporal Images, 4th Conf. on Image
Information Mining for Security and Intelligence (ESA-EUSC 2006), CD-ROM , Torrejon Air Base - Madrid,
Spain, November 2006, 6 pages.
Julea A., Méger N., Trouvé E., Sequential Patterns Extraction in Multitemporal Satellite Images, 10th European
Conf. on Principles and Practice of Knowledge Discovery in Databases (PKDD'06), Practical Data Mining
Workshop: Applications, Experiences and Challenges, Berlin, Germany, September 2006, pp. 94-97.
22
6. Conclusion
•
FGS-patterns allow to describe a SITS
spatially and temporally.
•
FGS-pattern extractions are tractable.
•
Future work include providing a single
clustering related to thematic classes:
–
land-use
–
snow, firn, ice …
23
ECOLE DE PRINTEMPS
SPRING SCHOOL
Extraction and Fusion of Information for
Displacement measurement from SAR Imagery
Extraction et Fusion d’Informations pour
la mesure
Mayde
1st Déplacement
- May 6th, 2011 par Imagerie Radar
Les Houches, Chamonix Mont-Blanc, FRANCE
http://www.efidir.fr
[email protected]
ECOLE DE PRINTEMPS
SPRING SCHOOL
• Program: Courses and practical work
– Ground electromagnetic response: from physical properties to SAR and
polarimetric SAR data
– InSAR D-InSAR processing: from raw data to displacement measures
– Multitemporal D-InSAR: from SAR time series to displacement monitoring
– Knowledge and information processing
– Inversion and data assimilation in geophysical models
+ Field experiment
•
Registration deadline: February 1st, 2011
•
Fee: 500€ - Students: 365€
http://www.efidir.fr
[email protected]