DLR R

1
DLR RD on IIM for high resolution images first
operational step - TerraSAR
Mihai Datcu
2
Past years Meta-data based file access
The meta - data
The data archive
The data
Data access
3
Today Interactive, user adapted, EO data content
access
Help to image classification
Suggest data
Mine Fields (Daedalus)
Access to information knowledge
Knowledge share
Help to image understanding
4
Whats new with high resolution EO High
resolution images transform 2D in 3D Cognitive
like interpretation shall encapsulate the
signal/instrument models (non visual data!) the
physical models Understanding shall
be adapted to the users context free of
prejudgements
5
Postulates Existing volume of unstructured
data prevents any systematic exploitation of its
information content Information extraction
depends critically on the descriptive or
predictive accuracy of the stochastic model
employed The augmentation of the data with
meaning, e.g. image understanding, can be
interpreted as a coding task which includes the
model of users conjecture Paradox People have
trouble in caching more than 7 items a time We
design systems to enable people to access 1000 TB

6
CONCEPT DE COMMUNICATION AVANCE
Sourcedinformation
Utilisateurs
Extraction de linformation
Représentation sémantique
Inférence du modèle de signal objectif
Inférence du modèle d information subjectif
Modélisation de la conjecture utilisateur
Modèles sémantiques
Modèles syntaxiques
7
CONCEPT DE COMMUNICATION AVANCE
Sourcedinformation
Utilisateurs
Extraction de linformation
Représentation sémantique
Inférence du modèle de signal objectif
Inférence du modèle d information subjectif
Modélisation de la conjecture utilisateur
Modèles sémantiques
Modèles syntaxiques
An observation, strictly, is only a sensation.
But as soon as we go beyond sensations we are
making inferences. Jeffreys
8
Spatial data modelling

Hierarchy of Information Representation
9
Spatio-temporal data modelling
10
Coding
11
Theoretical Premises

• Rate Distortion Based Analysis of Image
  Parameters Estimation for Information Mining
• Parameter estimation and theoretical accuracy
  bound Cramer/Rao
• Elements of R/D theory
• where is the squared error
  distortion
• Evaluation on algorithm libraries and sensor data
  sets
• GMRF algorithm
• Dyadic K-means
• Evaluation of the optimal size of the analyzing
  window for Landsat and Daedalus data

12
Concept

• R/D based concept for evaluation of data models
  and estimation algorithms
• Theoretical behavior for homogeneous texture
  (Brodatz data set)

• Visualization in the feature space of one sample
  image of the dataset
• For analyzing window of size 10 (left-plot)
• For analyzing window of size 60 (right-plot)

13
Results

• Results for assessment of non-stationarity data
• Landsat data set (WL30)
• Daedalus data set (WD20)
• Where W is the estimated optimal size of the
  analyzing window

14

• Incremental Clustering

• Vector quantization
• Grid-based density estimation
• Incremental compuations
• Adaptive to the new image data
• E.g. Mozambique coverage with Landsat images

15
Clustering and image complexity
16
MDL used to code object dynamic spatio-temporal
clustering

• Graph of trajectories

• Dynamic classification

17
INFORMATION BOTTLENECK PRINCIPLE - emerged from
Rate-Distortion theory. - a formalism to express
the trade-off between compression (short summary)
and the relevant information contained in the
summary.
18
Information Bottleneck can be viewed as a
Rate-Distortion problem based on KL divergence
19
Clustering and coding
20
Clustering and coding
21

• Kullback-Leibler (KL) similarity search

22

• Kullback-Leibler (KL) similarity search

KL between 2 images, 2 models each one
Probabilistic search results
local differences
local permutations
KL search results
23
DIRICHLET MODEL

• after Ni instances the likelihood is

24
DIRICHLET MODEL

• after a new training data set

25
Semantic coding 3
26
G0
GK
time

• On définit une transformation composée
  dopérations élémentaires
• f ?0 ? ?k? ?k U ?
• de coût égale à une somme pondérée de coût
  partiels relatifs aux similarités entre les
  différents attributs de graphes (différence,
  divergence de Kullback-Leibler)

• Apprentissage interactif de la distribution p(?
  T )
• a priori conjugué de Dirichlet p(?)
• p(?T(1))
  Dir(?1N1(1), ,1Nr(1)).

27
(No Transcript)
28
Advanced communication
Clusters
Semantic labels
Images
Ik
?j Model 1
p(? I )
p(L ? )
?k Model 2
29
Controlling the semantics
Signal classes

• Problem

Labels
In the system there may be labels with
different names and with the same information
content (the same meaning)
L1
. . . .
L2
Table LABEL from DataBase
30

• Solution similarity measure , Kullback-Leibler
  divergence

• Example

The initial label for two classfiles
The labels from database for the same classfiles
After Kullback-Leibler procedure the sorted
labels list is 2, 1, 6, 4, 5, 9, 7, 8. 3
31
KIM for MERIS SSE
Other mission Archive
MERIS RR Archive
User
FTP
Automatic Transfer Activation
Alga Bloom Service?
Ingestion
Subscription
e-mail
KIM Catalogue
Cloud-free Service
Orders
KIM
Information
Information Mining
SSE
User Authorisation
Expert
User
Service n
SSE Service Support Environment KIM
Knowledge-based Information Mining
32
TerraSAR sensor modes
33
Target discover and analysis FGAN PolSAR
Applications
Mapping by fusion of SRTM DEM and SAR image
34
Target analysis
35
TerraSAR Payload Ground Segment (GS)
Proposed Architecture
36
KEO - Knowledge-centered Earth Observation
Objectives a prototype system and environment
to foster the enlargement of EO data utilisation,
and in particular of the large archives of
multi-mission and multi-temporal images, provide
a better support to research, value-adding
industry, service providers and EO user
communities, like scientific investigations, risk
and disaster management, or in the GMES
programme Data TerraSAR, ENVISAT, ERS, Landsat,
SPOT (data volume 100GB/day.) Evaluators and
users ESRIN, DLR, EUSC, CNES, universities,
industry. Output operational system
37
KEO, PIMS TerraSAR X Payload GS
38
Research in the frame of the CNES DLR-ENST
Competence Center on Information Extraction and
Image Understanding for EO
Modelling and estimation of HR SAR data
(TerraSAR) Target analysis (TerraSAR) Geometrical
and topological description of HR SAR scenes
(TerraSAR) Multimodal SAR data analysis
(TerraSAR) KIM DIMS communication bridge
(TerraSAR) Mining image time series Joint image
compression and information extraction for
IIM Cognitive scene understanding Learning and
unlearning paradigms Adaptation to the user
conjecture and cognitive systems
39
RD in the frame of ESA projects PIMS and KEO
Feature extraction for EO application Feature
space analysis and data reduction Validation of
KIM for SAR data IIM for SRTM DEM data IIM for
SRTM SAR and TerraSAR images IIM scenario
development (TerraSAR) KIM DIMS interfacing
(TerraSAR) KEO architecture compatible with ESA
SSE Architecture design for future GS Mining
structured and un-structured data (TerraSAR)