KIM Validation for EO Archived Data Exploitation Support KIMV

1
KIM Validation for EO Archived
DataExploitation Support (KIMV)
Mihai Datcu DLR Oberpfaffenhofen
2
1997 first tests (ETH Zurich) whats that? ¼
scene 1999 MMDEMO (ETH Zurich) I2M exists and
works! 10 scenes 2002 KIM (ESA, DLR, ETHZ, NREC,
EUSC) system, appl. Use 20 GB 2003/4 use KIM in
projects SMART, PRESENCE 2003/4 Information
Theory for evaluation. 2005 KIMV operational
system, bugs, enhancements accuracy, application
scenarios1 TB, 1m 1Km, optical and
SAR. .more TBmore usersmore sensors TerraSAR
3

• KIM/KES system concept
• Knowledge-driven Information Mining (KIM)
• Knowledge enabled services (KES)
• KIM and KES are based on Human Centred Concepts
• Implements improved feature extraction
• search on a semantic level
• availability of collected knowledge
• interactive knowledge discovery
• share knowledge
• new visual user interfaces

4

• KIM and KES systems
• A library of algorithms which is used to extract
  the primitive features
• A machine learning (Bayesian network) algorithm
  to generate interactively image classifications
• A data base management system for the image
  content information catalogue and semantics and
  knowledge
• The systems are helping the user in his
  analytical task to extract the information the
  system records the knowledge and can reuse or
  communicate it. In addition, KIM and KES adapt to
  the user conjecture and are designed to operate
  very fast on large image volumes.

5
System complexity KIM/KES integrate natural
language text numerical records GIS spatial data
representation database and visual
capabilities analysis of multidimensional
pictorial structures computer vision pattern
recognition relational data models knowledge
representation and bases
6
System complexity Important complexity
factors is the unbalanced ratio between the
huge information volume of EO data (i.e. enormous
image archives) and the sequential, mainly
linguistic, and limited capacity of people to
access information perception of information
as signals-signs-symbols'' is generally not
dependent on the form in which the information is
presented but rather on the context in which it
is perceived, i.e. upon the intentions and
expectations of the perceiver.
7

• Validation procedure
• objective evaluation of system performance
• relevance in real applications with users in the
  loop, i.e. validation from the subjective
  perspective of the users interested in specific
  data and applications.

8
The expert evaluators KIM/KES systems
respond to existing and new requirements of a
very broad range of applications aerospace
agencies (ESA, CNES, NASA, DLR) satellite
centres (EUSC, ARCS) universities and research
units industry data providers
9
The tasks access to information in large EO
archives image interpretation understanding
phenomena target or objects detection informatio
n mining and knowledge discovery
10
The data sets
11
The operation modes Content Based Image
Retrieval (CBIR) CBIR is based on utilization
of semantic queries CBIR enables an operator to
see into a large volume Data/Information
mining explore the information content of the
images probabilistic image retrieval integrated
with interactive learning and image
classification Scene understanding derive
knowledge, interpret or understand the structures
and objects
12
The questionnaire evaluation for information
retrieval systems, man machine communication and
image classification rank the user satisfaction
on scale with 4 qualitative values (very good,
good, acceptable, uncertain) semantic
differentials for questionnaire-based
system validation (for characterization of the
task, search process, retrieved result and system
behaviour) evaluation of the man-machine
communication ( extent of system functionalities,
effects on the user, specific system like items,
and general score) guideline for a general
assessment of the validation results and
suggestions.
13
CBIR results analysis
14
I2M results analysis
15
SU/Classification results analysis
16
Man-Machine Communication
17
MERIS the classification
18
The method
19
Results
20
Results
21
Results
22
Conclusions            cloud and water the
classification given by training the system is
more than 90 similar to level 2 product in most
of the cases.            land classification is
not as similar as for cloud and water, this is
due to two facts          land could be
covered by cloud land is a very general
concept            ice classification a big
difference is detected. level 2 product
classification is considering ice over the water
and for as it is a classification of snow.
  snow classification level 2 product is
including the snow in cloud class, meanwhile KIM
can separate snow and cloud as two different
classes
23
Feature constancy (data models) Gemetry (HR vs.
LR) SPOT (CNES) data quality semantic labels
grows with higherresolution MERIS Landsat
ERS SPOT IKONOS 10
100 10 k100 1000