Geospatial Data Analysis for Security Applications at the JRC

1
Geospatial Data Analysis for Security
Applications at the JRC

• Challenges and Opportunities for Information
  Mining
• Frascati, 4 March 2008
• Katrin Molch
• katrin.molch_at_jrc.it

2
Overview

• Introduction to JRC-ISFEREA
• ISFEREA Interest in Image Information Mining
• Challenges of our work
• Expectations from IIM
• KIM ('Knowledge-Based Information Mining') Test
• Overview
• Results
• Conclusions

3
JRC-ISFEREA Activities
ISFEREA Information Support for Effective and
Rapid External Action Geospatial information
and analysis in support of crisis management,
e.g.

• Damage assessment
• Reconstruction monitoring

4
Example - Damage Assessment
Pre-event IKONOS
Post-event QUICKBIRD
Field photograph

• Rapid and accurate
• Count of buildings affected
• Assessment destroyed vs. damaged buildings

5
ISFEREA's Interest in IIM

• Our challenges
• Large area coverage high and very high resolution
  (optical) data
• Time constraints
• Labor-intensive manual interpretation
• IIM functionality of interest expectations
• Not to provide a classification
• To focus interpretation efforts on relevant image
  subsets
• Criteria
• Timely, reliable robust
• Low false negative rate -- none missed
• Integrate in in-house workflow
• ? Test of IIM system 'KIM' at ESA

6
KIM System - Overview
Earth Observation Image
Interactive Training
Image Query
Data Ingestion
Image Tiling
Feature Extraction
Clustering
Label Generation
Label Export
Image Classification
Ranking
7
Challenges for KIM

• High and very high resolution (optical) data
• Unpredictable data qualities
• Multi-temporal
• Different sensors
• Processing
• Semantic concepts - e.g. building, bridge
• Complex
• Possibly not well represented in generic features
  extracted
• Pushing KIM beyond its intended purpose
• Determine opportunities limitations
• Ideas for further development

8
KIM Test

• Collection
• Pre-event Ikonos product (2005)
• Post-event Quickbird product (2006)
• Features
• Spectral, texture, geometry
• Label generation
• Semantic class 'buildings'
• Two features at a time
• Ranking of tiles in collection
• 'Probability', 'separability', 'coverage'
• Simulate damage assessment scenario
• Quick overview of location of - potentially
  affected - buildings

9
Probabilistic Label Generation
IKONOS product pre-event (2005) Semantic class
Buildings
Spectral and texture at full-resolution
Spectral and Hu moments
Spectral and area
10
Ranking
Matched
Not matched
11
Ranking
Matched
Not matched
12
Summary of Results

• Several nodes required for ingestion
• Easy to generate coarse classification of large
  amounts of data
• System inner workings not transparent to user
• Ranking satisfactory for complex concept
  buildings
• Feature extraction and ranking sensitive to
  radiometric differences

13
Conclusions

• Ranking useful as indicator to focus
  interpretation efforts
• Difficulties in
• Handling of VHR data information density and
  complexity
• Associated user expectations
• Further IIM developments towards higher
  resolution data and more complex semantic concepts

14
Katrin Molch katrin.molch_at_jrc.it