Advanced Image Analysis Methods for Remote Sensing

1
Advanced Image Analysis Methods for Remote Sensing

• Joji Iisaka
• PFC,CFS, Natural Resources Canada
• contact jiisaka_at_pfc.forestry.ca

2
Outlines

• Introduction
• Data Management
• Intelligent Information fusion System for EOSD
• Problems of current image analysis methods
• Advanced Image Analysis Methods
• SVM, Wavelet, PSW, and FMM
• Conclusions

3
Introduction Remote Sensing in the 21st century
1

• Data volume
• More Satellites
• Resolution enhancement
• Spatial(30m-gt1-3m)
• Spectral Channel(7channels -gt 200channels)
• Time Resolution(Multiple satellites)
• 100GB/Day , 11,100TB in the year of 2010

4
Introduction Remote Sensing in the 21st century
2

• Wide variety of signals on terrain objects
• Ex. Polarimetric scattering, Interferometry vs.
  reflection
• Hyper ChannelsMore channels, sensor and
  information fusion
• Multi spatial resolution scale independent
  analysis
• Data normalization Temporal and Dynamic
  Analysis
• Automated , or less interactive Analysis

5
Data management System

• Automated data characterization and Cataloging
• Automated Feature Extraction,
• Feature Description
• Automated image Indexing
• Data Production Planning and Scheduling
• Data and demand driven Plan and Schedule
• Resource dynamic Resource Monitoring and
  Allocation
• Data Structure
• Non-Relational DBMS?
• Data Queries
• Natural language
• Content -oriented

6
Intelligent Information Fusion System for EOSD

• Planner/Scheduler
• Expected data rates
• System and Network load
• Task constrains

• Metadata Extraction
• Low level signal/image processing
• PNN/SVM decision trees
• Hierarchical image decomposition and
  transformation
• Expert system
• High -performance system( Real time Image
  Processing system, Parallel computing,
• high speed communication

Planner/Scheduler
Planner/Scheduler

• Product Generation
• Available Production Algorithms
• Data-Driven
• Shortage Control

Sensor Meta Data Reprocessing requests
Meta Data
Storage Meta Data
Raw Data

• Object Database
• Object-oriented DBMS (Not RDBMS)
• Catalogue of Earth and Forest metadata.
• Distribute over many Devices

• Mass Storage
• Distributed over many Devices(DVD,CD,)
• Semantic Modeling
• Storage Hierarchy

7
What is the measure of Machine Intelligence
Quotient?

• Learning Ex.Develop Procedures and Rules for
  Information extraction
• Automatic SummarizationEx.What objects or
  features are dominated in this scene?
• Automated De-ambiguity
• Understand Natural Language and Images
• Problem Solving and Solving strategy

8
Problems of Current Methods
9
Potential Approaches
10
Hyper Channel data processing(Hyperspectral
signal and information fusion)Problems of
Conventional Methods

• Computational loads
• Density estimation from high dimensional data
  sets Hough Effects(Less classification accuracy
  at fixed training data)
• Need to reduce dimensionality

11
Hyper Channel data processing(Hyperspectral
signal and information fusion) SVM(Support
Vector Machine)

• Design classification Algorithms from supervised
  learning
• Map the data with non-linear transform into
  higher dimensional space
• Seeks a separating hyper-surface through
  optimizing

12
Potential Solution for Multi-resolution images

• Essential requirements for hyper-spectral
  signature and image overlay among images of
  different spatial and spectral resolution
• Problems of current methods
• Ex. Fourier Transform Does not work for
  non-stationary images like remote sensing.
• Possible solution Wavelet Analysis

13
Wavelet Analysis.

• Wavelet vs. Fourier Analysis
• Non-stationary vs. Stationary signal
• Scale-translation
• Application 1Characterization of Hyper-spectral
  signature
• Application 2 Hierarchical Image decomposition

14
Spatial Information ExtractionPSW(Pixel
SwappingJ.Iisaka)

• Extension of Numeric Image Processing(Conventional
  Image Processing) and Logical Image
  Processing(Mathematical Morphology)
• Evaluation of the impact to image properties
  against reordering the pixel arraypixel pair
  exchange
• Use induced measures among image components
  divided in a hierarchical way Ex. Local Fractal
• geometrical Features point,line and area
  handling, corner and vertices
• Fuzzification of spatial attribute
  Point-likeliness
• Easy to implement and robust

15
Examples of PSW

• Road extraction from forest area using TM data

16
Geo-biological Information ExtractionEx NDVI

• NDVItan(theta)
• Sensitive to atmospheric condition ( Pass
  Radiance and attenuation, etc)
• Sensitive to Sun angle ( Not suitable for
  temporal analysis)
• FMM Less Sensitive to these variation

Ch4
X
Ch3Ch4
Ch4-Ch3
Ch3
17
Final Result
18
Summary and Conclusion

• New era of image computing for remote sensing
• Hyper channels
• Multi-scale
• Information fusion
• Intelligent and learning
• Keys for Success of EOSD
• Development of Intelligent Information Fusion
  system
• RD for
• Data management
• Information Extraction
• Automated Analysis