Automatic Target Detection from Hyperspectral Data

1
Automatic Target Detection from Hyperspectral Data
2
Objectives and Work Activities

• Objectives
• Develop an automated framework for hyperspectral
  target detection subject to environmental
  variability
• Lower the probability of false alarms

• Work Activities
• Trade Studies and Assessment
• Data Processing Analysis
• Algorithm Validation and Evaluation

3
Past Accomplishments

• Trade studies of spectral/spatial binning and
  their effect on detection results for a variety
  of target materials
• Processed data from several different
  collections, i.e., HYDICE Forest Radiance, AURORA
  at Aberdeen proving grounds, Hawaii coral
  mapping, drug interdiction, and submarine
  detection.
• Several detection algorithms were developed and
  compared -- Linear unmixing, CEM, OSP, RX, and
  PCA
• Several spectral comparison methods were
  evaluated, such as spectral fingerprinting, SAM,
  clustering. Simple atmospheric correction and
  de-shadowing techniques were tested. MODTRAN was
  run on several minerals to evaluate atmospheric
  variance.

4
Spatially Binned HYDICE Forest Radiance Subset
No Spatial Binning
4x4 Spatial Binning
8x8 Spatial Binning
16x16 Spatial Binning
5
ATR Procedure
Alarmed pixels
Classes
Target list with scores
6
ATR and Linear Unmixing

• Unmixing method projects spectral data into
  feature space, classifying each pixel by
  membership in endmember class
• Each class is sorted into either a background
  class (soil, foliage, water) or anomaly class
• First step of ATR identifies likely target
  pixels by degree of membership in anomaly class
• Next (optional) step identifies likely target
  pixels by similarity to spectra in library of
  targets (Bloodhounds). Similarity metrics
  include SAM and spectral fingerprints.
• Similarity metric must be robust to
  atmospheric/environmental changes between data
  collects
• Finally, contiguous target pixels are collected
  into objects and spatially filtered. Filters are
  upper and lower bounds on length, width, and
  aspect ratio

7
Bloodhound Detection of Unique Spectra Under
Different Atmospheric Conditions Using Spectral
Fingerprinting and SAM
Run 1 -- Bloodhound collected
Run 2 -- Bloodhounds detected
Run 2 -- Zoom on subdivision
8
Bloodhound Demonstration - APG
Run 1 -- Bloodhound collected
Run 2 -- Anomalies detected
Run 2 -- Bloodhounds detected
9
Impact on ATR technology

• Goal is the development and improvement of
  real-time, operational hyperspectral systems
• Spatial/spectral binning trade study shows much
  stronger degradation with spatial binning
• Several detection algorithms under study --
  Linear Unmixing, CEM, OSP, RX, PCA.
• Bloodhound comparison methods being developed
  -- spectral fingerprinting and SAM. Robustness
  of algorithms is being improved using simple
  atmospheric modeling, de-shadowing.
• Bloodhounds are a key technology for previous
  drug interdiction, sub detection, TUT
  collections, and for PREDATOR UAV hyperspectral
  system

10
Future Plans

• Continue further evaluations and refinements of
  detection methods
• Enhance algorithm discrimination with
  clustering
• Improve bloodhound methods using more
  intelligent target feature extraction and
  pattern comparison
• Optimize robustness of endmember selection
  process