Remote Sensing New Tools for Security

1
Remote Sensing New Tools for Security

• Harold Stone NEC Fellow, retired

2
Talk Outline

• Past uses of Advanced Sensors
• New Sensor Technology
• The Data Mining/Data Decision Problem
• Approaches to Image Data Mining

3
Past Uses of Advanced Sensors
4
Some Interesting Examples

• UAV video surveillance
• Iran nuclear tracking High resolution public
  satellite imagery
• The World Trade Center and Aerial hyperspectral
  imagery
• Hurricane Katrina, satellite imagery to assess
  contamination

5
UAV Technology
Source NOAA Magazine, 21 Mar 2006
6
UAV Technology
Payload 300 Kilos. Endurance 30 hours
Source NOAA Magazine, 21 Mar 2006
7
Public Satellite Imagery

• National Council of Resistance of Iran (NCRI) a
  clandestine organization that opposes the current
  Iranian government (www.ncri-iran.org)
• NCRI revealed detailed locations of Irans secret
  nuclear program. The locations could be viewed
  through public imagery
• The satellite imagery showed an active nuclear
  program that Iran denied existed.

8
IRAN Nuclear Program
9
IRAN Nuclear Program
Tehran, 16 November 2004
10
IRAN Nuclear Program
Tehran, 16 November 2004
11
IRAN Nuclear Program
Tehran, 16 November 2004
12
IRAN Nuclear Program
Arak, 26 September 2002
13
IRAN Nuclear Program
Arak, 27 February 2005
14
9/11 Hyperspectral Analysis
15 Sept. 2001
Before 11 Sept. 2001
15
WTC Contamination
16
WTC Contamination
17
WTC Contamination
AVIRIS Data, 17 Sept. 2001
Source Roberts and Herold, 2004
18
The Fluorocarbon Hazard

• 70,000 Kg of a fluorocarbon refrigerant were
  delivered to WTC a week before 9/11.
• It was stored in large tanks beneath the towers.
• Aerial imagery of the WTC site took multispectral
  images to analyze contamination from plume.
• Leaking fluorocarbons were discovered in the
  plume and traced to the refrigerants.

19
The Fluorocarbon Hazard

• The fluorocarbon in the presence of the fire
  could produce phosgene gas.
• The discovery of fluorocarbons in the plume
  prompted remediation by a specially equipped team
  who pumped the remaining refrigerant to safe
  container storage.

Source RT Magazine, Aug. 2002
20
Sensor Platforms
Source Roberts and Herold, 2004
21
Scope of the problem
Source Cahill et al.
22
Hurricane Katrina Aftermath
23
Hurricane Katrina Aftermath

• Hurricane Katrina flooded Petrochemical plants
  located in low lying areas of the Mississippi
  River delta.
• Resulting contamination
• Mississippi River
• Swamp (Bayou)
• Gulf of Mexico
• Sea, swamp, and river beds
• Satellite imagery was used to detect and
  remediate the contamination

24
RULLI

• Remote Ultra Low Light Imaging (Los Alamos
  National Laboratory)
• Remote sensing device using ultral low level
  intensity
• Single photon detector, gives X-Y-Time coordinate
  of each photon

25
RULLI
Source Albright et al, LANL
26
RULLI Example
Source Ho et al, LANL
27
RULLI Example
Source Ho et al, LANL
Source Priedhorsky et al, LANL
28
Data Mining Decision Problem
29
Too Much Data

• Satellite Imagery began in the late 1950s
• Image platforms have become numerous
• Sensor data per platform has increased enormously
• 4 channel imagery supplanted by 200 channel
  hyperspectral
• Resolution increase to a few meters
• Too much data collected per day to be analyzed by
  humans

30
Understanding Images

• Pixels are numbers
• Images show objects, regions, physical entities
• Problem Interpret the pixels. What do they
  show?

31
What to Seek

• Todays image may show nothing of interest.
• It may be very important for comparison with
  future changes
• Clouds, shadow, night occlude areas
• Difficult to detect occlusions automatically
• Areas of interest may be hidden or disguised
• Areas of interest may be similar to uninteresting
  areas difficult to detect automatically

32
The Genie Project

• One approach to data mining

33
The Genie Project

• Work done at Los Alamos Theiler, Perkins,
  Harvey and others
• Method
• A human gives examples of interesting regions
• The system has a toolbox of image processing
  functions
• A genetic algorithm tries many combinations of
  functions, rewarding successes, mutating
  combinations, until it discovers an algorithm
  that selects the interesting regions

34
Example Find Golf Courses
Training Image
Human Input
Not golf course
Golf course
No class
Source Harvey et al
35
The Synthesized Algorithm
4
10
7
Input Data
2
Mean
Define Region
Alt Seq. Open Close
Range
Mean
Variance
Close Open
Open Reconst
Intermediate Data
1
2
3
4
5
Fisher Linear Discriminator
Source Harvey et al
36
Result Find Golf Courses
Training Image
Human Input
Source Harvey et al
37
Result Find Golf Courses
Identified objects
Human Input
Source Harvey et al
38
Applications

• Genie has been used for security applications by
  Los Alamos National Lab
• Results are classified
• LANL has continued and grown funding for Genie
  based on its proven successes
• Genie has broad applicability, but is high risk
  because success is not assured
• Genetic algorithm is difficult to assess, but
  IT WORKS!

39
KIM

• Datcu et al

40
KIM Approach
Source Datcu and Seidel
41
KIM Examples
Find landing places for small aircraft
Find structures, trees, and roads
Source Datcu and Seidel
42
The Hard Problem

• Give meaning to pixels
• Mimic human image processing
• Find features, regions, remove noise and
  occlusions
• Genie compose graphics processes with genetic
  algorithm
• Kim detect features statistically
• Human guidance to build objects from pixels
• Genie goal and non-goal identification
• KIM semantic grouping of features

43
Prospects for the Future

• Massive archival data
• New instruments, new ways to collect data
• New ways to analyze data and to make sense of it
• Human interaction will be embedded in the system
  for the foreseeable future
• The major limitation is the inability to extract
  the information we need from the data we
  have.Connect the dots