A Comparison of Techniques for Discriminating Buried Unexploded Ordnance UXO

1
A Comparison of Techniques for Discriminating
Buried Unexploded Ordnance (UXO)

• Edwin Roger Banks, Edwin Núñez, Paul
  AgarwalCOLSA Corporation, Huntsville, AL
• Marshall McBride, Ron LiedelU.S. Army Space
  Missile Defense Command (USASMDC), Huntsville, AL

DISTRIBUTION A. Approved for public release
distribution unlimited.
2
The Problem

• Over 10 million acres of land in the US need to
  be cleared of buried unexploded ordnance (UXO)
  DoD report
• UXO injures thousands each year worldwide

3
Discrimination Problem Statement
What is the best way to interpret sensor data for
buried objects to produce a description of the
object?
In particular, does the sensor data indicate UXO?
4
Methods of Discrimination

• Mathematical model inversion
• Empirical (signature matching)
• Genetic Programming (GP)
• Others

This presentation focuses on GP
5
What is Genetic Programming?

• GP is an optimization or discovery technique
• GP works with a population of individual
  solutions
• GP evolves progressively better solutions over
  many generations using
• Selection based on fitness
• Recombination of different parents
• Mutation

6
Basic GP Algorithm
Create initial population of solutions
Evaluate fitness function for each solution
Select parents for recombination based on
fitness
Iterate for thousands of generations
Recombine parents
Mutate offspring
7
Recombination and Mutation
?
Mutation
? Recombination
8
GP Variations

• Three GP variations tried
• Evolve an expression over sensor data values that
  provides a confidence level that the object is
  UXO
• Also evolve expressions for objects depth,
  weight, and length
• Evolve a decision tree to decide whether object
  is UXO or not
• Evolve a genetic algorithm (GA) weighting of
  sensor data factors

9
GP Requirements

• Concise encoding
• Fitness function
• Computational resources

10
Sensor Data Source

• Jefferson Proving Grounds (JPG) had four
  technology demonstration phases (1994-1999)
• UXO and non-UXO deliberately buried on 160 acre
  site
• JPG published sensor data

11
Problem Formulation Choices

• UXO detection vs. UXO discrimination
• Scanned data
• Flagged targets
• Data source Data sets included ground
  penetrating radar (GPR), magnetic (M), and
  electro-magnetic (EM) sensors from various vendors

12
Problem Statement

• We chose
• Discrimination problem (harder)
• GeoPhex GEM-3 data (EM)
• JPG Phase IV (discrimination) data
• 160 target points (UXO and non-UXO)
• 50 ordnance (15 mortars, 35 projectiles) of
  various size
• 110 non-ordnance (fragments)

13


From Robitaille et al, JPG Tech Demo Program
Summary, 1999
14
GEM-3 JPG-IV Data Available

• Measurements of InPhase (I) and Quadrature (Q)
  response to EM pulse, based on GEM-3 sensor
• Eight frequencies, from 30 Hz to 23,970 Hz
• Measurements made in 5x5 grid with center point
  directly at the flag. Grid point spacing 0.23
  meters
• Total data per target 2 x 8 x 25 400 readings
  (64,000 numbers in all)

15
JPG-IV Ground Truth Data Available

• Target Number
• Northing
• Depth
• Size
• Declination
• Weight
• Width
• Diameter

• Survey Number
• Easting
• Type (Ord, non-Ord)
• Azimuth
• Class mortar, projectile
• Length
• Thickness
• Description, e.g. 60 mm Mortar (w/o fuze)

16
Specific GP Problem Statement

• Primary discriminate ordnance from non-ordnance
• Secondary estimate depth, length, and weight
• Produce a prioritized dig list

17
UXO Discrimination Difficulty Example
Complete sensor data available to GP is shown.
18
UXO Discrimination Difficulties

• Only 3-4 items of each type in target set
• Weight varied from 0.1 kilogram to 43 kilograms
• Burial depth, azimuth, and declination varied
• Ground truth changed (mainly declination)
• Gaps in sensor data
• Lack of background calibration
• Data not clean

19
UXO Solution Encoding

• Chromosome fitness weighted sum of the errors
  (compared to ground truth) in
• depth
• length
• weight
• ordnance/non-ordnance FO 1.0 FN 0.3
  FP
• where FP is number of false positives, FN is
  number of false negatives

Depth (2)
Chromosome
Length (2)
Weight (2)
Ordnance or Non-ordnance (4)
Fitness function F FO Fdepth Flength
Fweight
20
Encoding the UXO Problem

• Operator Set , -, , /, , If-then-else, log,
  etc.
• Operators applied to features such as maximum
  quadrature signal over all frequencies
• Several dozen features
• Interestingly, best features for GP were not
  those with highest correlation to ground truth
• Most useful features depended on length of
  computer runs

21
Alternative Formulations Explored

• Problem was encoded as
• Genetic Program (evolve an expression as
  discriminator)
• Genetic Algorithm (combine features linearly)
• Decision Tree (fastest evaluation)
• GP gave best overall results

22
Prior JPG-IV UXO Study Results
From Jefferson Proving Ground Technology
Demonstration Program Summary by Robitaille,
Adams, ODonnell, and Burr. May 1999.
23
Combined GP and JPG-IV Results
? False Positives
110
0
0
False Negatives ?
50
24
Conclusions

• Genetic Programming solution substantially
  improved object discrimination ability
• GP approach exceeded mathematical/physics-based
  and signature-file approaches
• Buried UXO discrimination using GP approaches
  ideal goals of the Army Corps of Engineers

25
Points of Contact

• Edwin Roger Banks
• Edwin Núñez
• Paul Agarwal

COLSA Corporation
256-964-5555
26
EXTRA SLIDES
27
Methodology

• 160 targets split into 3 sets training (60),
  test (30), validation (10)
• Scores computed for the previously unseen
  validation set
• Ten runs made so that the validation set covers
  all 160 targets (Jack Knife)
• Result is 10 chromosome solutions applicable to
  the field
• All above steps repeated 10 times for committee
  vote (bagging)
• Total of 100 chromosomes evolved for each
  committee run

28
ARCGPL Tool Used for Examining UXO Data
29
UXO Further Discrimination Features