Wednesday (Group 2) - SERDP-ESTCP - Strategic Environmental ...
Wednesday (Group 2) - SERDP-ESTCP - Strategic Environmental ...
Wednesday (Group 2) - SERDP-ESTCP - Strategic Environmental ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Munitions Management (MM)<br />
Poster Number 43 – <strong>Wednesday</strong><br />
Ground Based Detection and Discrimination — EMI & Magnetometers / Modeling & Processing<br />
T<br />
DISCRIMINATION PERFORMANCES OF COMMERCIAL SENSORS AT APG<br />
AND YPG<br />
DR. DEAN KEISWETTER<br />
SAIC, Inc.<br />
120 Quade Drive<br />
Cary, NC 27513<br />
(919) 454-3212<br />
keiswetterd@saic.com<br />
CO-PERFORMERS: Mr. Levi Kennedy (Signal Innovations <strong>Group</strong>, Inc.);<br />
Dr. Leslie Colins (Duke University)<br />
he objective of <strong>SERDP</strong> Project MM-1505 is to determine the level to which data acquired at<br />
Aberdeen Proving Ground and Yuma Proving Ground using commercial electromagnetic<br />
induction and magnetic sensors support feature-based discrimination decisions. As part of this<br />
study, we are evaluating discrimination performances for various models, classifiers, and training<br />
data. Three fundamental issues are being investigated: namely, the model used during<br />
characterization and the impact that classifier selection and training has on performance. The<br />
UXO Standardized Test Sites in Aberdeen Maryland and Yuma Arizona represent real world<br />
scenarios. They are rich in data, ground truth, near-surface conditions, and target diversity.<br />
Because of this, they present an unprecedented opportunity to systematically examine the<br />
importance of models, classifiers, and training data. The near-surface conditions vary and the<br />
buried targets of interest range from 20mm projectiles to 500 lb bombs.<br />
We selected five data sets based on data quality, type, signal-to-noise, and availability at<br />
appropriate intermediate processing stages. The datasets include time-domain EM61 (man-towed<br />
single sensor and vehicle-towed array), time-domain EM63, frequency-domain GEM-3 (vehicle<br />
towed array), and magnetic data (array). The models being investigated include a standard dipole<br />
(expressed using polarizabilities, an ellipsoidal representation, and empirical fits), joint<br />
frequency-time domain, and singular expansion models. The classifiers being investigated<br />
include: the Generalized Likelihood Ratio Test, Support Vector Machines, Relevance Vector<br />
Machine, and K-nearest neighbor. Various levels of training data are being extracted using<br />
ground truth labels from the calibration, blind grid, and portions of the open field areas.<br />
Results of our analysis indicate that the nominal dipole fit error is in excess of 20%. This<br />
combined with the facts that (i) the TOI range from 20mm projectiles to 500 lb bombs and<br />
(ii) the non-TOI target is typically medium to large in size and thick walled, significantly limited<br />
discrimination performances. If we assume that the objective is to discriminate all UXO from all<br />
non-UXO at these sites, very little if any discrimination capability is observed. If we restrict the<br />
TOI class to specific sizes of UXO, however, such as small, medium, or large UXO which is<br />
often the case for live sites, performance improves significantly.<br />
G-11