Continued Discrimination Demonstration Using Advanced EMI Models at Live UXO Sites: Data Quality Assessment and Residual Risk Mitigation in Real Time
Dr. Fridon Shubitidze | Dartmouth College/White River Technologies
Objectives of the Demonstration
The primary objectives of this project were to:
- Implement and demonstrate robust procedures and approaches for advanced electromagnetic induction (EMI) sensor data pre-processing, inversion and sub-surface target classification;
- Assess quantitatively the quality and utility of advanced EMI sensor data in terms geologic and background noise effects, and the use of multi-object inversion algorithms to de-couple the EMI response of targets in the presence of high-density metal contamination;
- Validate processing technology based on extracted intrinsic (effective dipole polarizability) and extrinsic (location) target parameters from measured data, and identify robust classification features in order to distinguish unexploded ordnance (UXO) targets from non-hazardous objects; and,
- Fully characterize the discrimination ability and limitations of the advanced models with regard to the number of objects, target size and material heterogeneity, geology, and background noise.
UXO classification procedures consist of the following sequential steps: background corrections, target detection/picking, data inversion and target feature parameter estimation, ranking, training, and finally classification, i.e., separating UXO from non-hazardous anomalies. Under this project the project team has developed and tested a user-friendly software package for advanced EMI sensor data preprocessing, inversion and classification. The software package, which supports both cued and dynamic survey datasets, allows the efficient execution of the following procedures:
- Background correction: During this process each EMI dataset is first normalized by a corresponding transmit (Tx) current. Next, in the case of cued datasets, background data files are selected for a process of subtracting background levels from the original EMI anomaly dataset, and in the case of dynamic survey datasets, a median filter approach is employed for removing background noise from target signals.
- Data inversion: After background EMI levels have been applied and corrupted channels removed, the combined Orthogonal Normalized Volume Magnetic Source – Differential Evolution (ONVMS-DE) algorithms are applied to the anomaly datasets using a multiple source inversion approach. The intrinsic and extrinsic parameters of the targets are then extracted and used for ranking.
- Targets picking using survey data set: Once background levels are removed from the survey data, two approaches are used to pick targets for cued interrogations: (1) The traditional method that utilizes signal amplitudes on a two dimensional map and identifies peaks of signals above a prescribed threshold level; and (2) A semi-supervised Gaussian clustering process which clusters the inverted extrinsic (source locations) parameters into a three dimensional space and identifies targets using cluster centers.
- Ranking: This process uses extracted intrinsic classification features of the targets, such as total ONVMS-effective polarizabilities, (via one, two and three sources), to rank anomalies.
- Training: Typically, target classification feature parameters are clustered and site-specific training target lists are used to support final classification. Specifically, these training data are used to assess background noise levels, validate inversion results, confirm preliminary target ranking results, and (more importantly) determine an optimal “stop-dig” point which optimizes classification performance. The stop-dig point is established through evaluation of training data derived from “uncertain anomalies,” which are located between targets which are definitely targets of interest (TOI) and those targets which are definitely clutter in the preliminary ranked list.
- Classification: Once the ground truth is obtained from the training targets, all anomalies are classified as TOI or clutter, and the optimal stop-dig point is defined.
During the course of this project, the project team processed multiple datasets collected at eleven ESTCP UXO Live Site demonstrations, including:
- Spencer Range, TN;
- Camp Edwards Massachusetts Military Reservation (MMR), MA;
- Camp Elis, IL;
- Fort Rucker, AL;
- New Boston Airforce Station, NH;
- Southwestern Proving Ground, AR;
- Waikoloa Maneuver Area (WMA), HI;
- Andersen AF base, Guam;
- Fort Bliss, TX;
- West Mesa, NM; and
- Fort Ord, CA.
The advanced EMI data inversion and classification technology was demonstrated at these eleven UXO live sites for identifying all TOI and eliminating more than 75% of the clutter. The technology has applied to cued and dynamic data sets collected by the next generation EMI sensors: such as MetalMapper, 5x5 and 2x2 Time Domain Electromagnetic Towed Array Detection System, man portable vector, Berkeley UXO Discriminator and one pass transient electromagnetic array. The demonstrations have showed that for most sites the advanced classification technology identified all TOI while correctly classifying 75% to 92% of the clutter at specified Stop-Dig points.
However, there were few sites where the algorithm did not correctly classify one or more TOI due to insufficient data quality, magnetic soil and inaccurately documenting the intrusive results. These illustrated the importance of a well-defined data collection procedure and accounting magnetic soil responses during intrusive investigations. A comparison of the classification results for different sensors showed that they perform equally well when data are analyzed using the advanced EMI models. The project team comes to a similar conclusion when compared between classification results for cued and survey data sets. The choice of which sensor in with mode to deploy on site can therefore be driven by cost and which system can most efficiently survey the terrain.
Points of Contact
Dr. Fridon Shubitidze
Dartmouth College/White River Technologies
SERDP and ESTCP