The objective of this project was to develop improved capabilities for classifying buried objects as unexploded ordnance or clutter that exploit all of the information available in the electromagnetic induction (EMI) response of the object. The project was later re-directed to look into better ways of processing the survey data being collected by the newly developed advanced EMI sensors to intelligently select possible target locations.
A detection filter has been implemented for dynamic survey data from advanced EMI sensors. By applying the standard dipole inversion model to a fixed grid of locations, the fit quality parameter, coherence between model and measured data, can be used as a single detection quantity. Contour plots of detection filter coherence can be used to pick peak locations as possible targets.
This approach has several advantages over the current technique of looking at only measured signal amplitude and picking regions of peak signals. Based on the dipole model, the detection filter makes use of all channels of data and peaks only at the target location. Similar to the signal amplitude approach, forward model signals can be used to set a detection filter threshold for a given target of interest (TOI) at its maximum depth of concern. Unlike the model-based transmit and receive (Tz-Rz) threshold however, the detection threshold is based on embedding the model signal in background areas of the actual survey data and applying the filter to both model and noise and only the measured noise. This process gives an estimate of both the filter signal output and the filter noise.
At Southwestern Proving Ground, Arkansas, the filter response was found to de-emphasize a large number of small, surface clutter items and place them below the filter detection threshold. At Spencer Artillery Range, Tennessee, the single dipole inversion result was found to be sufficient to eliminate two thirds of the detect filter peaks based on the inverted polarizations being much smaller than the TOI. Applying the N-dipole fit was found to flag multiple item cases and often matched the ground truth. Finding TOI at Fort Bliss, Texas also required the N-dipole fit. However, in concentrated regions of clutter, some TOI were still missed. Increasing the number of dipoles fit may improve this, but there were several other factors limiting the Fort Bliss results.
Stationary EMI noise levels fall off with time gate and do not correlate across receive channels. Unfortunately, in motion, the 2x2 sensor on a wheeled platform is dominated by induced voltages from the receive coils bouncing in the earth’s magnetic field. This noise is apparent in the MetalMapper on a ground dragged sled, but reduced when it is on a vehicle mount. This noise source could also be reduced by differencing the receive cubes with an extra one mounted above the sensor. Because the detection filter is based on the inversion model, the quality of the results is strongly affected by position errors. The current advanced EMI acquisition systems do not adequately match the position measurements in time to the EMI data.
The IDL code for processing 2x2 and MetalMapper dynamic data, running the detection filter, and inverting any EMI sensor data are available for distribution. Eventually, this code will be implemented in UX-Analyze. The detection filter process was applied to 2x2 dynamic data collected for the Former Camp San Luis Obispo, California Treatability Study by associates at Acorn SI. They have also used the algorithm on MetalMapper data collected for an ESTCP Live Site Demonstration at Twentynine Palms, California.
Overall, the detection filter output can be predicted for a given TOI to a specified depth along with the filter background noise. The filter exploits all channels of data from the EMI sensors and results in a higher signal to noise ratio output than simple consideration of signal amplitude. It was observed that the filter depth setting could be used to de-emphasis small surface clutter in the target selection process, but still select all targets of interest. Given the filter peak locations, inversions can be performed on data windows centered at each. In data sets with high target densities, it was found to be useful to apply N-dipole inversions at the filter peaks as well. The inverted polarizations were found to match those found from cued data collected with the same sensor.