Objective

Improvement in the detection and discrimination of unexploded ordnance (UXO) using the electromagnetic induction (EMI) regime has been consistent during the last two decades, in both instrument hardware and detection software. There is still a disparity between the often idealized environments considered (“golf course” situation) and the reality in the field. Working on diminishing this disparity, various SERDP projects focus on developing versatile (next generation) instruments as well as robust software for data processing in realistic environments, yielding ever improved classification performance.

The current protocol for surveying fields is typically organized in two stages. First the sensors are used in a survey mode, or dynamic mode, collecting data over large areas and flagging locations that may be contaminated by a subsurface UXO. The flagging process is typically based on magnetic field amplitude, based on the rationale that high amplitudes may correspond to buried UXO and should therefore be further investigated. Hence, at the second stage, the sensors return to the locations of interest for more in-depth data collection in order to feed the identification and classification algorithms with high quality and diverse data. Flagging based on field amplitude, however, has potential drawbacks: the magnetic field in the EMI regime decays fast and a deep UXO may not produce a strong response while still being potentially hazardous.

The purpose of this SERDP Exploratory Development (SEED) project was to investigate an alternative method to flag areas of interest at the first stage of data collection with the following objectives:

  • Work with dynamic data – Dynamic data are typically more noisy and less diverse than cued interrogation data (collected at the second stage). The method should therefore be tolerant to noise and able to provide inversion results with such a reduced dataset.
  • Provide position and polarizability estimates – The researchers wish to go beyond the field plot and better identify the targets in order to make more informed decisions on the probability of presence of targets of interest. The output of the method should therefore include target position in the (xy) plane for subsequent data collection if necessary, target depth to potentially flag deeper UXO, and target polarizabilities to have a first identification procedure and not miss deep UXO or small items.
  • Real-time processing – The method should not slow down the survey speed and should therefore work in real-time while still satisfying the two constraints above. For purposes of practicality, the researchers set the real-time limit to 100 ms, which is the rate at which dynamic data are currently acquired by sensors such as the MPV-II and the MetalMapper.

Technical Approach

The approach of this project was to implement an algorithm based on Kalman filters (KF) and Extended Kalman filters (EKF) for the sequential processing of data as they become available. Kalman filters have been developed as a stochastic signal processing technique using Bayes rule in order to filter a signal out of a well characterized noise. Kalman filters, however, only provide an optimal minimum mean square error estimator if some constraints are satisfied.

The researchers supposed that these constraints are satisfied: all noise sources can be drawn from a Gaussian distribution as well as prior probability density functions related to the unknown parameters to be estimated. Kalman filters also require the parameters to be linear with measurements: within the dipole approximation often used to process EMI data, this condition is satisfied by the dipole moment which can be directly related to the polarizability tensor. The position, however, appears as a non-linear parameter which the researchers estimate using the extended Kalman filter and a linearization of the equations. Finally, both algorithms (KF and EKF) are iterated to sequentially process the data and yield converged values of all parameters to be estimated.

Results

The main achievement of this SEED project is the implementation and validation of an iterative KF-EKF approach satisfying all conditions above: dynamic data processing, inversion of all the parameters of the dipole model, and real-time processing. In addition, the validation has been performed on various targets, including a small 20-mm item which remains a challenging target to identify. The validation of the KF-EKF iterative algorithm has been performed on all next generation sensors available to the researchers: TEMTADS, MPV, MPV-II, and MetalMapper. For practicality purposes, the researchers concentrated on the last two sensors since the TEMTADS is usually not used for dynamic surveys and since the MPV-II is currently replacing the MPV.

It is worth mentioning that the purposes of dynamic data collections with the MPV-II and the MetalMapper were different in the investigation. The MPV-II, meant for hand-held operation, was used to interrogate a limited area within which a single target was present, akin to being waved by an operator in a detection mode. The MetalMapper was driven along adjacent 60-m long lanes to flag target locations across a large geographical area. In this latter configuration, multiple targets were present in the underground and sequentially appeared and disappeared from the field of view of the sensor as it was driven across the area. In both cases, the sensors operated in dynamic mode, meant for rapid land survey at the expense of data quality. Despite the noisy data, the KF-EKF algorithm was able to converge to proper solutions as verified independently either using the Gauss-Newton reference algorithm, using ground truth information whenever available, or comparing with independently obtained results. In addition, processing times of each new data collection were measured to be less than 100 ms on a regular 2 × 2.8 GHz Quad-Core Intel Xeon computer, which is the real-time limit for the application. The inverted parameters include position, depth, and polarizabilities, albeit limited to the short data collection time range inherent to a dynamic survey. Yet, even if this dynamic data is not sufficient to perform classification, it is often sufficient to identify the subsurface anomaly as possible UXO or not, either based on simple signal correlation to a library or based on a volume estimate. The method can therefore be used in land survey for real-time target mapping, flagging those locations to which the instrument needs to return for more exhaustive, cued interrogation, data collection.

Benefits

As part of this work, a graphical user interface (GUI) was proposed to display the results to an operator, showing in real-time the evolution of position, depth, and polarizability estimates as new data become available during the measurement process. Combining this information, an indication on target aspect ratio, a volume estimate, and the likelihood of being a target of interest are also displayed on the GUI. Such an integrated tool could be a useful feature to implement on existing sensors and taken on-board during field surveys.

Future work in this direction would include:

1. Integration to existing on-board software with sensors of interest

2. Extensive validation in various terrain configurations

3. Optimization of feedback information (probability of target, library matching, volume estimation)

4. Extension to new sensors

5. Use of an 8 ms dynamic window (instead of the current 2.7 ms) to provide considerably improved identification possibilities since the KF-EKF algorithm achieves more than just detection

  • Kalman Filters,

  • Electromagnetic Induction (EMI),

  • Sensors,

  • UXO Detection and Classification,