Electromagnetic induction (EMI) and magnetometer systems have been the principal sensors deployed for digital geophysics-based unexploded ordnance (UXO) sensing. Improved EMI-sensor capabilities, in particular, have the potential to enhance UXO discrimination. SERDP and ESTCP funded development of multicoil EMI systems that provide significant capability and diversity with respect to the shape of the incident magnetic field as well as in how the induced magnetic fields are measured. However, the large number of sensor parameters (number of transmit/receive coils and the time/frequency sample rate) often necessitate hardware design tradeoffs, with the goal of achieving practical sensing costs. By making these sensor-design tradeoffs in hardware, one necessarily loses functionality, limiting the utility of the system (e.g., the system may have to be tailored in hardware to particular classes of UXO and UXO depths). This SERDP Exploratory Development (SEED) effort was directed toward developing the adaptive sensor-management architecture needed for next-generation EMI systems.

The objective of this project was to develop the basic Partially Observable Markov Decision Process (POMDP) sensor-management framework, which entailed defining the class of actions and observations of relevance to adaptive UXO sensing with next-generation EMI systems.

Technical Approach

The adaptive EMI sensor-management framework tailors the use of sensor assets to the target under test. Sensor functionality is preserved by making fewer compromises in hardware, with practical sensing costs achieved through optimal and selective use of sensor assets. The algorithm also adaptively determines when to terminate sensing (i.e., when the amount of data measured is sufficient for classification within user-defined risk constraints). The POMDP algorithm sequentially and optimally selects a sequence of EMI sensor parameters for distinguishing UXO from clutter. While the overall EMI system hardware may support a number of coils and temporal/spectral sampling rates that would be impractical if all possible measurements were performed, the POMDP algorithm sequentially tailors the measurements to the item under test. The particular item under test is, of course, unknown before excavation, but the POMDP algorithm infers through the measured data which type of item is most likely, thereby integrating the sensing and signal processing.


The researchers successfully developed the POMDP algorithm and tested it on data simulated to replicate the Berkeley UXO Discriminator (BUD), a next generation EMI system.  The researchers first designed an active-learning based information-theoretic technique that efficiently chooses a sequence of sensing actions to minimize the uncertainty on the unknown model parameters. Once approximate locations and model parameters of buried objects were obtained, the researchers employed the second phase of the strategy, where a POMDP-based policy dictates how a buried object needs to be illuminated by different transmitter coils in order to identify the “class” of the buried object.


By implementing the algorithmic techniques for optimal adaptive EMI systems, hardware designers will be relieved of many design constraints. Specifically, rather than having to design a system a priori (in hardware) to address time/energy constraints, a more complete EMI system can be designed, and the POMDP framework can guide efficient utilization of the sensor resources. In addition, for any existing or planned next-generation EMI system, the algorithms could yield an optimal use of sensor resources, only performing those measurements that are needed for an item under test and reducing sensing time/costs.

  • Analysis ,

  • Electromagnetic Induction (EMI) ,

  • Machine Learning ,

  • Physics-based ,

  • Classifiers