A significant challenge associated with new electromagnetic induction (EMI) systems is the time required to collect all data, and this limitation often necessitates a priori hardware compromises that undermine subsequent system performance. Adaptive approaches would allow one to retain full EMI hardware functionality, with sensing speed attained by adaptively tailoring the sensor use to the item under test. Under the SERDP Exploratory Development (SEED) project MR-1591, researchers developed two adaptive sensing techniques. The first approach assumes no a priori knowledge of the particular targets and clutter under test, and it adaptively performs measurements with the objective of best inferring the EMI model parameters of the buried item (without necessarily knowing its location a priori). The second approach performs adaptive sensing with the objective of making a final classification decision, assuming that one knows the types of clutter and UXO under test. The advantage of these approaches is that they allow optimal, adaptive use of sensor resources.
The objective of this project is to develop an adaptive sensor management architecture of interest for two applications: (1) optimal use of sensor assets for sophisticated next-generation (e.g., multi-coil) systems such as the TEM array, and (2) guiding the use of emerging portable (handheld) systems on where to collect new data and when to terminate sensing. The research constitutes active learning with the purpose of optimally using sensor assets. To support this objective, researchers have investigated the partially observed Markov decision process (POMDP) and related information-based sensing algorithms. In this project, they will develop a fully Bayesian approach while designing a sensing policy, incorporating both prior knowledge about the site and uncertainties about the model.
At the Former Spencer Artillery Range (Camp Spencer) in Tennessee, discriminations on datasets from two companies across two sites were performed: Open and Dynamic. Both companies, NAEVA and URS, used the MetalMapper sensor. For each company and site, a set of three discrimination methods was employed. These discrimination methods were designed to accommodate sites that had differing levels of intrinsic complexity and available information about the types of UXO present. These methods, labeled aggressive, intermediate and conservative, varied the amount of training data requested from 0 to around 50 labels. Classification for the conservative method utilized a generative Bayesian classifier while the intermediate and aggressive methods used a semi-supervised, parametric Bayesian classifier. To complement these discrimination methods, a complexity metric was developed that allows for a priori selection among the three modeling approaches for a new site.
At Camp San Luis Obispo (SLO) in California, specific core technologies were examined and validated during the follow-on analysis. These technologies fall broadly into the four analysis categories: the sensor/target model, feature selection, classification, and active label selection. A sensor/target model using a multi-anomaly dipole inversion was developed. Feature selection was performed using the Bayesian Elastic Net, which has the benefit of retaining correlated and informative features for classification. Classification was performed using two approaches: a supervised, sparse Bayesian classifier, and a semi-supervised, parametric Bayesian classier.
A main objective of the Camp Spencer discrimination was to differentiate between aggressive, intermediate, and conservative approaches. These approaches differ in the amount of training data they require. Some of the initial dig lists based on these approaches missed quality control (QC) seeds, which were subsequently included in the dig lists as training data. To keep the amount of training data consistent between the different modeling approaches, however, the missed QC were not included in any of the predictive models. Instead, the stop-dig thresholds were adjusted to capture the QC seeds. The final dig list for every modeling approach and every sensor captured all UXO in the Dynamic area. Some approach/sensor combinations missed UXO in the Open area.
At SLO the multi-anomaly dipole model was effective at producing features easily recognized as being UXO where a single anomaly model would not be effective. The semi-supervised classifier will outperform the supervised classifier when a difficult UXO is connected to other UXO by neighboring observations (in feature space). Both the semi-supervised and supervised classifiers are robust to changes in their user-defined parameters. Non-myopic active learning is effective, in general, at capturing difficult UXO and enhancing the quality of training data for feature selection. When active learning is used in conjunction with feature selection, appropriate features can be selected from few training data.
The adaptive sensor-management algorithms allow the sophistication of multi-coil systems to be retained, while still achieving fast sensor collections, since typically only a subset of the full sensor functionality is applied to a given target. Moreover, this research has the potential of optimizing the use of handheld systems, assuring appropriate data have been collected for parameter inversion.