Summary of Progress on SIG Ft. Ord ESTCP DemVal
Abstract
We report on progress under an ESTCP demonstration plan dedicated to demonstrating active learning- based UXO detection on an actual former UXO site (Ft. Ord), using EMI data. In addition to describing the details of the active-learning algorithm, we discuss techniques that were required when applying this method to field data, including a clustering algorithm that plays a key role in properly labeling the data (learning when to label a non-UXO item as UXO-like). The spatially-varying EMI response of each anomaly is first fit using a dipole model, from which each anomaly is characterized in terms of two dipole-moment magnitudes and two resonant frequencies. Information-theoretic active learning is then conducted on all anomalies to determine the labels for the small subset of most informative anomalies, defined by those that would be most beneficial for classification purposes if the associated label was available. The labels of these most important anomalies are obtained via object excavation. Each anomaly is also characterized by a size feature, obtained by fitting the EMI response to a bivariate Gaussian model. Using this size feature, all anomalies are clustered via a variational Bayesian Gaussian mixture model. Before designing the classifier, the dipole and size features are used to establish which non-UXO items are sufficiently UXO-like that they should be excluded when designing the classifier; we exclude those non-UXO items, since their inclusion during training would undermine subsequent classifier performance. A kernel matching pursuits (KMP) classifier (using the four dipole-model features) is then constructed. An optional but attractive (based on performance) postprocessing step is also considered, this again exploiting the size-feature clustering result.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 2007
- Accession Number
- ADA631423
Entities
People
- David R. Williams
- Lawrence Carin
Organizations
- Signal Innovations Group, Inc.