Bayesian Mitigation of Sensor Position Errors to Improve Unexploded Ordnance Detection

Tantum, Stacy L.; Yu, Yongli; Collins, Leslie M.

doi:10.1109/lgrs.2007.912088

Cited by 44 publications

(41 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In [5] and [6], methods are developed to account for positional error in inversion using minimax and Bayesian formulations, respectively. Lhomme et al developed metrics for measuring data quality in the figure of merit (FOM), which is an empirical measure of the expected reliability of the recovered model [7].…”

Section: Detecting Outliers To the True Positive Distributionmentioning

confidence: 99%

Additional Analysis of the ESTCP Discrimination Study Data at Camp Sibert, Alabama. Project 200504: Practical Discrimination Strategies for Application to Live Sites

Billings¹,

Pasion²,

Beran³

2008

View full text Add to dashboard Cite

Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA 22202-4302, and Geonics EM61 cart, MTADS EM61 and EM-63 cart data from Camp Sibert were investigated to determine if the number of "can't analyze" anomalies could be reduced and more objective stop-digging criteria selected. Many of the "can't analyze" anomalies in the MTADS EM61 were caused by cart-bounce along North-South transects and could be avoided by using only East-West transect data. However, poor data coverage increased the chances of generating false negative declarations. If North-South transects are retained, sensor motion relative to the ground can cause difficulties in obtaining good model fits to the data. A modeling framework developed elsewhere was used to determine if an anomaly was caused by sensor motion or a compact metallic target. The method has promise but had limited applicability due to the lack of accurate ground-clearance and topographic data. To determine an objective operating point, the training data must be representative of the test-data. In particular, outliers need to be avoided and here this was achieved by using multiple feature vectors, obtained by analysis of a depth versus misfit curve, in the classification. Using this technique a false-negative in the EM-63 data could be avoided. We investigated the cause of the can't analyze anomalies of geological origin whose amplitude exceeded 25 mV along the E-W transects. We hypothesized that many of these anomalies were due to sensor movement relative to ground. The background response was modeled by estimating a ground clearance from the elevation data and assuming that the background magnetic susceptibility was uniform in each cell. The accuracy of our modeling was limited due to filtering artifacts in the observed data, the accuracy of the ground clearance estimate, and small scale topography (i.e. depressions and bumps on the surface that would affect the measured data).In many cases we found that small scale anomalies could be predicted using our modeling techniques and that sensor movement was indeed the likely cause. We also found that there were a number of metallic anomalies whose response could not be properly modeled due to variations in the signal from the background due to sensor movement. There were also several anomalies that were caused by longer wavelength spatial variations in magnetic soil properties that were not suppressed by the detrend filters that were used to pre-process the sensor data. We conclude that sensor movement relative to th...

show abstract

Section: Detecting Outliers To the True Positive Distributionmentioning

confidence: 99%

Additional Analysis of the ESTCP Discrimination Study Data at Camp Sibert, Alabama. Project 200504: Practical Discrimination Strategies for Application to Live Sites

Billings¹,

Pasion²,

Beran³

2008

View full text Add to dashboard Cite

show abstract

“…Increasing error, bias, and feature variability subsequently negatively impacts discrimination performance (Bell 2005;Walker et al 2006;Steinhurst et al 2005;Shamatava et al 2004;Tantum et al 2003Tantum et al , 2007Tarokh et al 2007). Uncertainty in sensor positions arises from both measurement uncertainty (e.g., noise in the GPS data), as well as from issues that arise from the sensor deployment.…”

Section: Introductionmentioning

confidence: 99%

“…The other set of approaches is more statistically based, and incorporating knowledge that the sensor positions are uncertain into the inversion process (Tantum et al 2003(Tantum et al , 2007Tarokh et al 2007). Both Bayesian and minimax approaches have been considered, and fairly dramatic improvements in both parameter estimation and discrimination performance under conditions of sensor position uncertainty have been demonstrated (Tantum et al 2007;Tarokh et al 2007).…”

Section: Introductionmentioning

confidence: 99%

“…In particular, we demonstrate the impact on performance if the statistics are inaccurate. These studies were not performed in Tantum et al (2007), where only accurate prior information on the statistical parameters of the position uncertainty were considered. In Sect.…”

Section: Introductionmentioning

confidence: 99%

“…Generally speaking, the inversion procedure fits the measured data to several instantiations of the signatures predicted by the forward (phenomenological) model, and then uses some error metric to determine which set of parameters generated the predicted signature from the forward model that bests fits the data. It is common to utilize a least squares method to minimize the error between the measured data, r, and a signature predicted by the forward model, s. Following the notation in Tantum et al (2007) for the most general scenario, suppose we partition the feature space obtained through the inversion as [X,c,q,k] where X are the parameters that come from the model that will be passed to the discrimination algorithm, c are the parameters associated with the model but that will not be used for discrimination, q are parameters of the model assumed to be known, and k represents the dependence of the signal on the independent variable of time or frequency. In the problem posed here, X would be the dipole parameters (and potentially estimated depth), c would be the extrinsic parameters such as dipole x-, y-and z-location and dipole angles, and q would be the assumed known measurement positions.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Modeling position error probability density functions for statistical inversions using a Goff–Jordan rough surface model

Tantum

Zhu

Torrione

et al. 2008

Stoch Environ Res Risk Assess

Self Cite

View full text Add to dashboard Cite

Buried unexploded ordnance (UXO) continues to be a difficult remediation problem from both a sensing and a discrimination point of view. Modern approaches to both the sensing and discrimination problems utilize high bandwidth electromagnetic induction (EMI) sensors to collect geo-referenced data which is then inverted, or fit, using a forward model in order to obtain features that can be directly interpreted using the physics associated with electromagnetic induction-based sensing. These features are then used in a variety of classification architectures. One aspect of this process that has captured recent interest is that uncertainty in the positions at which data was collected can degrade the inversion performance and thus the subsequent classification. Several mechanisms to address this issue have been explored that range from filtering and prediction of actual positions to exploiting Bayesian approaches for uncertainty mitigation. In the Bayesian approach, a statistical model of the position errors is used as a prior for integrating over the uncertainty in the inversion process. In this study, we demonstrate that errors in the statistical priors used in this process can negatively impact subsequent classification performance, thus highlighting the need for an accurate statistical model for the position errors. Next, we propose a mechanism by which to obtain such models. Specifically, we utilize a Goff-Jordan rough surface model and simulate the sensor data collection system motion over the simulated ground or ocean surfaces to calculate errors and generate statistical models. Our results suggest that this approach can be used to develop the statistical models necessary for mitigating uncertain position information.

show abstract

Overview of Statistical Tests for Unexploded Ordnance Detection

Delic

NATO Science for Peace and Security Series B: Physics and Biophysics

View full text Add to dashboard Cite

Bayesian Mitigation of Sensor Position Errors to Improve Unexploded Ordnance Detection

Cited by 44 publications

References 26 publications

Additional Analysis of the ESTCP Discrimination Study Data at Camp Sibert, Alabama. Project 200504: Practical Discrimination Strategies for Application to Live Sites

Additional Analysis of the ESTCP Discrimination Study Data at Camp Sibert, Alabama. Project 200504: Practical Discrimination Strategies for Application to Live Sites

Modeling position error probability density functions for statistical inversions using a Goff–Jordan rough surface model

Overview of Statistical Tests for Unexploded Ordnance Detection

Contact Info

Product

Resources

About