This PDF file contains the front matter associated with SPIE Proceedings Volume 10182, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and Conference Committee listing.

Detection and classification of targets is a process that we perform every day, often without even realizing it; the ability to locate and recognize a specific face in a crowd, to be able to identify a bird by its song, or to know what is for dinner simply by the aroma. Much effort has been spent in trying to replicate these uncanny abilities of our senses through purpose-built sensors. For many of these sensing modalities, automatic target recognition algorithms have been developed to automate the process of locating and recognizing the occurrence of a specific state from within the sensed data. This survey is an overview of the metrics used to assess and quantify the quality or ‘goodness’ of the data, mainly for imaging sensors, from the perspective of automatic target recognition performance. Digital image data go through several transformations as they are moved from the recording sensor to personal electronic devices, resulting in distortions. There has been considerable research in the area of quality assessment of digital image and video media. Often, the image quality assessment approaches rely on having the original image or at least information about the original image available for the evaluation. These approaches may also rely on knowing the source of distortion. The accuracy is typically gauged using the image quality assessment made by the human visual system. As a consequence, the metrics measure aspects of the image that are tied closely to the human perception of quality. Evaluating image quality from an automatic target recognition perspective offers many challenges, as the original data is not likely to be available. Additionally, there may be multiple unknown sources of image quality degradation forcing the need for a general solution. Finally, the human perception of ‘good quality’ may not necessarily correlate to optimal image quality for automatic target recognition performance, in which case metrics modeled after the human visual system may not be the best choice for image quality assessment. The goal of this survey is to provide a high-level background and overview of how image quality assessment is currently performed for different sensing modalities. Additionally, the challenges that are being addressed within the various disciplines will be discussed and commonalities will be highlighted.

The detection of improvised explosive devices (IED) is still a challenging task. Important components of these IEDs are often thin pressure plate structures which connect the activator with the explosive device by wires. The detection of wires could therefore be useful to detect the IEDs, since the detection of the explosives itself by identifying specific characteristics is impossible for many sensor types, and quite expensive and time consuming for few being able to perform this task. In this paper investigations on the detection of thin wires using Multiple Input Multiple Output (MIMO) synthetic aperture radar (SAR) are discussed.

This paper aims to investigate different methods for the detection of underground concrete structures using as a case study an area of abandoned military bunkers. Underground structures can affect their surrounding landscapes in different ways, such as alter the moisture capacity of soil, its composition and the vegetation vigor. The latter is often observed on the ground as a crop mark; a phenomenon which can be used as a proxy to denote the presence of underground nonvisible structures. A number of vegetation indices such as the Normalized Difference Vegetation Index (NDVI), Simple Ratio (SR) and Enhanced Vegetation Index (EVI) were utilized for the development of a vegetation index-based procedure aiming at the detection of underground military structures by using existing vegetation indices or other in-band algorithms. One of the techniques examined is that of the C-Band Synthetic Aperture Radar (SAR), which provide information on the vegetation height based on the analysis of the difference between areas of buried structures and reference areas.

Current land mine detection methods predominately rely on the use of Ground Penetrating Radar (GPR) and metal detectors to scan the ground for disturbances in electromagnetic waves that would indicate a higher concentration of metal. For soldiers, these tools combined with trained observational skills, makes the rapid clearance of mine fields more of an art; one that has become a necessity with the presence of Improvised Explosive Devices (IEDs). Some of these detection methods are taught and trained to every soldier, as the principles behind IED detection are similar no matter the IED construction. Specifically, all IEDs require the use of wires in some way.

This research has expanded the use of Compton backscattering detection methods and pencil beam imaging, proven to be slow but accurate. Investigation of fan beam geometries for backscatter imaging is ongoing. The goal is to allow for a rapid scan of potential pressure plates, focused on detecting the signature of a wire in a mock IED, with efforts to improve imaging properties to aid soldiers. This has the potential to increase the accuracy of current interrogation and detection methods. This research has demonstrated some success over current interrogation methods, with the potential to allow military units to interrogate suspected IEDs through nonphysical means with greater image resolution than GPR. Ultimately, this could allow for the ability to clear suspected enemy obstacles faster, with greater accuracy, and providing more security to the soldiers on the ground.

Nuclear Quadrupole Resonance (NQR), a type of radio frequency spectroscopy, holds the promise of unambiguous detection of particular explosives; the associated resonant frequencies are virtually unique. This specificity is spoiled by natural and anthropogenic interference that can swamp the NQR signal. Fortunately, the spatial magnetic signature from the explosive differs significantly from that of interference and can be exploited to separate the signals. An array of coils, however, cannot provide truly independent measurements due to the inductive coupling between the coils. Single coil configurations can cancel out constant interference and retain the signal from an NQR coil, but the balance between arms of the coil is compromised by differential coupling to the environment. Atomic magnetometers, an emergent technology predicted to surpass the sensitivity of coil detection, do not suffer from such coupling and have no fundamental limitation to forming an array of independent sensors. We have demonstrated up to 94× interference rejection with a 4-sensor array spanning 25 cm. The array symmetry permits rejection of linearly varying interference. We discuss the prevention of reradiation from DC field coils used to set the magnetometer resonance frequency. Such reradiation leads to non-linear variation. The benefits of larger sensor arrays are also explored.

Data collected by hand-held and vehicular-mounted ground penetrating radar (GPR) devices can be viewed as a sequence of A-scans grouped into frames. Sequences of frames create a three-dimensional representation of the collected area. The ground structure within this area includes: the ground layer, sub-surface layers, explosive hazards, and non-explosive (clutter) objects. In previous work, we found a wireframe view of two-dimensional layers within the three-dimensional volume. In this work, we analyze how to use image segmentation techniques to identify and view the entire three-dimensional volume of layers and objects found in the data. First, image value and contour differentiate structure. Then, we employ a multi-stage process of image segmentation, clustering analysis using Competitive Agglomeration, and optimizing a Markov Random Field (MRF). The collection sequence of hand-held system data may not always fit into a grid representation. Therefore, segmentation techniques are modified from a frame by frame grid sequence to a nearest neighborhood of three-dimensional regions. Results are displayed in an interactive viewing tool that displays each stage of the process and represents the scene elements identified.

Data from multiple sensors has been collected using a handheld system, and includes precise location information. These sensors include ground penetrating radar (GPR) and electromagnetic induction (EMI) sensors. The performance of these sensors on different mine-types varies considerably. For example, the EMI sensor is effective at locating relatively small mines with metal while the GPR sensor is able to easily detect large plastic mines. In this work, we train linear (logistic regression) and non-linear (gradient boosting decision trees) methods on the EMI and GPR data in order to improve buried explosive threat detection performance.

Sensors which use electromagnetic induction (EMI) to excite a response in conducting bodies have been investigated for the purpose of detecting buried explosives. In particular, wide band EMI sensors which use a relatively low number of operating frequencies have been used to discriminate between types of objects, and to detect objects with very low metal content.¹ In this paper, Fourier features are extracted using the 2D Fourier transform from the complex data; both spatially, and across operating frequencies. Then, the Multiple Instance Adaptive Coherence Estimator (MI-ACE)² is used to learn cross-validated target signatures from these features. These signatures, as well as learned background, statistics are used with ACE to generate confidence maps, which are clustered into alarms. Alarms are scored against a ground truth and compared to other detection algorithms.

Intermediate electrical conductivity (IEC, 100-105 S/m) objects are increasingly important to properly detect and classify. For the US Military, carbon fiber (CF) "smart bomb" unexploded ordnance (UXO) are contaminating training ranges. Home-made explosives (HME) may also fit in this conductivity range. Objects in this conductivity range exhibit characteristic quadrature response peaks at high frequencies (100 kHz-15 MHz). Previous efforts towards electromagnetic induction (EMI) sensing of IEC targets have required single-turn, small-diameter transmitter (Tx) and receiver (Rx) loops. These smaller loops remain electrically short in the high frequency EMI (HFEMI) range (100 kHz-15 MHz), a necessary feature, but provide low signal-to-noise ratio (SNR), especially at low frequencies (1000 Hz-10 kHz). We propose a modification to our pre-production HFEMI instrument which has a hybrid low frequency/high frequency transmit coil. This hybrid system uses many turns in the traditional range, and a single wire turn at HFEMI frequencies, to maximize SNR across a wider EMI band. The turns which are not current carrying in high frequency mode must have negligible inductive coupling to the single current-carrying turn, low enough that any coupling is suitable for background subtraction. This is enforced by mutually disconnecting every turn from every other turn. The instrument uses the same calibration techniques as previously introduced,1 namely background subtraction and ferrite compensation. This paper dis- cusses engineering tradeoffs, compares results to numerical models and actual data from an advanced induction sensor, shows improvement in signal-to-noise ratio (SNR) at traditional EMI frequencies, and shows the same ability to detect IEC targets in the HFEMI band.

For Sonar Automatic Target Recognition problems, the number of mine-like objects is relatively small compared to the number non-mine-like objects available. This creates a heavy bias towards non-mine-like objects and increases the processing resources needed for classifier training. In order to reduce resource needs and the bias towards non-mine-like objects, we investigate selection methods for reducing the non-mine-like target samples while still maintaining as much of the original training information as possible. Specifically, we investigate methods for reducing sample size and bias while maintaining good classifier performance. Several methods are considered during this investigation that cover a wide range of techniques, including clustering and evolutionary algorithms. Each method is evaluated based on the classifier performance when trained on the chosen data samples and the execution time to select the new training set. Results on each method tested are presented using sonar data collected using a sidescan sonar system.

An approach to image labeling by seabed context based on multiple-instance learning via embedded instance selection (MILES) is presented. Sonar images are first segmented into superpixels with associated intensity and texture feature distributions. These superpixels are defined as the "instances" and the sonar images are defined as the "bags" within the MILES classification framework. The intensity feature distributions are discrete while the texture feature distributions are continuous, thus the Cauchy-Schwarz divergence metric is used to embed the instances in a higher-dimensional discriminatory space. Results are given for labeled synthetic aperture sonar (SAS) image database containing images with a variety of seabed textures.

Characteristics of underwater targets displayed in synthetic aperture sonar (SAS) imagery vary depending on their environmental context. Discriminative features in sea grass may differ from the features that are discriminative in sand ripple, for example. Environmentally-adaptive target detection and classification systems that take into account environmental context, therefore, have the potential for improved results. This paper presents an end-to-end environmentally-adaptive target detection system for SAS imagery that performs target recognition while accounting for environmental context. First, locations of interest are identified in the imagery using the Reed-Xiaoli (RX) detector and a Non-Gaussian detector based on the multivariate Laplace distribution. Then, the Multiple Instance Learning via Embedded Instance Selection (MILES) approach is used to identify the environmental context of the targets. Finally, target features are extracted and a set of environmentally-specific k-Nearest Neighbors (k-NN) classifiers are applied. Experiments were conducted on a collection of both high and low frequency SAS imagery with a variety of environmental contexts and results show improved classification accuracy between target classes when compared with classification results with no environmental consideration.

Some synthetic aperture imaging techniques have strict path requirements and need to be processed in overlapping chunks. Path adjustments applied for images can decorrelate and alter the appearance of overlapping regions, which can make mosaicing the images chunks for post mission analysis challenging. In this paper, we investigate the application of automatic change detection techniques to the problem of high accuracy, multi-band mosaicing. We will describe and demonstrate the aspects of synthetic aperture sonar (SAS) change detection, in particular automated change detection (ACD) co-registration processing that enables the production of good quality of seabed maps known as mosaicing where individual SAS runs from different traverses are pieced together in adjacent positions to form a complete image.

Change Detection is the process of detecting changes from pairs of multi-temporal sonar images that are surveyed approximately from the same location. The problem of change detection and subsequent anomaly feature extraction is complicated due to several factors such as the presence of random speckle pattern in the images, variability in the seafloor environmental conditions, and platform instabilities. These complications make the detection and classification of targets difficult. In this paper, we propose a change detection technique for multi-temporal synthetic aperture sonar (SAS) images, based on independent component analysis (ICA). ICA is a well-established statistical signal processing technique that aims at decomposing a set of multivariate signals (in our case SAS images) into a base of statistically independent data-vectors with minimal loss of information content. The goal of ICA is to linearly transform the data such that the transformed variables are as statistically independent from each other as possible. The changes in the imaging scene are detected in reduced second or higher order dependencies by ICA and the correlation among the multi-temporal images is removed. Thus removing dependencies will leave with the change features that will be further analyzed for detection and classification. Test results of the proposed method on SAS images (snippets) of declared changes from automated change detection (ACD) process will be presented. These results will illustrate the effectiveness of ICA for reduction of false alarms in the ACD process.

Automated change detection (ACD) is a technique that automatically discerns any area of change when comparing two images of the same geographic location over different moments in time. Within the ACD processing stream, co-registration ensures the areas depicted in two images coincide. The difficulty in co-registering sonar images of the sea floor can arise from a difference in vehicle trajectories, low resolution, and the presence of noise. Moreover, the changing features of the sea floor can further add to the difficulty. The successful co-registration of sonar images is important when comparing images, and is thus required in areas such as change detection and mosaicing. In this effort, a three-step co-registration process is used: co-registration by navigational alignment, fine-scale co-registration using SIFT, and local co-registration that corrects navigational differences. In this paper, we focus on the final step where phase alignment occurs. To eliminate unreliable unwrapped phase data, we introduce a novel histogram based adaptive thresholding technique which rejects errors in phase alignment occurring in the across-track direction of the vehicle. Further, an adaptive thresholding technique is applied to the change-map generated following the co-registration stage. To isolate pixels of interest related to anomalies or targets, a thresholding method is applied in conjunction with principal and independent component analysis (PCA and ICA).

We will demonstrate the effectiveness of these adaptive thresholding techniques in sub-pixel co-registration and target detection.

Search planning for UUV missions involves a complex optimization problem over multiple degrees of freedom. The detection performance of individual searches is governed by the local environment and the sensor settings that determine the rate at which target objects of interest can be identified. This paper addresses some of the computational aspects of collaborative search planning when multiple search agents seek to find hidden objects (i.e. mines) in adverse local operating environments where the detection process is prone to false alarms. For these conditions, we apply a receiver operator characteristic (ROC) analysis to model detection performance and develop a Bayesian risk objective that enumerates the required number of detections for validation as part of the modeling paradigm. In this paper we consider an expanded optimization process whereby ROC operating points are optimized simultaneously for all validation criteria with the best performing combinations selected. Details of its application within a broader search planning context are discussed. An analysis of this optimization process is presented. Numerical results are provided to demonstrate the effectiveness of the approach.

Ground penetrating radar (GPR) based detection systems have used a variety of different features and machine learning methods to identify buried hazards and distinguish them from clutter and other objects. In this study, we describe a new feature extraction method based on Kolmogorov complexity and information theory. In particular, a three dimensional subset of GPR data centered at alarm location is partitioned into two-dimensional non-overlapping cells. Then, each cell is compressed using gzip and a feature vector is formed from the file sizes of the compressed cells. Finally, an SVM classifier is trained on compression features. The proposed method is applied to data acquired from outdoor test sites containing over 3800 buried hazards, including nonmetal and low-metal targets. The performance is measured by use of ROC curves and compared against four algorithms. These algorithms are based on geometric features and are fused with a powered geometric mean method. The compression-based algorithm outperforms the other individual methods. We also tested different fusion algorithms involving combinations of these five algorithms. The best combination, the product of compression algorithm and two of the others, dominates the current state of the art solution by a significant margin.

Ground penetrating radar (GPR) systems have emerged as a state-of-the-art remote sensing platform for the automatic detection of buried explosive threats. The GPR system that was used to collect the data considered in this work consists of an array of radar antennas mounted on the front of a vehicle. The GPR data is collected as the vehicle moves forward down a road, lane or path. The data is then processed by computerized algorithms that are designed to automatically detect the presence of buried threats. The amount of GPR data collected is typically prohibitive for real-time buried threat detection and therefore it is common practice to first apply a prescreening algorithm in order to identify a small subset of data that will then be processed by more computationally advanced algorithms. Historically, the F1V4 anomaly detector, which is energy-based, has been used as the prescreener for the GPR system considered in this work. Because F1V4 is energy-based, it largely discards shape information, however shape information has been established as an important cue for the presence of a buried threat. One recently developed prescreener, termed the HOG prescreener, employs a Histogram of Oriented Gradients (HOG) descriptor to leverage both energy and shape information for prescreening. To date, the HOG prescreener yielded inferior performance compared to F1V4, even though it leveraged the addition of shape information. In this work we propose several modifications to the original HOG prescreener and use a large collection of GPR data to demonstrate its superior detection performance compared to the original HOG prescreener, as well as to the F1V4 prescreener.

In recent years, the Ground Penetrating Radar (GPR) has successfully been applied to the problem of buried threat detection (BTD). A large body of research has focused on using computerized algorithms to automatically discriminate between buried threats and subsurface clutter in GPR data. For this purpose, the GPR data is frequently treated as an image of the subsurface, within which the reflections associated with targets often appear with a characteristic shape. In recent years, shape descriptors from the natural image processing literature have been applied to buried threat detection, and the histogram of oriented gradient (HOG) feature has achieved state-of-the-art performance. HOG consists of computing histograms of the image gradients in disjoint square regions, which we call pooling regions, across the GPR images. In this work we create a large body of potential pooling regions and use the group LASSO (GLASSO) to choose a subset of the pooling regions that are most appropriate for BTD on GPR data. We examined this approach on a large collection of GPR data using lane-based cross-validation, and the results indicate that GLASSO can select a subset of pooling regions that lead to superior performance to the original HOG feature, while simultaneously also reducing the total number of features needed. The selected pooling regions also provide insight about the regions in GPR images that are most important for discriminating threat and nonthreat data.

The ground penetrating radar (GPR) is a popular remote sensing modality for buried threat detection. In this work we focus on the development of supervised machine learning algorithms that automatically identify buried threats in GPR data. An important step in many of these algorithms is feature extraction, where statistics or other measures are computed from the raw GPR data, and then provided to the machine learning algorithms for classification. It is well known that an effective feature can lead to major performance improvements and, as a result, a variety of features have been proposed in the literature. Most of these features have been handcrafted, or designed through trial and error experimentation. Dictionary learning is a class of algorithms that attempt to automatically learn effective features directly from the data (e.g., raw GPR data), with little or no supervision. Dictionary learning methods have yielded state-of-theart performance on many problems, including image recognition, and in this work we adapt them to GPR data in order to learn effective features for buried threat classification. We employ the LC-KSVD algorithm, which is a discriminative dictionary learning approach, as opposed to a purely reconstructive one like the popular K-SVD algorithm. We use a large collection of GPR data to show that LC-KSVD outperforms two other approaches: the popular Histogram of oriented gradient (HOG) with a linear classifier, and HOG with a nonlinear classifier (the Random Forest).

The Ground Penetrating Radar (GPR) is a remote sensing modality that has been used to collect data for the task of buried threat detection. The returns of the GPR can be organized as images in which the characteristic visual patterns of threats can be leveraged for detection using visual descriptors. Recently, convolutional neural networks (CNNs) have been applied to this problem, inspired by their state-of-the-art-performance on object recognition tasks in natural images. One well known limitation of CNNs is that they require large amounts of data for training (i.e., parameter inference) to avoid overfitting (i.e., poor generalization). This presents a major challenge for target detection in GPR because of the (relatively) few labeled examples of targets and non-target GPR data. In this work we use a popular transfer learning approach for CNNs to address this problem. In this approach we train two CNN on other, much larger, datasets of grayscale imagery for different problems. Specifically, we pre-train our CNNs on (i) the popular Cifar10 dataset, and (ii) a dataset of high resolution aerial imagery for detecting solar photovoltaic arrays. We then use varying subsets of the parameters from these two pre-trained CNNs to initialize the training of our buried threat detection networks for GPR data. We conduct experiments on a large collection of GPR data and demonstrate that these approaches improve the performance of CNNs for buried target detection in GPR data

Tohoku University has developed dual sensor ALIS for humanitarian demining. ALIS is a system combining a GPR and a metal detector, and 2 sets of ALIS are used by CAMC (Cambodian Mine Action Center) since 2009 and one team is working with them and detected more than 80 mines. Data acquisition and interpretation under strong clutter conditions is very challenging for the use of GPR and metal detector in mine fields. We use SAR (Synthetic Aperture Radar) processing for GPR imaging and it is quite successful. Based on the experience of ALIS in Cambodia, we used sensor technology for survey activities in natural disasters. In September 2014, Ontake-san volcano in Nagano, Japan erupted and 58 were killed by fallen volcanic ash. We introduced a metal detector used for humanitarian demining to the local police and one victim could be found. We could prove that the function of a metal detector reducing the response from magnetic and metallic soil worked well. In 2015, we also introduced GPR for survey under the ash. In March 2011, the east part of Japan was attacked by a huge scale Tsunami and more than 15,000 were killed and still more than 2,500 are missing. We expect that many missing victims were flashed away by Tsunami and could be buried in sand in sand beaches or areas near the costs. We have worked together with local police stations and used metal detectors and multistatic GPR system “Yakumo” for the survey of Tsunami victims. In many beaches, we found that the semimetal layer of the Tsunami in 2011 is about 50-70cm thick, and could detect many objects buried under this sand layer. We think SAR processing in GPR is important for detecting small buried objects. We also demonstrate visualized metal detector sensors can be effectively used for survey in landslides.

Ground penetrating radar (GPR) has the ability to detect buried targets with little or no metal content. Achieving superior detection performance with a hand-held GPR can be very challenging due to the quality of the data, inconsistency of target signatures, variety of target types, and effects of a human operator. In this paper, we investigate the use of a local binary patterns (LBP) feature vector for target versus non-target discrimination from hand-held GPR data. First, a prescreener algorithm is applied to the GPR data. Then, a GPR B-scan is gathered at each prescreener alarm location and separated into several spatial and depth regions. LBP processing is applied to each spatial and depth cell individually and the LBP features from each cell are group together to form a feature vector. The resulting LBP features are invariant to amplitude scaling and represent the texture in the data. Using this feature vector, a classifier is trained to perform target versus non-target discrimination at each prescreener declaration. Experimental results illustrate the ability of the LBP features to improve detection of buried targets, especially low-metal and non-metal anti-tank and anti-personnel targets.

Buried threats such as Improvised Explosive Devices (IEDs) and UneXploded Ordnance (UXO) can be composed of different materials including metal, carbon fiber, carbon rods, and nonconducting material such as wood, rubber, fuel oil, and plastic. Electromagnetic induction (EMI) instruments have been traditionally used to detect high electric conductivity discrete targets such as metal UXO. The frequencies used for this EMI regime have typically been less than 100 kHz. To detect intermediate conductivity objects like carbon fiber, higher frequencies up to the low megahertz range are required in order to capture characteristic relaxation responses. Nonconducting voids in an otherwise conducting background medium like soil channel currents around the void. These channeling currents exhibit relaxation responses similar to conducting targets but with a much higher frequency response. Nonconducting plastic landmines can be considered a void plus small metallic parts such as the firing pin, and a characteristic relaxation response due to both the void and the metal parts can be obtained which can reduce false alarms from EMI instruments that detect only the metal. To predict EMI phenomena at frequencies up to 15MHz, we modeled the response of conducting and nonconducting targets using the Method of Auxiliary Sources. Our high-frequency electromagnetic induction (HFEMI) instrument is able to acquire EMI data at frequencies up to that same high limit. Modeled and measured characteristic relaxation signatures compare favorably and indicate new sensing possibilities in a variety of scenarios including the detection of voids and landmines.

Broadband electromagnetic induction sensors are effective at detecting and classifying buried metal, such as landmines, by using an incident magnetic field to induce eddy currents in buried metal and measuring the secondary magnetic field they produce. A target’s magnetic polarizability is a frequency dependent tensor that expresses the relationship between the incident magnetic field and the scattered field created by the eddy currents. Viewed as a singularity expansion, the magnetic polarizability can be decomposed into contributions from a set of eddy-current modes which relax at different frequencies. This decomposition provides a unique signature for targets of interest, allowing them to be effectively distinguished from clutter. In this paper, an improved, streamfunction-based numerical approach will be presented that can accurately compute the eddy-current modes that flow in thin conducting shells and their contribution to the singularity expansion of the magnetic polarizability.

This work describes a target location estimation procedure for single channel 2-D electromagnetic induction (EMI) scan data for both a rotating and non-rotating sensor head. The location estimation technique is based on a new low-rank model for electromagnetic induction data and and addresses difficulties in how to efficiently construct the dictionary for inverting the single channel EMI data. The location estimation technique is tested on simulated data for a single z-directed wire loop for both a rotating and non-rotating sensor head, and on non-rotating sensor head laboratory measurements for four different mine types. The technique is shown to accurately estimate the target location for simulated data and experimental data, and is able to estimate the location of strong mine targets at depths of 20 – 30 cm with ±2 cm error.

Many different coil head configurations are currently used in electromagnetic induction (EMI) systems for sensing buried targets. It is difficult to compare the performance of different coil head designs to one another. Comparing two particular implementations is not particularly challenging, but fairly comparing general coil head designs is nontrivial. This work details both a target sensitivity metric and soil sensitivity metric that account for differences such as overall coil size, length, and wire diameter. The metrics are then used to compare the performance of a range of concentric EMI coil head designs.

A broadband electromagnetic induction (EMI) sensor array is developed to help discriminate between types of metallic targets. The sensor uses a single dipole transmit coil and an array of four quadrapole receive coils. The sensor operates in the frequency domain and collects data at 15 logarithmically spaced frequencies from 1 kHz to 90 kHz. The system is designed to be very sensitive while making very accurate measurements. The accuracy is very important to be able to filter out soil effects and to effectively classify targets. Experimental results are presented to show the accuracy of the system.

Handheld electromagnetic induction (EMI) sensors used for detection of buried landmines typically employ planar sensor head geometries comprised of a single transmitter and one or more receiver coils. Such configurations have dominant sensitivities oriented in a single direction and are therefore well-suited for metal detection. However, when one tries to characterize the target with a planar EMI sensor, e.g. reconstruct the parameters of an induced dipole model, the problem becomes challenging, since the target needs to be interrogated by all three orthogonal field components in order to obtain unambiguous results. Sweeping the sensor over a target helps with this issue, since the target gets magnetically illuminated from different directions, however a downside is that the sensor head position needs to be known in a ground reference frame, at each scan location. Although various solutions for sensor head position tracking have been reported, they are all more or less obtrusive when used in field conditions. Also, such systems introduce additional uncertainties, often with non-normal distributions, thus complicating inversion procedures. In this paper, we investigate an alternative approach to dipole-based target characterization, where the problem is analyzed in a moving, sensor-based reference frame by dynamically processing sensor data at each scan location using the extended Kalman filter. For a chosen planar coil geometry and scan pattern, we first analyze the basic stability of a filter, with an emphasis on its local observability. We then show how the target's magnetic polarizabilities and target-to-sensor distance can be simultaneously estimated using EMI data only.

Depleted uranium (DU) is a byproduct of the uranium enrichment process and contains less than 0.3 % of the radioactive U-235 isotope. Since, the natural uranium has about 0.72 % of the uranium U-235 isotope, the enrichment produces large quantities of low-level radioactive DU. The non-fissile uranium U-238 isotope constitutes the main component of DU and makes it very dense. With 19.1 g/cm³ density, the DU is about 68.4 % denser than lead. Because of its high density, the DU has been used for as armor-piercing penetrators by the U.S. army. There are at least 30 facilities where munitions containing DU have been evaluated or used for training. These evaluation studies have been conducted with and without catch-boxes and have left a legacy of DU contamination. Thus, there are needs for rapid and cost-effective approaches to detect and locate subsurface DU munitions and to assess large contaminated areas. In this paper, a new ultra-wideband (from 10s of Hertz up to 15 Megahertz) geophysical instrument is evaluated for sensing subsurface DU munitions and DU materials related to contaminations in soil. Namely, full electromagnetic induction (EMI) responses are investigated using computational and experimental data for a DU rod, dart, and three samples of Yuma Proving Ground (YPG) soils. Numerical data are obtained via the full 3D EMI solver based on the method of auxiliary sources. The EMI signals sensitivity with respect to DU size, orientations, and material composition are illustrated and analyzed. Comparisons between computational and experimental studies are demonstrated. The studies show that the new ultra-wideband EMI sensor measures the complete polarization relaxation response from the DU rod and dart, and is able to sense relative DU contamination levels in soil.

Substantial interest resides in identifying sensors, algorithms and fusion theories to detect buried explosive hazards. This is a significant research effort because it impacts the safety and lives of civilians and soldiers alike. Herein, we explore the fusion of different algorithms within and across ground penetrating radar (GPR) and electromagnetic induction (EMI) sensors on a U.S. Army NVESD furnished experimental handheld demonstrator (EHHD) platform. Fusion is not trivial as these sensors have different sampling rates, resolutions, they observe different spatial areas and they span different portions of the electromagnetic spectrum. Herein, we investigate and compare the use of different approaches for co-registration and decision-level fusion using the Choquet integral (ChI). With respect to the ChI, we explore the impact of using a "global" (single) aggregation strategy (operator) versus tailoring different ChIs to subsets of algorithms and sensors. Receiver operating characteristic (ROC) curve results are shown for data from a U.S. Army test site containing multiple target and clutter types, burial depths and times of day.

Wideband electromagnetic induction (WEMI) sensors can be used to detect, classify, and locate metallic targets buried underground. By examining the WEMI model from a new perspective, a low-rank equivalence for the WEMI data is obtained. The low-rank physical WEMI model directly leads to a new “filterless” processing framework that differs radically from traditional approaches to WEMI processing. The new framework enables target signature recovery independent of location estimation. Processing of field data with the new filterless WEMI framework is presented to show classification results based on the low-rank target signature.

This paper discusses the application of several image formation techniques to forward looking ground penetrating radar (FLGPR) data to observe if they improve target-to-clutter ratio. Specifically, regularized imaging with 𝐿1 and total variation constraints and coherence-factor filtered images are considered. The technical framework and software implementation of each of these image formation techniques are discussed, and results of applying the techniques to field collected data are presented. The results from the different techniques are compared to standard backprojection and compared to each other in terms of image quality and target-to-clutter ratio.

Buried targets pose a serious threat to modern soldiers and civilians alike, thus detecting them from a safe standoff distance is an important step in their remediation. Many successful vehicle-based detection systems have been designed to utilize forward-looking ground penetrating radar (FLGPR) for buried target detection at a distance, however, FLGPR has an inherently low signal-to-clutter-ratio (SCR) so its performance is limited. To address this limitation, suites of sensors have been added to some of these vehicle-based systems. In this work we utilize data from these various sensors to improve the buried target classification accuracy. Specifically, we present features extracted from FLGPR, lidar, and thermal- and visible-spectrum camera data, then fuse the various features using a kernel-based classifier. Our results indicate that fusing these multimodal features yields a higher classification performance than utilizing data from the FLGPR alone. We also analyze each sensor's incremental improvement of classification accuracy by performing numerous experiments with different permutations of the sensors.

Detection of improvised explosive devices (IEDs) presents a significant challenge for stand-off sensors. IEDs can be constructed from a wide variety of materials and take many different forms. Forward-looking and side-looking sensor systems attempt to detect targets at significant stand-off but typically observe a potential threat area over a wide range of aspect before detection and classification is attempted. Furthermore, in practical detection scenarios IEDs are obscured preventing unhindered line-of-sight interrogation. Both of these characteristics can lead to target signatures which become incoherent over time. In this paper we investigate an advanced imaging technique designed to improve image quality of obscured targets utilizing a wide-band RF array. Specifically, we consider buried targets and obscured roadside targets. Results are presented in terms of improved signal to clutter ratio of image reconstructions based on simulated and experimental GPR collections.

In an application of side-attack explosive hazard (SAEH) detection, one challenge is detecting occluded targets. Forward-looking ground penetrating radar (FLGPR) is a sensor that has the capability to handle this type of problem since this frequency can penetrate occlusion and detect the explosive. However, the explosive can be concealed by multiple types of objects along the side of the road, such as rocks, vegetation, trash, etc. In FLGPR imaging space, these clusters of occlusion come in many shapes and sizes. The log Gabor descriptor, which has the advantage of separating different types of textures, is a potential solution for the aforementioned issue. However, in order to fully utilize the log Gabor filter bank, there are many parameters that need to be handcrafted for each specific problem. Since the possibility of designing a filter bank is endless, brute-force search is not possible. An additional element to be considered is the size of the filter bank, since a large filter bank requires more computation time. In this paper, we propose the use of multi-objective optimization based on an evolutionary algorithm to streamline the process of designing a log Gabor filter bank. The objectives of the optimization are to design a filter bank that can separate the explosive hazards from false alarm hits while keeping the size of the filter bank to a minimum. The data we use to validate our proposed method was collected on arid lanes by a MIMO FLGPR system on board of the U.S. Army prototype vehicle. The experiments are designed to measure the improvement of the optimization over the manual tuning and to demonstrate the impact of the optimization’s parameters on the performance of the filter bank.

Forward-looking ground penetrating radar (FLGPR) is a remote sensing modality that has been investigated for buried threat detection. The FLGPR considered in this work consists of a sensor array mounted on the front of a vehicle, which inspects an area in front of the vehicle as it moves down a lane. The FLGPR collects data using a stepped frequency approach, and the received radar data is processed by filtered backprojection to create images of the subsurface. A large body of research has focused on developing effective supervised machine learning algorithms to automatically discriminate between imagery associated with target and non-target FLGPR responses. An important component of these automated algorithms is the design of effective features (e.g., image descriptors) that are extracted from the FLGPR imagery and then provided to the machine learning classifiers (e.g., support vector machines). One feature that has recently been proposed is computed from the magnitude of the two-dimensional fast Fourier transform (2DFFT) of the FLGPR imagery. This paper presents a modified version of the 2DFFT feature, termed 2DFFT+, that yields substantial detection performance when compared with several other existing features on a large collection of FLGPR imagery. Further, we show that using partial least-squares discriminative dimensionality reduction, it is possible to dramatically lower the dimensionality of the 2DFFT+ feature from 2652 dimensions down to twenty dimensions (on average), while simultaneously improving its performance.

Substantial interest resides in identifying sensors, algorithms and fusion theories to detect explosive hazards. This is a significant research effort because it impacts the safety and lives of civilians and soldiers alike. However, a challenging aspect of this field is we are not in conflict with the threats (objects) per se. Instead, we are dealing with people and their changing strategies and preferred method of delivery. Herein, we investigate one method of threat delivery, side attack explosive ballistics (SAEB). In particular, we explore a vehicle-mounted synthetic aperture acoustic (SAA) platform. First, a wide band SAA signal is decomposed into a higher spectral resolution signal. Next, different multi/hyperspectral signal processing techniques are explored for manual band analysis and selection. Last, a convolutional neural network (CNN) is used for filter learning and classification relative to the full signal versus different subbands. Performance is assessed in the context of receiver operating characteristic (ROC) curves on data from a U.S. Army test site that contains multiple target and clutter types, levels of concealment and times of day. Preliminary results indicate that a machine learned CNN solution can achieve better performance than our previously established human engineered Fourier-based Fraz feature with kernel support vector machine classification.

Objects imaged in synthetic aperture acoustic data have a unique appearance. Due to this, we propose that examining the texture between targets and non-targets will prove more descriptive and improve classification performance. A few common texture feature extraction methods are those derived from grey-level co-occurrence matrix (GLCM), local binary patterns (LBP), and local directional patterns (LDP). LDP uses a set of filters to measure the local directional response around each pixel and then builds a binary code like LBP. The feature vector is a histogram of those binary codes. However, the set of filters used may not be the optimal set needed to achieve the best classification accuracy and a binary coding may not be the best aggregation method. In this paper, we apply known sets of two-dimensional filters, not necessarily directional, as well as develop a new approach to aggregation. Different filter sets provide the algorithm with a broader description beyond the direction of edges and thus better representation of texture. A more complex aggregation method allows more information retention in the feature vector. These modifications, to the existing LDP algorithm, will allow classifiers to more accurately distinguish between the textures of targets and non-targets. A support vector machine (SVM) helps evaluate the performance of the new feature extraction method and compare its performance to other common extraction methods on data collected at a US Army test site. This will be used to build an online classifier system for testing on lane-based data.

Recently, the Stalker system has been developed as a high-resolution three-dimensional radar imaging system for the detection of concealed roadside explosive hazards. This system has shown considerable capability in distinguishing between true targets and false alarms using conventional processing techniques such as RX filtering on 2D projections of the data. In this paper, we develop an extension of these methods for use with 3D radar imagery. We show several different prescreening approaches for automatically marking potential target locations and describe an evaluation program called the Tiger scorer. We tested our approach on data collected at an arid U.S. Army test site.

Explosive hazards in current and former conflict zones are a serious threat to both civilians and soldiers alike. Significant effort has been dedicated to identifying sensors, algorithms and fusion strategies to detect such threats. However, a challenging aspect of the field is that we are not necessarily at war with the threats (objects). Instead, we are at conflict with people who are constantly evolving their strategies of attack along with their preferred threat. One such method of threat delivery is side attack explosive ballistics (SAEB). In this article, we explore different 3D voxel-space radar signal processing methods for SAEB detection on a U.S. Army provided vehicle-mounted platform. In particular, we explore the fusion of a matched filter (MF) and size contrast filter (SCF). Clustering is applied to the fused result and heuristics are used to reduce the systems false alarm rate. Performance is assessed in the context of receiver operating characteristic (ROC) curves on data from a U.S. Army test site containing multiple target and clutter types, levels of concealment and times of day.

One promising technique for improving tunnel detection is the use of spotlight synthetic aperture radar (SL-SAR) in conjunction with focusing techniques. Still, clutter arises from surface variations while severe attenuation of the target signal occurs due to the dielectric properties of the soil. To combat these ill-effects, this work aims to improve imaging and detection of underground tunnels by examining the feasibility of matched illumination waveform design for tunnel detection applications. The tunnel impulse response is incorporated in an optimum waveform derivation scheme which aims to maximize the signal-to-interference and noise ratio (SINR) at the receiver output. Numerical electromagnetic simulations are used to consider wave propagation in realistic soil scenarios which include uniform and non-uniform moisture profiles. It is demonstrated that by considering matched illumination waveforms for transmission in SL-SAR systems, an improvement in the detection and imaging capabilities is achieved through enhanced SINR.