WHILE diplomats work to restrict the manufacture, sale, and use of land mines worldwide, a massive cleanup effort

is needed to find and destroy the estimated 100 million land mines still buried in 65 countries. Land mines left behind from wars worldwide are one of the century's main unsolved problems of war and remain the focus of humanitarian mine detection and removal primarily in Europe, Africa, Asia, and Central and South America. A combination of technologies from Lawrence Livermore National Laboratory is being directed toward the most daunting challenge presented by land mines--quickly determining the location of each individual land mine in an area so all of them can be removed. The Laboratory's patented micropower impulse radar and advanced imaging technologies are being combined in a practical system called the Land-Mine Detection Advanced Radar Concept, or LANDMARC, that is making pivotal advances in meeting the challenge of land-mine detection.

The Detection Dilemma Effective solution of the problem posed by land mines means that close to 100% of the mines in any area must be detected at the fastest rate possible and with few false alarms (i.e., mistaking a buried object, such as a rock, for a mine). The United Nations, for example, has set the detection goal at 99.6%, and the U.S. Army's allowable false-alarm rate is one false alarm in every 1.25 square meters. No existing land-mine detection system meets these criteria. And the reasons for this failure have as much to do with the mines themselves and the variety of environments in which they are buried as with the limits or flaws in the current technology. Land mines are of two basic types--antitank and antipersonnel. Antitank mines are larger and more powerful than antipersonnel mines. However, antipersonnel mines are the most common type of mine, yet the most difficult to find because they are small and often made of plastic. Antitank mines generally contain more metal than do antipersonnel mines and are thus more easily detectable by simple metal detectors. Both types are buried as close to the surface as possible and are found in a variety of soils and terrain-rocky or sandy soil, open fields, forested areas, steep terrain, jungle. For both types of mines, detonation is typically caused by pressure, although some are activated by a tripwire or other mechanisms. Thus, a land-mine detector must do its job without having direct contact with a mine. It also must be able to locate all types of mines individually in a variety of environments. Other Detection Technologies Various detection technologies are currently used, each with limits or flaws. Dogs and other "sniffers" have high ongoing expenses, are subject to fatigue, and can be fooled by masked scents. Metal detectors are sensitive to metal mines and firing pins but cannot reliably find plastic mines. Infrared detectors effectively detect recently placed mines, but they are expensive and limited to certain temperature conditions. Thermal neutron activation detectors are accurate but are large for field use, slow, and expensive. In early attempts, ground-penetrating radar was sensitive to large mines, had good coverage rate at a distance, and with signal processing, could discriminate antitank mines from clutter such as rocks beneath the ground surface. This type of radar, however, remains expensive, cannot detect antipersonnel mines because its resolution is too low, and frequently records false alarms from clutter sources. Livermore's ongoing LANDMARC project addresses all of these problems and stands a good chance of solving them, especially the problems of detecting small plastic antipersonnel mines and reducing the false-alarm rate. Livermore's Systems Approach The LANDMARC system's enabling technology is micropower impulse radar (MIR), which was invented at Livermore in 1993 as an outgrowth of the Nova laser program.1 The invention, which won an R&D 100 Award in 1993 and an Excellence in Technology Transfer award from the Federal Laboratory Consortium in 1996, led directly to a battery-operated pulsed radar that is remarkably small and inexpensive, has a wide frequency band, and works well at short ranges2--all necessary attributes of land-mine detection systems. MIR's small size, light weight, and low power requirements make it superior to any

previous attempts to use ground-penetrating radar to detect land mines. MIR's ultrawide bandwidth is the source of the high-resolution imaging capabilities that differentiate LANDMARC from similar land-mine-detection technologies. Furthermore, the ability to group individual MIR units in arrays increases the speed and coverage area of LANDMARC's detection work. Livermore's LANDMARC team has combined MIR units with a high-performance imaging system that uses sophisticated computer algorithms to convert large amounts of raw waveform data from the MIR units to high-resolution two- and three-dimensional images of the subsurface. The prototype systems enable users to visualize both antitank and antipersonnel mines and to differentiate them from rocks and other clutter of similar size and shape by the reflected MIR signal. Once the mines can be "seen" and identified, they can be recovered and destroyed. LANDMARC prototypes have multiple MIR units that are either configured in a handheld wand, much like that used for simple metal detectors, or mounted on a small robotic cart (Figure 1). In either configuration, the MIR array is passed over the ground with the antennas of the units about 10 centimeters above the surface. The units rapidly emit microwave impulses with very short risetimes (100 trillionths of a second) that radiate from transmitting antennas and penetrate the ground. These impulses strike and penetrate buried objects, bounce back to a receiving antenna, and are sampled and processed by an onboard computer to measure changes in the dielectric and conductivity properties of the subsurface. In a few seconds, the data reconstruction algorithms convert the raw radar data into high-resolution two- and three-dimensional tomographic images of the subsurface (Figure 2). On the system currently under development, the images will appear on either a laptop computer or the operator's headset screen.

LANDMARC Innovations One of LANDMARC's chief contributions to land-mine detection technology is combining MIR units with a high-performance imaging system.3 LANDMARC's MIRbased imaging software, which was originally developed for radar inspection of steelreinforced concrete bridge decks, provides a great improvement over previous land-mine detection technology in sorting out clutter--the most difficult of the imaging tasks--and lowering the false-alarm rate. Central to perfecting LANDMARC's imaging capabilities are the comprehensive signal and noise models being developed by the Livermore team. These models are based on the contributions from temperature differences, inhomogeneity in the soil, increased noise resulting from multiple reflections in MIR arrays, surface reflections, and subsurface clutter such as rocks, roots, and voids. They identify terrain and soil conditions where radar is likely to work well and other situations where different types of sensors would be more appropriate. More important, the models are used to design algorithms to help reduce the false-alarm rate and increase the positive identification rate in laboratory and field tests, both of which, in turn, improve LANDMARC's ability to discriminate between mines and clutter.

Results from Field Testing Preliminary experiments identified the operational requirements of the prototype systems. The LANDMARC team developed the reconstruction algorithms that generate a three-dimensional image and is using them to investigate design trade-offs such as array size, sampling rate, and overall speed. In laboratory tests, the prototype clearly distinguished plastic antipersonnel mines from surrounding soils. In field tests at Fort Carson in Colorado and Fort A. P. Hill in Virginia, funded by the U.S. Defense Advanced Research Projects Agency (DARPA), the system performed well, though at a slow pace. The images it produced indicated that much progress has been made in removing the strong ground-surface reflection and other noise sources--that is, improving the signal-toclutter ratio. Field tests also indicated areas for additional refinement, among them using higher frequencies (that is, wider bandwidth) to improve resolution and better distinguish mines from clutter, and providing the system with a means of communicating a more accurate field position of the imaged mines. Future Plans When field tests with the prototypes are complete, the LANDMARC team plans to conduct blind tests at U.S. Army mine fields to measure detection probabilities under realistic conditions. In addition, plans to speed up the scan rate with advanced arrays are under way. Already experienced in industrial licensing of the MIR technology, the team will then direct LANDMARC toward external sponsorship for deployment in actual mine fields. The Department of Defense, U.S. industries, nongovernmental organizations such as Operation USA and the World Bank, and foreign governments have all shown interest in using Livermore's land-mine detection technology.

New Land Mine Detection Tool Called Major Breakthrough

Hand-held device combines metal detector with ground-penetrating radar

Washington -- The U.S. State Department announced August 29 what it calls a major breakthrough in land mine detection and removal technology. The U.S. Department of Defense has begun to field a hand-held device that combines metal detection with ground-penetrating radar, called a Handheld Standoff Mine Detection System (HSTAMIDS). What makes the new device so important, according to the announcement, is that HSTAMIDS can screen out the many bits and pieces of metal found in mined areas and on former battlefields that give a "false positive" signal to metal detectors. International mine action standards require land mine removal personnel to dig up every piece of metal found by their detectors to ensure that no land mine has been missed, the State Department says. This metallic clutter now can be ignored and not unearthed, saving time and, by discovering land mines faster, many innocent lives. The new dual system has undergone extensive testing and rigorous evaluation in the United States and elsewhere, the announcement said. The system was used in humanitarian land mine removal field evaluations and demonstrations in Southeast Asia and Africa in 2004-2005. Overall, the system was field tested with more than 10,000 mine targets and more than 50,000 pieces of metallic debris in widely varied environmental conditions in nine test arenas around the world. It was tested by five different humanitarian land mine removal teams from Afghanistan, Angola, Cambodia and Thailand before beginning humanitarian mine action operations, according to the announcement. Canada, the Netherlands, Sweden and the United Kingdom were among governments that joined with the United States in the testing effort. During this particular series of field tests, the newly trained dual-system operators significantly outperformed experienced metal detector operators. The metal detectors employed in these tests currently are used throughout the mine removal community. The Defense Department's Humanitarian Demining Research and Development Team has fielded 2,000 new detection systems with the U.S. military since the spring of 2006. By the year's end, that number should rise to 3,100 systems delivered to the field. Additionally, the manufacturer is supplying the new system to humanitarian mine removal operations in Cambodia, Afghanistan and Thailand, the announcement states.

REAL-TIME LANDMINE DETECTION USING HMM 4.1. Data preprocessing Preprocessing is an important step to enhance the mine signatures for detection. In general, preprocessing includes ground-level alignment and signal and noise background removal. In the proposed system, we assume that noise is much smaller than the background signal. In fact, the GPR hardware component performs waveform averaging to reduce background noise so that it is small enough to be ignored. The remaining background signals include the selfantenna coupling and ground reflection. Our preprocessing technique is designed to suppress this background signal. Our preprocessing follows the shift-and-scalemodel proposed by Brunzell [9], where the current background vector sample is assumed to be a shifted and scaled version of a background reference. Background removal then requires the estimation of both shift and scale parameters, which are changing from sample to sample. The original shift-andscale model assumes the shift is an integer. In practice, this assumption may not be accurate as the shifting can be in the order of subpixels, and interpolation is needed to perform subpixel shifting. Unfortunately, interpolation increases the computation significantly, and prohibits real-time processing. In this paper, we propose an efficient subpixel shift and scale preprocessing in frequency domain so that subpixel shifting can be realized in a simple manner to reduce computation. The subpixel shift and the scale factor are obtained using the maximum-likelihood (ML) approach. Experimentally, we have found that subpixel shifting reduces the number of false alarms by at least a factor of 1.5 when compared to Brunzells integer shifting. EXPERIMENTS 5.1. Experimental data GPR data collected from three different sites in the United States was used in our experiments. We will refer to these as collection 1, collection 2, and collection 3, respectively. Collection 1 contains mines and false alarms/background signatures extracted from data collected at site 1. These signatures were selected using a combination of ground truth and visual examination, and consist of 2 945 mines and 5 737 false alarms/background observation sequences. A more detailed description of this signature library and the extraction process can be found in [12]. This collection is used to learn the initial mine and background HMM model parameters using the basic Baum-Welch algorithm. Collections 2 and 3 contain GPR lane data collected from sites 2 and 3, respectively. Some of these lanes were used for corrective training, and the remaining for testing. Testing and adapting the model parameters on lanes is the most representative of real-world operational mode. Site 2 has 4 lanes labeled lane 4, 6, 8, and 9. Site 3 has 3 lanes labeled lane 51, 52, and 56. Each lane is 3m wide, and its length

varies from 50 to 200 m. The ground truth, that is, location of the mines, of these calibration lanes is available, and we use it only during the corrective training or in the evaluation of the algorithm, and not in the HMM testing algorithm itself. Multiple data files were collected from each lane. Each file, referred to as one pass, corresponds to running a vehicle-mounted GPR system over the lane. A total of 83 passes were collected from site 2, and only one of these passes (from lane 4) was used for corrective training to adapt the HMMs to this site. Collection 3 includes 18 passes, and one of the passes (from lane 51) was used for training. Table 1 summarizes the data collections used in our experiments. We should note here that although the lanes used in our experiments do not have explicitly emplaced clutter, they were set up to represent real on-road conditions that a vehicle-mounted mine detection system will encounter. Thus, they do include clutter objects such as void, rocks, and probably some metal pieces. 5.2. Basic HMMtraining The signature library (i.e., collection 1) was used to train the basic HMMmodels. Using this collection, BWalgorithm was used to estimate the HMM parameters = (m, b), where m is estimated from the mine signatures and b is estimated from the false alarms/background signatures.We will refer to these as the baseline models, or simply base. 5.3. CorrectiveHMMtraining using the signature library Using base as initial model parameters, we applied the corrective training procedure to the signature library and adjusted the parameters to reduce the number of misclassified signatures. A total of 10 iterations were needed for the parameters to stabilize.We will refer to these models as Sig. 5.4. CorrectiveHMMtraining using lane data The rational of the proposed training approach is to adapt the model parameters to different geographical sites that have different characteristics due to variations in soil type and other conditions. Thus, we have selected one pass from collection 2 and used it (along with its ground truth) to adapt the baseline HMMsto this site. The training pass was selected from lane 8 as this lane contains more mines with more variety. In all experiments reported in this paper, a total of 20 corrective training iterations were performed, and intermediate model parameters were saved after each iteration. We will refer to these models as S2-L8(k) (i.e., models adapted to site 2 using lane 8 after k iterations). Similarly, we have adapted the baseline HMMs to site 3 by applying the corrective training using one pass from this collection (from lane 51). For this site, we obtained the intermediate models S3-L51(k). 5.5. Experimental results In this section, we present, analyze, and compare the results obtained using the corrective training procedure with those obtained using the baseline model. Our results could not be compared with the basic system introduced in [12], as these

systems use different GPR prototypes, and thus, have preprocessed the data differently. However, we should point out that the two systems are very similar. The major differences are the additional pre- and postprocessing steps outlined in Section 4.3. These steps do not affect the detection results significantly. They are added mainly to reduce the execution time, and thus, enable real-time implementation and testing. This is because (1) the sequences are no longer required to have a constant length of 15 (maximum length); (2) only 3 not 5 states are needed for the mine model; and (3) several sequences are rejected without testing them with the background models using the Viterbi algorithm. The performance of the different model parameters was scored in terms of probability of detection (PD) versus false alarm rate (FAR). Confidence values were thresholded at different levels to produce receiver operating characteristics (ROC) curves. For a given threshold, a mine is detected if there is an alarm within 0.25 meters from the edge of the mine with confidence value above the threshold. Given a threshold, the PD for a lane or a set of lanes is defined to be the number of mines detected divided by the total number of mines. The FAR is defined as the number of false alarms per square meter. Figure 9a shows the overall ROCs for all of the 83 passes of collection 2. The ROCs are displayed for the confidence values generated using the baseline model, base (no corrective training), the Sig model (corrective training using signature library), and the S2-L8(k) models (k iterations of corrective training using one pass from collection 2). As can be seen, when compared to the base ROC, the Sig and S2-L8(k) ROCs are shifted left (i.e., lower FAR for the same PD) and shifted up (higher PD for the same FAR). Thus, one can conclude that corrective training improves the overall detection rate. To quantify the improvement in the detection rate, we compute the area under the ROC in the [0, 0.02] FAR range.1 For the rest of this paper, we will use the term performance to refer to this area. A perfect detector (detects all the mines with no false alarms) would have an area of 0.02. Figure 9b displays the values of these areas for the ROCs of the different models. As can be seen, the area jumps from 0.013 (for the baseline parameters) to 0.0134 for the Sig parameters, and to 0.014 S2-L8(20), achieving an overall performance gain of about 10%. As can be seen, the performance curve starts to flatten after 15 iterations. Performing the corrective training for more than 20 iterations does not improve the performance any further. Figure 9 displays the results averaged over 83 passes. The performance (measured by the area under the ROC) for the individual passes varies significantly. For instance, for some passes, the performance remains constant. These passes are usually either clean, with all mines easily detectable, or very noisy with several weak mine signatures. Thus, the performance is either close to optimal, and cannot be improved any further, or poor and cannot be improved,

as many of the mines could not be detected using the current preprocessing and feature representation techniques. For most passes, the performance improves. Figures 10 and 11 display the results for two different passes over lanes 4 and 6, respectively, where the performance improves by about 50%. For Figure 10, the improvement occurs after the initial corrective training with the signature library, and more training with the lane data does not affect the results. On the other hand, for Figure 11, no significant improvement was obtained after the corrective training with the signature library. The largest performance gain was obtained using S2-L8(10).