The United Nations has estimated that about 100 million mines lie buried around the world, claiming 10,000 deaths annually and at least twice as many seriously injured. Many victims are small children and elderly villagers in poor nations. Finding buried landmines is a difficult task, because landmines have different variations in size, shape, composition and depth. Figure 1 shows some typical landmines. Metal detectors (MDs) are the most widely used device for landmine detection. MDs have great difficulty detecting landmines that are made of plastic or have low metal content. Ground Penetrating Radar (GPR) is an emerging technique for landmine detection that can detect plastic and/or low metal content landmines.

Prototype GPR systems are usually vehicle mounted or hand held. Vehicle mounted systems are aimed at detecting anti-tank mines rapidly while the vehicle is driving down the road. Hand-held systems are designed to operate in a variety of terrain to detect both anti-tank and anti-personnel mines. However, GPR return signals are very susceptible to environmental conditions including soil type, soil moisture, and ground bounce interaction. As a result, novel and adaptive signal processing techniques are indispensable to reduce false alarm rates and improve overall detection.

The Center for Geospatial Intelligence (CGI) at the University of Missouri-Columbia (UMC) is one of only a few academic institutions funded by the US Army to support activities on subsurface ordinance and landmine detection research. CGI researchers collaborate with other institutions including the University of Florida (UF) and Duke University in a variety of landmine detection research. Below we summarize the GPR techniques for mine detection in both hand-held and vehicle based systems developed at the UMC. Fusion of GPR and MD sensor data is also described.

Vehicle Mounted GPR

Vehicle mounted systems have a GPR attached to the front of a vehicle that transmits and receives the radar signal at equally spaced positions as the vehicle moves. Figure 2 shows a prototype vehicle mounted GPR mine detection system. The vehicle speed and the height of GPR above the ground surface are relatively constant. Landmines in this case often produce stable signatures, such as hyperbolic shapes, for detection. The challenge in a vehicle-based system is pre-processing to align ground bounce and reduce the background so that features can be extracted from the hyperbolic signature for mine detection.

UMC and UF have derived an efficient technique to align the ground surface and perform ground response removal through linear prediction pre-processing (LPP) to enhance the hyperbolic signature for mine detection. Figure 3(a) shows a section of B-scan (depth vs. distance traveled) GPR data that contains a difficult to detect plastic mine buried at 2 inches deep. Figure 3(b) shows the results after ground alignment and LPP processing. The hyperbolic mine signature appears very well after LPP. Figures 3(c) and 3(d) are the raw data and LPP results for another plastic mine buried at 4 inches deep, and Figures 3(e) and 3(f) are for the case of a metal mine. It is clear that the LPP is very effective and enhances the mine signature for detection.

Hand-Held GPR

A second type of mine detection system uses the GPR in a hand held unit. The GPR sensor is attached at the tip of the unit. An operator carries the unit and sweeps it across the ground, as is illustrated in Figure 4. Because of human factors, the sweeping speed and the sensor-to-ground distance are time varying. The position and velocity of the sensor are not measured and are not available to any detection algorithm. Hence, the GPR landmine signatures are inconsistent and the hyperbolic signatures as observed in the vehicle based system are no longer present. All these factors make landmine detection using a hand held GPR is a very difficult task. Furthermore, the detection needs to be causal and no future look ahead is allowed.

UMC and UF have developed a Correlation Detector (CorrDet) algorithm for hand-held GPR mine detection. The CorrDet algorithm uses statistical signal processing techniques and models the background response using a time-varying linear-prediction model. Any deviation from the background indicates a potential mine target. Sub-band processing and depth processing are also used to improve further the detection performance. Various versions of the CorrDet algorithm have been adopted in some hand-held mine detection units and are being operationally used in Afghanistan and Iraq by U.S. Army units.

Figure 5(a) shows the raw background data. It contains 4 back and forth sweeps across the ground. The data is measured in the frequency domain and the return signal strength at different frequencies is shown as colors of different intensity. Figure 5(b) shows the raw data containing 4 sweeps over the same mine target as indicated by the marker in the upper sub-window. Due to the large ground response, it is impossible to determine if there are any mine signals present in Figure 5(b). Figures 5(c) and 5(d) give the results after the CorrDet algorithm to remove the background, and the mine signals appear quite well in Figure 5(d). Figures 5(e) and 5(f) show the final detection values from the CorrDet algorithm in which the red-lines indicate the adaptive threshold. It is clear that the CorrDet algorithm is able to detect the mines very well and suppress false alarms from variations due to background.

Figure 6 gives the Receiver Operating Characteristics (ROC) curve of the CorrDet algorithm generated over 2300 mine targets and an equivalent amount of background data. Compared to the baseline technique that used only differential energy detection, the CorrDet algorithm reduces the false alarm rate (FAR) by 50% at a 90% Probability of Detection (Pd).

GPR and MD Fusion

GPR is able to detect low-metal content and plastic mines where a Metal Detectors (MD) fails. On the other hand, GPR responds to many clutter objects such as rocks or pieces of wood that do not have any metal content. Typically, GPR has a higher number of false alarms near the ground surface, and MD detects a mine better if it is clear to the ground. UMC and UF have developed sensor fusion techniques that combine the advantages of the two sensors to improve performance.

Figure 7 shows the ROC curve of the fusion results, and compares it with the detection results from the best human expert operating the hand-held mine detection unit. The MD results for fusion are from Duke University. Note that the fusion algorithm is automatic and does not involve any human expert decision. The figure clearly demonstrates the significance of our research results since we are able achieve comparable results with the best human operator. Also it is important to note that in practice, human operator results will be worse since it will not be used by well-trained experts, but instead by lesser-trained soldiers who must operate the hand-held units in the field. Thus, the automatic target detection algorithm provides consistent results that are much less reliant on expert human operation and decisions.