Detection of Shallowly Buried Landmines from Ground Penetrating Radar Signals

A method for detecting shallowly buried landmines using sequential GPR data is presented. After removing dominant coherent component of ground surface reflection from GPR data, three kinds of target features related to wave correlation, energy ratio, and signal arrival times are extracted. Since the detection problem treated here is reduced to a binary hypothesis test, an approach based on a likelihood ratio test is employed as a detection algorithm. In order to check the detection performance, the Monte Carlo simulation is carried out for data generated by the two-dimensional finite difference time domain (FDTD) method. The results given in the form of ROC curves show that good performance is obtained in spite of simple three-dimensional feature vector. INTRODUCTION Detection of small and shallowly buried landmines is a very challenging problem. As compared with a metal detector that is widely used for landmine detection, ground penetrating radar (GPR) system seems to offer the promise of this problem, especially for detection of plastic landmines with little or no metal content [1]. However, the GPR performs inadequately due to the ground clutter because returns from the shallowly buried landmines and that from ground surface overlap in time. Furthermore, the GPR also receives returns from other subsurface objects such as rocks, tree roots, or metal fragments in the ground, which leads to high levels of false alarms [2][3]. In this paper, we present a method for detecting shallowly buried landmines using sequential GPR data. First, we remove the strong coherent component of ground surface reflection from sequential GPR data using correlation between the GPR signal and deformed incident pulse. After the removal, we extract three kinds of target features related to wave correlation, energy ratio, and signal arrival time from the residual signals. Since this detection problem is reduced to a binary hypothesis test (H1 : landmine, H0 : no landmine), we employ here a classical detection theory based on likelihood ratio test [4]. In order to check the detection performance, the Monte Carlo simulation is carried out for data generated by the two-dimensional finite difference time domain (2D-FDTD) method [5]. The results are shown in the form of receiver operating characteristics (ROC) curves [4], which quantify the probability of detection as a function of the false alarm rate (FAR). The results show that good performance is obtained in spite of simple three-dimensional feature vector. SIGNAL PROCESSING AND FEATURE EXTRACTION Figure 1 shows a typical configuration of the GPR measurement system for detecting shallowly buried landmines. The GPR measurements are made at multiple observation points above the rough ground surface using transmitting and receiving antenna system. The transmitting antenna sends out a short duration pulse and the receiving antenna samples the returned signal that includes target response together with reflection from the rough ground surface. Because the ground surface reflection is very strong compared to the response from plastic landmines, a pre-processing step of ground clutter removal from the GPR signals is required. However, complete removal of the ground clutter from the GPR signals is impossible. Therefore, we simply reduce the ground clutter contribution by subtracting dominant coherent component of ground surface reflection. Since the coherent component is the reflection from a flat ground surface without any buried target under it, we can approximately express it in terms of the Fresnel reflection coefficients R( ) and incident pulse ) (t q . Taking into account of this fact, we decompose the GPR signal ) (t pm measured at the observation point ) , , 2 , 1 ( M m m into a dominant coherent term and a residual term as follows: M m q p R t r t q R t p a m a m m a m m , , 2 , 1 , max ), ( ) ( ) ( , , , , (1) where a t q t q a ) ( ) ( , is the scaled and shifted incident pulse, m R is a constant that corresponds to the reflection coefficient of the flat ground surface, and ) (t rm is the residual term that includes reflection from the target and incoherent component of the reflection from the rough ground surface (and also additive noise). Since the coefficient m R can be estimated by maximizing the inner product a m q p , , with respect to scale and shift parameters ( , a), the residual term ) (t rm is obtained by subtracting ) ( , t q R a m from the signal ) (t pm . Note that, in Eq.(1), the coefficient m R is assumed to be a constant because the surface reflection is