Numerical modelling and neural networks for landmine detection using ground penetrating radar

Abstract

Abstract-A numerical modelling case study is presented aiming to investigate aspects of the applicability of artificial neural networks (ANN) to the problem of landmine detection using ground penetrating radar (GPR). An essential requirement of ANN and machine learning in general, is an extensive training set. A good training set should include data from as many scenarios as possible. Therefore, a training set consisting of simulated data from a diverse range of models with varying: topography, soil inhomogeneity, landmines, false alarm targets, height of the antenna, depth of the landmines, has been produced and used. Previous approaches, have employed limited training sets and as a result they often have underestimated the capabilities of ANN. In this preliminary study, a 2D Finite-Difference TimeDomain (FDTD) model has been used as the training platform for ANN. Although a 2D approach is clearly a simplification that cannot directly translate to a practical application, it is a computationally efficient approach to examine the performance of ANN subject to an extensive training set. The results are promising and provide a good basis to further expand this approach in the future by employing even more realistic, but computationally expensive, 3D models and well-characterised, real data sets.

A Machine Learning-Based Fast-Forward Solver for Ground Penetrating Radar With Application to Full-Waveform Inversion

“…A training set should include a wide spectrum of possible scenarios in order to adequately map the feature space of a given problem [5]. To achieve this we used numerical modeling to obtain synthetic, but nonetheless realistic and most importantly well-labeled data [41], [42]. Our forward solver used to generate the training data is gprMax [1], which is an open source electromagnetic simulation software based on the FDTD method [43].…”

Health Monitoring of Tree Trunks Using Ground Penetrating Radar

“…In this paper two approaches have been chosen based on their performance and computational simplicity. The first one is the singular value decomposition (SVD) filter [36], [37] and the second one is the linear combination filter [38]. The image is then manually thresholded [45], [29] and a particle-swarm optimization (PSO) is used to fit the resulting anomaly, which can no longer be approximated with a hyperbola.…”

“…The processing framework proposed in this paper consists of two parts. Initially an SVD [36] and a linear filter [38] are used in order to reduce the ringing noise that is present to the measured B-Scans due to the layered nature of the tree-trunk [35]. The post-processed data are then manually thresholded and a PSO is used in order to estimate the origins and the size of the targets based on their reflection patterns.…”

Realistic FDTD GPR Antenna Models Optimized Using a Novel Linear/Nonlinear Full-Waveform Inversion

“…Models using simplified excitation sources, such as Hertzian dipoles, especially for high frequency applications, can produce significantly different responses from real measurements [10]. Therefore, these models cannot be easily employed either as a forward solver for inversion purposes, or for generating synthetic training sets for machine learning applications [13].…”