Multi-scaled power spectrum based features for landmine detection using Ground Penetrating Radar

Abstract

Landmines are a humanitarian challenge because they do not discriminate between soldiers and civilians. One potential method for detecting unexploded landmines in post-conflict zones is the use of ground penetrating radar (GPR). However, detecting shallowly-buried landmines under a variety of soil and surface conditions is a challenging task in terms of signal processing. The present paper proposes a new approach based on multi-scale power spectrum features. The power spectrum of GPR response is processed using multi-scale triangular filter banks in order to produce a set of feature coefficients, which are then ranked by their significance with respect to the classification using the feature selection algorithm ReliefF. Using these ranked features, a support vector machine based classifier is used to evaluate the feasibility of landmine detection using this approach. The soil conditions considered are homogeneous (dry), homogeneous (moist) and heterogeneous, with different surface roughnesses. The results indicate that the proposed method can successfully detect landmines in a variety of environments using a relatively small number of spectral features. In addition, the GPR frequency bands that are most effective for landmine detection tasks are identified.