Landmine detection improvement using one-class SVM for unbalanced data

Abstract

Ground Penetrating Radar (GPR) has been a precious tool for humanitarian demining. The GPR scans the ground and delivers a three-dimensional matrix representing three types of data: Ascan, Bscan and Cscan. The Ascan data represents the response from a reflection signal of a pulse emitted by the GPR at a given position. In the proposed landmine detection method, the Ascan data is normalized and then classified using Kernel based One Class Support Vector Machine (OSVM). In fact, OSVM has the main advantage of handling unbalanced data, where is not the case for multiclass SVM. Our landmine detection method was tested and evaluated on the MACADAM database which is composed of 11 scenarios of landmines and 3 scenarios of inoffensive objects (wood stick, SodaCan, pine, stone). Experimental results have shown the superiority of the RBF kernel OSVM over others kernel functions based multiclass SVM in term of classification accuracy especially, as landmine data is unbalanced.