Online dictionary learning method for extracting GPR diffractions

Abstract

Diffractions in a Ground-Penetrating Radar (GPR) data carry significant responses from near-surface small-scale fractures or karsts. However, this geological information is generally difficult to extract because of the shielding effect of strong reflections from subsurface layers. In order to solve this problem, a GPR diffraction extraction method is proposed for individually separating and imaging of GPR diffractions that incorporates a local plane-wave destruction filter with an online dictionary learning algorithm. The strong reflections are estimated and eliminated by the local plane-wave destruction method and the weak GPR diffractions are extracted by a sparse coding algorithm. In solving this model, a trust-region algorithm is used for accelerating the sparse coding procedures that can scale up gracefully to a large GPR data processing. A numerical experiment demonstrates the good performance of the proposed method in destroying strong reflections and enhancing weak diffractions from small-scale void holes. Real data application further verifies its potential value in resolving fine details of subsurface small-scale buried targets, such as pipes or void holes.