Synchrosqueezing-Based Shallow Neural Network for Near-Surface Void Recognition by Ground Penetrating Radar

In this letter, wavelet synchrosqueezing transformation (WSST)-based technique is introduced for ground penetrating radar (GPR) near-surface void recognition. WSST can extract the signal features effectively because the ground surface reflection or antenna coupling suppresses the target echo. However, these features are distorted by mathematical procedures and clutter. A shallow neural network (SNN) is applied to estimate the possible targets in this case. The SNN only has full connection layers, and its training sets are simulated data under different levels of SNR. After the network is trained, the result scalar can be used to identify the near-surface targets. Simulation and experimental results show that the proposed method achieves 1/10 range resolution binary classification. Compared with recent studies, the detectable thickness improves (at least two times smaller). Besides, the method has a limited-size network and can be trained in situ, which makes it more suitable for engineering uses.