VAE-ResNet Cascade Network: An Advanced Algorithm for Stochastic Clutter Suppression in Ground Penetrating Radar Data
Xiangyu Wang, Hai Liu
Abstract
Ground Penetrating Radar (GPR) is a subsurface sensing technology extensively utilized in various applications. However, the inhomogeneous distribution of the subsurface medium results in a stochastic clutter and drastically degrades the precision of subsurface anomalies imaging. To suppress the stochastic clutter, we introduce a deep learning-based cascade network, termed the Variational Autoencoder (VAE)-Residual Network (ResNet). It employs a unique architecture that integrates the VAE and ResNet, and is further enhanced by a Residual Feature Distillation Block (RFDB) module. This configuration capitalizes on the RFDB module’s sophisticated feature extraction capabilities, thereby enhancing its proficiency in suppressing stochastic clutter. Consequently, it facilitates an end-to-end processing of stochastic clutter in GPR recordings. Finally, Reverse Time Migration (RTM) is utilized to image the processed GPR data, thereby emonstrating the enhanced accuracy in subsurface anomaly imaging and detection. Comparative analyses with established algorithms, including Non-local Means (NLM) and Block-matching and 3D Filtering (BM3D) are conducted through numerical, laboratory and field experiments. The results have not only demonstrated the proposed algorithm’s effectiveness but also established its superiority over traditional methods. The field experimental results validate the practicality and wide applicability of the proposed algorithm.