This study presents a deep learning framework, based on wave responses and a three-dimensional convolutional neural network (3D-CNN), for detecting subsurface cavities in a 3D homogeneous semi-infinite domain. In the data generation process, the displacement fields are obtained using simulation software, with a square grid of surface sensors capturing wave responses from various cavity configurations. We present the overall architecture of the 3D-CNN and a proximity function–based method to tackle the class imbalance, known to degrade the performance of CNN, raised in the grid-wise classification problem because cavity points are vastly outnumbered by no-cavity points. The CNN results in, as an output feature, a 3D probability map indicating whether each grid point is a void or not. To investigate the robustness and sensitivity, five different 3D-CNN architectures are investigated in terms of varying sensor density, temporal resolution, and the presence of noise in the input measurement data. Blind tests on previously unseen targets examined the performance of the proposed CNN models to localize cavity structures with high fidelity. Furthermore, our method is examined using displacement-field wave measurement from a third-party simulator. Numerical results indicated that dense spatial and temporal resolution in input measurement is critical for optimal performance. The inclusion of noise in the measurement in the training data enables better performance. Overall, this work highlights the potential of 3D-CNNs as a powerful tool for nondestructive subsurface anomaly detection, offering accuracy, adaptability, and robustness for applications in geophysics, underground monitoring, and nondestructive testing (NDT) purposes.
Tajmiri et al. (Fri,) studied this question.