Deep Learning in Seismic Tomography
Seismic Tomography algorithms are designed to recover Earth’s interior from observations of seismic waves. Where seismic wave is a type of shock wave that appears in earthquakes. These algorithms are useful for detecting gas and oil, as well as anomalies, like underground tunnels and air pockets.
The recorded seismic waves travel through the ground in unknown path-patterns, which result in unique attenuation and delay effects. Each material in the ground, corresponds to a unique set of propagation velocity values of seismic waves. Based on the arrival times of all seismic waves recorded, each seismic tomography algorithm attempts to estimate the propagation velocity profile of the medium.
The recent breakthrough in seismic tomography research, comes from deep-learning based algorithms. Using large data-set of instances, researchers have trained deep neural networks to preform seismic tomography imaging. The received images are more accurate than before.
In our work, we attempt to improve the accuracy even further, by designing more suited network architectures for seismic tomography task. We discuss limitations of current deep neural network models, and propose measures to negate them.
The first proposition is to replace traditional convolution kernel with a transformer kernel (multi-head self-attention). From analysis of the seismic measurements’ characteristics, we concluded that the new kernel is preferable in the case of seismic tomography.
The second proposition is incorporating multi-resolution hierarchical structure into the model. We show that with this proposed structure, we are able to overcome problems assigned with large-scale deep neural networks, and achieve superior results.
We evaluate our proposed network architectures and proposed implementations in two cases. The first case is the case of undersea gas detection. And the second case is the case of underground tunnels detection. In both cases, we show marked improvements for our proposed architectures with respect to reconstruction error metrics, and with comparison to alternative popularly used architectures.
* M.Sc. student under the supervision of Professor Arie Feuer.