Encoding spatio-temporal coordinates in convolutional neural networks for seismic data processing

Seismic data processing plays a vital role in transforming raw seismic recordings into interpretable images of the subsurface. While Convolutional Neural Networks (CNNs) have demonstrated potential in treating seismic processing as an image-to-image translation task, their conventional implementations, originally tailored for natural image analysis, often overlook the inherent spatio-temporal characteristics of seismic data and consider only seismic amplitudes. This limitation restricts their effectiveness in fully capturing seismic signal behavior. In this work, we propose a novel methodology that explicitly integrates spatio-temporal information into CNN-based processing frameworks. Specifically, we focus on the tasks of multiple attenuation and residual moveout correction. We evaluate the impact of incorporating spatio-temporal coordinates into CNN architectures using a field data example, demonstrating improved performance in seismic demultiple and alignment tasks. The proposed approach has the potential to improve the applicability of CNNs to a broader range of seismic processing tasks by treating seismic data as signals rather than considering purely visual patterns extracted only from amplitudes.