Implicit Neural Representations for Unsupervised Seismic Data Processing.

Ji Li, Daniel O. Trad

Implicit Neural Representation (INR) has recently gained prominence across various domains within computer vision and signal processing. INR involves learning a continuous function over a set of points, representing a significant advancement in seismic signal analysis when contrasted with conventional methods. Notably, the network’s intrinsic low-frequency bias encourages the acquisition of self-similar features ahead of high-frequency,incoherent ones during training. This unique characteristic can effectively mitigate coherent and incoherent noise in seismic data. Furthermore, the capacity to represent signals continuously, unlike discrete forms, facilitates the interpolation of seismic signals on irregular grids. In this paper, we present an approach utilizing the Siren (a specific INR architecture with a sine activation function) to address various seismic processing challenges,including denoising for both coherent (e.g., ground roll) and incoherent (random and erratic) noise as well as seismic data interpolation. Notably, our method stands out because it obviates the need for paired training datasets, rendering it a zero-shot unsupervised learning approach. Evaluation results demonstrate its superior performance compared too ther state-of-the-art unsupervised deep learning techniques and traditional methods.