Deep Learning–Based Denoising Improves Receiver Function Imaging Using Dense Short-Period Teleseismic Data

Abstract

Abstract Receiver function (RF) imaging using seismic data from dense short-period arrays has gained increasing importance in recent years in investigating fine-scale structures of the crust and uppermost mantle. A crucial step in such studies is to remove the instrument response (IR) to enhance teleseismic signals at ∼0.01 to 5 Hz, thereby simulating broadband records. However, this procedure also amplifies noise within the same frequency band. For weak signals, distinguishing them from noise is often challenging and in some cases is even impossible with traditional denoising methods such as filtering. To address this challenge, we develop a new convolutional neural network model, NodalWaden, using decades of high-quality global broadband teleseismic body waves for training. The broadband data exhibit the characteristics we target to achieve by removing the IR from the short-period records. The applicability of NodalWaden is justified by denoising the three-component short-period records of more than 18 months from 155 nodes deployed in northern Sumatra. We find that NodalWaden substantially improves the signal-to-noise ratio (SNR), upgrading ∼50% of the teleseismic data from the “very-low-SNR” (∼1) to “very-high-SNR” (>10) categories. RFs calculated from the denoised dataset show better separation of merged phases and noticeable enhancement of weak signals, resulting in improvement in the quality of structure imaging. In particular, a positive phase is consistently detected at ~2 s throughout the dataset and interpreted as the Conrad discontinuity, which is unresolvable in the original RFs. This denoising technique would be particularly useful for short-duration (e.g., one month) deployment with limited teleseismic data, both from the past and in the future.