Deep-learning-based sferics recognition for audio magnetotelluric data processing in the dead band

Abstract

In audio magnetotelluric (AMT) sounding data processing, the absence of sferic signals in some time ranges results in a lack of energy in the AMT dead band, causing unreliable resistivity estimations. To address this issue, we develop a deep convolutional neural network (CNN) to automatically recognize sferic signals from redundantly recorded data over a long time range and use these sferic signals to accurately estimate resistivity. The CNN is trained using field time-series data with different signal-to-noise ratios (S/Ns) acquired from different regions of mainland China. To solve the potential overfitting due to the limited number of sferic labels, we develop a training strategy that randomly generates training samples with random data augmentations while optimizing the CNN model parameters. The training process and data generation are stopped when the training loss converges. In addition, we use a weighted binary cross-entropy loss function to solve the sample imbalance problem to optimize the network better, use multiple reasonable metrics to evaluate the network performance, and perform ablation experiments to optimize the model hyperparameters. Extensive field data applications indicate that our trained CNN can robustly recognize sferic signals from noisy time series for subsequent impedance estimation. The results indicate that our method can significantly improve the S/Ns and effectively solve the lack of energy in the dead band. Compared with the traditional processing method, our method can generate smoother and more reasonable apparent resistivity-phase curves and depolarized phase tensors, correct the estimation error of the sudden drop in high-frequency apparent resistivity and abnormal behavior of phase reversal, and better estimate the real shallow resistivity structure.