Applying Feature Transformation-Based Domain Confusion to Neural Network for the Denoising of Dispersion Spectrograms

Abstract Ambient noise tomography has been widely used to estimate the shear-wave velocity structure of the Earth. A key step in this method is to pick dispersions from dispersion spectrograms. Using the frequency–Bessel (F-J) transform, the generated spectrograms can provide more dispersion information by including higher modes in addition to the fundamental mode. With the increasing availability of these spectrograms, manually picking dispersion curves is highly time and energy consuming. Consequently, neural networks have been used for automatically picking dispersions. Dispersion curves are picked based on deep learning mainly for denoising these spectrograms. In several studies, the neural network was solely trained, and its performance was verified for the denoising. However, they all learn single-source data in the training of neural network. It will lead the regionality of trained neural network. Even if we can use domain adaptation to improve its performance and achieve some success, there are still some spectrograms that cannot be solved effectively. Therefore, multisources training is useful and could reduce the regionality in training stage. Normally, dispersion spectrograms from multisources have feature differences of dispersion curves, especially for higher modes in F-J spectrograms. Thus, we propose a training strategy based on domain confusion through which the neural network effectively learns spectrograms from multisources. After domain confusion, the trained neural network can effectively process large number of test data and help us easily obtain more dispersion curves automatically. The proposed study can provide a deep insight into the denoising of dispersion spectrograms by neural network and facilitate ambient noise tomography.