Denoising contaminated seismic signals for later processing is a fundamental problem in seismic signals analysis. Neural network approaches have shown success denoising local signals when trained on short-time Fourier transform spectrograms. One challenge of this approach is the onerous process of hand-labeling event signals for training. By leveraging the SCALODEEP seismic event detector, we develop an automated set of techniques for labeling event data. Despite region specific challenges, training the neural network denoiser on machine curated events shows comparable performance to the neural network trained on hand curated events. We showcase our technique with two experiments, one using Utah regional data and one using regional data from the Korean peninsula.
Deep neural networks (NNs) typically outperform traditional machine learning (ML) approaches for complicated, non-linear tasks. It is expected that deep learning (DL) should offer superior performance for the important non-proliferation task of predicting explosive device configuration based upon observed optical signature, a task which human experts struggle with. However, supervised machine learning is difficult to apply in this mission space because most recorded signatures are not associated with the corresponding device description, or “truth labels.” This is challenging for NNs, which traditionally require many samples for strong performance. Semi-supervised learning (SSL), low-shot learning (LSL), and uncertainty quantification (UQ) for NNs are emerging approaches that could bridge the mission gaps of few labels and rare samples of importance. NN explainability techniques are important in gaining insight into the inferential feature importance of such a complex model. In this work, SSL, LSL, and UQ are merged into a single framework, a significant technical hurdle not previously demonstrated. Exponential Average Adversarial Training (EAAT) and Pairwise Neural Networks (PNNs) are chosen as the SSL and LSL methods of choice. Permutation feature importance (PFI) for functional data is used to provide explainability via the Variable importance Explainable Elastic Shape Analysis (VEESA) pipeline. A variety of uncertainty quantification approaches are explored: Bayesian Neural Networks (BNNs), ensemble methods, concrete dropout, and evidential deep learning. Two final approaches, one utilizing ensemble methods and one utilizing evidential learning, are constructed and compared using a well-quantified synthetic 2D dataset along with the DIRSIG Megascene.
Denoising contaminated seismic signals for later processing is a fundamental problem in seismic signals analysis. The most straightforward denoising approach, using spectral filtering, is not effective when noise and seismic signal occupy the same frequency range. Neural network approaches have shown success denoising local signal when trained on short-time Fourier transform spectrograms (Zhu et al 2018; Tibi et al 2021). Scalograms, a wavelet-based transform, achieved ~15% better reconstruction as measured by dynamic time warping on a seismic waveform test set than spectrograms, suggesting their use as an alternative for denoising. We train a deep neural network on a scalogram dataset derived from waveforms recorded by the University of Utah Seismograph Stations network. We find that initial results are no better than a spectrogram approach, with additional overhead imposed by the significantly larger size of scalograms. A robust exploration of neural network hyperparameters and network architecture was not performed, which could be done in follow on work.
With the continuing development of more capable data gathering sensors, comes an increased demand on the bandwidth for transmitting larger quantities of data. To help counteract that trend, a study was undertaken to determine appropriate lossy data compression strategies for minimizing their impact on target detection and characterization. The survey of current compression techniques led us to the conclusion that wavelet compression was well suited for this purpose. Wavelet analysis essentially applies a low-pass and high-pass filter to the data, converting the data into the related coefficients that maintain spatial information as well as frequency information. Wavelet compression is achieved by zeroing the coefficients that pertain to the noise in the signal, i.e. the high frequency, low amplitude portion. This approach is well suited for our goal because it reduces the noise in the signal with only minimal impact on the larger, lower frequency target signatures. The resulting coefficients can then be encoded using lossless techniques with higher compression levels because of the lower entropy and significant number of zeros. No significant signal degradation or difficulties in target characterization or detection were observed or measured when wavelet compression was applied to simulated and real data, even when over 80% of the coefficients were zeroed. While the exact level of compression will be data set dependent, for the data sets we studied, compression factors over 10 were found to be satisfactory where conventional lossless techniques achieved levels of less than 3.