Deep learning for quality control of receiver functions

C. Gong, Ling Chen, Zhuowei Xiao, Xu WangFrontiers in Earth Science 2022 · DOI

TL;DR

This paper introduces a deep learning approach for automatically quality controlling seismic receiver functions, a crucial step in understanding Earth's subsurface structure. By employing a CNN-BiLSTM architecture, the method achieves human-level accuracy and significantly outperforms traditional automated methods, enabling the use of data from smaller earthquakes that were previously discarded. A key innovation is a training schema that requires no manual labels, making the process fully automated.

What Problem Does This Paper Address?

Receiver functions are powerful tools for imaging discontinuities in the Earth's crust and upper mantle, such as the Moho. However, the quality of these functions directly impacts the accuracy of structural models. Manually quality controlling receiver functions is accurate but extremely time-consuming, especially with the ever-increasing volume of seismic data. Existing automated methods often struggle with noisy data and require extensive manual tuning of hyperparameters, limiting their reliability and generalizability. This paper aims to bridge this gap by developing a deep learning model that can automatically and accurately perform quality control, even on noisy data and from smaller seismic events.

Methods

The researchers explored four deep learning architectures:

  • Fully Connected (FC) Model: A baseline model using only fully connected layers.
  • Convolutional Neural Network (CNN) Model: Utilizes convolutional layers for feature extraction, followed by fully connected layers for classification. This leverages the success of CNNs in image and time-series processing.
  • CNN-BiLSTM Model: Builds upon the CNN model by adding a Bidirectional Long Short-Term Memory (BiLSTM) layer. BiLSTMs are adept at capturing temporal dependencies, making them suitable for sequential data like seismic waveforms.
  • CNN-BiLSTM-Trans Model: Further enhances the CNN-BiLSTM by incorporating a Transformer layer, known for its attention mechanisms.

The models were trained on receiver functions windowed from 50 seconds before to 50 seconds after the P-wave arrival. The output was a binary classification: "good" or "bad" receiver function. To address data imbalance, a balanced sampling strategy was employed. Early stopping was used to mitigate overfitting.

A significant contribution is the proposal of a fully automatic training schema. Instead of relying on meticulously hand-labeled data, this approach uses receiver functions labeled by a traditional automatic method (HiPerSeis) as training data. The authors hypothesize that deep learning's inherent robustness to noise allows it to learn effectively even from these less-than-perfect labels.

Datasets

The study utilized seismic data from three broadband seismic stations: AK.SAW, IC.QIZ, and CB.CAD.

  • Training and Validation Data: 12,487 receiver functions from earthquakes with magnitudes greater than 5.5 were used to compare model designs. These were manually labeled as "good" or "bad" based on waveform similarity across back-azimuths.
  • Testing Data: An additional 40,806 receiver functions from earthquakes with magnitudes between 5.0 and 5.5 were used to evaluate the model's performance on smaller events.

The paper also mentions using data from the IRIS (Incorporated Research Institutions for Seismology) earthquake catalog.

Key Insights and Ideas

  • CNN-BiLSTM Dominance: The combination of convolutional layers for feature extraction and BiLSTM layers for temporal context proved to be the most effective architecture for receiver function quality control, outperforming simpler CNNs and more complex Transformer-based models.
  • The Power of BiLSTM: Adding BiLSTM layers to a CNN significantly improved performance, highlighting the importance of capturing sequential dependencies in seismic waveforms for accurate quality assessment.
  • Transformer Limitations: The Transformer layer, while powerful in other domains, did not further improve performance in this specific task, suggesting it might require larger datasets or different architectural adaptations for seismic applications.
  • Early Stopping is Crucial: The study emphasizes that deep learning models are prone to overfitting on this task, making early stopping a vital technique for achieving reliable generalization.
  • No Manual Labels Needed: The development of a training schema that uses labels from traditional automatic methods (like HiPerSeis) is a game-changer. This bypasses the laborious manual labeling process, making the entire workflow fully automated and scalable.
  • Leveraging Smaller Earthquakes: The model's ability to effectively process data from smaller earthquakes (M 5.0-5.5) is a major breakthrough. This unlocks a vast amount of previously underutilized seismic data, potentially improving the resolution and coverage of subsurface imaging.

Key Results

  • Accuracy: The best-performing model, CNN-BiLSTM, achieved an accuracy of approximately 91% on manually labeled validation sets.
  • Comparison with Traditional Methods: The CNN-BiLSTM model retrieved approximately 5 times more reliable receiver functions from smaller earthquakes (M 5.0-5.5) compared to the traditional HiPerSeis method.
  • Qualitative Validation: The average waveforms and H-κ stacking results from receiver functions identified as "good" by the CNN-BiLSTM model (from smaller earthquakes) were comparable to those obtained from manual quality control of larger earthquakes (M > 5.5). This demonstrates the high quality of the automatically selected data.
  • Performance Metrics: Table 2 shows that the CNN-BiLSTM model achieved the highest True Positive (TP) and True Negative (TN) rates and the lowest False Positive (FP) and False Negative (FN) rates across the tested datasets, indicating superior classification performance.

Limitations and Future Work

  • Noisy Label Training: While the fully automatic training schema (using noisy labels from traditional methods) achieved comparable performance to manual labeling, it was slightly lower. The authors suggest that constructing larger, high-quality regional or global datasets could further improve this.
  • Receiver Function Calculation Consistency: The study notes that different methods and parameters used to calculate receiver functions can lead to variations in input waveforms. Consistency in these preprocessing steps is crucial for training and applying the deep learning model.
  • Future Directions:
    • Extending the method to S-wave receiver functions and P-wave transverse receiver functions.
  • Developing regional or global-scale datasets to further enhance the performance of models trained on noisy data.
  • Investigating the impact of different receiver function calculation parameters on model performance.

Summary updated

Abstract

Receiver function has been routinely used for studying the discontinuity structure in the crust and upper mantle. The manual quality control of receiver functions, which plays a key role in high-quality data selection and accurate structural imaging, has been challenged by today’s booming data volumes. Traditional automatic quality control methods usually require tuning hyperparameters and fail to generalize to low signal-to-noise ratio data. Deep learning has been increasingly used to deal with extensive seismic data. However, it generally requires a manually labeled dataset, and its performance is highly related to the network design. In this study, we develop and compare four different deep learning network designs with manual and traditional quality control methods using 53293 receiver functions from three broadband seismic stations. Our results show that a combination of convolutional and long-short memory layers achieves the best performance of ∼91% accuracy. We also propose a fully automatic training schema that requires zero manually labeled receiver function yet achieves similar performance to that using carefully labeled ones. Compared with the traditional automatic method, our model retrieves ∼5 times more reliable receiver functions from relatively small earthquakes with magnitudes between 5.0 and 5.5. The average waveforms and H-κ stacking results of these receiver functions are comparable to those obtained by manual quality control from earthquakes with magnitudes larger than 5.5, which further demonstrates the validity of our method and indicates its potential for making use of smaller earthquakes in the receiver function analysis.
BibTeX
@inproceedings{gong2022deep,
  author = {C. Gong and Ling Chen and Zhuowei Xiao and Xu Wang},
  title = {{Deep learning for quality control of receiver functions}},
  booktitle = {Frontiers in Earth Science},
  year = {2022},
  doi = {10.3389/feart.2022.921830},
}
Figure