PhaseNet-TF Alaska - iter2

Model Description

PhaseNet-TF is an advanced deep learning model for automatic seismic phase picking (P-wave, S-wave, and PS-wave detection) using spectrogram-based image segmentation approaches. The model leverages DeepLabV3Plus architecture to detect seismic arrivals with high accuracy, especially for weak and noisy signals from ocean-bottom seismometers and weak phases such as slab interface refracted PS and SP waves. This Alaska version is specifically trained on the PS_Alaska dataset for P and S phases. For more details, please refer to the paper and the PhaseNet-TF repository.

Model Architecture

Backbone: DeepLabV3Plus with ResNet34 encoder
Input: 3-component seismic waveforms converted to 6-channel spectrograms (real + imaginary)
Output: Probability maps for P, S, PS phases and noise
Sampling Rate: 40 Hz (dt_s = 0.025s)
Window Length: 4800 points (120 seconds)
Spectrogram Size: 64 × 4800 (frequency × time)
Input Channels: 6 (3 real + 3 imaginary spectrogram channels)
Output Classes: 4 (noise, P, S, PS)

Load the checkpoint

checkpoint = torch.load("pytorch_model_iter2.bin", map_location="cpu")

Citation

If you use this model in your research, please cite:

@article{jie2025background,
  title={Background Seismicity and Aftershocks of the 2020-2021 Large Earthquakes at the Alaska Peninsula Revealed by a Deep-learning-based Catalog},
  author={Jie, Yaqi and Wei, Songqiao Shawn and Zhu, Weiqiang and Freymueller, Jeffrey Todd and Elliott, Julie},
  journal={Authorea Preprints},
  year={2025},
  publisher={Authorea}
}

License

This model is licensed under the MIT License.