Cold Diffusion Model for Seismic Signal Denoising

Some Results

We apply diffusion models for denoising seismograms. Utilizing the STEAD Dataset, we demonstrate the distinct performance of various denoising models:

Example 1: Qualitative Picker Analysis

This figure showcases a superior performance of the sampling method in comparison to the direct model. Notably, the direct model exhibits a tendency to retain a certain level of noise prior to the P-wave arrival, which is not observed in the sampling method. This retention of noise in the direct model can lead to premature P-wave picks, as evidenced in the data. Conversely, the sampling method demonstrates a more accurate noise reduction, resulting in a clearer delineation of the P-wave arrival. This comparison underlines the enhanced capability of the sampling method in accurately identifying seismic events, thereby reducing the likelihood of early P-wave detection errors inherent in the direct model approach.

Example 2: Qualitative Amplitude Analysis

In Example 2, we demonstrate the Cold Diffusion Model's superiority in preserving seismic signal amplitudes through a sampling strategy ('sampling 300'). This method aligns more accurately with the original waveform, maintaining amplitude integrity, as contrasted with the U-Net model ('direct 300') which exhibits amplitude attenuation and increased residuals, particularly in higher amplitude segments.

Some Possible Applications

Enhance Automatic Picking Performance

In our evaluation, we applied PhaseNet to waveforms for assessing denoiser impacts on P and S wave arrivals. Histograms comparing "direct," "sampling," and "deep denoiser" methods highlighted the "sampling" method's superior accuracy in aligning with manual picks, especially at higher parameter settings. The "direct" method showed improved P-wave accuracy with parameter increases, while the "deep denoiser" displayed moderate recall rates. Overall, S-wave detections were consistently precise across methods, but P-wave picks varied, with the "sampling" method showing the least discrepancy from manual picks. This study underscores the significance of denoising in automated seismic analysis, with the "sampling" approach being notably effective.

Dataset

Distribution of STEAD Dataset

You can find the dataset used for training, evaluation, and testing on Zenodo: [STEAD subsample 4 CDiffSD](https://zenodo.org/record/10972601)

Data Loader Usage

To create a DataLoader, use the following parameters:

• feature_columns: List of feature columns (e.g., 'Z_channel', 'E_channel', 'N_channel').
• target_columns: List of target columns (e.g., 'p_arrival_sample', 's_arrival_sample').
• trace_name_column: Column containing trace names.

The DataLoader has three elements:

• Index for mapping to trace names.
• Normalized or non-normalized channels ('Z_channel', 'E_channel', 'N_channel') based on the normalize flag.
• 'p_arrival_sample' (P-wave arrival) and 's_arrival_sample' (S-wave arrival) for target columns.

This DataLoader setup allows us to train our model using the prepared data.

Example of how to use utils.create_data_loader:

import pandas as pd
import utils.utils_diff as u

df_path = your_path + 'df_train.csv'
df_path_noise = your_path + 'df_noise_train.csv'

df = pd.read_pickle(df_path)
df_noise = pd.read_pickle(df_path_noise)

feature_columns = ['Z_channel', 'E_channel', 'N_channel']  
target_columns = ['p_arrival_sample', 's_arrival_sample'] 
trace_name_column = 'trace_name' 

train_loader, val_loader, test_loader, index_to_trace_name = u.create_data_loader(df, feature_columns, target_columns, trace_name_column, batch_size= ... , shuffle=True)
train_noise_loader, val_noise_loader, test_noise_loader,index_to_trace_name = u.create_data_loader(df_noise, feature_columns, target_columns, trace_name_column,is_noise = True, batch_size= ..., shuffle=True)

How to Start Training

To replicate the results and start training the model, follow these steps:

1. Setup the Environment:

   • Ensure you have conda installed. If not, download and install it from Conda's official website.
   • Download the environment.yaml file from this repository.

   conda env create -f environment.yaml
   conda activate cold-diffusion
   

2. Download the Dataset:

   • Download the STEAD dataset from Zenodo.

3. Run the Training Script:

   • Ensure the dataset is placed in the correct directory as expected by the script.
   • Execute the training script with the following command:

   python train.py --dataset_path path/to/STEAD_dataset
   

4. Monitor Training:

   • Training logs and checkpoints will be saved in the specified directory. Monitor the training process using these logs.

Configuration Details

To start training, you need to configure the arguments in the config_parser.py file. Here are some important settings:

• Set training=True for training mode or training=False for testing mode.
• Specify the model path using path_model if you want to use a pre-trained model for testing.
• Set the channel type using channel_type=0 for E, channel_type=1 for N, and channel_type=2 for Z.
• During training, you can set tuning=True to enable hyperparameter tuning, or tuning=False to disable it.
• Adjust the T parameter to set the number of timesteps for the model.

Denoising ObsPy Waveforms

The denoise_obspy_waveform folder contains a script to denoise seismic waveforms loaded via ObsPy. Usage:

• Download the checkpoint from Zenodo and place it in the appropriate directory.
• Load your ObsPy waveform and run the denoising script.
• Adjust the T parameter in the script to control the number of denoising steps (fewer steps may work better for high SNR signals).

Related Paper

• Seismic Signal Denoising and Decomposition Using Deep Neural Networks
• Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise
• Seismic Signal Denoising and Decomposition Using Deep Neural Networks

Citation

If you use this code or dataset in your research, please cite our paper:

@article{trappolini2024cold,
  title={Cold diffusion model for seismic denoising},
  author={Trappolini, Daniele and Laurenti, Laura and Poggiali, Giulio and Tinti, Elisa and Galasso, Fabio and Michelini, Alberto and Marone, Chris},
  journal={Journal of Geophysical Research: Machine Learning and Computation},
  volume={1},
  number={2},
  pages={e2024JH000179},
  year={2024},
  publisher={Wiley Online Library}
}