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

Example 2: Qualitative Amplitude Analysis

Enhance Automatic Picking Performance

Distribution of STEAD Dataset

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)

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.

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.

• 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

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}
}