Graph Neural Network Daily GNSS Denoising

Repository for the daily GNSS time series denoising using a graph neural network

Preprocessing

In the preprocessing folder, you will find all the scripts and notebook to prepare the data:

• 2 download script for UNR and CWU
• 2 raw to netcdf to convert and concatenate all stations together for both processing centers
• station maintenance correction notebook to remove trend using the station maintenance logs, and remove outliers
• add tremor notebook to co-locate tremors in time and space with the stations (possibility to tune the parameters)
• resource folder with the necessary files to download and add info to the stations

Scalers and trends are saved and important to keep for rescaling the results

Graph construction

• create_graph: main script to launch the graph construction
• GNSS_daily_clas: pytorch geometric dataset class for the GNSS dataset construction
• graphs utils: utility functions for the graph construction

training and prediction

• NN_architecture: pytorch geometric GNN class
• GNN training script: automatically trains on the first n graphs of the dataset and saves the last 400 for testing, equivalent to saving 2022 and 2023 to testing in our use case.
  • Command Line Arguments:
    • 'lr' (float): Learning rate for the optimization algorithm.
    • 'hidden_layer' (int): Size of the hidden layer in the neural network.
    • 'nb_epoch' (int): Maximum number of epochs for training.
    • 'dataset_id' (str): Identifier for the dataset being used.
    • 'lambda_center' (float): Lambda parameter for the center part of the loss function.
    • 'post' (str): Additional string to add at the end of the output filename.
    • Example: python GNN_training.py 0.001 64 100 my_dataset _experiment_1 This example sets lr=0.001, hidden_layer=64, nb_epoch=100, dataset_id='my_dataset', lambda_center=0.1, and post='_experiment_1'.
  • Then predicts for all the graphs in the dataset

Results

• results_to_nc_daily: Notebook to read the graph results from the GNN and generate a netcdf
  • read graphs
  • average the overlapping windows
  • rescale and add trend
  • some simple verification figures
  • final cleanup and save

References

TODO Add paper ref TODO Add result dataset zenodo