LEO SC-SatCom

May 28, 2026 · View on GitHub

Overview

This repository is the official implementation of the paper “Doppler-Adaptive Digital Semantic Communication for Low Earth Orbit Satellite Systems”, which includes the training and evaluation framework for a C-VQ-VAE-based model using switching batch normalization (SBN), the training of an estimator to mitigate Doppler effects, and the training of a reinforcement learning agent to optimize the transmission processes. The codebase includes pre-training scripts, evaluation tools, and a GUI-based visualization module.

Pretrained Model

A pretrained model (.h5 file) and saved data are available in Google Drive.
You can access it here: Download from Google Drive

After downloading, place the folder in your current directory before running the evaluation or training script.

Code Structure

1. Pretraining: train_c_vq_vae_model.py

  • Trains a C-VQ-VAE model for digital-compatible reconstruction tasks.

  • The model is trained on an AWGN-fixed channel, which results in significant performance degradation when applied to Rician or Rayleigh fading channels.

  • The total number of training epochs is set to be 100.

  • Key training consideration:

    • During the training iterations, the recovered indices must not be used; instead, training should directly utilize the quantized vectors.

2. Save PSNR and symbol data for efficient training of PPO agent and post equalizer: save_psnr_data.py and save_symbol_data.py

  • Executes two saving codes for efficient training of the PPO agent and the post-equalizer.
  • This is a tricky implementation technique, as running without stored code would take an extremely long time to train.
  • Data is stored inside doppler_data folder.

3. Training: train_tx_agent_ppo.py

  • Trains a PPO agent applied to the transmitter for adaptive modulation selection.
  • The total number of training epochs is set to be 200.

4. Training: train_post_model.py

  • Trains a post-equalizer applied to the receiver for mitigation of residual Doppler effects.
  • The total number of training epochs is set to be 50.

Running the GUI

python app_tk.py
  • UI Execution:
    • Launching the script opens a graphical user interface, allowing users to adjust parameters and visualize performance metrics.

Parameter Configuration Guide

Dataset Options:

  • cifar10: 32x32 RGB images, 10 classes
  • eurosat: 64x64 RGB satellite images, 10 classes

Modulation Types:

  • SC_auto: Smart coding with automatic modulation
  • SC_none: Smart coding without modulation optimization
  • TN_auto: Traditional method with automatic modulation
  • BPSK, QPSK, 16QAM, 64QAM, 256QAM: Specific modulation schemes

Channel Coding:

  • true: Enable channel coding (LDPC codes)
  • false: Disable channel coding
  • both: Compare both scenarios

Compensation Methods:

  • lmmse: Linear Minimum Mean Square Error

  • mrc: Maximum Ratio Combining

  • none: No compensation

  • SNR Configuration:

    • Users can manually adjust minimum and maximum SNR values (in dB).

  • Execution:

    • After configuring all parameters, clicking "Run Simulation" will generate and display the performance plots automatically.

License

This project is licensed under the MIT License