speckcn2 2.0.0

Estimate Cn2 from Speckle patterns using machine learning

SpeckleCn2Profiler:

Improving Satellite Communications with SCIDAR and Machine Learning

Overview

Optical satellite communications is a growing research field with bright commercial perspectives. One of the challenges for optical links through the atmosphere is turbulence, which is also apparent by the twinkling of stars. The reduction of the quality can be calculated, but it needs the turbulence strength over the path the optical beam is running. Estimation of the turbulence strength is done at astronomic sites, but not at rural or urban sites. To be able to do this, a simple instrument is required. We want to propose to use a single star Scintillation Detection and Ranging (SCIDAR), which is an instrument that can estimate the turbulence strength, based on the observation of a single star. In this setting, reliable signal processing of the received images of the star is most challenging. We propose to solve this by Machine Learning.

Statement of Need

SpeckleCn2Profiler addresses the challenge of atmospheric turbulence profiling for optical satellite communications using machine learning-based analysis of SCIDAR data. The primary target users include:

• Aerospace engineers working on free-space optical communication systems
• Atmospheric scientists studying turbulence profiles and optical propagation
• Astronomers characterizing atmospheric seeing conditions
• Researchers in adaptive optics and atmospheric characterization

While traditional SCIDAR instruments require complex signal processing, SpeckleCn2Profiler provides a machine learning-based approach that enables simpler, more accessible turbulence profiling. Unlike classical methods that rely on analytical inversion techniques, our ML-based approach can handle noisy data more robustly and provides uncertainty quantification through ensemble predictions.

Repository Contents

This repository contains the workflow to implement and train machine learning models for turbulence strength estimation from SCIDAR data. Extensive Documentation is available to explain the methodology, algorithms used, and guidelines for using the code.

Getting Started

To get started with the project, follow these steps:

Prerequisites

To correctly install speckcn2, you need Python 3.9 or higher. If you don't have it installed, you can download it from the official website.

You will also need header files required to compile Python extensions (contained in python3-dev):

Ubuntu/Debian:

sudo apt-get install python3-dev

Other systems:

# Fedora/RHEL
sudo dnf install python3-devel

# macOS
xcode-select --install

# Windows: Headers are included with the official Python installer

Installation

Option 1: Install from PyPI (Recommended)

python -m pip install speckcn2

We strongly recommend using a virtual environment:

# Create virtual environment
python -m venv speckcn2-env

# Activate virtual environment
# On Linux/macOS:
source speckcn2-env/bin/activate
# On Windows:
# speckcn2-env\Scripts\activate

# Install the package
python -m pip install speckcn2

Verify installation:

python -c "import speckcn2; print('Installation successful!')"

Option 2: Development Installation

For advanced users and developers who want to modify the code:

git clone https://github.com/MALES-project/SpeckleCn2Profiler.git
cd SpeckleCn2Profiler
git submodule init
git submodule update
pip install -e .

Usage

To use the package, you run the commands such as:

python <mycode.py> <path_to_config.yml>

where <mycode.py> is the name of the script that trains/uses the speckcn2 model and <path_to_config.yml> is the path to the configuration file.

Here you can find a typical example run and an explanation of all the main configuration parameter. In the example submodule you can find multiple examples and multiple configuration to take inspiration from.

Quick Example

Here's a minimal working example to train a model on sample data:

import speckcn2 as sp2
import torch

# Load configuration
config = sp2.load_config('path/to/config.yaml')

# Prepare data
all_images, all_tags, all_ensemble_ids = sp2.prepare_data(config)

# Normalize tags
nz = sp2.Normalizer(config)

# Split data
train_set, test_set = sp2.train_test_split(all_images, all_tags,
                                            all_ensemble_ids, nz)

# Setup model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model, last_model_state = sp2.setup_model(config)
model = model.to(device)

# Define loss and optimizer
criterion = sp2.ComposableLoss(config, nz, device)
optimizer = sp2.setup_optimizer(config, model)

# Train
model, avg_loss = sp2.train(model, last_model_state, config,
                            train_set, test_set, device,
                            optimizer, criterion, criterion)
print(f'Training complete. Final loss: {avg_loss:.5f}')

Sample data and configuration files are available in the examples/ directory.

What can we predict?

A machine learning model trained using speckcn2 can predict:

1. Instantaneous turbulence strength

Given a speckle pattern, the model can predict the instantaneous turbulence strength and also provide an uncertainty estimate if more patterns are available.

2. Parameters estimation

The model can also estimate important parameters that are useful for the analysis of the speckle pattern. At the moment we support:

• Fried parameter r0
• Isoplanatic angle θ0
• Rytov Index σ

We also provide histograms of the estimated parameters and the error of the estimation.

Running Tests

To verify your installation and ensure everything works correctly, you can run the test suite using pytest:

pytest

Note: Some tests may fail on first run because test data needs to be set up. If this happens, run:

python ./scripts/setup_test.py
pytest

For more details on testing and development, see our Contributing Guidelines.

Contribution Guidelines

We welcome contributions to improve and expand the capabilities of this project. If you have ideas, bug fixes, or enhancements, please submit a pull request. Check out our Contributing Guidelines to get started with development.

Generative-AI Disclaimer

Parts of the code have been generated and/or refined using GitHub Copilot. All AI-output has been verified for correctness, accuracy and completeness, revised where needed, and approved by the author(s).

How to cite

Please consider citing this software that is published in Zenodo under the DOI 10.5281/zenodo.11447920.

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.