Mapping Morphological Signatures with Rotation-Invariant Scattering Networks

This repository contains the official codebase for the Bsc dissertation: "Mapping Morphological Signatures: Using Rotation-Invariant Morlet Scattering Networks to Identify and Interpret Galaxies." This project implements an unsupervised, $SE(2)$ rotation-invariant scattering network to extract geometrically stable morphological features from astronomical surveys. It evaluates the critical trade-off between explicit mathematical invariance (Rigid Motion) and learned invariance (ResNet50V2 CNNs), utilizing the Galaxy10 DECals dataset as a benchmark for the Hubble Flow continuum.

The Dataset: Galaxy10 DECals

This pipeline is built to process the Galaxy10 DECals Dataset, an astrophysical dataset containing 17,736 images across 10 broad morphological classes.

Example samples from the Galaxy10 DECals dataset demonstrating the morphological continuum from smooth ellipticals to disturbed mergers.

---

Installation & Setup

It is highly recommended to run this pipeline within an isolated Python Virtual Environment to prevent dependency conflicts (especially regarding GPU-accelerated libraries like CuPy and TensorFlow).

1. Clone the repository:

2. Create and activate a virtual environment:

On macOS and Linux:

python3 -m venv venv
source venv/bin/activate

On Windows:

python -m venv venv
.\venv\Scripts\activate

3. Install dependencies:

pip install --upgrade pip
pip install -r requirements.txt

(Note: For GPU acceleration, ensure you have the correct CUDA toolkit installed for your system to support CuPy and TensorFlow/PyTorch).

Repository Structure

The codebase is organized into sequential stages of the analytical pipeline: Pre-processing, Embedding Extraction, and Topological Analysis.

.
├── .gitignore                  # Standard Git ignore file
├── requirements.txt            # Python dependencies (NumPy, SciPy, CuPy, UMAP, HDBSCAN, etc.)
│
├── Pre-processing & Utilities
│   ├── pre_process_data.py     # Cosmological scaling (7.35/z), sky subtraction, and Arcsinh stretching
│   ├── sky.py                  # Custom Sigma-clipping algorithms for sky background calculation
│   ├── sky_subtract.ipynb      # Jupyter notebook for prototyping/visualizing background noise removal
│   └── view_data.ipynb         # Interactive HDF5 dataset explorer and visualization prototype
│
├── Wavelet & Scattering Core
│   ├── create_wavelets.py      # Generates the Morlet wavelet filter banks (phi and psi filters)
│   ├── 1dwavelets.py           # Foundational 1D wavelet exploration and testing utilities
│   ├── rigid_motion_embedding.py # Core $SE(2)$ invariant joint rigid motion scattering implementation
│   └── kymatio_embedding.py    # Baseline extraction using the standard (covariant) Kymatio library
│
├── Deep Learning Baseline
│   ├── cnn_embedding.py        # Extracts 2048-d feature vectors from the ResNet50V2 baseline
│   └── modelCNNpaper.py        # CNN architecture from A post-merger enhancement only in star-forming Type 2 Seyfert galaxies: the deep learning view
│
└── Manifold Analysis
    └── cluster_pipeline.py     # UMAP projection, HDBSCAN clustering, Silhouette sc