https://github.com/techiesamosa/lunaxplore
LunaXplore enhances low-light images from the PSR regions of lunar craters to improve signal-to-noise ratio (SNR). By applying advanced deep learning and image processing techniques, the project creates high-resolution image maps from Chandrayaan-2's OHRC, aiding lunar landing site selection and supporting geomorphological studies of the lunar pole
https://github.com/techiesamosa/lunaxplore
deep-learning gan hpc image-processing keras opencv python retinex-theory signal-to-noise-ratio tensorflow
Last synced: 5 months ago
JSON representation
LunaXplore enhances low-light images from the PSR regions of lunar craters to improve signal-to-noise ratio (SNR). By applying advanced deep learning and image processing techniques, the project creates high-resolution image maps from Chandrayaan-2's OHRC, aiding lunar landing site selection and supporting geomorphological studies of the lunar pole
- Host: GitHub
- URL: https://github.com/techiesamosa/lunaxplore
- Owner: TechieSamosa
- License: other
- Created: 2024-08-30T12:43:05.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2025-01-08T09:39:47.000Z (over 1 year ago)
- Last Synced: 2025-04-09T04:11:14.035Z (about 1 year ago)
- Topics: deep-learning, gan, hpc, image-processing, keras, opencv, python, retinex-theory, signal-to-noise-ratio, tensorflow
- Homepage:
- Size: 1.3 MB
- Stars: 3
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# 🌌 AETHER: Illuminating the Unseen 🌗
-----
## 🚀 Overview
**AETHER** is a deep learning framework designed to solve one of the most significant challenges in planetary science: visualizing the **Permanently Shadowed Regions (PSRs)** of the Moon. These areas, which haven't seen sunlight in billions of years, are prime locations for water ice but are notoriously difficult to image. Using a rich suite of data from ISRO's **Chandrayaan-2** mission, AETHER employs **Self-Supervised Learning** to enhance faint signals, transforming noisy, low-light patches into scientifically valuable maps.
-----
## ✨ Key Features
* **Self-Supervised Enhancement**: Leverages a novel pretext task to teach the model lunar topology from unlabeled data, eliminating the need for hand-annotated training sets.
* **Multi-Modal Data Fusion**: Intelligently integrates optical (OHRC), radar (DFSAR), elemental (CLASS), and spectral (IIRS) data to create a holistic and information-rich view of the lunar surface.
* **High-Fidelity Generative Model**: Utilizes a fine-tuned Generative Adversarial Network (GAN) to reconstruct shadowed terrain with plausible details, maximizing the Signal-to-Noise Ratio (SNR).
* **Mission-Ready Outputs**: Produces high-resolution, georeferenced polar maps critical for landing site selection, resource prospecting, and geomorphological studies.
-----
## 🛠️ Technical Architecture
AETHER is built with a modern, research-focused tech stack designed for high-performance computing and reproducibility.
* **Core Framework**: **PyTorch** for its flexibility and dynamic computation graphs.
* **Data Handling**: **GDAL** & **Rasterio** for geospatial data processing, **Pandas** for metadata management.
* **Image Processing**: **OpenCV** and **Scikit-Image** for preprocessing and classical enhancement algorithms.
* **Models & Training**: **Hugging Face Accelerate** for distributed training and **Weights & Biases** for experiment tracking.
* **Hardware**: Designed for **NVIDIA GPUs** using **CUDA** for accelerated model training.
-----
## 🧠 Methodology
Our pipeline transforms raw, heterogeneous data into clear, scientifically-valuable maps. The cornerstone of AETHER is its ability to fuse diverse data sources, providing our model with a rich, multi-faceted understanding of the lunar surface.
### 1\. Multi-Modal Data Ingestion & Fusion
We treat the lunar surface as a multi-layered dataset. Each instrument provides a unique layer of information, and by combining them, we create a "super-image" that offers far more context than a simple photograph.
* 🛰️ **Primary Optical Data (OHRC):** This is the high-resolution visual canvas we aim to enhance. It provides the base imagery, which suffers from low signal in shadowed regions.
* 🪨 **Structural & Texture Data (DFSAR):** The radar data acts as our structural blueprint. Unaffected by shadows, it reveals surface roughness and physical topography, telling the model the *shape* of the ground.
* 🧪 **Elemental Composition Data (CLASS):** This provides the chemical fingerprint. By mapping elements like Silicon, Iron, and Magnesium, CLASS tells the model *what the ground is made of*, allowing it to connect chemistry to texture.
* 💧 **Mineralogical & Volatile Data (IIRS):** The infrared spectrometer is key for identifying specific minerals and, crucially, the spectral signature of water ice and hydroxyl molecules. This layer guides the model toward areas of high scientific interest.
* 🗺️ **Geometric & Navigational Data (SPICE):** The essential Rosetta Stone. SPICE kernels provide the precise position, orientation, and timing data required to take all the above layers and align them perfectly into a single, cohesive coordinate system.
> **The Fusion Process:** All data layers are georeferenced using SPICE and stacked into a multi-channel tensor. Instead of a 3-channel (RGB) image, our model processes a rich data cube, enabling an informed, context-aware enhancement.
### 2\. Self-Supervised Pre-training (The Core Engine)
The model learns the fundamental patterns of lunar terrain without labels.
* **Pretext Task**: We employ a "masked autoencoder" approach. The model is fed a multi-channel data tensor with random patches masked out and is trained to reconstruct the missing information. To succeed, it must learn the intrinsic relationships between radar texture, optical shadows, and chemical composition.
### 3\. Generative Fine-Tuning
The pre-trained encoder from the SSL phase is used as the backbone of a custom Generative Adversarial Network (GAN).
* This GAN is specifically fine-tuned on the task of **low-light enhancement**. It takes a noisy OHRC image as input and generates a clean, high-SNR version, guided by the deep contextual understanding from the fused data.
### 4\. Map Synthesis & Visualization
The enhanced image tiles are stitched together and projected into a polar stereographic map.
* The final output is a high-resolution map highlighting previously unseen details within the Moon's Permanently Shadowed Regions.
-----
## 🌍 Impact & Applications
AETHER directly contributes to the next era of lunar exploration by:
* **Enabling Safer Landings**: Providing mission planners with clear, detailed views of potential landing sites for robotic and crewed missions.
* **Accelerating Scientific Discovery**: Unlocking vast, unexplored regions for geological analysis and the search for water ice.
* **Advancing Planetary AI**: Pioneering the use of self-supervised techniques for low-data, extreme-environment scenarios in space exploration.
-----
## 📂 Repository Structure
```
aether/
│
├── data/ # Scripts for downloading and preprocessing public data
├── notebooks/ # Jupyter notebooks for exploration and visualization
├── models/ # Pre-trained model weights and configurations
├── src/ # Main source code
│ ├── data_loader.py # Data ingestion and fusion pipeline
│ ├── ssl_trainer.py # Self-supervised pre-training module
│ ├── gan_enhancer.py # GAN fine-tuning and inference module
│ └── projection.py # Geospatial mapping and output generation
│
├── tests/ # Unit and integration tests
├── scripts/ # Helper scripts for training, inference, etc.
└── requirements.txt # Project dependencies
```
-----
## 🚀 Getting Started
### **Prerequisites**
* Python 3.10+
* Anaconda or Miniconda
* NVIDIA GPU with CUDA 11.8+
### **Installation**
1. Clone the repository:
```bash
git clone https://github.com/TechieSamosa/AETHER.git
cd AETHER
```
2. Create and activate a conda environment:
```bash
conda env create -f environment.yml
conda activate aether
```
3. Run inference on a sample image:
```bash
python scripts/enhance.py --input /path/to/sample.img --output /path/to/enhanced.png --model models/aether_v1.pth
```
-----
## 🎯 How to Contribute
We welcome contributions from the global community of researchers and developers\!
1. **Fork the repo** and create your branch from `main`.
2. **Open an issue** to discuss the change you wish to make.
3. **Make your changes** and ensure all tests pass.
4. **Submit a pull request** with a clear description of your contribution.
-----
## 📚 References & Acknowledgements
This work builds upon the state-of-the-art in deep learning and planetary science. We gratefully acknowledge the data provided by the **ISRO Science Data Archive (ISDA)**.
* **"Masked Autoencoders Are Scalable Vision Learners"** (He et al., 2021)
* **"SinGAN: Learning a Generative Model from a Single Natural Image"** (Shaham et al., 2019)
* **Chandrayaan-2 Mission Data Handbook**, ISRO
* **"Analysis of Permanently Shadowed Regions of the Moon using LRO and Chandrayaan-2 Data"**
-----
## 🏅 Contributors
A huge thank you to all the contributors who have dedicated their time and expertise to making AETHER possible!
-----
## 📄 License
This project is licensed under the **MIT License**. See the `LICENSE` file for details. This permissive license allows for wide use and collaboration.