Asteroid_detection_CNN

Convolutional neural networks for detecting faint asteroid trails in wide-field astronomical images, developed in the context of Rubin Observatory / LSST data analysis.

This repository contains a PyTorch-based production implementation for pixel-level detection of trailed moving objects, together with tools for dataset generation, training, evaluation, and post-processing.

---

Scientific context

Detecting faint asteroid trails in modern sky surveys is challenging due to:

• low surface brightness of trails,
• variable observing conditions,
• strong background contamination,
• limitations of classical detection pipelines.

This project explores deep-learning-based pixelwise segmentation to recover trailed sources that are often missed by traditional algorithms, with a focus on:

• LSST / Rubin Observatory–like data,
• realistic injected datasets,
• reproducible training and evaluation.

---

Features

• PyTorch U-Net–based architectures with residual and attention blocks
• Robust preprocessing (MAD normalization, sigma clipping)
• HDF5-based tiled datasets for large focal-plane images
• Multi-stage and curriculum training strategies
• Pixel-level and object-level evaluation metrics
• Two-stage detection concepts (connectivity + scoring)
• Compatibility with Rubin Butler–based data products

---

Typical workflow

1. Dataset creation

   • Inject synthetic asteroid trails into single-visit images
   • Store images, masks, and metadata in HDF5 / CSV format

2. Training

   • Train segmentation networks on tiled image data
   • Use class-imbalanced losses and staged training

3. Evaluation

   • Pixelwise ROC / F1 / AUC
   • Object-level detection via connected components
   • Comparison with LSST stack detections

---

Requirements (indicative)

• Python ≥ 3.9
• PyTorch
• NumPy, SciPy, pandas
• h5py
• matplotlib
• Rubin Science Pipelines (for data generation and injections)

Exact environments depend on whether you are running on:

• Rubin USDF / SDF
• local workstation
• HPC cluster (SLURM)

---

Status

Active research and development.
The codebase evolves alongside:

• Rubin commissioning data (ComCam / LSSTCam)
• improved injection realism
• new detection post-processing strategies