Asteria 🌌

Asteria is a lightweight, high-performance deep learning inference engine and instance deblender for stars in astronomical imaging, powered by boundary-aware UNet++ (Nested U-Net).

Designed for the Nancy Grace Roman Space Telescope, Asteria provides promptless, end-to-end star detection in a single neural forward pass.

_{Example of Asteria star detection on a simulated Roman WFI exposure.}

Key Features

• Lightweight & Self-Contained: Pure PyTorch UNet++ architecture without bulky dependencies or heavyweight vision foundation models.
• Hardware Acceleration with Automatic Fallback: Auto-detects and optimizes execution across NVIDIA CUDA (with TF32 and AMP), Apple Silicon MPS (Metal Performance Shaders on macOS), and universal CPU.
• Physical Boundary-Aware Deblending: Trained with 3-class supervision:
  • 0: Sky Background
  • 1: Source Interior / Core
  • 2: Deblending Boundary / Contact Interface
• Seamless Tiled Dense Inference (predict_dense): Sliding-window 2D Hann/cosine window blending to process arbitrary image sizes (from local cutouts up to full $4088 \times 4088$ Roman WFI exposures) with bounded memory ($< 350$ MB VRAM).
• Photometric LeakySB Stretch: Non-linear surface brightness transformation that preserves stellar point-spread functions (PSFs) while lifting faint, diffuse galactic wings.
• Unified Astronomical I/O: Direct support for 2D astronomical FITS (.fits, .fits.gz), Roman ASDF (.asdf Level 2 calibrated exposures and Level 3 mosaics via roman_datamodels), and raw NumPy arrays. Automatically converts Level 2 exposures to surface brightness units ($\text{MJy/sr}$) using calibrated photometry and enables $\text{ZP}=20.472$.
• Instant Deblending: Marker-controlled watershed expansion turning dense neural predictions into discrete celestial catalogs with centroids, bounding boxes, and photometry.

Installation

From PyPI

pip install asteria-roman

From Source

git clone https://github.com/PabloMSanAla/asteria.git
cd asteria
pip install -e .

Quickstart

1. Python API

import asteria

# 1. Load pre-trained model (auto-downloads weights on first run, auto-selects device)
segmentor = asteria.load_model("latest")

# 2. Run dense 3-class segmentation inference on a FITS or ASDF exposure
pred_mask, probs, stretched = segmentor.predict_dense(
    "scene_0001_F106.fits",
    tile_size=512,
    overlap=64,
)

# 3. Render publication-ready side-by-side 2-panel figure
fig = asteria.plot_two_panel(
    stretched=stretched,
    pred_mask=pred_mask,
    save_path="prediction.png",
)

# 4. Extract discrete, deblended astronomical sources (marker-controlled watershed)
catalog = segmentor.segment("scene_0001_F106.fits")
print(f"Extracted {len(catalog.masks)} distinct celestial objects!")
for idx, (cx, cy) in enumerate(catalog.centroids[:5]):
    print(f"Source #{idx + 1}: centroid=({cx:.1f}, {cy:.1f}), area={catalog.areas[idx]} px")

2. Command-Line Interface (CLI)

Asteria includes the asteria-predict CLI command:

# Basic prediction and display
asteria-predict scene.fits

# Full-field tiled inference with saving and FITS export
asteria-predict scene.fits \
    --tile_size 512 \
    --overlap 64 \
    --output results.png \
    --export_fits segmap.fits \
    --deblend \
    --no_display

Model Architecture & Weights History

For comprehensive architectural details and benchmark comparisons, see the Asteria Documentation.

License

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