psy125/mm2025

DeepSpace Dataset

Dataset Description

The DeepSpace dataset contains multi-spectral astronomical images from NASA's James Webb Space Telescope (JWST) CEERS (Cosmic Evolution Early Release Science) program. This dataset is specifically designed for unsupervised denoising and astronomical image analysis applications, preserving the complex noise characteristics inherent in space telescope observations.

Dataset Details

Source Data

We utilize data from NASA's CEERS program, a JWST initiative investigating galaxy evolution in the early universe. The CEERS team provides calibrated data following pixel-alignment across different frequency bands, which are widely adopted in the astronomical community for scientific analysis.

Spectral Bands

The dataset consists of multi-spectral images captured across five Near-Infrared Camera (NIRCam) frequency bands:

• F115W: 1.15 μm filter
• F150W: 1.50 μm filter
• F200W: 2.00 μm filter
• F356W: 3.56 μm filter
• F444W: 4.44 μm filter

Data Processing

• Source Extension: SCI extension from CEERS FITS files containing calibrated pixel values
• Patch Size: 360×360 pixels (non-overlapping patches)
• Normalization: Z-Scale normalization using AstroPy's implementation

Dataset Structure

Data Splits

Split	Samples	Percentage	Purpose
Train + Validation	4200	80.0%
Test	1,050	20.0%

Total: 5,250 image patches

Split Strategy:

• 80% train+validation (4,200 patches), 20% test (1,050 patches)

Data Instances

Each instance contains:

• image: 360×360 pixel astronomical image patch
• nircam: NIRCam designation (nircam4, nircam5, nircam8, nircam9, nircam10)
• filter_band: JWST filter band (f115w, f150w, f200w, f356w, f444w)
• subfolder: Original subfolder identifier (1, 2, 3)
• patch_index: Sequential patch number within original image
• split: Dataset split assignment

Scene-to-NIRCam Mapping

• nircam4: scene1-4
• nircam5: scene5-6
• nircam8: scene7-8
• nircam9: scene9-12
• nircam10: scene13-16

Noise Characteristics

The dataset preserves JWST's distinctive noise patterns, including:

• Cross-shaped artifacts: Wavelength-dependent noise patterns visible in frequency domain
• Correlated noise: Spatially correlated noise structure from detector characteristics
• Realistic observations: Authentic noise from cosmic rays, detector effects, and background

These characteristics make the dataset particularly valuable for developing and evaluating astronomical image denoising algorithms.

Usage

from datasets import load_dataset

# Load complete dataset
dataset = load_dataset("username/deepspace-jwst")

# Load specific split
train_data = load_dataset("username/deepspace-jwst", split="train")
val_data = load_dataset("username/deepspace-jwst", split="validation")
test_data = load_dataset("username/deepspace-jwst", split="test")

# Example: Process by filter band
for sample in train_data:
    if sample['filter_band'] == 'f200w':
        image = sample['image']
        nircam = sample['nircam']
        # Your analysis here...

Citation

If you use this dataset in your research, please cite:

@dataset{deepspace_2025,
  title={DeepSpace: Toward Interpretable Space Image Denoising},
  year={2025},
  publisher={Hugging Face Datasets},
  url={https://huggingface.co/datasets/psy125/acmmm2025},
  note={Derived from NASA CEERS program data}
}

@misc{ceers_program,
  title={CEERS: The Cosmic Evolution Early Release Science Survey},
  author={CEERS Team},
  url={https://ceers.github.io},
  note={NASA JWST Early Release Science Program}
}