onboard-coop/emit-test-dataset

Dataset Card for EMIT Test Dataset

Dataset Description

This dataset contains a comprehensive test data sample derived from the Earth Surface Mineral Dust Source Investigation (EMIT) instrument, an imaging spectrometer operating aboard the International Space Station (ISS).

It provides multiple EMIT Level 1B Radiance granules alongside ground truth labels, visualisations, and reduced-band subsets for evaluating different band selection strategies.

• Original Data Source: NASA JPL EMIT Data Portal
• Total Size: ~174 GB

---

Dataset Structure

The dataset consists of a main data/ directory containing the full raw hyperspectral data, alongside three additional directories that contain subsets reduced to 50 spectral bands based on specific band selection strategies.

Each subdirectory within these main folders represents a distinct EMIT overpass/granule (e.g., EMIT_L1B_RAD_001_20230203T171446_2303412_007, EMIT_L1B_RAD_001_20240120T154618_2402010_008, etc.).

1. Full Hyperspectral Data (data/)

Within each of the granule directories in the data/ folder, you will find the following file structure:

Raw Hyperspectral Data

• EMIT_L1B_RAD_...nc: The original Level 1B Radiance NetCDF file containing the raw hyperspectral data cube from the EMIT instrument.

Visualisations

• 40_rgb.png: A standard RGB composite rendering of the scene for visual reference.
• Mag1c Renders: Visualisations of the Mag1c algorithm outputs at different maximum threshold values for easy inspection.
  • 40_mag1c_1750max.png
  • 40_mag1c_3500max.png
  • 40_mag1c_5250max.png
  • 40_mag1c_7000max.png

Machine Learning Labels & Inference

• label.npy: Ground truth labels (array format) for the scene.
• label.png: Visual representation of the ground truth labels.
• linknet_inference.npy: The stitched predictions from the LinkNet model.
• mag1c_sas.npy: Mag1c-SAS output.
• mask.npy: Binary array used to filter invalid regions of the image.

2. Reduced-Band Subsets

These folders contain data that has been downselected to exactly 50 bands to simulate and evaluate accelerated onboard processing (such as Mag1c-SAS). We tested three strategies to find the optimal balance between accuracy and processing time:

• selected_50_bands_evenly_spaced/: Selects bands in the ~2122-2488 nm range with even spacing between them. (Note: This strategy can be scaled up to 72 bands, at which point all available bands within this range are used, matching the original Mag1c).
• selected_50_bands_highest_transmittance/: Selects the bands with the highest absolute transmittance, as these bands possess the strongest methane signal.
• selected_50_bands_highest_variance/ (Variance Increase): Starts with the band that has the highest transmittance, and selects each subsequent band to maximize the variance in CH₄ transmittance. This enables a more accurate approximation of the methane transmittance function.

Granule Directory Structure (Reduced-Band Subsets)

Within each granule folder in the three selected_50_bands_* directories, the data is provided as individual files rather than a single NetCDF cube to allow for modular testing. You will find the following file structure:

• Selected Spectral Bands: [wavelength]nm.tif (e.g., 2123nm.tif, 2130nm.tif, ..., 2485nm.tif). Exactly 50 individual TIFF files, each representing a single 2D spatial array of radiance values for that specific spectral wavelength.
• RGB Context Data: red.tif, green.tif, blue.tif for standard optical reference.
• Ground Truth & Validity Masks:
  • label.npy: Ground truth labels (array format).
  • valid_mask.tif: Binary mask used to filter out invalid regions of the whole image.
• Algorithm & Model Outputs:
  • mag1c_sas.tif: Output from the accelerated Mag1c-SAS algorithm run on the full scene.
  • mag1c_sas_tiling.tif: Output from the accelerated Mag1c-SAS algorithm run on tiles.
  • inference.tif: LinkNet model predictions where both Mag1c-SAS and inference were performed on the full scene (this is possible because the model is fully convolutional).
  • inference_tiling.tif: LinkNet model predictions where both Mag1c-SAS and inference were performed on tiles.
  • inference_tiling_inference_only.tif: LinkNet model predictions where Mag1c-SAS was performed on the full scene, but the LinkNet inference was performed on tiles.

---