SamTr7/01-emit-dataset-canary

01-emit-dataset canary

Compact, ML-split-ready slice of the 01-emit-dataset pipeline: 5 EMIT scenes already processed to ortho rgb/enh/mask_bin arrays, plus 512×512 plume/no-plume crops with masks.

The canary is the inputs side for figure regeneration and downstream-pipeline testing without NASA Earthdata credentials. The full corpus (raw L1B/L2B, multi-scene catalog, train manifest) lives elsewhere.

Scenes

split	n	timestamps
test	1	20220814T051412
train	3	20220810T064957, 20220811T042630, 20250922T204933
val	1	20220810T065132

All scenes are stored in the scenes/<timestamp>/ directory with three files:

• rgb.npy — float32, ortho-projected RGB rendering (3 channels)
• enh.npy — float32, EMIT L2B CH4ENH (methane enhancement, ppm·m)
• mask_bin.npy — uint8, plume binary mask derived from L2B CH4PLM

Crops

crops/{train,val}/{crops,masks}/*.npy — 512×512 patches, 4-channel (R,G,B,enh), plus single-channel uint8 masks. Counts: pos=4, neg_hard=4, neg_easy=4. Augmentation = none, strict_ok = true (originals only). Metadata in crops/metadata.jsonl.

Catalog

catalog.json — subset of the full pipeline catalog with the original EMIT product filenames per scene (RAD / OBS / CH4ENH / CH4PLM). Useful for citing the source granules; not needed to run the canary downstream pipeline.

Usage

from huggingface_hub import snapshot_download
local = snapshot_download(
    repo_id="SamTr7/01-emit-dataset-canary",
    repo_type="dataset",
    local_dir="data/canary",
)

Or via the project script: python scripts/setup_canary.py.

Source

Built from 01-emit-dataset processed-ortho outputs. EMIT data is courtesy of NASA/JPL under the EMIT Open Data License.