HLS-LST-CrossSensor (HLC)

277,808 calendar-month bundles across 2,490 MGRS tiles: one HLS condition per bundle and 2,892,007 usable coarse/fine LST patch-month pairs.

Release packaging: lossless-hub-v1-20261005. Release completeness is recorded in release.json. This is an archival, lossless release with an explicit Python reader; automatic Dataset Viewer support is not claimed.

Data and splits

Original compressed NPZ files are grouped into approximately 4 GiB uncompressed tar shards, without array conversion or numeric changes. The restore script verifies every member and restores the exact timestamps required by the frozen reader. Months in a bundle may come from different years; use available-month slots and actual dates, not a continuous-time assumption. Native coarse LST comes from Terra/MODIS MOD11A1 Collection 6.1. New synthetic-degradation experiments use a shared [16:240,16:240] center crop with no random flips; raw fine grids remain 256x256.

API split	HLS / records	fine LST samples
train	250,449	2,608,772
val	13,694	141,711
test	13,665	141,524

Train, val and test membership is fixed. MGRS tile IDs are disjoint within this dataset; this is not proof of zero geographical-footprint overlap. HLP and HLC share some tile IDs and are not automatically independent from one another.

Download and read

hf download dsaaf/HLS-LST-CrossSensor --repo-type dataset --local-dir ./HLS-LST-CrossSensor
cd HLS-LST-CrossSensor
pip install -r eo_data/requirements.txt
python scripts/restore.py

Private repositories require authentication. HLC restoration requires additional space for the extracted files (about 404 GiB); keep the downloaded shards until verification completes. HLP is read directly from its original ZIP.

from pathlib import Path
from eo_data import open_dataset
root = Path('.').resolve()
ds = open_dataset('lstsr_tb', 'lst', 'val', root=root,
                  prepared=root / 'processed_data/v1')
sample = ds[0]
temperature_c = ds.denormalize(sample['image']) / 100.0
valid = sample['reference_mask']
ds.close()

LST is stored in Celsius times 100. Apply the valid mask. Normalized images must first be denormalized. Fine LST is a satellite-retrieved reference product. Six HLS channels are Blue, Green, Red, NIR, SWIR1, SWIR2.

The frozen reader exposes individual modalities for reconstruction; it is not a ready-made coarse LST + HLS to fine LST training interface. Retain masks and pairing metadata when constructing paired tasks.

Files and reproducibility

• DATASHEET.md: links to the current evidence and detailed data report.
• processed_data/: frozen indices, normalization and fingerprints, with original paths preserved.
• eo_data/: the compatible reader.
• checksums/ and release.json: file integrity and packaging manifests.
• scripts/restore.py: verified restoration after download.

The original layout is retained inside this repository so the frozen readers can resolve their paths. Never run preprocessing to replace the supplied splits. HLP v2 must always be selected explicitly.

Provenance, limitations and licensing

See the full current report for sources, dates, masks and limitations. HLP pairs use HLSL30 v2 and Landsat Collection 2 surface temperature; HLC additionally pairs MODIS and Landsat LST. Optical conditions are not guaranteed to be synchronous with each temperature reference.

Dataset-specific author attribution, license and citation are not yet finalized in the source project. No license is inferred from TIDE or from a model name. No DOI or paper citation is invented for this release. The full report discusses remaining provenance and registration checks.