ZiyunPOLYU/hls-s2opt-v4-20pct-train

HLS S2-opt v4 20pct training data

This public dataset contains the complete train.nc HLS (Harmonized Landsat and Sentinel-2) training archive as one NetCDF4 file.

Dataset contents

• 2,679,120 pixels concatenated from 80 HLS tiles
• 1,220 three-day time bins, from 2016001 through 2025364 (YYYYDDD)
• 13 native HLS bands: B01 B02 B03 B04 B05 B06 B07 B08 B8A B09 B10 B11 B12
• 99,770 entries in the per-file granule table
• Pixel-first layout: data(pixels, time, bands)

Variables

• data: int16, HLS surface-reflectance digital numbers. Missing values are -9999. Convert valid values to reflectance with data * 1e-4.
• fmask: uint8, original HLS Fmask bits. Bit 0=cirrus, 1=cloud, 2=adjacent, 3=shadow, 4=snow, 5=water; 255 is fill.
• sensor: uint8; 0=none, 1=S30 (Sentinel-2), 2=L30 (Landsat).
• source_idx: int32 index into the granules table; fill is -1.
• obs_time: int32 observation date encoded as YYYYDDD; fill is 0.
• lon, lat: float32 pixel coordinates.
• tile_idx: uint16 index into tile_names.
• granule_id, granule_sensor, granule_datetime, granule_json, granule_version, granule_cloud_coverage, granule_spatial_coverage: source-granule metadata.

The scale factor is stored as the custom hls_scale_factor attribute. It is intentionally not a CF scale_factor attribute, so readers should not apply automatic NetCDF packing.

Reading the file

from netCDF4 import Dataset

with Dataset("train.nc") as ds:
    dn = ds["data"][:]              # pixels, time, bands; int16
    reflectance = dn.astype("float32") * 1e-4
    mask = ds["fmask"][:]

For large files, read a pixel slice instead of loading the whole data array:

with Dataset("train.nc") as ds:
    batch = ds["data"][0:1024, :, :]

Provenance

The source file describes this archive as HLS S2-opt v4 20pct train.nc, with one selected source observation per pixel and three-day time bin. Cloud information remains in fmask; it is not baked into data. The 80 tile blocks follow the source ingest order, with raster order preserved inside each tile.