ZiyunPOLYU/hls-classification

HLS v2 classification point series

Point time series sampled from NASA HLS v2 for three classification benchmarks. One compressed NumPy archive per task. Inside each archive, one .npz per MGRS tile and year.

Task	Points	Tile-years	Years
CDL	11000	523	2023
CropHarvest	15802	170	each point's sample_year, 2019–2022
GLaNCE	2188	22506	2014–2019, every point, every year

Archives: cdl.tar, cropharvest.tar, glance.tar.

Observation product

Each .npz keeps every HLS granule that has a real reflectance value at the point. Cloudy observations stay in. Reflectance is the original int16 digital number (reflectance = DN * 0.0001). Fill is -9999. Fmask is the raw byte; fill is 255.

Bands, in order: B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B10, B11, B12. Landsat-30 has no red-edge or B08/B8A/B12 slot, so those positions are -9999. Sensor code: 1 = S30, 2 = L30.

Arrays:

• point_index, lon, lat, label, sample_id: the points in this tile-year
• obs_point: row into point_index for each observation
• obs_time: yyyyDDDTHHmmss
• obs_sensor, fmask, data with shape (observations, 13)
• bands, year, tile

A tile-year with no clear-enough observation is still present, with zero rows in data. CropHarvest also stores coarse_label.

Labels and coordinates match the source tables used to choose the points. CDL labels are the 2023 class. CropHarvest labels are the plot label. GLaNCE labels are class_ids, and sample_id is the site id.