Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Server error while post-processing the rows. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

TED downstream Earth observation inputs

Point archives used for the TED downstream evaluation. Each task keeps its own HLS observations, Sentinel-2 observations, and, where a local product exists, the Google AlphaEarth annual embedding. Pretraining cubes are not in this dataset.

Source licenses stay with the original products: NASA HLS, Copernicus Sentinel-2, and the Google Satellite Embedding product. This repository only repackages the extracted point series.

Tasks

Classification (5): LCMAP, GlanCE, GlobalTree, CDL, CropHarvest.

Change (4): Hansen, Wildfire, LCMAP-C, LandTD. These are point-level change tasks, stored under change/.

Added classification: CityRep (8 cities, 2024, 160,000 points) and LUCAS land cover (203,569 points).

Regression: OpenET monthly ensemble ET, 35,683 points, years 2017–2023.

What each archive contains

hls_common_processed.nc is the composited HLS series used by the original five classification and four change evaluations. The array has six reflectance bands plus NDVI. It is not the raw scene stack.

hls_l30.tar and hls_s30.tar for CityRep and LUCAS are per-point common-band series. Band names in those files are Blue, Green, Red, NIR, SWIR1, SWIR2. CityRep files are year_2024.npz. LUCAS files are observations.npz.

hls_extra_l30.tar and hls_extra_s30.tar are the supplemental bands that were not stored in the common-band files.

Sensor Common six, stored separately Supplemental bands in the extra tar
L30 B2, B3, B4, B5, B6, B7 B1, B9, B10, B11
S30 B2, B3, B4, B8A, B11, B12 B1, B5, B6, B7, B8, B9, B10

L30 B10 and B11 are temperature. Extra files keep band_valid; a missing channel is NaN and is not filled. For the original five classification and four change tasks, these sidecars are annual cloud≤90 candidate scenes. They are not row-aligned to the composited NetCDF. CityRep and LUCAS extra files follow the source image-id order.

OpenET HLS is stored per point under hls_l30.tar and hls_s30.tar. The L30 sample uses B1, B2, B3, B4, B5, B6, B7, B9, B10, B11. The S30 sample uses the 13-name set B1–B12 including both B8 and B8A. band_valid says which channels exist on each date. Masks official_support_mask, valid_month_mask, and annual_mask select the support year, the target month, and the calendar year.

Sentinel-2:

  • Original nine tasks, sentinel2.tar: 10 bands in order B04, B02, B03, B08, B8A, B05, B06, B07, B11, B12, from COPERNICUS/S2_SR_HARMONIZED, cloud ≤ 90. Coverage is the points that were downloaded, not every NetCDF row. CDL and CropHarvest match their full point counts. LCMAP 17,780, GlanCE 10,100, GlobalTree 4,000, Hansen 1,135, Wildfire 840, LandTD 1,879, LCMAP-C 110.
  • CityRep and LUCAS: s2_bands plus s2_extra_bands. The CityRep sample is the same 10 bands plus B01 and B09.
  • OpenET: 12 L2A bands, B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, B12. No cirrus B10.

AlphaEarth is GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL, 64 dimensions.

Task File Rows Valid
CityRep alphaearth_annual_v2_native10m.npz 160,000 160,000, year 2024
LUCAS alphaearth_annual_v1.npz 203,569 203,569, years 2018 and 2019
OpenET alphaearth_annual_v1.npz 35,683 35,683, years 2017–2023

The original five classification and four change tasks also have hls_daily_with_extra.npz. It keeps the daily six-band-plus-NDVI archive and adds supplemental bands on that same time axis. A day is filled only when the original reflectance is finite. S30 is used when any S30 scene falls on that day; otherwise L30 is used. Overlapping tiles from the same pass are separated by distance from the point to the MGRS tile center, and the nearest center is kept. extra_bands_aligned is always (sample, time, band). chosen_sensor is 1 for S30 and 2 for L30. hls_daily_source.npz is the untouched daily archive. alphaearth.npz is the local GSE embedding for the five classification tasks. The four change tasks have no local AlphaEarth file.

Layout

classification/{lcmap,glance,globaltree,cdl,cropharvest}/
change/{hansen,wildfire,lcmap_c,landtd}/
new_classification/{cityrep,lucas}/
regression/openet/
ready_3day/
catalog/hls_extra_bands_v1_20261025/
examples/

ready_3day/{task}_fullband_3day.nc is one scene per 3-day bin, 18 bands. Reflectance is DN x10000, NDVI is unitless, and L30 B10/B11 are Celsius. data is (time, band, sample). Tasks: cdl, cropharvest, glance, globaltree, lcmap, cityrep, lucas, openet, ethct2020, ameriflux. ethct2020 is crop type with split train/test. ameriflux stores annual GPP as label (GPP_NT_VUT_REF) and gpp_dt_vut_ref; split is empty.

Extract a tar with tar -xf <file>.tar. Paths inside a tar are the original dataset directory or point-id directory.

Examples

examples/openet_readout.py scores the official OpenET split. It reads regression/openet/alphaearth_annual_v1.npz and regression/openet/sample_index.parquet (label, split, partition). --features hls_year_mean uses the calendar-year cube ready_3day/openet_fullband_3day.nc. The paper support window and the target month are the masks in the OpenET HLS tars, not that cube.

examples/classification_cube.py reads a classification cube. CityRep and LUCAS have split in the file. The five older classification cubes do not.

Downloads last month
317