Datasets:

ONERA
/

S1_GRD_LC

license: other
pretty_name: Sentinel-1 GRD + Land Cover
tags:
  - sar
  - sentinel-1
  - land-cover
  - esa-worldcover
  - remote-sensing
  - webdataset
task_categories:
  - image-segmentation
  - image-to-image
size_categories:
  - 10K<n<100K

ONERA/S1_GRD_LC — Sentinel-1 GRD + Land Cover

For each high-resolution SAR imagette (UMBRA) from SARLO-80 dataset, we provide the co-located Sentinel-1 GRD 10 m backscatter (VV/VH) and the corresponding ESA WorldCover 10 m land-cover map.

Source

Each sample is anchored on a high-resolution UMBRA SAR imagette. From the SICD metadata we recover the imagette's 4 WGS84 corners (image-to-ground projection) and derive its geographic bounding box.

Adding Sentinel-1 GRD (method + format)

Method. From the 4 corners we build a north-up UTM grid at 10.0 m. We then query Sentinel-1 GRD (IW mode, VV + VH) from the Copernicus Data Space Ecosystem through the Sentinel Hub Process API, sampled directly on that grid. Backscatter is orthorectified (Copernicus DEM), expressed as GAMMA0_TERRAIN, and converted to dB. The composite is built over 2020-01-01 → 2021-01-01.

Because the S1 GRD is orthorectified (north-up product), there is no warp to the SAR geometry at this stage: S1 and land cover share the same grid (crs + transform + dimensions) and are therefore co-registered.

Format.

File	Type	Description
`{id}.map.s1_grd.npy`	float32 `(2, H, W)`	[VV, VH] in dB, UTM grid
`{id}.map.s1_vv.png` / `s1_vh.png`	uint8 `(H, W)`	dB visualization (VV∈[-25,0], VH∈[-30,-5])
`{id}.map.s1_rgb.png`	uint8 `(H, W, 3)`	RGB = [VV, VH, VV−VH]

Collection: SENTINEL1_IW. Units of the .npy: dB. PNGs are simple 8-bit renderings with a fixed dynamic range (for visual inspection only).

Land cover

Source: ESA WorldCover v100 (2020), sampled on the same grid as the S1 GRD.

File	Type	Description
`{id}.map.landcover.npy`	uint8 `(H, W)`	WorldCover class IDs
`{id}.map.landcover.png`	uint8 `(H, W, 3)`	official colorization

Class table:

ID	Class	Color
10	Tree cover	`#006400`
20	Shrubland	`#ffbb22`
30	Grassland	`#ffff4c`
40	Cropland	`#f096ff`
50	Built-up	`#fa0000`
60	Bare / sparse vegetation	`#b4b4b4`
70	Snow and ice	`#f0f0f0`
80	Permanent water bodies	`#0064c8`
90	Herbaceous wetland	`#0096a0`
95	Mangroves	`#00cf75`
100	Moss and lichen	`#fae6a0`

Without / with affine transform

Every sample is available in two co-registered frames:

map_grid (prefix map.) — north-up UTM 10.0 m map grid, without any affine transform.
sar_frame (prefix sar.) — the same S1 and land cover resampled into the native UMBRA SAR image frame using an affine transform estimated from the 4 corners (S1 bilinear, land cover nearest-neighbor). Aligned with the original SAR amplitude ({id}.umbra_sar.png). The 2×3 affine matrix is stored in meta.json (affine_map_to_sar_2x3).

Sample contents

{id}.map.s1_grd.npy       {id}.sar.s1_grd.npy
{id}.map.s1_vv.png        {id}.sar.s1_vv.png
{id}.map.s1_vh.png        {id}.sar.s1_vh.png
{id}.map.s1_rgb.png       {id}.sar.s1_rgb.png
{id}.map.landcover.npy    {id}.sar.landcover.npy
{id}.map.landcover.png    {id}.sar.landcover.png
{id}.umbra_sar.png        (original UMBRA SAR amplitude, SAR frame)
{id}.meta.json            (corners, bbox, grid, affine, S1/LC info)

Directory structure

train/chunk_XXX/shard-YYYYY.tar   (WebDataset)

Loading (WebDataset)

import webdataset as wds, numpy as np, io, json

url = "https://huggingface.co/datasets/ONERA/S1_GRD_LC/resolve/main/train/chunk_000/shard-00000.tar"
ds = wds.WebDataset(url)
for s in ds:
    vv_vh = np.load(io.BytesIO(s["map.s1_grd.npy"]))     # (2,H,W) dB
    lc    = np.load(io.BytesIO(s["map.landcover.npy"]))  # (H,W) classes
    meta  = json.loads(s["meta.json"])
    break

map_grid = without affine (UTM), sar_frame = with affine (SAR frame).