calebrob6/offshore-wind-s1

Offshore wind turbines in Sentinel-1 time series

This dataset contains Sentinel-1 radar image time series of offshore wind farms in the English Channel and off the coast of England, with masks that mark the pixel of each wind turbine.

Contents

There are 367 patches. Each patch is 512 x 512 pixels at 10 m resolution (5.12 km across), and the patches do not overlap. The patches were placed to cover 3,275 selected wind turbines from OpenStreetMap. They also contain 44 OpenStreetMap turbines from outside the selection (at Robin Rigg in the Solway Firth and in Liverpool Bay), so the masks mark 3,319 turbines.

Each patch has two time series, one from ascending orbits and one from descending orbits. Each time series is a float32 array with shape (T, C, H, W) = (5, 3, 512, 512):

• T is the 5 most recent acquisitions up to 25 October 2026 that come from a single relative orbit and fully cover the patch. They are in chronological order. Consecutive images are 6 days apart (one pair is 12 days apart), and the earliest image is from 20 September 2026.
• C is VV, VH and VV/VH. VV and VH are terrain-corrected backscatter (gamma0) in linear units, taken unchanged from the Microsoft Planetary Computer sentinel-1-rtc collection. The third channel is VV divided by VH.

Missing pixels are NaN. At least 99.99% of the pixels in every image are valid.

Each patch also has two masks with shape (512, 512). A pixel is 1 if it contains a turbine and 0 otherwise.

• all includes every turbine.
• constructed leaves out the 607 turbines that OpenStreetMap tags as under construction. 102 patches contain only these turbines (for example in East Anglia THREE and Hornsea 3), so their constructed mask is empty. Some of these tags are out of date: all 62 turbines at Dieppe-Le Tréport are tagged as under construction but are visible in the images. At Hornsea 3 and East Anglia THREE about a third of the turbines are visible, and at Courseulles-sur-Mer only weak returns are visible.

Every image also has metadata: the acquisition time, the satellite (Sentinel-1C or Sentinel-1D), the relative and absolute orbit, the STAC items and asset URLs it was read from, and the fraction of valid pixels.

Turbines were selected from OpenStreetMap nodes tagged as wind generators. A turbine is included if it is at sea and is either inside the English Channel or closer to England than to any other coast. This includes the French wind farms in the Channel and leaves out turbines in Welsh and Scottish waters.

Files

File	Contents
s1_patches.zarr	Zarr store (format 2) with the images, masks, metadata and a turbine table
offshore_wind_turbines.gpkg	The 3,275 turbine points with their OpenStreetMap tags

The main arrays in the Zarr store are listed below. Patches are in the same order in every array.

Array	Shape	Contents
ascending/image, descending/image	(367, 5, 3, 512, 512)	Images, with dimensions (patch_id, time, band, y, x)
ascending/datetime, descending/datetime	(367, 5)	Acquisition times. The other per-image metadata arrays in these groups have the same first two dimensions.
masks/all, masks/constructed	(367, 512, 512)	uint8 turbine masks
patch_id, epsg, transform, bounds	(367, ...)	Patch ID, UTM zone (EPSG:32630 or EPSG:32631), affine transform and bounds in that zone
turbines/*	(3319,)	OpenStreetMap ID, longitude, latitude, patch index, pixel row and column, whether the turbine is under construction, and whether it is one of the 3,275 selected turbines

Loading

xarray can open the store directly from the Hub and only reads the data you select:

import xarray as xr

tree = xr.open_datatree("hf://datasets/calebrob6/offshore-wind-s1/s1_patches.zarr", engine="zarr")
image = tree["ascending"]["image"].sel(patch_id="p0201")  # (5, 3, 512, 512)
dates = tree["ascending"]["datetime"].sel(patch_id="p0201").values
mask = tree["masks"]["constructed"].sel(patch_id="p0201")

To work with a local copy, download it with hf download calebrob6/offshore-wind-s1 --repo-type dataset --local-dir offshore-wind-s1.

Processing

The patch grids are aligned with the 10 m grid of the Sentinel-1 RTC scenes, so most images are copied from the scenes without resampling. 80 of the 3,670 images come from scenes stored in a neighbouring UTM zone and were reprojected with bilinear resampling. 140 images span two consecutive scenes from the same pass and were combined into one image. The read_method and stac_item_id arrays record this for each image.

Licence

The images are derived from the Planetary Computer sentinel-1-rtc collection, which is licensed under CC BY 4.0. They contain modified Copernicus Sentinel data, and the terrain correction was done by Catalyst. The turbine points and masks are derived from OpenStreetMap data, © OpenStreetMap contributors, which is available under the Open Database License (ODbL).