TC-SAR-Wave: An event-based Sentinel-1 SAR-derived significant wave height dataset for tropical cyclones

The primary estimator uses SAR-derived descriptors and acquisition geometry at inference; background SWH and wind are not its inference inputs. Ancillary/background variables must not be confused with inputs to the primary estimate.Native-resolution Sentinel-1 imagery, raw satellite-altimeter and in situ matchup tables, training-reference fields, and model checkpoints are not redistributed in this product ZIP. The included window-mean VV/VH variables are not native-resolution SAR images.

ERA5 (doi.org/10.24381/cds.adbb2d47) and WAVERYS (doi.org/10.48670/moi-00022) contribute wave-reference information during model development; their roles differ from that of the released SAR-only estimate.2 File namingThe archive filename `TC-SAR-Wave_v1.0.zip` identifies the dataset and release version. Scene files follow the relative path `data/events/<EVENT_NAME_YEAR>/<track_id>/L3_WAVE/<scene_id>_l3_wave.nc`.

The event directory contains the inherited cyclone-event name and year, and `track_id` identifies the parent archive's within-event track group. It is not a replacement for the Sentinel-1 orbit identifier. The `scene_id` retains the source Sentinel-1 identifier, including satellite platform, acquisition mode/product identifier, UTC acquisition start and end timestamps, and orbit/product identifiers.

The suffix `_l3_wave.nc` identifies the combined scene wave-product file; the file includes both selected L2 descriptors and L3 estimates. The event-year label does not replace the acquisition timestamp. File paths, scene identities, times, sizes, and quality counts are indexed in `metadata/scene_inventory.csv`.3 Spatial and temporal informationThe acquisition coverage extends from 2015-09-26T15:59:47Z to 2026-02-26T18:21:08Z.

Each file represents an individual Sentinel-1 scene, with the representative scene time stored in UTC; acquisition start and end timestamps are retained in the scene identifier. Sampling is irregular and does not provide continuous temporal coverage of each cyclone.The released scene grids span approximately 172.524 degrees W to 179.815 degrees E and 46.880 degrees S to 68.620 degrees N across the North Atlantic, western North Pacific, eastern North Pacific, North Indian, South Indian, and South Pacific basins.

These are bounding limits of discontinuous scene coverage, not a continuously observed global rectangle. Each cell is geolocated by two-dimensional `latitude` and `longitude` arrays in decimal degrees. The `x` and `y` dimensions index the scene grid rather than defining a regular projected coordinate system.

Approximately 1 km posting must not be interpreted as 1 km independent spatial resolution: estimates use overlapping 512 by 512 GRD-pixel windows, with physical support described in the files.4 Missing data and quality controlsThe products preserve observation gaps and rejected retrievals rather than constructing a gap-free ocean field. Primary SWH is unavailable where input support, ocean fraction, radiometry, mandatory quality criteria, or applicability requirements are not satisfied.

Missing floating-point values are represented by NaN and must not be replaced with zero. Users should jointly inspect `swh_valid_mask`, `swh_product_status`, and `swh_qc_flag`. Status 0 denotes accepted in-domain retrievals; status 1 denotes accepted challenge-domain retrievals requiring caution.

Statuses 2-5 are not accepted quantitative primary estimates. Missing feature values can occur separately from missing SWH and should retain their individual meanings.Across the released files, 87,163,308 of 125,600,117 scene-grid postings contain accepted primary SWH estimates. The remaining 38,436,809 postings (30.60%) are not accepted primary estimates.

This proportion refers to all grid postings within the distributed scenes, including masked and unsupported areas; it is not a fraction of the global ocean, a measure of independent observations, or evidence of incomplete downloads. At the scene level, no-product and unsupported-polarization cases are explicitly retained in `metadata/coverage_ledger.csv`, without fabricated empty quantitative products. Users must account for irregular sampling, overlapping spatial support, quality exclusions, and limited high-wave validation when interpreting coverage and uncertainty diagnostics.