Omer-Sela/upper-level-trough-ridge-detection-data

Detection of Upper-Level Troughs and Ridges Using Deep Learning — Dataset

Application in the Mediterranean

Expert annotations, model-ready ERA5 fields, and the derived climatology accompanying Detection of Upper-Level Troughs and Ridges Using Deep Learning – Application in the Mediterranean by Ofir Ariel, Omer Sela, Hadas Saaroni, and Baruch Ziv.

Project resources

Resource	Link
Interactive project page	Explore every benchmark scene and the seasonal climatology
Source code	Sela-Omer/upper-level-trough-ridge-detection
Dataset	Omer-Sela/upper-level-trough-ridge-detection-data
Model checkpoints	Omer-Sela/upper-level-trough-ridge-detection-models

Dataset summary

The expert-labelled benchmark contains 600 trough scenes from 2018–2020 and 200 ridge scenes from 2018. Every sample uses a regular 1° grid spanning 20–60° N and 20° W–50° E (41 × 71 grid cells). The spline catalog contains 2,579 expert trough axes and 683 expert ridge axes.

Ridge timestamps are a subset of the 2018 trough timestamps. They remain separate task rows because their expert annotations and stored atmospheric fields are task-specific.

Repository contents

Path	Description
data/training_data.nc	Z500, U500, V500, and training targets for all 800 task rows.
data/expert_annotations.nc	Thin raster representations of the expert-labelled axes.
metadata/samples.parquet	Sample identifiers, timestamps, task labels, and cross-validation folds.
metadata/axes.parquet	GeoParquet 1.1 expert B-splines with WKB LineStrings, spatial bounds, knots, control points, and evaluated curves.
metadata/dataset_manifest.json	Dimensions, counts, and SHA-256 checksums.
metadata/INTERFACE.md	Field-level data interface and conventions.
climatology/seasonal_climatology_1979_2024.nc	Annual and seasonal model-derived axis-frequency fields.
climatology/provenance.json	Climatology periods, sampling, model revisions, and processing provenance.
archive/manifest.json	Machine-readable index for the 1979–2024 scene archive and its monthly row groups.
archive/fields/year=YYYY.parquet	Annual Z500, U500, and V500 fields at 00 and 12 UTC.
archive/axes/task=TASK/year=YYYY.parquet	Annual production-model axes as GeoParquet LineStrings.
archive/provenance/year=YYYY.json	Per-year file hashes and pinned checkpoint provenance.

Data interface

data/training_data.nc contains float32[sample, latitude, longitude] inputs z500, u500, and v500, together with the derived targets axis_target, supervision_mask_r5, side_label_r5, and signed_split_r5.

Coordinates include sample_id, task (0 for trough and 1 for ridge), valid_time, latitude, and longitude. Z500 is geopotential height in metres; wind components are in m s⁻¹.

Expert axes in metadata/axes.parquet are directly readable by GeoPandas and QGIS as LineString geometry in OGC:CRS84. The original grid-coordinate spline fields are retained for exact reproduction of the paper's curve-level metrics.

For fold k, rows whose cv_fold equals k form the validation split; all remaining rows of the same task form the training split. Fold assignment is deterministic and uses seed 12345.

Loading the dataset

from huggingface_hub import hf_hub_download
import xarray as xr

path = hf_hub_download(
    repo_id="Omer-Sela/upper-level-trough-ridge-detection-data",
    repo_type="dataset",
    filename="data/training_data.nc",
)
dataset = xr.open_dataset(path)

The code repository also provides a task-aware dataset loader and command-line workflows that download and cache the required artifacts automatically.

Historical scene archive

The historical archive covers all 33,604 ERA5 scenes at 00 and 12 UTC from 1979 through 2024. It provides 149,853 trough axes and 145,957 ridge axes from the task-specific production checkpoints, together with the corresponding Z500, U500, and V500 fields. Field arrays use the same 41 × 71 domain and units as the benchmark.

Annual files contain monthly Parquet row groups. The root archive/manifest.json records paths, sizes, SHA-256 hashes, grid conventions, model revisions, and the UTC extent of each row group. This supports selective HTTP range reads from research software and the interactive project page.

import geopandas as gpd

axes = gpd.read_parquet(
    "https://huggingface.co/datasets/"
    "Omer-Sela/upper-level-trough-ridge-detection-data/"
    "resolve/main/archive/axes/task=trough/year=2024.parquet"
)

Pin a release tag or commit instead of main for reproducible use. The scene-level archive uses the production checkpoints; the seasonal climatology below is the paper's separately generated ensemble aggregate.

Seasonal climatology

climatology/seasonal_climatology_1979_2024.nc covers 1 January 1979 through 31 December 2024 at 00 and 12 UTC. Its task coordinate contains trough and ridge; its period coordinate contains ANN, DJF, MAM, JJA, and SON.

The principal variables are intersection_count, intersection_frequency, and n_times. weighted_frequency additionally weights detected axis pixels by non-negative curvature and wind speed. z500_mean provides the corresponding ERA5 annual and seasonal mean geopotential-height fields. See climatology/provenance.json for source-cache checksums and processing parameters.

Source data and attribution

The atmospheric fields are derived from ERA5 hourly data on pressure levels, produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) and retrieved through the Copernicus Climate Data Store. ERA5 is distributed under the Creative Commons Attribution 4.0 International licence.

For this release, the source fields were spatially and temporally subsetted, provided on a regular 1° grid, restricted to 500 hPa and 00/12 UTC, converted to the documented variables and units, and repackaged as NetCDF and Parquet.

ERA5 reanalysis data, © European Centre for Medium-Range Weather Forecasts (ECMWF), retrieved through the Copernicus Climate Data Store. Contains modified Copernicus Climate Change Service information (1979–2024). Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains.

Scientific users should cite both the ERA5 dataset DOI above and Hersbach et al. (2020), The ERA5 global reanalysis.

Limitations

The benchmark samples the Mediterranean and adjacent North Atlantic and European region; it is not globally representative. Axis annotations encode expert synoptic interpretation rather than a unique physical ground truth. Ridge labels cover one year and therefore provide less temporal diversity than the trough labels.

Citation

If you use this dataset, please cite the accompanying paper and the ERA5 source identified above.

License

The authors' original expert annotations, model outputs, metadata, and dataset compilation are licensed under Creative Commons Attribution 4.0 International. The ERA5-derived atmospheric fields are distributed under the upstream CC BY 4.0 licence and retain the ECMWF and Copernicus attribution given above. No authorship or ownership claim is made over the underlying ERA5 data.