This dataset provides yearly binary atmospheric-river (AR) masks derived from MERRA-2 reanalysis, produced by the DARK detection algorithm. Although the method was developed with a polar (Antarctic) emphasis, the implementation operates globally and scans the full MERRA-2 grid. Output files follow the ARTMIP Tier-1 conventions (variable name, attributes, dimensions, CF metadata, and filename pattern).

Method 

Input field. The algorithm uses MERRA-2 total IVT directly from ARTMIP yearly files (DOI: 10.5065/D62R3QFS).

Overview of the steps :

1. Rolling seasonal threshold (98th percentile).
   For each leap day-of-year (1…366) we compute a 15-day rolling window climatology of IVT at each grid cell and take the 98th percentile across all selected days and years. Non-leap years are mapped to the leap DOY by shifting days ≥ 29 Feb (+1).

2. Exceedance mask.
   A grid cell is marked 1 if the local IVT of a given day exceeds its local and day-of-year threshold (98th percentile as defined in 1), else 0.

3. Area filtering (< 20, 000 km²).

   We treat longitude as periodic so features crossing that seam aren’t split and drop components with area < 20,000 km². Grid-cell areas are computed on WGS-84 using geocentric radius and the local Δlat/Δlon.

4. Length filtering ( < 2000 km) with seam-free centerlines.
   For each remaining component (AR candidate) and time:

   • We compute a skeleton of each AR candidate, build a pixel-adjacency graph, enumerate endpoint-to-endpoint paths, and select the one most aligned with the component’s PCA axis. The path is extended to the object edges and smoothed with a B-spline. Geodesic segment lengths are summed using WGS-84 (geopy) to get a sub-grid AR length. Components shorter than 2000 km are discarded.

5. Output assembly.
   We write a yearly byte (i1) mask ar_binary_tag(time,lat,lon), with 1 where an AR is detected and 0 elsewhere.

Files & Format

• One NetCDF per year:
  MERRA2.ar_tag.Buffet.<freq>.<YYYYMMDD-YYYYMMDD>.nc
  where <freq> is inferred (e.g., 3hourly, 6hourly).

• Variables

  • ar_binary_tag(time, lat, lon) — byte; 1 = AR, 0 = no AR.
    Attributes:

    • description = "binary indicator of atmospheric river"

    • scheme = "Buffet"

    • version = "1.0"

    • credits = "Developed by Victoire Buffet, Vincent Favier, and Benjamin Pohl"

  • time(time) — CF time with standard_name, long_name, units = "minutes since <year>-01-01 00:00:00", calendar = "standard", axis = "T".

  • lat(lat) — degrees_north; CF/ARTMIP attributes (standard_name, long_name, axis).

  • lon(lon) — degrees_east; CF/ARTMIP attributes (standard_name, long_name, axis).

• Compression: data variable uses NetCDF-4 deflate (zlib=True, complevel=1, shuffle=True). time is unlimited.

Spatial/Temporal Coverage

• Spatial: Global MERRA-2 native 2D grid (the algorithm targets polar/Antarctic ARs but evaluates the globe). 0.5° in latitude, 0.625° in longitude.

• Temporal: 1980–2017, using the time resolution provided by the input MERRA-2 IVT files (3-hourly).

Parameters (defaults)

• Rolling window: 15 days

• Percentile: 98th

• Minimum area: 20,000 km²

• Minimum length: 2,000 km

Reproducibility

Inputs

• Yearly MERRA-2 files with a 2D IVT variable:
  ARTMIP_MERRA_2D_{year}_3h.nc (or equivalent pattern)

Two stages

1. Build rolling‐DOY percentiles (once):
   percentile_98_IVT_rolling_15D_merra2.nc

2. Tag each year to produce ar_binary_tag yearly files.

Software

• Python ≥ 3.9; key libraries: xarray, numpy, scipy, scikit-image, geopy, networkx, matplotlib.

• Geodesics computed with geopy (WGS-84).

• Files written CF-compliant with netCDF4.

Credits

Developed by Victoire Buffet, Vincent Favier, and Benjamin Pohl at Institut des Géosciences de l'Environnement (Grenoble) and Biogéosciences (Dijon).

Keywords

Atmospheric Rivers; MERRA-2; ARTMIP; IVT; Antarctic; Polar Meteorology

Related Material

• ARTMIP format and Tier-1 conventions.

• MERRA-2 reanalysis documentation.