FrontFinder

Keras models that detect synoptic and mesoscale weather features (fronts, troughs, drylines) on a global 0.25° grid from numerical weather prediction fields.

Files

File Description
v2.0.0.keras Current model (served at fronts.espr.ai). Previously named nbs_square_dilation_trof_tt_no_inst.keras.
uv_q_theta_e.keras Earlier model, kept for reference.
geopotential_at_surface_era5_0p25deg.npy Static surface geopotential (terrain) input required by v2.0.0.keras.

v2.0.0

A 2D UNet3+ that takes a flat 28-channel input and outputs per-pixel class probabilities.

Inputs

Fields on the ECMWF IFS open-data 0.25° grid (721 x 1440). The model contains its own normalization (Rescaling) layer, so inputs are given in physical units.

Pressure-level variables, each at 1000, 925, 850, 700, 500 and 300 hPa (6 levels x 4 variables = 24 channels):

  • equivalent potential temperature (theta-e), derived from IFS fields
  • u component of wind
  • v component of wind
  • specific humidity

Single-level variables (4 channels):

  • 10 m u wind
  • 10 m v wind
  • 2 m dewpoint temperature
  • surface geopotential (m² s⁻², ERA5 convention, not height in meters)

Surface geopotential is not part of the IFS open-data real-time stream. It is supplied by geopotential_at_surface_era5_0p25deg.npy: float32, shape (721, 1440), latitude 90 to -90, longitude 0 to 359.75. It was taken from the public ARCO-ERA5 archive (geopotential_at_surface, a time-invariant field).

Outputs

Class index order (index 0 is background):

Index Class
0 background
1 cold front
2 warm front
3 stationary front
4 occluded front
5 dryline
6 trough
7 tropical trough
8 instability line

The model was trained with an instability-line class ("no_inst" in its original filename), but that class is not used in the served product. The other seven classes are.

Usage

Load with Keras (keras.saving.load_model). Large domains are run in overlapping patches and stitched with blending. The serving pipeline (tiling, IFS ingest, derived variables, GeoZarr output) is described at fronts.espr.ai.

Attribution

Surface geopotential comes from ERA5 (Hersbach et al., ECMWF / Copernicus Climate Change Service). Please acknowledge ERA5 and ECMWF open data when using these inputs.

Downloads last month
116
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support