1 degree track¶
This page documents how OceanBench constructs the 1 degree evaluation track. It is intended to remove ambiguity for both benchmark users and technical reviewers. We will go through how reference datasets are selected for the 1 degree track, the 1 degree GLORYS dataset exposed for training and how challenger datasets are interpolated.
How reference datasets are selected for the 1 degree track¶
For the 1 degree track, OceanBench uses dedicated 1 degree reference datasets for:
GLORYS reanalysis
GLO12 analysis
Then, when the challenger resolution is “1_degree”, OceanBench opens the precomputed 1 degree references. The related public dataset helpers are:
1 degree GLORYS dataset for training¶
For training outside the official evaluation workflow, OceanBench also exposes the 1 degree GLORYS dataset through:
oceanbench.datasets.reference.glorys_reanalysis_1_degree()
This public API is documented in:
How challenger datasets are interpolated¶
The 1 degree challengers exposed in oceanbench.datasets.challenger are:
glo12_1_degree()glonet_1_degree()wenhai_1_degree()xihe_1_degree()
Those are base challengers on a higher resolution (1/4 degree for GLONET, 1/12 degree for the others) that are interpolated to the 1 degree resolution. The corresponding public challenger dataset loaders are documented in:
Each of them applies the same interpolation logic:
Rename dataset dimensions and variables to the OceanBench standard names.
Infer the target 1 degree latitude and longitude grid from the native domain bounds.
Build a regular grid with 1.0 degree spacing and cell centres at
n + 0.5.Interpolate the dataset with
xarray.Dataset.interp.Mark the resulting dataset source with
resolution=\"one_degree\"so downstream staging and caching logic can distinguish it from the native-resolution challenger.