Field U-Net¶

ERA5ResidualRegressor reconstructs one surface wind-speed field. With ERA5, it predicts a physical correction to the reanalysis anchor:

\[ \hat v = v_{\mathrm{ERA5}} + f_\theta(x, m). \]

Without ERA5, the same model family predicts absolute wind. Scalar intensity and radii diagnosed from this image provide the field-only baseline in the paper comparison.

Inputs and architecture¶

The ERA5 configuration receives 23 data channels: ten GEO, nine ERA5 source/derived, storm-center distance, and three solar-time fields. The model appends condition validity, target-normalized ERA5 wind, and ERA5 validity, giving 26 U-Net inputs. The GEO-only model receives 14 data channels plus the condition mask.

The residual U-Net uses GroupNorm, SiLU, strided downsampling, bilinear upsampling, and encoder skips. Its zero-initialized residual head initially reproduces ERA5. Predictions are clamped nonnegative with no default upper cap.

Objective and diagnostics¶

The field-only presets use Huber loss in physical m/s, with a 2 m/s transition. ERA5 runs supervise the common SAR/ERA5-valid footprint. High-wind weights increase smoothly from 1× to 8× across 25–50 m/s. A robust top-0.5% inner-core peak term has weight 0.05. A separate 0.05-weighted off-swath term discourages unsupported corrections where ERA5 is valid; it is disabled without ERA5. Radial-profile and exceedance-area loss weights are zero in the default.

Validation reports physical errors, high-wind errors, structural diagnostics, and ERA5 skill where applicable. The default scheduler monitors val/peak_structure_score, while the experiment selects checkpoints by val/eye_structure_score. These are different from the training objective; inspect the resolved config when comparing runs.

Train with ERA5¶

uv run geo2wf-train experiment=intensity_comparison_unet \
  data.root=/path/to/paired \
  data.ibtracs_file=/path/to/ibtracs.ALL.list.v04r01.csv

This experiment uses the joint data module to keep the cohort aligned with scalar comparisons; it still learns only the field.

Train without ERA5¶

Use experiment=intensity_comparison_unet_no_era5 with the same local paths. This disables ERA5 inputs, residual addition, off-swath anchoring, and ERA5 comparison metrics, while retaining the cohort-availability filter.

For local checkpoint use, see commands; for metric interpretation, see evaluation.