Skip to content

Ambient Occlusion Objective

Ambient training learns from sparse EN4/ARGO observations without converting unobserved ocean cells into supervised targets. The model is conditioned on a further-corrupted version of the observed field, while loss is evaluated only where the original sparse field and the paired target are valid over ocean.

Notation

  • x: sparse depth-aligned EN4/ARGO values.
  • m: original sparse observation mask (x_valid_mask).
  • m': further-corrupted mask, with m' ≤ m.
  • y: paired dense target used to establish valid depth/ocean support.
  • v: target-valid mask (y_valid_mask).
  • l: ocean support (land_mask in the batch contract, where ocean is valid).

The model condition contains x ⊙ m', m', the configured dense surface channels, and coordinate/date context. In the current x0 objective, the supervision mask is

\[ M_{loss} = m \cap v \cap l, \]

and the supervised target on that support is the original sparse field x. Unobserved pixels are not assigned pseudo-labels by this objective.

Corruption policy

The local pixel preset uses:

model:
  parameterization: x0
  clamp_known_pixels: false
  ambient_occlusion:
    enabled: true
    further_drop_prob: 0.25
    apply_to_noisy_branch: true
    shared_spatial_mask: true
    min_kept_observed_pixels: 50
    require_x0_parameterization: true

The drop mask is shared spatially across field channels when configured. A minimum-support guard restores observations when random dropping would leave too few supervised pixels. With apply_to_noisy_branch=true, the further-corrupted support also masks the noisy target branch seen by the denoiser.

Constraints

  • Ambient mode requires x_valid_mask in each batch.
  • With require_x0_parameterization=true, any other parameterization is rejected.
  • Loss is normalized over valid weighted support, not the full patch area.
  • clamp_known_pixels=false means sampling does not overwrite predictions with observed x after each reverse step.
  • Exporters restore geospatial invalid support and final land nodata after model prediction; that output masking is separate from the training objective.

Relationship to other losses

The ambient diffusion term has weight 1.0 in the local preset. Optional sparse observation, profile-increment, GLORYS structure, and spectral terms are disabled. Auxiliary timestep weighting is therefore dormant until an auxiliary term is enabled. See Auxiliary losses.

The HPC synthetic-target and direct-GLORYS presets disable ambient mode and use ordinary dense supervision instead.