Data Overview

DepthDif learns to densify sparse in-situ ocean profiles into gridded temperature fields. The active GeoTIFF workflow can also train a salinity-only or joint temperature + salinity pixel model when --scenario salinity or --scenario joint is selected. The training data combines satellite surface context, profile observations, ocean reanalysis targets, and auxiliary sea-surface height plus sea-surface salinity and density on one shared global grid.

Use Data Sources for product-specific details, Dataset Downloads for raw and packaged download commands, Depth Alignment for vertical profile projection, Data Export for the GeoTIFF training store, and Data Contract for exact model tensor shapes.

Modalities

The current training workflow is built around these modalities:

Modality	Source	Variables	Role
Sparse profile temperature	EN4 / ARGO	TEMP, projected to GLORYS depths	Sparse subsurface conditioning signal.
Sparse profile salinity	EN4 / ARGO	PSAL_CORRECTED, projected to GLORYS depths	Auxiliary profile context stored with the training data.
Surface temperature	OSTIA	analysed_sst	Dense surface conditioning signal.
Target ocean temperature	GLORYS	thetao	Dense 3D supervision target.
Reanalysis salinity	GLORYS	so	Dense aligned ocean state variable stored alongside temperature.
Sea-surface height	Sea Level L4	adt	Dense surface-height context stored on the same grid.
Sea-surface salinity/density	SSS MULTIOBS	sos, dos	Dense surface salinity and density context stored on the same grid.
Land/ocean mask	Rasterized world polygons	output_land_mask	Defines patch candidates and provides final output cleanup support; model-facing land_mask is derived from finite GLORYS target support.

Temperature is kept physically in Kelvin in the exported GeoTIFF dataset, then converted or normalized by the loader as needed for model training. Salinity is stored in PSU and normalized only when the GeoTIFF dataloader is configured with --scenario salinity or --scenario joint.

Shared Axes

All dense exported rasters use the committed global 0.1 degree land-mask grid:

src/depth_recon/data/dataset_creation/data_download_raw/get_world/world_land_mask_glorys_0p1.tif

This grid gives a single spatial reference for GLORYS, OSTIA, sea-level, SSS fields, and ARGO profile locations. The vertical axis is the 50-level GLORYS depth coordinate. ARGO profiles are interpolated onto those same depths before they are rasterized into model patches.

Temporal Cadence

GLORYS weekly dates define the training timeline. For every GLORYS target date:

• GLORYS thetao and so are saved for that weekly date.
• OSTIA analysed_sst is saved as a centered 7-day mean around that date.
• Sea-level adt is saved as a centered 7-day mean around that date.
• SSS sos and dos are saved as centered 7-day means around that date.
• ARGO profiles are assigned to the nearest GLORYS weekly date inside the same centered temporal window.

Training View

The model-facing training sample is a patch cut from the shared grid. By default it contains sparse ARGO temperature observations, dense OSTIA surface temperature, the dense GLORYS temperature target, and masks that tell the model which values are observed, supervised, ocean, or missing. With --scenario salinity, the loader skips temperature tensors and returns normalized x_salinity and y_salinity plus salinity-specific masks. With --scenario joint, it returns both temperature and salinity tensors, and PixelDiffusionConditional stacks them at the model boundary.

The precise tensor contract is intentionally separated from this overview. See Data Contract for shapes, normalization, masks, and loader assembly rules.