README.md

---
license: cc-by-4.0
pretty_name: DepthDif GeoTIFF raster and aligned ARGO dataset
tags:
- oceanography
- argo
- glorys
- ostia
- sea-level
- sea-surface-salinity
- geotiff
- zarr
configs:
- config_name: profile-index
  data_files:
  - split: profiles
    path: indices/profiles.parquet
  - split: variables
    path: indices/variables.parquet
---

<p align="center">
  <img src="assets/figures/banner_depthdif.png" width="85%" alt="Banner Image" />
</p>

<p align="center">
  <a href="https://depthdif.donike.net/">
    <img src="https://img.shields.io/badge/Project-Website-0b2e4f?style=for-the-badge" alt="Open Documentation" />
  </a>
  <a href="https://github.com/simon-donike/DepthDif/">
    <img src="https://img.shields.io/badge/Project-GitHub-0b2e4f?style=for-the-badge" alt="Open GitHub" />
  </a>

</p>


# DepthDif GeoTIFF Raster and Aligned ARGO Dataset

This dataset package contains the model-ready Ocean variables (ARGO submarine data, sea surface height, sea surface temperature and
salinity, as well as GLORYS reanalysis information for 50 depth levels. The ARGO data has been projected onto the GLORYS grid in order
to build a ML-ready dataset. The intention is that users can create tensors easily for CV-inspired ML approaches to ocean-variable
reconstruction. While traditionally this problem has been approached point-wise, this dataset enables easy training of CV models such
as inpainitng-inspired approaches. 

## Dataset Preview
Rasterized Argo-profiles aligned with GLORYS product, examples:

<p align="center">
  <img src="assets/data/geotiff_dataset_random100_surface.png" width="85%" alt="Random surface-level training dataset patches" />
</p>

## Related Dataset

[OceanTACO](https://huggingface.co/datasets/nilsleh/OceanTACO) is a concurrent
development of a related global ocean dataset built from overlapping source
products. We focus on L4 products and make some assumptions for the users in order to enable
comparative benchmarks. If you want easy access to the actual underlying data, check out this dataset.


## Layout

The `rasters/` directory is intentionally at the repository root. It contains
the aligned uint8 GeoTIFF products used by the pixel-space dataloader. The
compact `argo/argo_profiles_on_grid.zarr` store is the grid-indexed ARGO input
used by that dataloader.

The package intentionally contains two ARGO Zarr stores with different roles.
`argo/argo_profiles_on_grid.zarr` is the compact grid-indexed store meant to be
used together with the GeoTIFF raster dataset. `data/argo_glors_ostia_ssh.zarr`
is the full enriched profile-level store and holds the complete ARGO
collocation dataset, including the sampled GLORYS, OSTIA, sea-level, and
sea-surface-salinity context.


## Raster Products

All GeoTIFF rasters are exported on the GLORYS 0.1 degree global grid
(`EPSG:4326`, 3600 x 1800 pixels, west-to-east longitudes from -180 to 180 and
north-to-south latitudes from 90 to -90). The current package contains 1283
weekly target dates per raster product, from 2000-01-01 through 2024-07-26.
Files are named `<variable>_YYYYMMDD.tif`.

The GLORYS variables are depth-resolved 50-band GeoTIFFs:

- `rasters/glorys/thetao/`: potential temperature, encoded as Kelvin.
- `rasters/glorys/so/`: salinity, encoded as PSU.

The surface products are single-band GeoTIFFs aggregated to the same weekly
target dates with a centered 7-day mean window:

- `rasters/ostia/analysed_sst/`: OSTIA analysed sea-surface temperature in Kelvin.
- `rasters/sealevel/adt/`: absolute dynamic topography in meters.
- `rasters/sss/sos/`: sea-surface salinity in PSU.
- `rasters/sss/dos/`: sea-surface density in kg/m3.

Raster pixels are stored as `uint8` with `255` reserved for nodata. Valid codes
`0..254` are linearly decoded using the stretch ranges in `manifest.yaml`;
per-file statistics, source filenames, compression, target dates, and the full
depth axis are also recorded there.

## PyTorch Dataset and DataLoader

The repository includes a standalone loader package in `depthdif_dataset/`.
It is designed to work directly from a local Hugging Face dataset checkout
without installing the full DepthDif training repository.

Install the loader dependencies in your environment:

```bash
pip install -r requirements-loader.txt
```

Minimal PyTorch usage from the dataset repository root:

```python
from depthdif_dataset import ArgoGeoTIFFGriddedPatchDataset, build_dataloader

dataset = ArgoGeoTIFFGriddedPatchDataset(
    geotiff_root_dir=".",
    split="all",
    tile_size=128,
    patch_stride=128,
    max_dates=1,
    metadata_cache_dir=None,
)
loader = build_dataloader(dataset, batch_size=2, num_workers=0)
batch = next(iter(loader))
print(batch["x"].shape, batch["eo"].shape, batch["land_mask"].shape)
```

The default sample contains normalized tensors for sparse ARGO temperature
input (`x`), dense GLORYS temperature target (`y`), dense surface context
(`eo`), validity masks, the ocean `land_mask`, target `date`, patch
coordinates, and a small `info` dictionary. Set `include_salinity=True` to
add `x_salinity`, `y_salinity`, and their masks. Use `eo_source="sss"` with
`eo_var_name="sos"` to use sea-surface salinity as the surface context
instead of OSTIA SST.

A runnable smoke test is included:

```bash
python examples/torch_dataloader.py --root . --date-start 20000101 --max-dates 1 --batch-size 2
```

ARGO-support filtering for train/validation splits requires counting profile
overlap per patch/date. The loader skips that scan for `split="all"` unless
`count_argo_support=True` or `--require-argo` is set in the example script.

### Patch Grid, Overlap, and Land Filtering

The loader builds square patches on the fixed 0.1 degree GLORYS grid. With the
default `tile_size=128`, one sample covers a 12.8 x 12.8 degree region and has
50 depth bands. `patch_stride` controls how far the next patch starts:

- `patch_stride=128`: non-overlapping global tiles.
- `patch_stride=96`: 32-pixel overlap, or 3.2 degrees at 0.1 degree resolution.
- `patch_stride=32`: 96-pixel overlap, or 9.6 degrees at 0.1 degree resolution.

Smaller strides create more samples and make neighboring patches share more
context, but they also increase row-index size and training time. If you use
overlap for train/validation splits, prefer a temporal validation split such as
`val_year=2018`; spatial random splits with overlapping patches can leak nearly
identical context between train and validation.

<p align="center">
  <img src="assets/data/patch_grid/patch_grid_global_overview.webp" width="78%" alt="Global patch grid overview" />
</p>

<p align="center">
  <img src="assets/data/patch_grid/patch_overlap_regional_example.webp" width="72%" alt="Regional example of overlapping patch windows" />
</p>

`max_land_fraction` filters out patches that are mostly land. The default
`max_land_fraction=0.30` keeps patches with at least 70 percent ocean pixels.
Increase it for coastal finetuning, or decrease it for open-ocean training.
The returned `land_mask` tensor uses `1` for ocean/support pixels and `0` for
land or unavailable support.

<p align="center">
  <img src="assets/data/patch_grid/land_fraction_filter_examples.webp" width="72%" alt="Examples of patch filtering by land fraction" />
</p>

### Common Loader Recipes

Use a small date slice for quick inspection without building a large metadata
cache:

```python
dataset = ArgoGeoTIFFGriddedPatchDataset(
    geotiff_root_dir=".",
    split="all",
    date_start=20000101,
    max_dates=1,
    tile_size=128,
    patch_stride=128,
    metadata_cache_dir=None,
)
```

Use overlapping patches for training:

```python
dataset = ArgoGeoTIFFGriddedPatchDataset(
    geotiff_root_dir=".",
    split="train",
    tile_size=128,
    patch_stride=32,
    val_year=2018,
    metadata_cache_dir="depthdif_cache",
)
```

Use salinity targets and ARGO salinity inputs:

```python
dataset = ArgoGeoTIFFGriddedPatchDataset(
    geotiff_root_dir=".",
    split="all",
    include_salinity=True,
    output_fields=("temperature", "salinity"),
    metadata_cache_dir=None,
    max_dates=1,
)
```

Use sea-surface salinity as the surface context instead of OSTIA SST:

```python
dataset = ArgoGeoTIFFGriddedPatchDataset(
    geotiff_root_dir=".",
    split="all",
    eo_source="sss",
    eo_var_name="sos",
    include_salinity=True,
    metadata_cache_dir=None,
    max_dates=1,
)
```

Use synthetic sparse observations sampled from the dense GLORYS target instead
of real ARGO profiles:

```python
dataset = ArgoGeoTIFFGriddedPatchDataset(
    geotiff_root_dir=".",
    split="all",
    synthetic_mode=True,
    synthetic_pixel_count=250,
    require_argo_for_all=False,
    metadata_cache_dir=None,
    max_dates=1,
)
```

Filter samples to patch/date rows that contain ARGO support. This does extra
index work and is worth caching when used repeatedly:

```python
dataset = ArgoGeoTIFFGriddedPatchDataset(
    geotiff_root_dir=".",
    split="all",
    count_argo_support=True,
    require_argo_for_all=True,
    metadata_cache_dir="depthdif_cache",
    date_start=20000101,
    max_dates=1,
)
```

The ARGO support map below shows why this filter changes the row distribution:
many open-ocean patches have dense support, while other valid ocean patches may
have no profiles for a given weekly target date.

<p align="center">
  <img src="assets/data/argo_valid_pixels_per_patch.webp" width="78%" alt="ARGO valid pixels per training patch" />
</p>

### Sample Dictionary

The default dataset returns normalized tensors:

- `x`: sparse ARGO temperature input, shape `(50, tile_size, tile_size)`.
- `y`: dense GLORYS temperature target, shape `(50, tile_size, tile_size)`.
- `eo`: dense surface context, shape `(1, tile_size, tile_size)`.
- `x_valid_mask` and `y_valid_mask`: boolean temperature masks.
- `x_valid_mask_1d`: depth-collapsed sparse-input support mask.
- `land_mask`: ocean/support mask, shape `(1, tile_size, tile_size)`.
- `date`: target date as `YYYYMMDD`.
- `coords`: patch center latitude and longitude when `return_coords=True`.
- `info`: patch/date metadata when `return_info=True`.

When `include_salinity=True`, samples also include `x_salinity`,
`y_salinity`, `x_salinity_valid_mask`, `y_salinity_valid_mask`, and
`x_salinity_valid_mask_1d`.

Temperatures are normalized from Celsius using the DepthDif training statistics;
salinity is normalized from PSU. To recover physical units:

```python
from depthdif_dataset import salinity_normalize, temperature_normalize

temperature_c = temperature_normalize(mode="denorm", tensor=batch["y"])
salinity_psu = salinity_normalize(mode="denorm", tensor=batch["y_salinity"])
```

## ARGO Alignment Examples

ARGO profiles are projected onto the fixed 50-level GLORYS depth axis before
spatial rasterization. The examples below show the grid-indexed ARGO
representation and profile-level alignment quality.

<p align="center">
  <img src="assets/data/argo_on_glorys_grid_3D.gif" width="70%" alt="Depth-aligned ARGO values on the GLORYS grid" />
</p>

<p align="center">
  <img src="assets/data/profile_comparison_good_alignment.png" width="72%" alt="Example of ARGO-to-GLORYS profile alignment" />
</p>

The full enriched profile-level ARGO collocation dataset is available at:

```python
import xarray as xr

ds = xr.open_zarr("data/argo_glors_ostia_ssh.zarr", consolidated=None)
```

The lightweight Parquet indices are included for preview and filtering:

```python
import pandas as pd

profiles = pd.read_parquet("indices/profiles.parquet")
variables = pd.read_parquet("indices/variables.parquet")
```

Coverage:

- Raster target dates: 2000-01-01 to 2024-07-26
- Raster target date count per product: 1283
- Enriched ARGO profiles: 9485977
- Enriched ARGO profile dates: 2000-01-01 to 2024-07-31
- Compact grid-indexed ARGO profiles: 9451644
- GLORYS depth levels: 50

The package is released as CC BY 4.0. Upstream product licenses and citation
requirements for EN4/ARGO, GLORYS, OSTIA, sea-level, and
sea-surface-salinity products still apply; see `LICENSE` for the attribution
notice.