OceanVision
PS 26066 MoES / INCOIS Smart India Hackathon 2026

Subsurface temperature, reconstructed from orbit

Ocean temperature at fifteen standard depths from the surface to 1000 m across the North Indian Ocean, daily at 0.25°, inferred from satellite surface fields alone. Every field on this page is computed on request. The page holds no precomputed data; each date you enter runs the model.

Headline result

Scored once, on a window nobody had looked at

Jul–Dec 2025 was left untouched while the model was developed, then scored a single time with the model frozen. Delayed-mode Argo, … levels, … days, satellite inputs only. The metric is profile-mean RMSE °C, meaning per-depth RMSE at the 15 mandated depths, then an unweighted mean over depths. Lower is better.

We are not the best product in this region. Two baselines beat us, Armor3D and GLORYS2V4, and both ingest the Argo profiles they are then scored against. See the full baseline table.

Per-depth detail, all fifteen levels

Proof of concept · PS bullet 6

Enter any date in the archive and watch the model run

Nothing here is cached in the page. Choosing a date sends that day's surface stack through the ensemble on the server. The first request for a date takes several seconds, after which the server holds that date in memory and depth changes are immediate. The elapsed time, and whether the answer came from cache, is printed under the map every time.

Step
Framing
Model input, satellite surface fields
    Depth m
    move over the map
    loading the archive bounds…
    –
    –°C

    Arrow keys step the date sideways and the depth up and down while the explorer is on screen.

    Temperature profile
    OceanVision

    Click anywhere on the map for a modelled column. Red rings mark delayed-mode Argo floats within ±3 days of the chosen date. Click one to compare against what it actually measured.

    Satellite embedding · PS bullet 2

    …

    The encoder's own representation of that pixel, read live: the activation immediately before the network's 1×1 output convolution. Bars above the axis are positive, below negative.

    Reference

    The evidence behind the number

    2023–24, delayed-mode Argo, … days, … identical rows. The GREP ensemble is published on the exact 101 × 241 PS grid, so there is no horizontal regridding. Lower is better.

    HYCOM GOFS 3.1 was scored on its own separate …-day sample: HYCOM …, GLORYS12 …, OceanVision …. This is a different row set and must not be ranked directly against the table above.

    Sealed Jul–Dec 2025, … rows: Armor3D …, OceanVision …, GLORYS12 ….

    The fair framing

    OceanVision is the only product in that table that never sees a subsurface observation. Every other one ingests the Argo profiles it is then scored against. Armor3D does so most directly: its declared Copernicus sources are ['In-situ observations', 'Numerical models', 'Satellite observations'] and its method (Guinehut et al. 2012) is an optimal interpolation of in-situ profiles onto satellite fields.

    What that leaves us

    Among the six products that do ingest in-situ data, OceanVision beats four: GLORYS12, HYCOM, ORAS5 and C-GLORS. The next real bar is GLORYS2V4 at …, and the whole remaining gap sits at 50 to 150 m.