Skip to content
bryjudyPublic

About

Open global deep-sea minerals prospectivity map (0.05° cube, 125 layers), nodule detector and resource-model pipeline — MIT/CC-BY

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

Repository files navigation

Ocean Map — open deep-sea minerals prospectivity, nodule detector and resource-model pipeline

Everything here was built in one weekend (18–21 Sept 2026) from open data, and is released so anyone can use it for free. Code is MIT; our derived data, models and documents are CC BY 4.0; upstream data keep their own licences (see ATTRIBUTION.md).

Live atlas: https://storage.googleapis.com/oceanmap-atlas/index.html — 16 raster layers, EEZ boundaries, Cook Islands licences, top target hexes. Detector results: …/detector.html. Worked example (images + cores → block model → draft report): …/example.html.

What is in the open data bucket gs://oceanmap-open (public, no login)

path what size
oceanmap.zarr/ the 0.05° global feature cube, ~125 layers (terrain, geology, bottom water, productivity, legal flags, model outputs) ~2.7 GB
h3/h3_r5.parquet, h3/h3_r5_scored.parquet 1.49 M ocean hexes × all layers; scored version adds access/logistics/targeting 0.5 GB each
truth/ nodule presence/absence (Dutkiewicz 2020 compilation), NOAA-MMS chemistry (34k samples), SO268 box cores 3 MB
detector/coverage_v3_best.pt ResNet-34 coverage model (image → % nodule coverage), 1120-px input 85 MB
detector/coverage_v3_all_images.csv per-image coverage for the 34k SO268 frames 2 MB
tiles/ XYZ PNG tiles z0–6 for the atlas layers 0.5 GB
https://storage.googleapis.com/oceanmap-open/<path> downloads any object. The SO268 imagery itself (50 GB) is not mirrored — get it
from PANGAEA (datasets 957259–957273, CC-BY).

Headline results (all measured, none tuned on the test data)

  • Nodule occurrence model (LightGBM on ~100 layers, labels = 10.5k public samples): random-CV AUC 0.91, leave-one-basin-out AUC 0.77 (the honest number). Blind checks: P = 0.68 inside ISA nodule contract areas vs 0.17 ocean mean; 0.76–0.82 inside the Cook Islands licences.
  • Sedimentation rate is the dominant predictor, then bottom O₂, productivity, sediment thickness, benthic biomass, crustal age/spreading, ruggedness.
  • Grade (Ni+Cu+Co) does NOT transfer across basins (R² ≈ 0). Crust occurrence transfers at AUC 0.85.
  • Coverage detector v3: MAE 2.95 % coverage, R² 0.65 on 9,096 held-out images from 4 unseen profiles; 1.2–2.1 % on sparse/medium fields, ~8 % on packed facies. Classical thresholding baseline: 3.4 % / 0.53.
  • Coverage → kg/m²: within one rich facies, imagery gives ±2.5–3.5 kg/m² around a 19 kg/m² mean; box cores remain the anchor.

Repository layout

  • oceanmap/ — grid definition, source registry, ingest/derive scripts (GEBCO, GlobSed, Seton, Bio-ORACLE, GLODAP, polygons…), tile renderer, VM startup scripts.
  • models/ — occurrence/crust/grade models, transfer experiments, uncertainty ensemble, targeting, dossiers, paper outline, findings.
  • detector/ — coverage regressor (local, Beam, GCE variants), calibration tests, block model, report generator, results.
  • atlas/ — the web atlas source and 130+ layer quicklooks.
  • research/ — the three research reports (data catalogue, market/regulatory state Sept 2026, AUV + nodule ML).
  • docs/ — plan, running status log, weekend report.
  • data/truth/ — small truth tables; data/*blocks*.csv — Cook Islands licence blocks digitised from public SBMA PDFs.

Reproduce

cd oceanmap && uv venv .venv && uv pip install -p .venv/bin/python -e . lightgbm scikit-learn pykrige mercantile
# pull the cube instead of rebuilding: gcloud storage cp -r gs://oceanmap-open/oceanmap.zarr data/
python models/predict_global.py v3      # occurrence model + global prediction
python detector/block_model_v0.py       # worked example block model (needs data/rescue + SO268 CoMoNoD tabs; see sources.yaml)

Caveats

Global model at 5 km hexes with legacy, margin-biased labels; occurrence not grade; ISA geological data are confidential so there is no public ground truth of true resource statements; the detector's labels are an algorithm's output (CoMoNoD), not human annotation. Nothing here is a resource statement or investment advice.

Bryce Judy — bryce.judy@tarifflo.com

About

Open global deep-sea minerals prospectivity map (0.05° cube, 125 layers), nodule detector and resource-model pipeline — MIT/CC-BY

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages