ohw26_proj_eutrophos

This is the project repository from OceanHackWeek 2026 (Bamfield, BC).

• README.md — this file, the project overview.
• contributor_folders — personal working folders; the project's development history.
• data — shared datasets. Large regenerables (raw archives, hourly ERA5, fetch caches) are gitignored; see the Datasets table for what ships in-repo.
• final_notebooks — polished, re-runnable notebooks that reproduce the headline claims.
• scripts — the pipeline: fetch → derive → classify → model → bake the dashboard.
• docs — the dashboard itself (GitHub Pages serves this folder).

Project Name and one-line description

VI Near-Bottom O₂ — mapping, classifying, and modeling near-bottom dissolved oxygen around Vancouver Island, with every claim reproducible from this repository.

Live dashboard: https://oceanhackweek.org/ohw26_proj_eutrophos/

Collaborators

Name	Role
Ernest Ohandja Nomo	Ocean Data Collector
Ashmeet Singh	Prospective PHD Student
Ejay Aguirre	Undergraduate Student in DS

Planning

• Initial idea: map oxygen status around Vancouver Island from every source we could reach, then use a model to fill the space between measurements.
• Working notebooks: final_notebooks/ (01 → 03).
• Dashboard build and deployment notes: DASHBOARD.md.

Background

Near-bottom oxygen around Vancouver Island is set by a mix of natural deep-Pacific (oxygen-minimum-zone) water, wind-driven upwelling on the outer shelf, and stratification and renewal cycles in the inside waters and fjords. Some basins (Saanich Inlet, Alberni Inlet, Barkley Canyon) are naturally hypoxic or anoxic for much of the year; others flirt with thresholds seasonally. We wanted one map that shows the observed status honestly — and a model that extends it, with its uncertainty on display.

Goals

1. Assemble every reachable near-bottom O₂ record (DFO, ONC, Community Fishers, OOI) into one consistent, mL/L, near-bottom dataset.
2. Classify every site under three lenses (typical low / worst case / seasonal) with confidence grades.
3. Train and honestly validate a model predicting O₂ from weather, river flow, depth, and season — with calibrated uncertainty.
4. Ship it all as a fetch-free static dashboard anyone can open.

Datasets

Dataset	Source	Licence / note
dfo_casts.csv (17,399 casts)	DFO IOS via CIOOS Pacific ERDDAP	OGL-Canada
cf_casts.csv (846 casts)	ONC Community Fishers	ONC data policy
site_daily.csv (16 cabled sites)	Ocean Networks Canada	ONC data policy
dfo_moorings_daily.csv (21 clusters)	DFO BC Shelf Mooring Program via CIOOS	OGL-Canada
ooi_oxygen_daily.csv (WA line; training only)	NSF OOI Data Explorer	OOI policy
bc_lighthouses_daily.csv (1914–2019)	DFO BCSOP via CIOOS	OGL-Canada
GEBCO 2026 grid subset (.nc)	GEBCO/BODC	public domain
training_table.csv + forcing_features_daily.csv	this repo (committed)	MIT; the model retrains from these alone
model_predictions.csv	regenerated locally by scripts/train_oxygen_model.py	not committed (~50 MB); bands ship baked in data.js

Full citations, licences, and required acknowledgments: DATA_SOURCES.md.

Workflow/Roadmap

data/derived/*.csv + GEBCO .nc
        │
        ├── scripts/build_training_table.py ──▶ training_table.csv
        ├── scripts/train_oxygen_model.py ──▶ model_predictions.csv (bands)
        ├── scripts/predict_grid.py ──────▶ docs/model_grid/ frames
        ▼
scripts/build_dashboard.py  ──▶  docs/   (baked, fetch-free static site)
        │                          ▲
        ├── scripts/test_dashboard.js — 92 headless checks (jsdom+Leaflet)
        └── final_notebooks/ 01→03 — provenance, classification
                                      reproduction, band verification
GitHub Pages serves docs/ on every push to main.

Reproduce this work

pip install -r requirements.txt
python scripts/build_training_table.py     # or skip: table is committed
python scripts/train_oxygen_model.py       # exams + station bands (~3 min)
python scripts/predict_grid.py             # modeled surface frames (~3 min)
python scripts/build_dashboard.py          # bake the site
python -m http.server -d docs 8000         # open localhost:8000

Refreshing the observations additionally needs credentials (ONC token as ONC_TOKEN, Copernicus ~/.cdsapirc) and the scripts/fetch_*.py chain — everything is cached and resumable.

Results/Findings

• An interactive dashboard: 43 classified sites, ~17.6k individual casts, GEBCO relief/isobaths, model bands at 656 stations, and a 5-frame modeled surface with confidence shown as transparency.
• Model skill (station-blocked validation): typical error 0.46 mL/L, correct 4-class status 80% at never-seen stations, vs 1.08 mL/L for a region×month climatology; prediction bands calibrated (×1.55) to 80% held-out coverage.
• 43/43 sites reproduce their published classification from the raw committed data (final_notebooks/02).
• The dashboard's numbers are verified in committed notebooks and a 92-check headless UI suite — not slides.
• SEVIP anoxia: the Saanich sill sensor logged 14 days at or below 0.1 mL/L in fall 2024 — surfaced by the classification, visible on the dashboard, reproduced in final_notebooks/02.

Lessons Learned

• Bake, don't fetch: a static data.js outlives every API hiccup.
• Verify claims in committed notebooks, not slides.
• A fetch box is a hard wall — empty map regions can mean "never asked."
• Same filename ≠ same file: print provenance (row/station counts) at every pipeline stage.
• Show model uncertainty as transparency; fade to nothing far from data.

References

Full source citations, licences, and required acknowledgments: DATA_SOURCES.md.

Contains modified Copernicus Climate Change Service information (2026).

To cite: "OHW26 eutrophos team (2026), VI Near-Bottom O₂, github.com/oceanhackweek/ohw26_proj_eutrophos"