Fáilte · Welcome
Live Dinophysis Bloom Map · Ireland
A weekly look at bloom risk for Irish shellfish sites, built on Marine Institute open data and rigorously tested against standard baselines.
Garry Lohan and Felix Sproll
Use ← → or swipe · N speaker notes · O all steps · F full screen
An fhadhb · The problem
Toxic plankton can close shellfish farms at short notice
- Dinophysis is a microscopic dinoflagellate (a plankton alga). Mussels and other shellfish filter it from the water and build up its toxins, which cause diarrhetic shellfish poisoning (DSP).
- To protect consumers, harvesting areas are closed when toxins in shellfish flesh exceed the regulatory limit. For growers, closures are costly and often come with little warning.
- A count of 100 cells/L of Dinophysis is the trigger level in the national monitoring programme: a flag for possible toxin risk, not a closure.
Can we say, a week or two ahead, where that level is likely to be reached?
Na sonraí · The data
Built on open, weekly Marine Institute counts
- Every count comes from the Marine Institute’s open data service (ERDDAP), with a publication timestamp on every sample.
- Each Sunday’s forecast uses only counts that were already public by that Sunday. No peeking at the future.
How to read it
One dot per monitoring site
The colour is the chance that Dinophysis reaches 100 cells/L there in the next two weeks. Bigger dots mean a higher chance.
- Low under 5%
- Watch 5–15%
- Elevated 15–30%
- High 30% or more
- Paler, dashed no sample yet this week
Click a site for its history. Try Bays for weekly notes and Season replay to watch a past season unfold.
Torthaí · How good is it?
Modestly better than the usual pattern, and we can say how much
In plain words: the score (PR-AUC, 0 to 1) measures how well the forecast puts the real bloom weeks at the top of its list. The comparison is “what normally happens at this site in this week of the year”. Blooms are rare (about 4% of site-weeks), so both numbers are far above chance. The point is the gap.
Caveat: for the same number of alarms it does not catch more closures than a simple trigger-level rule, ≥100 cells/L at the site last week (41 vs 46 of 62 South-West closures), but it raises far fewer false-alarm runs (41 vs 124–130). It is a surveillance aid, not a safety system.
Tástálacha · The test log
125 tests run, 64 null results reported
- Every experiment is pre-registered where possible and logged with its verdict; all results are reported, including null results.
- Recent counts are the forecast. What was in the water at and near a site in recent weeks carries almost all the skill; a simple logistic model gets within about 0.01.
- Environmental extras did not help: sea temperature, heatwaves, wind, chlorophyll, currents and rivers added nothing once recent counts were known.
An Fhrainc · France transfer
The method travels
- Retrained on France’s far longer Ifremer REPHY record, the same recipe beats the French seasonal norm. The Irish result is not specific to one coastline.
- A nugget: when mussels are over the toxin limit, oysters are too 15.3% of the time in France vs 3% in Ireland.
Tonnta teasa mara · Marine heatwaves
More heatwaves, but no detectable effect on blooms
- Heatwaves happen in summer, and so do blooms. Compare the same site in the same month, and the apparent “+40%” disappears.
- June 2023 brought a severe shelf heatwave, yet national Dinophysis counts and closures were below normal. A heatwave is not a bloom.
- Inside sheltered shellfish bays, temperature loggers record a median 82% more heatwave days than satellites see.
Tocsainí eile · Other toxins & species
One promising lead, and several clear “no”s
- Amnesic toxin (ASP) in king scallops: a toxin-history forecast scored 0.748 vs 0.443 for “same as last time” in Ireland. Promising, but it did not replicate in France.
- “Hidden bloom” flag: a toxin rise with no preceding high count. Flagged sites had a 23% chance of a ≥100 count within 21 days vs 5.1% otherwise, and it replicates in France.
- Paralytic (PSP), azaspiracid (AZP) toxins and Karenia mikimotoi: no gain over simple baselines.
- Mesodinium rubrum, the ciliate Dinophysis feeds on: adding its counts gave only a small gain (0.318 vs 0.312) that does not pass the pre-registered 95% rule, and its rises did not clearly come before blooms.
Samplaí · Sampling & turnaround
The day of sampling matters more than turnaround time
Share of samples public before the Sunday forecast
- Even with zero lab delay, skill would rise only from 0.312 to 0.324.
- Knowing this week’s counts is the biggest lever we found (up to +0.063). The missing information sits in late-week samples.
- Ideas (untested hypotheses): sample earlier in the week, issue on Tuesday, sample twice weekly at South-West sentinel sites in peak season, and add fast on-site screening.
What’s next
From hindcasts to a genuine, locked-in test
- 4 Oct 2026First locked-in-advance weekly forecast issued
- 7 Oct 2026Model v2 frozen after a red-team review found no leakage bug
- 25 Oct 2026First genuine scoring of a live forecast
- 2027Confirmatory season, pre-registered against a simple comparator
Candidates being tested, not adopted: French pre-training, a scallop ASP forecast, a harvest-risk product, and sampling-schedule trials.
Go raibh maith agaibh · Thank you
Garry Lohan and Felix Sproll
sliothar1.github.io/PA-Marine-Demo
Built on the Marine Institute’s national phytoplankton and biotoxin monitoring and its open ERDDAP service, with E.U. Copernicus Marine Service information and Ifremer REPHY data. Research prototype; not an official product or official advice. Déanta i nGaillimh · Made in Galway.
Built as a prototype by Garry Lohan and Felix Sproll