Knowing where it is tells you almost nothing about where it goes.
Detection and prediction are separate disciplines that share a vocabulary, which is how monitoring budgets end up buying one and being billed for both.
The distinction
Two different products, one word.
Detection answers a question about the present tense: is there something in this water now, where, and how much. It is validated by going and looking.
Prediction answers a question about a state that has not happened. It cannot be validated by looking; it is validated statistically, over many events, against what a naive guess would have achieved. Those are different epistemics and they need different procurement language.
The confusion is expensive. A programme buys a system on the strength of a compelling hindcast, then discovers in its second season that a hindcast is a curve fitted to events already known, and that the operational product has never been scored against anything.
So the four questions below are the ones we would want answered before signing. They are also the ones we expect to be asked in return.
What a forecast owes you
Four things, all of which should be in writing.
Lead time, stated as a decision window
Not "early warning". A number of hours or days, tied to a decision somebody actually takes — reduce abstraction, cancel a harvest, staff a night, mobilise a vessel. A forecast that arrives after the decision point has zero operational value, whatever its accuracy.
Skill against a named baseline
The baseline is usually persistence: assume tomorrow looks like today. It is embarrassingly hard to beat over short horizons. A model that does not report skill relative to persistence and climatology has not been evaluated, it has been demonstrated.
A false-alarm rate you can live with
Every threshold trades missed events against false alarms, and there is no setting that avoids both. The right point depends on what each costs you. A plant that can throttle cheaply wants a sensitive trigger; a resort that must not close twice in a summer does not.
Verification, published, after the fact
Each season, scored against what happened, in a document somebody outside the supplier can read. Better still where that outside party held the raw observations all along rather than receiving them attached to the scores. Systems that publish their misses improve. Systems that publish only their hits are marketing.
Where the physics ends
Transport is predictable. Growth is not, yet.
This is the single most useful thing to understand about the state of bloom forecasting, and it explains most of what forecasts get right and wrong.
Moving an existing patch of water around is a fluid dynamics problem, and fluid dynamics is mature. Given a decent circulation model, wind forcing and a known starting distribution, predicting where a surface bloom drifts over the next two or three days is genuinely tractable.
Predicting whether a population initiates, blooms, produces toxin or collapses is a biology problem with nutrient supply, light, temperature, grazing, strain variability and competition in it. Those processes are parameterised from limited data, and the parameters are not stable between regions.
So forecasts of drift and arrival tend to be useful, and forecasts of initiation and toxicity tend to be weak. A supplier who is clear about which one they are selling is being straight with you.
Two other constraints bind hard. Models are only as good as their initial conditions, and initial conditions come from the sparse, cloud-interrupted observations described on the remote sensing page. And most coastal regions lack the long, consistent, quality-controlled record that a statistical model needs to learn from at all.
Building the chain
What sits behind an operational bloom forecast.
Five components. A programme missing any one of them can still produce a map, which is why so many programmes produce maps.
- A physical model of the water body Circulation, stratification, exchange and residence time, forced by a meteorological model. In an enclosed basin or an engineered coastline this is where local bathymetry and structures matter more than any global product will capture.
- An observation stream to initialise from Satellite fields for extent, moored instruments for rate of change, samples for identity. The assimilation step is where most of the engineering effort actually goes, and it is rarely what a demonstration shows you.
- A biological component, honestly bounded Growth and decay terms with parameters that were fitted somewhere. Ask where. Parameters carried from a temperate estuary into a warm, hypersaline gulf are a known source of confident nonsense.
- A decision layer Thresholds, alert levels and a defined recipient, agreed with the people who act. A forecast with no named recipient is a research output, and there is nothing wrong with that as long as nobody is billing for an operational service.
- A verification loop Scores computed every season and fed back into the model and the thresholds. Without it the system cannot improve, and neither can the people using it.
The same discipline, applied to treatment
Measure the effect. Do not infer it from a mechanism.
Analytics people find this instinct easy and vendors of every kind find it inconvenient, which is why it is worth stating in a section of its own.
Take the technology this network operates. Some of it is measured and not seriously disputed: oxygen nanobubbles in water carry zeta potentials of roughly minus thirty-four to minus forty-five millivolts, which is what keeps them from coalescing.
Then there is a quantity that is measured but not settled. Gas–liquid mass transfer certainly improves — the published enhancement factors against conventional bubbles at equal gas volume run from around one and a half times to nine, with one study reporting about eleven. We cannot read that study's method, because it is paywalled. So the honest statement is a range with a wide error bar, and anyone quoting the top of it as the number has done what a forecaster does when they publish only the good season.
And then there is a mechanism that is genuinely contested. A 2020 study from Moleaer and Arizona State University reported that injected nanobubbles produce reactive oxygen species including hydroxyl radicals. A controlled 2023 study by Chae and colleagues in ACS ES&T Engineering found that generation minimal, if it occurred at all, under the ambient conditions tested.
Three tiers, three different treatments. Build on the first, bound the second, and hold the third open — which is exactly how a forecast should handle a well-constrained variable, a loosely constrained one and a parameter nobody has pinned down.
Why this belongs on a forecasting page
A programme that routes an intervention with analytics inherits responsibility for saying what the intervention did. If the treatment case rests on an unresolved mechanism, the verification case rests on it too, and the whole chain becomes unfalsifiable.
If you are evaluating a system
Six questions, in the order we would ask them.
Which decision is this forecast for, and who takes it. What is the lead time at that decision point. What skill does the product show against persistence over the last two seasons on our own water. What is the false-alarm rate at the threshold you propose. Where were the biological parameters fitted. And who publishes the verification.
Six answers. Any supplier who has run an operational service will have them to hand, and any supplier who has not will offer a demonstration instead.
Routing an intervention is the point where this stops being an analytical exercise. Alarivean's control stack exists to turn a risk surface into a vessel's next task, and HABguard covers what happens on a site when that task arrives. What the treatment itself is, and under what permit, is on HABslayer.
Sources cited on this page
- Science of the Total Environment — mass transfer of nanobubble aeration and its effect on biofilm growth. The abstract is paywalled and we have not read the method, which is why we quote a range rather than its headline figure.
- Springer — nanobubble stability and zeta potential.
- Chae, Kim, Kim and Fortner — reactive oxygen species generation from nanobubbles, ACS ES&T Engineering, 2023. The 2020 Moleaer and Arizona State University result reaching the opposite conclusion was issued as a press release rather than in the peer-reviewed record.
Aquatic Prosperity as a Service
Score us against your own record.
Send two seasons of observations and the events you remember. Alarivean will say what could have been anticipated from that data, what could not, and what an additional layer would have bought you.