Detection, sensing, forecasting
A satellite measures reflectance. A fluorometer measures fluorescence. A microscope measures what one analyst saw settle in a chamber one afternoon. Everything a monitoring program publishes is an inference, and knowing which one you are reading is most of the skill. It becomes the whole job once a number is going to decide something with money behind it.
The inference chain
What anybody cares about sits at the end of that chain: harm to a person, a fish, a membrane or a beach. Nothing in orbit and very little in the water measures harm, so every operational product chains inferences and hopes the couplings hold.
An ocean-color sensor records radiance at the top of the atmosphere. Corrected for the atmosphere, that becomes water-leaving radiance; a band ratio turns it into a chlorophyll-a estimate. Chlorophyll is a proxy for phytoplankton biomass, biomass is a weak proxy for the abundance of any single species, and species abundance is a weak proxy for toxin.
Each arrow is defensible on its own. Multiply the uncertainties and a chlorophyll anomaly buys you a sampling run and nothing more. Closing a beach takes a toxin number off a laboratory bench.
The observation stack
A monitoring program is a portfolio, not a purchase. The commonest failure is buying one layer well and leaving unfunded the layer that would have made it interpretable.
| Instrument class | What it physically measures | What you infer from it | What it cannot resolve |
|---|---|---|---|
| Satellite ocean color | Radiance in a handful of visible bands, over a wide swath | Chlorophyll-a, turbidity, surface bloom extent and drift | Species, toxin, anything under cloud, anything below the first optical depth |
| Hyperspectral imaging | Radiance in many narrow, contiguous bands | Pigment assemblage, and sometimes a dominant group | Toxin concentration, or anything under cloud |
| Airborne and drone survey | The same optics, flown under the cloud, on your schedule | High-resolution extent along a specific stretch of coast | Continuity — you get the day you paid for and nothing between |
| Moored sensor and buoy | Temperature, salinity, dissolved oxygen, fluorescence, turbidity | Condition and rate of change at one point, continuously | Anything a kilometer away; and fluorescence is not cell count |
| Glider and profiler | The same variables, but through depth and along a track | Structure — where the layer sits, how thick it is, where it is going | Real-time coverage of a wide area; endurance is finite |
| Sample, microscope, assay | Cells in a settling chamber; toxin in a matrix | Species identity, cell density, toxin concentration | Everywhere and everywhen you did not take a bottle |
Nothing in the last column is a defect. It is the specification, and better software will not fix it.
Limits
Chlorophyll-a is common to almost all phytoplankton, so a strong anomaly tells you biomass and nothing about whether the dominant organism is harmless or produces a neurotoxin. Toxin production also varies with strain and conditions within one species. The bottle is not optional.
Optical sensors do not see through cloud, and a polar-orbiting instrument returns when its orbit brings it back, which no contract moves. A monsoon or dust season can remove a fortnight of imagery exactly when you need it. Procurement coverage statistics rarely mention clear-sky coverage, the only kind that counts.
Near shore, the signal carries suspended sediment, dissolved organic matter and bottom reflection, and a pixel spanning water and shore mixes all of it. Standard open-ocean algorithms were not built for that, and they are weakest exactly where blooms hurt most.
No single instrument class answers all three at once.
What instrumentation revealed
In 2018 a University of East Anglia team working with Sultan Qaboos University published a survey of the Gulf of Oman made with seagliders — slow, autonomous, and cheap enough to lose. Two worked the area for eight months, reaching water ships had not sampled in nearly half a century because of piracy and geopolitics.
Where they expected some oxygen they found an area larger than Scotland with almost none. The lead author called it worse than feared, the zone vast and growing. Nothing on the shoreline would have shown it.
The depleted layer also rises and falls between seasons, squeezing fish into a thin band near the surface. That is structure, which a satellite cannot give you.
Source: a research summary; the journal paper could not be retrieved. The square-kilometer and micromole figures that circulate with it remain unchecked against the original.
The other half of the problem
Detection gets the attention. Verification is where monitoring programs and treatment suppliers both tend to go quiet.
If a water body is treated, somebody has to establish what changed, against what baseline, over what area and for how long. That is harder than detecting the bloom, because the counterfactual is unobservable and the water keeps moving.
Alarivean's oxidation duty runs as a contained process: water is drawn in, run through an enclosed instrumented stage at high flow, and discharged inside a specification agreed for that outlet, so verification has a pipe and an assay to point at. Measuring a pipe is a solved problem. Measuring a mixing zone is not. Nothing is stored, and a verifier receives a running trace off an outlet on a hull under way.
Because intake and outlet share one body of water, a whole class of question drops out: where a treated parcel went, what it was mixed into, whose receiving water carries it now. What has to be measured is the trace, the receiving water around the work, and the difference between them.
Verification splits by mission risk. Routine low-demand work runs continuously against a validated control envelope on real-time inline instruments, with periodic confirmatory samples to an independent laboratory. High-risk work — toxin blooms above all — runs to a tighter control envelope, on more conservative source-term settings and under closer independent oversight, and should be scoped as slower work. For the monitoring clause, the control is preventive: an inline reading cuts the oxidant or stops the intake before an out-of-spec discharge can form, and the client holds contractual authority to halt an active deployment on suspicion. Nobody can run a laboratory-turnaround assay on every parcel of a continuous stream; a supplier who implies otherwise is describing a system that does not exist.
NOAA's National Centers for Coastal Ocean Science validated an ozone nanobubble aeration system on an eight-acre pond near Fort Myers Beach in 2018, reporting complete elimination of algae within forty-eight hours with proper reoxygenation and no apparent harm to aquatic life. A separate 2020 evaluation of a nanobubble ozone system on ship ballast water found it effective against algae, bacteria and motile zooplankton without a statistically significant adverse residual-toxicity effect in the receiving water.
Both are bounded environments with a controllable boundary, and neither names the vessel Alarivean operates. No open-water result has been published for this class of treatment by anybody yet. Both studies are sound; neither was asked this question.
Methodology, in the sense the word carries elsewhere
Skill against a stated baseline. A false-alarm rate. Verification somebody else can repeat. And a reader whose fee holds steady whichever way the number falls.
Skill comes first, and it is comparative, so it needs its comparator written down. A forecast is scored against persistence or against climatology on the same water, over a stated run of seasons, by whoever holds the archive. A skill figure with no comparator named is a number about nothing.
The false-alarm rate comes second, and it is the one most often missing. Tune a system to catch the faint signals and it will call events that do not arrive. Tune it to stay quiet and it will sit through some that do. Both settings are defensible and both have to be stated, because whoever acts on the signal wears the cost of each kind of error, and the two errors cost different amounts.
Verification is third. Hindcast against the observed record, published with the misses left in, and refreshed on a stated cycle so a score describes the system as it runs today.
Custody is fourth. A score produced by the supplier of the forecast is a sales document. The same score produced by a party with no fee riding on it is a measurement, and the difference is the entire difference.
Put the four together and the shape that appears is an index. Layers and bands. A floor below which the signal does nothing and a ceiling above which it saturates. A trigger set down before the season it governs. Adjustment on a published cycle as the archive lengthens. And a residual gap between what the index reads and what happened at the asset itself, which is why the in-situ layer keeps earning its budget however good the satellite gets.
No instrument of that kind is proposed here and none is claimed to exist. The observation is narrower and it belongs to monitoring: the specification that makes a program worth acting on is the same specification that would make its numbers worth settling against, and a program that cannot answer those four questions is not yet either.
What a published verification carries
Six lines. A supplier who can produce all six is describing a system somebody can rely on. One who can produce two is describing a map.
Custody of the record
On a slide the two look identical. What separates them is whether anybody outside the company doing the work could see the raw stream while it was still arriving.
Provenance is harder to repair afterward than any instrument limit, and almost free to get right at the start.
A time series that reaches a reviewer as a processed export has already passed through decisions nobody can now inspect. Which stations were in. Where a gap was filled, and how. Which calibration applied on which day, and when the sonde was last wiped. None of those choices has to be made in bad faith for the result to be useless. They only have to be invisible.
Alarivean's programs engage local scientific institutions continuously, with access to the live feeds the operator works from and the freedom to run their own sensors beside them and keep whatever those produce. Alarivean writes that in as a requirement on itself: a treatment record with a single custodian cannot be audited, only believed.
If the operator's servers went dark tonight, would anyone else still hold the record? Put the answer in the schedule, with the institution named and engaged for as long as the program runs.
Adjacent work
Immediate, Significant, Scaled
Station records, chlorophyll time series, cell counts, the years the bulletins were wrong. Back comes a read on what your current stack will and will not support, where an extra instrument would add no new information, and which decisions your lead time can reach.
A day of warning is a day of work where treatment already runs, and a better-documented loss where it does not. So in-water treatment is contracted as an ongoing subscription. It runs through the year because the stressors do, keeping them below the state a forecast would otherwise have to catch in time.
