acquamArIna at the Marina di Trieste

The pilot site of the water quality platform. It is ours, so we can tell the whole of it.

The problem

Water quality in a marina is checked by sampling: someone goes out, takes a sample, sends it to a laboratory, and a few weeks later a report comes back. Between one sample and the next nothing is known, and when the report does arrive it describes a situation that has since changed.

There is a second problem, less technical and just as real: the report is a number in a PDF that, outside the laboratory, almost nobody can read. Whoever manages the water cannot use it to decide, and whoever lives around it cannot use it to form a view.

The constraint

A marine environment does not forgive: salinity, biofouling, wave motion and intermittent power are the four things that sink sensor projects in water, and they sink them in the third month, not the first week.

The second constraint was one of method. Eleven of the things we wanted to measure — heavy metals, COD, BOD — cannot be read by a probe: they are determined in a laboratory. Estimating them continuously needed a model, and training a model needed laboratory samples paired with probe readings, collected in the same place at the same moment, for months.

What we did

Eight probes in the water, measuring pH, redox potential, conductivity, turbidity, oil in water, polycyclic aromatic hydrocarbons, nitrates and chlorophyll continuously. The data arrives over MQTT, buffered locally when connectivity drops: outages are the norm here, not the exception.

A model trained on site, pairing the continuous readings with laboratory samples, to estimate the eleven parameters the probes do not read: COD, BOD, total suspended solids, lead, cadmium, mercury, copper, zinc, nickel, arsenic and chromium. That pairing period is the first job of every installation: without it, nobody can honestly give you a precision figure.

A single index from 0 to 100, calculated with the CCME formula, so the result is readable without asking anyone to interpret nineteen parameters.

Three statistical methods in parallel for anomaly detection, and each parameter linked to the regulatory reference that applies to it: Italian Legislative Decree 152/2006 Table 1B, Directive 2008/105/EC on priority substances, the MARPOL convention, and Directive 2006/7/EC on bathing water.

Four surfaces to read it from: the operations console for whoever runs the site, the public dashboard designed to WCAG 2.2 level AA, the programming interface documented to OpenAPI 3.1 and released under a Creative Commons BY 4.0 licence, and a citizen reporting channel that strips image metadata before storage.

The result

More than ninety days of historical telemetry processed across the eight active probes, and nineteen parameters available continuously — eight measured, eleven estimated — where before there were periodic samples and nothing in between.

The platform reached TRL 6, a prototype demonstrated in a relevant environment: the level at which a technology stops being research and starts being a product. Bringing it to catalogue is under way, and we follow the first installations ourselves.

⚠️ What this result is not. The model-estimated parameters have no legal standing and do not claim any: legal standing belongs to the laboratory report, which the platform manages in its own module with chain of custody. The continuous estimate exists to decide when an official analysis is worth running, and to avoid finding out about a problem three weeks later.

Go to the acquamArIna page → · Let’s talk about your site →