Fusing optical data for clearer water-quality decisions

Turbidity, chlorophyll and coloured dissolved organic matter (CDOM) sensors each reveal a different part of the aquatic environment. Turbidity indicates how strongly suspended particles scatter light, chlorophyll provides an optical proxy for phytoplankton biomass, and CDOM describes light-absorbing dissolved organic compounds often associated with tannins, wetlands and catchment runoff. Used together, they can distinguish a muddy flood plume from an algal bloom or from dark, tea-coloured water carrying dissolved carbon.

The value of combining these measurements depends on more than placing three instruments on the same frame. Sensor outputs must be synchronised, corrected for optical interference, checked against field conditions and interpreted with hydrology, depth and weather data. For Australian monitoring programs, this can mean adapting a fusion workflow to tidal estuaries, tropical rivers, coral-reef waters, inland storages or construction-site runoff.

Why the three measurements complement one another

Turbidity sensors commonly estimate particle concentration from light scattering, although the relationship between nephelometric readings and suspended solids varies with particle size, mineral composition and colour. Chlorophyll fluorescence responds primarily to phytoplankton pigments, while CDOM fluorescence or absorption responds to dissolved organic molecules. Their signals can change independently, which makes their combination useful for separating water-quality processes that look similar in a single channel.

A rise in turbidity accompanied by stable chlorophyll may indicate sediment resuspension, dredging or bank erosion. A chlorophyll increase without a comparable particle response may point towards phytoplankton growth. When turbidity and CDOM rise together after rainfall, catchment wash-off or wetland drainage becomes more plausible. These interpretations are indicators rather than automatic diagnoses, because sensor geometry, fouling and changing water chemistry can alter the observed relationships.

The combined dataset is especially valuable when paired with suspended-solids measurements, conductivity, temperature, depth, dissolved oxygen, rainfall and flow. A high-frequency optical record can then be compared with laboratory samples and hydrodynamic observations, providing context that a single proxy cannot supply.

Preparing sensors for a shared measurement system

Before fusion, each instrument needs a traceable calibration and a clear description of its measurement units. Turbidity may be reported in NTU or FNU, suspended solids in milligrams per litre, chlorophyll as a fluorescence signal or estimated micrograms per litre, and CDOM in quinine sulphate equivalents, relative fluorescence units or absorption coefficients. These values should not be merged as though they were directly comparable concentrations.

Sensor placement also matters. Optical heads should view the same water mass without obstructing one another or creating shadows and bubbles. A frame mounted in a river should account for velocity, depth variation and bed proximity; an estuary deployment must consider tides and salinity gradients. In Australia, equipment installed near Brisbane or Sydney may experience heavy storm-driven sediment pulses, while tropical deployments around Cairns can face rapid biofouling and warm-water growth.

Documentation should record optical path length, wavelength bands, sampling interval, firmware, cleaning schedule and calibration standards. Product specifications and supporting technical material can be checked through the technical downloads before designing a multi-sensor deployment or translating legacy D&A Instruments data into a current Campbell Scientific workflow.

Aligning time, depth and location

Data fusion begins with temporal alignment. Instruments may use different clocks, internal averaging periods and logging intervals, so timestamps should be converted to a common standard, typically UTC for the data system with local Australian time retained for reporting. Clock drift can create artificial leads and lags, particularly when a short sediment pulse passes a monitoring station.

A practical workflow resamples the measurements to a shared interval while retaining the original observations. For example, one-minute sensor values can be summarised into five-minute medians, percentiles and valid-sample counts. Median values reduce the influence of bubbles and transient spikes, whereas maximum values may be important when compliance depends on short-lived plume peaks. The aggregation rule should match the monitoring objective rather than being selected for convenience.

Depth alignment is equally important in stratified water. A surface chlorophyll maximum may occur above a denser sediment layer, while CDOM can be concentrated in freshwater entering below or beside saline water. Fixed-depth instruments, profiling packages and autonomous vehicles therefore require a common vertical reference. GPS position, tide stage and platform motion should be retained so that readings from separate transects can be compared spatially.

Correcting optical interference and sensor drift

Optical channels are not perfectly independent. Suspended particles can scatter excitation and emission light in fluorescence measurements, causing chlorophyll or CDOM readings to be biased. High CDOM can absorb light used by another optical channel, while dense algal material may contribute to apparent turbidity. The size and colour of particles further influence these effects, so a universal correction equation rarely performs well across every water body.

Correction can begin with laboratory or field experiments covering the expected range of turbidity, chlorophyll and CDOM. Reference samples analysed by gravimetry, pigment extraction or spectrophotometry can be paired with simultaneous sensor readings. A multivariable regression, partial least-squares model or machine-learning algorithm may then estimate the interference terms. The calibration set should include clean water, storm runoff, resuspended sediment and biologically productive conditions where possible.

Drift detection requires repeated checks rather than a single calibration at deployment. Stable reference periods, dark readings, wiper diagnostics and comparison with duplicate instruments can identify fouling or lamp degradation. A sudden change in the relationship between turbidity and suspended solids may be a real particle-composition shift, but it may also indicate a dirty optical window. Quality flags should preserve this distinction instead of silently replacing questionable data.

Combining data at feature level

Feature-level fusion converts raw time series into interpretable indicators before they enter a model. Useful features include rolling median turbidity, chlorophyll rate of change, CDOM-to-turbidity ratio, fluorescence peak width, lagged rainfall, tidal stage and vertical gradients. Derivatives can identify rapid plume fronts, while moving averages describe persistent conditions associated with algal growth or catchment discharge.

Normalisation is important when channels have different scales. Z-scores based on a stable baseline can show unusual behaviour, while robust scaling using median and interquartile range is less affected by extreme flood events. Ratios should be used carefully: a small denominator can create unstable values, and fluorescence readings near the detection limit may make a seemingly large change meaningless.

Classification models can use these features to label observations as sediment-dominated, algal-dominated, CDOM-rich or mixed. Random forests, support-vector machines and gradient-boosting methods are suitable when labelled examples exist. Unsupervised clustering can reveal recurring water types when labels are limited, but clusters still require interpretation through field samples and local hydrology. For Australian catchments, a model trained in a clear coastal lagoon should not automatically be applied to a tannin-rich northern river or a turbid Murray–Darling tributary.

Using physical and probabilistic models

Physical models improve interpretation by linking sensor changes to transport and water movement. A rainfall-runoff model can explain why turbidity and CDOM rise after catchment rainfall, while a sediment-transport model can connect turbidity peaks with dredging, river discharge or bed shear stress. Chlorophyll dynamics may require light availability, residence time, nutrient supply and temperature. Combining these processes helps distinguish an observed signal from a plausible cause.

Probabilistic methods are useful when measurements are incomplete or uncertain. A Bayesian network can represent relationships between rainfall, discharge, sediment, CDOM and phytoplankton, producing probabilities for different water-quality states rather than a single rigid label. Kalman filters and related state-space models can estimate an underlying condition from noisy observations, update it as new sensor values arrive and continue operating when one channel is temporarily unavailable.

Sensor fusion should retain uncertainty at every stage. A result based on three clean, synchronised channels deserves more confidence than one based on a fouled chlorophyll probe and intermittent CDOM data. Confidence scores can incorporate calibration error, missingness, signal-to-noise ratio, range checks and agreement with laboratory samples. This is particularly important when data support decisions around sensitive marine habitats such as the Great Barrier Reef.

Managing deployments in Australian conditions

Australian monitoring programs often combine regulatory reporting with operational decisions. A dredging project in a tidal harbour may need rapid warnings when a suspended plume moves beyond a defined zone, while a regional water authority may focus on seasonal trends in reservoirs and raw-water intakes. Construction-site monitoring around Melbourne, Perth or the Gold Coast can require robust sensors that tolerate intermittent flows, sediment pulses and long periods between site visits.

For erosion programs, fusion becomes more informative when optical readings are linked to rainfall intensity, discharge and site activity. A turbidity increase during earthworks may be separated from a chlorophyll-related rise in a downstream pond, while CDOM can indicate organic-rich runoff from disturbed soil or vegetation. Practical deployment considerations and the relationship between suspended-solids sensors and erosion events are discussed in this erosion monitoring guidance.

Field teams should plan for Australian conditions including intense summer rainfall, remote access, high solar exposure, marine growth and biosecurity requirements when moving equipment between catchments. Instruments used in Sydney Harbour, the Swan River or northern Queensland may need different antifouling approaches and maintenance intervals. A deployment log should record cleaning, servicing, weather, nearby works, unusual flows and any changes in sensor orientation.

Validating results and delivering useful outputs

Validation should combine independent measurements with event-based inspection. Grab samples collected during low, medium and high sensor readings can test whether the fused interpretation reflects actual suspended solids, pigment concentration and CDOM. Sampling immediately before and after storms, tidal changes or dredging shifts often provides more information than evenly spaced samples that miss the key transitions.

Performance metrics should match the intended use. Regression error may be appropriate for estimating suspended-solids concentration, while precision, recall and false-alarm rates matter for event classification. A model that detects every possible plume but generates constant alarms may be unsuitable for site management. Thresholds should therefore be evaluated against the cost of missed events, unnecessary inspections and delayed response.

The final data product may include cleaned sensor values, quality flags, derived features, water-type classifications, confidence intervals and event summaries. Dashboards can display aligned turbidity, chlorophyll and CDOM traces beside rainfall, flow and tide information, while machine-readable files support laboratory review and long-term modelling. Clear metadata ensures that future users understand whether a value is a direct optical measurement, a corrected estimate or an inferred environmental state.

A well-designed fusion system preserves the strengths of each optical sensor without pretending that they measure the same property. Turbidity reveals particle behaviour, chlorophyll indicates biological activity, and CDOM adds information about dissolved organic inputs and light climate. When calibration, timing, interference correction and uncertainty are handled carefully, their combined record can support more reliable decisions across Australian rivers, estuaries, reservoirs, construction sites and coastal waters.