Reading the image without mistaking it for a sample
A satellite image of suspended matter may appear to show exactly how much sediment is in the sea. It does not. It begins with light arriving at a satellite sensor, passes through atmospheric and optical corrections, and ends with a numerical estimate governed by an algorithm. To translate satellite ocean-colour suspended matter correctly, we need to identify which quantity was retrieved, which reflectance bands supported the estimate, and how independent water samples tested its reliability. Those steps turn a persuasive image into an interpretable measurement.
The most important distinction is between the satellite’s observation and the quantity written on the map. The sensor detects radiance in spectral channels. The processor estimates water-leaving radiance or remote-sensing reflectance. A separate empirical or semi-analytical relationship converts that optical information into suspended particulate matter, total suspended matter, or turbidity. Each arrow in this chain adds assumptions. A smooth-looking map does not erase those assumptions or turn calculated concentration into material collected on a filter.
A second distinction concerns language. Turbidity is an optical measure of how water scatters light under a defined instrument method, typically reported in units such as FNU. Suspended matter concentration describes particle mass per volume, commonly milligrams per litre. Both often respond to increasing particles, but they are not synonyms. A report that silently replaces ten FNU with ten milligrams per litre has altered the scientific evidence even if the two numbers look equally plausible.
The practical objective is therefore not to memorise a complicated algorithm. It is to ask: what was actually observed, what was inferred, what was sampled independently, and what uncertainties survive? A translator, editor, scientist, student or coastal manager can use that sequence before making claims about a plume, dredging activity, river discharge or long-term trend. Every worked number in this guide is explicitly illustrative unless tied to a cited methodology; none represents a real unreported field observation.
Decide which suspended-matter variable is on the map
Several labels travel through the literature: TSM, TSS, SPM, particulate matter concentration, non-algal particulate matter, and turbidity. Some projects use TSM and SPM in nearly equivalent ways, but their laboratory protocols and operational definitions still matter. Total suspended solids commonly involves filtration, drying and weighing a retained residue. Total suspended matter is also frequently an operational gravimetric concentration. A paper must say which material and procedure produced its reference values before an optical algorithm can inherit the same units.
A mass concentration of 25 mg/L means twenty-five milligrams of particulate material in each litre of water under the stated sampling and analytical definition. It does not mean a quarter of the water is solid, twenty-five particles were counted, or a single visible object was detected. Particle number, particle volume fraction, dry mass, organic fraction and mineral fraction are separate quantities. One sea sample can contain many tiny particles while another contains fewer larger grains yet gives similar dry mass.
Turbidity, often measured with a nephelometric sensor, is influenced by particle size, shape, refractive index, colour and measurement geometry. International and manufacturer methods can use different wavelengths and geometries; their units are not freely interchangeable. Some natural waters show a strong local correlation between turbidity and TSS. That is useful when explicitly calibrated, but it is not a universal chemical conversion. Even a locally strong regression can shift after a storm changes particle composition.
Look next for the satellite product’s data dictionary. Does the pixel represent a measured band, remote-sensing reflectance in inverse steradians, an algorithmic concentration in mg/L, or a quality flag? Is the pixel an instantaneous scene, daily composite, monthly median, or model-assisted estimate? The answer controls what can be said about a date, location and change. Missing data are not zero concentration, and a cloudy pixel is not evidence of clear water.
Follow photons from the surface to the detector
The satellite does not point a laboratory balance at the ocean. Sunlight first passes through the atmosphere, interacts with the sea surface, penetrates the water and is scattered or absorbed. Only a small, variable part of the light registered above the atmosphere can be attributed to the water. That is why estimating water-leaving radiance is a demanding inverse problem rather than a simple brightness-to-mass conversion. Aerosols, molecules and sun glint can dominate the signal before water information is isolated.
Within the water, absorption removes light from certain paths and scattering redirects light. Phytoplankton pigments, coloured dissolved organic matter, water molecules and particles each contribute to the spectral result. Suspended mineral material can increase backscattering, often lifting red or near-infrared water reflectance. But dark organic material and complex water mixtures modify that relationship. The same nominal suspended mass can produce different optical signals if mineral type, aggregation or particle-size distribution changes.
Remote-sensing reflectance is usually written Rrs(λ) and has units sr⁻¹. Conceptually, it is water-leaving radiance divided by downwelling irradiance just above the surface, with consistent spectral and geometrical conventions. Radiance and irradiance are not interchangeable. A numerical value in reflectance units cannot be treated as concentration until a retrieval relation has been specified, calibrated and applied. When translating a formula, preserve the wavelength, band definition and units rather than replacing them with an intuitive colour label.
The apparent colour in a display is also a design choice. An RGB image may be a true-colour composite, a false-colour composite or a concentration product stretched into a palette. Deep red on a legend can mean a high algorithm output rather than red light physically emitted by particles. Read the legend, masks, scale transformation and processing level before describing any physical meaning. A logarithmic colour bar can make moderate changes appear dramatic if its numerical ticks are ignored.
Why red and near-infrared channels matter
Blue and green channels can be useful for many ocean-colour questions, yet high particle concentrations create a special challenge: water may become optically complex, and some shorter wavelengths saturate or respond strongly to other constituents. Red wavelengths near the mid-six-hundreds of nanometres are frequently sensitive to suspended matter over useful ranges. In very turbid coastal and estuarine water, near-infrared channels may extend usable retrievals. There is no single preferred band for every site, season and sensor.
The 2010 Nechad, Ruddick and Park study developed and evaluated a multisensor approach for total suspended matter using reflectance in turbid waters. The 2015 Dogliotti and colleagues work examined red and near-infrared turbidity retrieval across coastal and estuarine conditions. Those are substantive methodological contributions, not permission to copy a coefficient from one sensor into an unrelated product. Each algorithm specifies its radiometric convention, wavelengths, calibration and applicable range.
A near-infrared band is not automatically better merely because the water is muddy. Atmospheric correction in those bands can be difficult, detector performance varies, and extremely high concentration can make an algorithm lose sensitivity. Some processors switch or blend red and NIR relations across an intermediate regime. Such a switch must be documented because it can introduce discontinuities or change error behaviour. It should not be described as though the sensor directly weighed sediment.
Check the band centre and its spectral response function. A nominal 665 nm band and a nominal 670 nm band are related but not identical observations, especially when a water spectrum changes rapidly. Band-integrated reflectance depends on the instrument’s response and the target spectrum. For a rigorous comparison, harmonise spectral definitions or quantify the difference. Reporting three decimals on a concentration estimate does not compensate for a band mismatch hidden in the method.
Atmospheric correction is often the limiting inference
Satellite radiance is influenced by molecular scattering, aerosol scattering, absorption by atmospheric gases, and reflections from the surface. A processing chain estimates those contributions so that the remaining signal can be attributed to the water. In turbid coasts, conventional assumptions that water-leaving radiation is negligible in selected near-infrared bands may fail. If a processor incorrectly assigns genuine water signal to aerosols, the retrieved reflectance can be biased low; other correction failures can create spurious high reflectance.
The issue becomes especially acute near land. A satellite footprint includes light from a finite neighbourhood and receives optical influence from nearby bright land through adjacency effects and the sensor’s point-spread response. Thin clouds, cloud shadows, foam, boats and shallow bottoms can also contaminate the data. A bright strip along a coastline might reflect a real suspended plume, imperfect land masking, or both. No honest interpretation chooses one simply because it produces a more interesting headline.
Rather than declaring one algorithm universally superior, examine validation in waters resembling the intended application. A 2021 comparison for Belgian turbid coastal waters evaluated several atmospheric correction approaches for Sentinel-3 OLCI using independent radiometric measurements. It showed that performance could differ by wavelength and processor under the study conditions. This supports a method-specific evaluation; it does not license transferring those error figures to every coastline or generation of software without checking.
Quality masks deserve their own audit. Cloud, high glint, stray light, negative water-leaving radiance, detector saturation and land adjacency are not merely cosmetic exclusions. A mask changes which locations and times enter an average, so it can bias a trend if missingness is correlated with storms or high sediment events. Keep the original flags and explain rejected observations. Filtering a dataset more aggressively may improve optical quality while simultaneously reducing the representativeness of the remaining time series.
Know your processing level and pixel geometry
Level-1 measurements normally concern calibrated, geolocated sensor radiances, while Level-2 ocean-colour files commonly provide geophysical variables and quality information on the sensor’s swath. Gridded Level-3 products summarise observations using a temporal and spatial aggregation procedure. The exact definitions are mission dependent, so inspect the product specification rather than assuming every provider makes identical choices. A gridded monthly product is not interchangeable with a single clear-sky overpass on a particular afternoon.
Geolocation errors matter at river mouths, dredging boundaries and narrow channels. A pixel centre may lie in open water while part of its optical footprint overlaps an island, boat wake or shallow bar. The pixel value represents an integrated measurement, not an infinitesimal point. Cropping the map more closely cannot restore lost spatial detail. A small bright feature seen at one display zoom level may disappear or shift when appropriate resampling is applied.
Spatial resolution also complicates field matchups. A water bottle samples a local volume at a specific depth; the satellite senses a spatially integrated, near-surface optical layer. Patchy suspended matter means the bottle may be unrepresentative of a whole pixel even with flawless laboratory measurements. Sampling several locations, recording GPS accuracy and summarising spatial variability can reveal this mismatch. A large discrepancy is not automatically an instrument failure; it may reflect differing sampling supports.
When readers compare products, insist on matching units, masks, spatial support, local time, observation geometry and aggregation methods. A three-day median with forty percent valid observations cannot be read like a daily snapshot with full coverage. A high-resolution image is not necessarily more accurate. Its finer pixels may have different signal-to-noise behaviour or more exposure to adjacency artifacts. Greater visual detail is valuable only when the uncertainty and processing conventions are explained alongside it.
Build a local retrieval relation without inventing universal coefficients
An optical retrieval usually requires a map from one or more reflectance observations to a variable such as TSM or turbidity. The relation may be linear within a limited calibration interval, nonlinear across a larger concentration range, or switch between spectral bands. Training data must include both accurate radiometry and independently measured particle or turbidity values. The choice of mathematical form should follow the physics, observed behaviour and declared domain, not simply the fit with the largest coefficient of determination.
A simple illustrative linear model is C = aR + b, where C has units mg/L, R is Rrs in sr⁻¹, a has units mg·sr/L, and b has units mg/L. The formula has no claim to universal applicability; it is useful for learning dimensional analysis. If a calibrated intercept produces a negative concentration at low reflectance, that is a warning about range and residual uncertainty, not a reason to pretend negative mass exists. Avoid extrapolating that line into clear offshore water without evidence.
More realistic semi-empirical forms may grow sharply as reflectance approaches a parameterised asymptote. The benefit is the ability to represent changing sensitivity; the hazard is enormous error amplification close to a denominator approaching zero. Algorithm limits should be written out and enforced. When a computed result falls outside the calibrated range or retrieval flags indicate saturation, report the condition rather than clipping silently to an attractive maximum that looks measured.
Calibration and validation must be independent in a meaningful sense. Splitting highly correlated samples from one cruise randomly can exaggerate performance because the same sediment conditions and atmospheric environment appear on both sides. A stronger design can hold out different days, stations, seasons or campaigns, then test transfer. Report sample counts, concentration distribution, range restrictions and failure modes. A low average error in calm water is not evidence of performance during a river flood.
Sampling and weighing suspended particles properly
Field concentration measurements form the physical anchor of many suspended-matter algorithms. A sampling record should include collection time, depth, coordinates, preservation, filtration volume, filter type, handling, drying method and balance precision. The measurement often depends on the dry residue retained on a filter after controlled preparation. Missing procedural details make apparent agreement with a satellite product difficult to interpret, because two laboratories may not capture exactly the same material or handle salts in the same way.
For an illustrative filter calculation, suppose a prepared filter has a dry mass of 0.15230 g and the same filter after sampling, drying and conditioning weighs 0.16130 g. The gained mass is 0.00900 g, equivalent to 9.00 mg. If 0.500 L of water passed through it, the uncorrected concentration is 9.00 mg divided by 0.500 L, or 18.0 mg/L. That is an example arithmetic chain, not a claim that a specific sea sample was collected.
Now suppose a handling blank prepared using the same protocol gains 0.00050 g, or 0.50 mg. An appropriate blank correction would leave 8.50 mg for the sampled water, giving 17.0 mg/L. The example teaches why small-looking masses matter. A blank correction should only be applied when justified by the method and paired controls; it must not be borrowed from unrelated batches. Record whether the published data are blank-corrected and whether the uncertainty incorporates filter and volume errors.
Other difficulties include filter clogging, subsampling heterogeneity, material lost during transfer, dissolved salts left after drying, volatile matter changes and balance instability. Particularly in saline water, rinsing practice and correction conventions are important. A method can be precise yet biased if all filters contain a systematic residual salt mass. Conversely, a carefully handled filter can still sample a water parcel unrepresentative of the satellite pixel. Laboratory quality and field representativeness need separate evaluation.
Turbidity is an optical proxy, not a hidden mass unit
Nephelometric turbidity is often easier to measure at high frequency than suspended dry mass. That makes it attractive for calibrating a continuous sensor in a port, estuary or river. The relation must be evaluated with paired local samples, because optical scattering depends on material properties as well as mass. Two particle populations of equal mg/L may produce markedly different FNU readings. Organic flocs and mineral silt can scatter light differently even at similar bulk concentration.
Suppose a clearly labelled local teaching model, based only on hypothetical paired observations, is TSS = 1.8 times turbidity, with 1.8 having units mg L⁻¹ FNU⁻¹ and zero intercept within the tested range. A value of 12 FNU would then correspond to an estimated 21.6 mg/L. The unit-bearing coefficient is essential. Writing ’12 FNU equals 21.6 mg/L’ without naming a local conversion would incorrectly imply a physical unit identity.
If the local slope has an estimated standard uncertainty of 0.2 mg L⁻¹ FNU⁻¹ and the sensor has a standard uncertainty of 1 FNU, a first-order independent-error calculation gives a concentration uncertainty of approximately square root of (12×0.2)² plus (1.8×1)². The result is about 3.0 mg/L before adding model residuals. This numeric exercise assumes independent errors and an appropriate linear model. Real calibration residuals and covariance may substantially change the interval.
Translation is most accurate when it preserves ‘measured turbidity’ and ‘estimated TSS’ as separate clauses. A real-time station may measure 12 FNU directly and infer 21.6 mg/L from a regression. A satellite may infer turbidity from reflectance, then a report may estimate TSS from turbidity, introducing another transfer step. Each conversion is an additional model with a provenance and uncertainty. Calling all three quantities ‘sediment level’ prevents readers from examining the chain.
Validation requires a deliberate matchup protocol
A satellite-field matchup is not simply two numbers copied from the same date. The field sample has an exact time and location; the satellite observation has its own overpass time, footprint, quality flags and atmospheric correction. A protocol sets allowed temporal differences, distance windows, pixel homogeneity tests and exclusion criteria before looking at the residuals. Conveniently relaxing the window only for unfavourable results defeats the purpose of validation.
Start with a field sample collected close enough to an overpass for the intended water dynamics, then inspect the surrounding satellite pixels. Do not automatically choose the single pixel with the smallest difference from the bottle. That is a form of selection bias. A local mean or median may be useful when stated consistently, but it must not mix land-contaminated or poor-quality pixels. Record valid-pixel counts and spread so that users can distinguish a stable matchup from a patchy plume.
The NASA SeaBASS archive and validation resources illustrate the importance of independently collected bio-optical and radiometric observations. A high-quality matchup dataset should preserve raw measurements, calibration information, collection metadata and processing versions. A field programme developed only from the same satellite product cannot provide wholly independent confirmation. The causal chain must not loop back on itself while being advertised as external validation.
Performance reporting should include more than correlation. Consider mean bias, median residual, mean absolute error, root mean square error, relative error where denominators are meaningful, sample count and performance by concentration regime. A strong correlation can coexist with a large systematic bias. Relative errors become unstable when the true concentration approaches zero. Scatterplots, residual-versus-concentration plots and held-out event diagnostics expose failure mechanisms hidden by one headline accuracy number.
Understand denominators, signs and summary statistics
Let estimated concentration be Ĉ and reference concentration be C. The signed residual can be defined as Ĉ minus C in mg/L. Positive means the retrieval overestimates relative to the laboratory reference under this convention; negative means it underestimates. If another publication defines residual with the opposite order, its sign interpretation reverses. A translation that swaps ‘bias’ and ‘absolute error’ hides this distinction, because absolute error discards the direction.
For example, take three independent reference values of 10, 20 and 40 mg/L and corresponding retrievals of 12, 18 and 45 mg/L. Residuals are +2, −2 and +5 mg/L. The mean signed bias is (2−2+5)/3, or +1.67 mg/L. Mean absolute error is (2+2+5)/3, or 3.00 mg/L. Root mean square error is square root of (4+4+25)/3, or approximately 3.32 mg/L. These are illustrative values checked directly from the definitions.
If absolute percentage error is computed as absolute residual divided by reference concentration, the three values are 20%, 10% and 12.5%, giving a mean of about 14.17%. Dividing by the retrieval rather than the reference would produce a different statistic. Dividing by a reference near zero can produce wildly inflated percentages that describe the denominator more than overall quality. Report the formula and how zeros or below-detection observations were handled.
Agreement within a specified tolerance is another distinct statistic. If a product is judged against a fixed limit of ±5 mg/L, all three examples are within or on the limit; this says nothing about performance at 0.5 mg/L versus 500 mg/L. If the tolerance is instead a percentage, different samples may pass or fail. A report should explain its decision rule, intended concentration regime and consequences of error before calling a satellite product ‘validated’ in a blanket sense.
Worked example: turn reflectance into an estimated concentration
Imagine two illustrative local calibration anchors: Rrs at one fixed red band is 0.010 sr⁻¹ when the reference mass concentration is 10 mg/L, and 0.020 sr⁻¹ when it is 30 mg/L. A linear relation through those two points has slope (30−10)/(0.020−0.010), or 2,000 mg·sr/L. Substituting the first point gives an intercept of −10 mg/L. The resulting teaching equation is C = 2,000R − 10, with the declared units understood.
For a quality-controlled reflectance R = 0.016 sr⁻¹, multiply 2,000 by 0.016 to obtain 32 mg/L and subtract 10 mg/L. The estimated C is 22 mg/L. Reverse-check it: C=22 implies R=(22+10)/2,000=0.016 sr⁻¹. The arithmetic is consistent, but consistency alone cannot validate the physical model. Two anchors define a line; they do not establish its error distribution, curvature or behaviour in different particle populations.
For a reflectance of 0.006 sr⁻¹, the same line yields just 2 mg/L, but that point is below the illustrative interval of 0.010–0.020 sr⁻¹. It is extrapolation, not an equally supported retrieval. At R=0.004 sr⁻¹ the line would yield −2 mg/L, which is physically impossible as a mass concentration. A defensible pipeline should flag such conditions, evaluate an alternative low-concentration model where available, or state that the estimate is outside validity.
Now suppose R=0.016 has standard uncertainty 0.002 sr⁻¹, and take the slope as fixed for the exercise. The reflectance contribution to concentration uncertainty is 2,000×0.002=4 mg/L. If separate model scatter is represented by 6 mg/L and the two contributions are independent, the combined standard uncertainty is square root of 4²+6², about 7.2 mg/L. The output is then 22 mg/L with illustrative standard uncertainty approximately 7.2 mg/L, not an exact 22.000.
Worked example: propagate errors through a real sampling choice
Suppose a coastal programme samples 0.500 L for a gravimetric determination and obtains a corrected retained mass of 8.50 mg. Its nominal concentration is 17.0 mg/L. If the corrected mass has standard uncertainty 0.60 mg and sampled volume has standard uncertainty 0.010 L, then for C=m/V the mass contribution is 0.60/0.500=1.20 mg/L. The volume contribution is C×0.010/0.500=0.34 mg/L. These terms describe different errors.
Assuming the mass and volume uncertainties are independent, the first-order combined standard uncertainty is the square root of 1.20 squared plus 0.34 squared, or about 1.25 mg/L. Rounding sensibly, the laboratory example gives 17.0 ±1.3 mg/L standard uncertainty, before accounting for spatial sampling mismatch or any additional systematic effects. A one-standard-uncertainty statement is not an automatic ninety-five-percent interval. Coverage requires further assumptions about distributions and degrees of freedom.
Suppose a nearby satellite pixel estimates 22 mg/L at the same stated band, while the sample yields 17 mg/L. The raw discrepancy is +5 mg/L using satellite minus reference. With the illustrative satellite standard uncertainty 7.2 mg/L and laboratory standard uncertainty 1.25 mg/L, their independent combined uncertainty is approximately square root of 7.2²+1.25²=7.3 mg/L. This arithmetic alone does not demonstrate agreement because sampling representativeness and shared calibration errors may still matter.
This is why a single good matchup cannot prove universal performance. We need many observations across distinct concentrations, particle types, seasons, atmospheric conditions and water states. And if the satellite algorithm was calibrated using that very laboratory sample, its apparent closeness cannot also count as wholly independent validation. Preserve the calibration-versus-validation distinction in any scientific summary. A detailed uncertainty budget is more informative than declaring a discrepancy ‘small’ without a denominator.
What to do when a plume seems to expand
An estuarine plume may change within hours after rain, wind or a tidal reversal. Satellite images show a subset of those conditions: times when a usable overpass occurred and quality flags permitted retrieval. Compare the same product version, compatible geometry, similar processing rules and meaningful temporal windows. A map before a storm and another three days later cannot isolate storm sediment transport from the tides, currents and missing scenes without additional evidence.
First define a threshold on the retrieved variable, such as an illustrative concentration above 20 mg/L, and label it explicitly as an operational threshold rather than a universal ecological hazard line. Count valid water pixels above that threshold and multiply by their appropriate pixel areas. If only 60% of the study domain is valid on one date, reporting an area fraction without stating that denominator can make an apparent expansion or contraction misleading. Missing observations must remain missing.
Next inspect whether a sensor or algorithm changed during the time series. A processing update can alter atmospheric correction, masks or calibration. A newly bright plume may partly reflect that change. Cross-check available in situ observations and, if possible, another independent sensor or hydrodynamic context. Agreement between two satellite products sharing training data and similar assumptions is useful but not fully independent evidence of concentration changes.
Finally write the claim at the strength the evidence supports. ‘The retrieved suspended-matter signal increased in valid pixels after the event’ is stronger scientifically than ‘the entire coast accumulated extra sediment’ when no mass budget was measured. A concentration map does not tell us sediment deposition, erosion rate or ecological damage by itself. Such outcomes require depth-integrated transport, time, hydrodynamics, sediment properties and habitat response to be evaluated separately.
Depth, mixing and vertical structure
Ocean colour primarily senses an optically weighted near-surface layer, not the complete water column. In very turbid water the penetration depth may be shallow; in clearer water it can be greater. That depth depends on wavelength and the optical properties of the water itself. A bottle drawn several metres below the surface may therefore be physically different from what the satellite retrieval represents. Record sample depth and stratification rather than assuming a vertical average is the right comparison.
Near the seabed, resuspension can create a dense layer of particles while the surface remains comparatively clear. A satellite may underrepresent that bottom-heavy distribution because light cannot penetrate deeply enough. Conversely, a buoyant river plume can carry suspended material at the surface above a saltier layer, producing a high near-surface signal while deeper water differs. Neither observation contradicts the other. The sampling supports and vertical circulation explain why their values need not match.
Bathymetry contributes another complication in optically shallow water. Light reflected from a bright sandy or vegetated bottom may reach the sensor and alter apparent colour. Some shallow-water algorithms explicitly model bottom reflectance; ordinary suspended-matter retrievals may not. Masking or treating shallow zones separately may be necessary. A bright shallow bank is not proof of particles floating in the water. The map must be interpreted together with depth and bottom type.
When deciding whether to use a satellite product for a river load estimate, remember that concentration alone is insufficient. Mass transport depends on concentration multiplied by water volume transport and integrated over the relevant section and time, with compatible spatial and vertical distributions. Multiplying a single surface pixel by an arbitrary discharge can generate a numerical value without an adequate physical basis. Explain the integration domain, flow information and stratification assumptions before describing flux.
Particle composition and changing optical sensitivity
Suspended particles vary in grain size, aggregation state, composition, porosity and organic content. Their mass-specific backscattering can therefore change without the total mass changing. A calibration based on fine mineral silt may behave differently when a seasonal plankton bloom adds organic particles or river turbulence forms large flocs. The retrieval algorithm is not necessarily wrong in a computational sense; its mapping from optical property to mass may have shifted because the underlying material changed.
A useful research design pairs mass measurements with additional observations: particle-size distribution, organic versus inorganic fraction, mineral information, optical backscattering or microscopic characterization. These data can explain why concentration residuals depend on place or season. If two groups of observations form separate branches in an Rrs-versus-TSM plot, averaging them into one line may conceal the cause of error. The best remedy can be a better physical classification rather than a higher-order polynomial.
Flocs are particularly interesting. Small primary particles can aggregate into larger structures whose optical response, settling velocity and breakage behaviour depend on turbulence and chemistry. Two samples with equal dry mass but different floc structure can be very different for navigation, ecological exposure or sediment transport. Satellite optical retrieval cannot automatically identify settling behaviour from mass concentration. Do not translate a reflectance increase into ‘faster sinking sediment’ without measurements supporting that conclusion.
High chlorophyll, coloured dissolved organic matter and suspended minerals can also covary near river mouths. One coloured component may be falsely credited for a signal driven by another. Multiband algorithms can help disentangle constituents under suitable conditions, but they introduce more parameters and possible instability. Independent field observations remain essential. More spectral bands create opportunities for discrimination; they do not guarantee a unique physical solution when several constituents produce similar spectral patterns.
Calibration transfer across sensors and missions
Moving an algorithm from one satellite mission to another is not merely copying a familiar product name. Sensors differ in band centre, spectral response, radiometric noise, spatial resolution, viewing geometry and processing chain. Calibration constants derived for one instrument may not be valid for another, even if both channels are casually called ‘red’. A defensible transfer evaluates spectral compatibility, atmospheric correction consistency and performance against independent field or radiometric reference data.
Suppose one sensor provides Rrs around 665 nm and another at 674 nm. If a coastal reflectance spectrum rises across that interval, a numerical difference could arise even if suspended matter was unchanged. A correction may require hyperspectral measurements or a validated spectral adjustment. Simply assuming that the two reflectances are identical can create an artificial step in a multi-year time series. Preserve the band centres in the documentation and in translated axis labels.
Cross-sensor maps should also use comparable temporal aggregation. One mission may observe in the morning and another in the afternoon, when river plumes, wind-driven resuspension or tides have changed. A scene-to-scene offset may represent real environmental variability rather than calibration bias. Pairing observations within a short window and stratifying by water conditions is more defensible than aligning two monthly means with different cloud-screening characteristics and declaring the residual an instrument correction.
A practical harmonisation report should separate radiometric harmonisation, algorithm harmonisation and statistical trend estimation. Radiometric harmonisation checks how sensors measure light. Algorithm harmonisation checks how that light becomes turbidity or mass concentration. Trend estimation assesses the physical changes after measurement conventions are aligned. Skipping the middle step is a common path to false trend confidence. A time series can be long without being internally comparable.
Quality-control decisions for a working coastal dataset
Begin an audit by checking units and variable names. Confirm that Rrs is not mislabeled as reflectance without geometry, that TSS is not substituted for turbidity, and that zero values are distinguishable from fill values. Check longitude and latitude order, geodetic datum, timestamps and time zone. A longitude sign error or daylight-saving conversion error may match an ocean sample with the wrong pixel. These errors are small in code but large in scientific interpretation.
Next check masks and flags by plotting the original quality layer alongside concentration. Look for impossible negative retrievals, isolated saturated pixels, coastline halos, cloud edges and abrupt seams between processing tiles. Do not automatically delete every extreme point; some true events are extreme. Instead trace suspect values back to radiance, atmospheric correction, band choice and coincident observations. Good quality control distinguishes implausible measurements from uncommon but credible physical conditions.
Then check the statistical design. Was the calibration range stated? Does the validation set contain independent dates and locations? Does the error metric become unstable near zero? Are residuals stratified by water type, concentration and viewing conditions? A product that passes on average may fail exactly where a harbour operation requires accurate high-concentration values. Relevance to the decision is as important as overall skill.
Finally maintain a provenance note with the provider, collection name, product identifier, processor version, retrieval algorithm, access or processing date, flags applied and transformation formulas. Readers should be able to reproduce a table from the same data and choices, although underlying archives may be revised. A clear method protects future reanalysis. A beautiful map with no version information is difficult to verify after the product is reprocessed.
Error patterns, diagnoses and repairs
If a coastal image shows an implausibly bright fringe hugging land, first suspect adjacency, stray light, land mask geometry or bright shallow bottom effects. Compare nearby quality flags, more offshore pixels and independent observations. A plausible repair is stricter edge masking or a processor validated for that coastline, not manually painting over the bright fringe. Any revised area estimate must use the changed valid-pixel denominator and disclose the lost coastal coverage.
If retrievals flatten at the highest field concentrations, the model may be saturating or the selected optical band may have insufficient sensitivity. Plot reference concentration against reflectance and residuals across the range. A better-suited band or validated blended algorithm may help, but it must be calibrated without sacrificing low-concentration behaviour. Writing ‘high turbidity’ for every saturated pixel is an honest categorical label; inventing precise numbers beyond the model range is not.
If satellite TSM is consistently above field gravimetric TSS, consider differences in particle composition, sample depth, mass protocol, time offset, atmospheric correction and model bias. Resist treating the laboratory value as infallible or automatically adjusting a single intercept. Check whether the discrepancy increases with concentration or changes seasonally. A correction fitted after viewing the validation results must be revalidated on new data; otherwise the supposed independent test has become training data.
If a monthly series shows a sudden drop exactly at a sensor transition, check data versions, band definitions, masks and temporal coverage before inferring ecological improvement. If a storm appears invisible, ask whether cloud screening removed the event and whether the surviving observations represent its aftermath. If a field-versus-pixel scatterplot has many low reference values, be cautious with percentage errors. Each pattern has a different repair; none is solved by making the legend more colourful.
Three field interpretation cases
Consider a hypothetical estuary receiving a fresh river pulse. Surface satellite reflectance rises in red and NIR channels after heavy rain, and field filters confirm increased suspended mass at several near-surface stations. That combination supports a local increase in surface suspended matter during the sampled window. It does not alone quantify sediment deposition downstream. To progress toward a transport estimate, we would need water velocity, section geometry, time integration and possibly vertical concentration profiles.
Consider a hypothetical sandy shallow shelf where a false-colour image shows broad bright areas on calm days. Field samples show little suspended mass, and the bright patches track shallow bathymetry rather than river discharge. Bottom contribution or other optical interference becomes a candidate explanation. A competent analyst would test bottom sensitivity and masks instead of forcing the observations through a turbidity relation. The correct communication may be that the suspended-matter product is unreliable in those shallow pixels.
Consider a hypothetical port where field turbidity increases during dredging, yet a satellite product returns missing data for most operating hours because of clouds. The operational record and satellite record answer different questions. The satellite’s absence cannot refute the field trend. An evidence summary could combine high-frequency station measurements with occasional clear-sky spatial coverage while keeping their measurement methods separate. Neither series should be disguised as a complete replacement for the other.
These cases illustrate a transferable discipline: formulate a physical explanation, identify an observation that could falsify it, and check whether independent evidence exists at the correct scale. The map is one member of an evidence system, not its judge. Terms such as ‘sediment plume’, ‘resuspension’ and ‘water quality impact’ each require their own supporting inference. Accurate translation preserves those causal boundaries.
Independent transfer problems with checked answers
Problem one: a gridded image contains 250 valid ocean pixels, each covering 0.09 km², and 80 of those valid pixels exceed a stated TSM threshold. The estimated mapped high-concentration area is 80×0.09=7.2 km². The fraction of observed valid area above threshold is 80/250=32%. If the region contains 400 ocean pixels in total, it would be incorrect to call 32% the fraction of the entire region. Its full-domain mapped fraction is 80/400=20%, while 150 pixels remain without valid estimates.
Problem two: a new local linear example uses C=1,500R−5 mg/L with R in sr⁻¹ and the coefficient carrying corresponding units. For R=0.014 sr⁻¹, C is 1,500×0.014−5=16 mg/L. If calibration covered R from 0.009 to 0.020 sr⁻¹, the computed point falls within that numerical interval. But it still needs uncertainty and water-type checks. At R=0.003 sr⁻¹ the equation gives −0.5 mg/L, a sign that the model cannot be treated as universally physical.
Problem three: a field filter gains 0.0060 g after 0.300 L of seawater, and its appropriate handling blank gains 0.0006 g. Corrected retained mass is (0.0060−0.0006) g, or 5.4 mg. The concentration is 5.4 mg / 0.300 L =18 mg/L. The important mistake to avoid is treating grams as milligrams without multiplying by one thousand. The resulting answer also depends on whether the blank is appropriate to that laboratory batch.
Problem four: reference concentrations are 8, 16 and 32 mg/L; satellite estimates are 10, 14 and 29 mg/L. Residuals are +2, −2 and −3 mg/L. Mean signed bias is −1 mg/L; mean absolute error is 7/3 or approximately 2.33 mg/L; RMSE is square root of (4+4+9)/3 or about 2.38 mg/L. The dataset is too small for credible global performance claims, but sufficient to practise the statistics and their different meanings.
Problem five: an apparent coastal plume covers 12 km² in one image and 15 km² in another, but valid coverage falls from 95% to 60%. Does the three-square-kilometre difference prove expansion? No. The mapped area may change with cloud masks, algorithm support and the times sampled. Compare only like-for-like domains and conditions, or give a coverage-aware range. The observed change is a hypothesis to investigate, not automatically a physical growth rate.
Problem six: a translator encounters ‘turbidity was 14 FNU and TSM was 28 mg/L’. Can this become ‘sediment concentration was 14 mg/L’? No. The first figure is a measured or derived optical turbidity, and the second is a mass concentration. Even if a local calibration gave exactly two milligrams per litre per FNU, its method and uncertainty would be required. Preserve both quantities, units and the relationship between them rather than compressing the sentence into a false equivalence.
Questions readers often ask
Does a brighter satellite pixel necessarily mean more sediment? Not necessarily. Brightness can depend on atmospheric correction, bottom reflectance, optical composition and display stretching. A validated concentration retrieval provides a stronger basis, but it still has a defined range. Read the numeric legend and quality flags. If a scene displays a concentration product rather than raw light, its value is an estimate generated through processing, not directly observed particle mass.
Can satellite data replace filter samples? They can extend observations across areas that field sampling cannot cover densely, and they can improve spatial context. But field measurements remain important for calibration, validation and detecting changes in particle composition. The most effective programme often combines well-documented in situ observations with carefully processed satellite scenes. Replacement language is misleading when the two methods sample different depths, times and spatial supports.
Is turbidity always converted to TSS by one constant? No. An empirical relation can be useful for a particular site and concentration range, but scattering efficiency changes with particle type and instrument method. The conversion should be named as a model, not a definition of units. A new storm season or major sediment-source change may require renewed calibration. Blindly applying a published factor to unrelated water can give a precise-looking and inaccurate result.
Why do neighbouring satellites disagree? They can differ in wavelength, viewing angle, spatial footprint, atmospheric correction, processing version, quality flags and observation time. Some of the discrepancy may be true water variability. A fair comparison aligns these factors and uses independent field evidence where available. Agreement is also not perfect proof when algorithms share assumptions and training data; dependence between methods needs to be considered.
What does an uncertainty of seven milligrams per litre mean? Its interpretation depends on whether it is a standard uncertainty, RMSE, confidence interval or another statistic. These are not interchangeable. A standard uncertainty is commonly a one-standard-deviation-like estimate under a specified model, while an RMSE summarises deviations over a validation set. Preserve the statistic name, denominator, unit, assumptions and coverage before communicating numerical reliability.
Can the map establish ecological harm, a dredging violation or a sediment transport budget? Not on its own. Such conclusions depend on thresholds or standards, exposure duration, habitat sensitivity, measurement uncertainty and often water-column dynamics. A remote-sensing concentration map may contribute important evidence, but the causal and regulatory interpretation requires additional observations. An article should never turn a colour ramp into a legal or ecological verdict without the relevant criteria and an appropriate assessment.
A decision workflow readers can reuse
When handed a suspended-matter product, begin by writing one plain-language sentence naming the observation and the inference separately. For example: ‘The satellite measured spectral radiance; a processor estimated remote-sensing reflectance; a calibrated algorithm estimated surface suspended matter.’ Then inspect the variable name, units, wavelengths, flags and algorithm description. If any link in that chain is undocumented, state the gap rather than improvising the missing methodology.
Next identify the decision that the product is meant to support. Monitoring a broad estuary for seasonal change demands different validation from checking a dredging plume near a short channel. Match the temporal resolution, spatial footprint and acceptable error to that decision. Seek independent sampling that spans the relevant water types and concentrations. Use several complementary metrics so that an apparently excellent average cannot conceal failures at the extremes.
Finally communicate what the evidence demonstrates and what remains uncertain. Give representative figures with correct units and denominators, state the validation range, and distinguish missing observations from low concentration. If an interpretation uses an additional conversion from turbidity to mass, show it. The reward is not a more dramatic map. It is a trustworthy chain from photons to calibrated estimates to a conclusion that another careful reader can examine and challenge.
Original methodological sources and connected reading
NASA Ocean Biology Processing Group, product status and algorithm descriptions; NASA SeaBASS validation resources; NASA NOMAD bio-optical dataset description. These first-party resources explain product definitions, radiometric evidence and independent validation rather than substituting a universal conversion constant.
Nechad, Ruddick and Park (2010), Calibration and validation of a generic multisensor algorithm for total suspended matter; Dogliotti and colleagues (2015), red and near-infrared turbidity algorithm; Vanhellemont (2021), atmospheric correction for OLCI in turbid Belgian waters. The cited methods have particular datasets and sensor assumptions; consult the originals before implementing their coefficients.
Connected eduKate reading: GHRSST sea-surface temperature and cloud-quality interpretation; SMAP salinity, land spillover and radio interference; Sea-ice concentration, footprint mixing and weather filters; Meaning and precision in vocabulary; How English works.