LEARN / OCEAN WAVES / CHAPTER 29

Wave-data Quality Control

Full original scientific lesson and experiments, reorganized into a dedicated chapter.

Original physics and annotationsInteractive scientific demonstrationsScientific teaching examples

Scientific scope: This lesson retains the original scientific explanations, diagrams, equations and interactive demonstrations. All numerical examples are educational and subject to their stated assumptions; they are not independently field-validated engineering design tools or operational forecasts.

CHAPTER 29 / WAVE-DATA QUALITY CONTROL

Keep the evidence.
Question the result.

Quality control makes observations interpretable and their limitations visible. It combines instrument knowledge, reproducible tests and scientific review. The objective is to identify unreliable information without erasing real extremes, unusual sea states or slow motions merely because they are unexpected.

Quality assurance prevents; quality control evaluates.

Calibration, deployment design, maintenance and configuration establish the conditions for reliable measurement. QC then evaluates the observations and derived products. A check cannot reconstruct a clipped crest, establish an unknown sensor elevation or compensate for an undocumented clock error merely by assigning a colour.

A passing test means that its particular criterion was satisfied. It does not prove that a record is correct. A suspect flag means that further attention is needed; an unusual storm wave may deserve that attention just as much as a sensor fault.

Evaluate the right level.

Separate individual samples, complete bursts, spectral bins and successive sea-state summaries. A missing sample is not the same as a missing sea state. An unreliable high-frequency bin need not invalidate every lower-frequency component, but it may prevent release of a parameter that depends on the full band.

Preserve the raw observations, individual test results and processing history. Real-time screening and later scientific review answer different questions with different information. Keep both versions rather than rewriting the original evidence.

01 · AcquireCalibration, geometry, clock, raw packets
02 · Inspect samplesIntegrity, missing values, waveform faults
03 · Inspect productsResponse, spectra, moments, direction
04 · Review and releaseContext, uncertainty, flags, audit trail

A reference framework, with an explicit scope.

The principal wave-specific crosswalk below follows IOOS QARTOD In-Situ Surface Wave Data, version 2.1 (2019). Its hierarchy applies within that framework; it is not a declaration that every instrument everywhere must run every test. Tests 17–18 concern supplied spectra. Instrument-specific tests need their corresponding channels.

Reference roles: standards, guidance and educational extensions
ReferenceRole in this chapterBoundary of the claim
IOOS QARTOD waves v2.1, 2019Wave-test index and flag vocabularyOperator-defined thresholds and sensor applicability remain essential.
IOC Manuals and Guides 26, 1993Historical international oceanographic validation guidance; wave-data appendixExamples require adaptation to present instrumentation and local conditions.
WMO-No. 8, 2024 editionObservation and quality-management framework; calibration and intercomparisonThe 2026 material identified as provisional is not treated as a finalized replacement here.
ISO/IEC 17025:2017International standard concerning testing/calibration laboratory competenceIt does not supply a universal wave-spike threshold. This website claims no accreditation or clause-by-clause compliance.
CF conventions 1.10, §§3.4–3.5Published data-exchange conventions for ancillary data and flagsMetadata conformance does not establish physical accuracy; the chosen convention version must be declared.

The 32-method guide is an original engineering synthesis, with practical implementation notes and cautions. It extends beyond the QARTOD test index to spectral mathematics, directional consistency, processing reproducibility and delayed review. The interactive algorithms are deliberately small teaching examples, not a certified implementation of all referenced procedures.

All 21 QARTOD wave tests: reference crosswalk

Numbers and priority labels refer to v2.1, Table 3-2 and §§3.3.1–3.3.3. Names below are abbreviated. Follow each link to the original implementation guidance in this lesson. ST means a short sampled record; LT means a series of wave products.

TestTopicManual hierarchy
1Echo strengthStrongly recommended
2Echo correlationStrongly recommended
3Acoustic noiseStrongly recommended
4Signal/noiseStrongly recommended
5Pressure/surface trackingStrongly recommended
6Acoustic velocity rangeStrongly recommended
7Acoustic velocity statisticsStrongly recommended
8Usable sample countStrongly recommended
9Short-record gapsStrongly recommended
10Short-record spikesStrongly recommended
11Short-record rangeStrongly recommended
12Segment shiftsSuggested
13AccelerationStrongly recommended
14Check factorsStrongly recommended
15Long-term mean/SDStrongly recommended
16Long-term flat linesRequired
17Operational frequency bandRequired
18Low-frequency energyRequired
19Bulk-value rangeRequired
20Bulk-value changeRequired
21Neighbour comparisonSuggested

Tests 1–8 are acoustic-instrument diagnostics. Their presence here does not add a wave-driven-current lesson. The laboratory below implements its own limited sample tests, not this entire crosswalk.

Fault laboratory: a flag is evidence, not a repair.

The clean reference is a prescribed two-component signal sampled at 2 Hz for 128 s. Choose an injected fault and inspect the unchanged received record, diagnostic residuals and spectral consequences. The reference is known only because this is a synthetic experiment; field QC rarely has such a truth signal.

A · Received record and known reference
Navy: received values. Dashed grey: clean reference. Red crosses: hard range failures; orange circles: local-test suspects; grey bands: missing samples. Markers are not replacement values.
B · Local residual and threshold
ei = |xi − (xi−1 + xi+1)/2|. Dashed coral: selected threshold. Missing neighbours and record endpoints are not evaluated.
C · Variance-density consequence
Blue: clean reference. Coral: received record or an explicitly labelled gap-filled copy. Both are demeaned, rectangular-window, one-sided periodograms. No flagged sample is silently deleted.

Teaching rules: |η| > 1.5 m → fail; local residual > threshold → suspect; a run of ≥8 samples with span ≤10⁻6 m → suspect. A half-record mean difference >0.30 m triggers a separate burst-level review. These limits are not published operational standards.

Exactly what this laboratory computes—and what it misses
Synthetic reference record
ηref(t)=0.40cos⁡(2π⋅0.125t)+0.15cos⁡(2π⋅0.1875t+0.7)
Sampling and record duration
Δt=0.5 s,N=256,D=NΔt=128 s
Spectral-bin resolution
Δf=1D
Reference variance
Vref=0.402+0.1522=0.09125 m2
Discrete Fourier transform
Xk=∑j=0N−1(xj−x―)exp(−i2πjkN)
One-sided variance-density periodogram
Sk=ck|Xk|2fsN
Endpoint convention
ck=1 (DC and Nyquist),ck=2 (elsewhere)
Variance conservation
∑kSkΔf=1N∑j(xj−x―)2

Samples occur from 0 to 127.5 s; the DFT record duration is NΔt = 128 s. The displayed spectrum uses the complete record, not only passing points. This makes contamination visible without compressing the time axis by deleting observations. A direct DFT is used for this small example; an FFT computes the same coefficients more efficiently.

Missing samples receive 9; a hard range failure receives 4; otherwise a detected local anomaly receives 3. A valid point with an unavailable three-point test receives 2; other points receive 1 for this limited test set. A segment shift is a separate record-level review, not automatically assigned to all samples. This is an explicit teaching aggregation policy, not a complete operational flag algorithm.

The flat-run detector flags an entire run retrospectively. The centred residual needs a future neighbour. These are delayed-inspection methods, unlike a strictly causal sample-by-sample service. The optional linear fill is allowed only for bounded gaps of at most eight samples; there is no extrapolation. Raw values and flags stay unchanged. A four-second interpolation can erase much of a short wave even though the curve looks plausible.

Clipping can pass all three local tests. A mean offset may also escape a spike threshold. Lowering a threshold can flag ordinary wave curvature. These are deliberate demonstrations of missed faults and false alarms, not defects hidden by the lesson. The spectral panel is omitted when a complete analysis copy is unavailable.

Read indicators in their scientific context.

A spike raises broadband variance; clipping can lower it. A shift can add low-frequency content. A gap cannot be made harmless simply by removing its missing timestamps. Compare the two spectra and the variance budget, then return to the received waveform to understand the cause. The largest change in a plot is not necessarily the most consequential error for a particular product.

A comprehensive method guide.

Each method below identifies its inputs, an implementation approach and a reason to avoid automatic overconfidence. Together they cover the observation lifecycle. They are a planning and review guide, not evidence that all methods have been run on an uploaded field dataset.

01 · Purpose, coverage and acceptance plan

Input level: Before deployment

Method. Define the quantity, frequency band, location, averaging interval, latency and acceptable uncertainty. Specify which tests operate on samples, bursts, spectral bins or successive sea states. Assign an owner to each exception and a deadline for review.

Interpretation and limitations. Document a test as not applicable when its inputs do not exist; do not silently call it passed. Requirements for storm detection and routine climatology may differ. A single checklist cannot establish fitness for every use.

02 · Calibration and traceability

Input level: Before and after deployment

Method. Keep calibration coefficients, dates, sensor serial numbers, reference instruments and uncertainties. Check zero, span and response across the intended operating range; compare post-recovery results with the original calibration before proposing a correction.

Interpretation and limitations. A new offset does not prove that a constant correction applies throughout a deployment. Drift may be gradual, abrupt or caused by fouling that changes frequency response. Preserve the original measurements and the correction model separately.

03 · Deployment geometry and health

Input level: Every deployment and service event

Method. Record position, water depth and datum, sensor elevation, orientation, mooring configuration, firmware and sampling settings. Review tilt, battery, temperature, diagnostics, fouling and service logs alongside wave products.

Interpretation and limitations. A moved sensor can remain within a numerical range while representing a different water depth or location. A pass flag from software cannot replace inspection of mounting, orientation and environmental exposure.

04 · Message integrity and provenance

Input level: Every received message

Method. Validate packet length, checksum where provided, schema, encoding, station identity, units and sequence numbers. Detect duplicate or truncated messages before joining them into an observation. Retain receipt time separately from measurement time.

Interpretation and limitations. A checksum supports transmission integrity, not physical accuracy. Retransmitted observations should be reconciled by identifier and version rather than counted as additional independent samples.

05 · Time stamps, order and clock drift

Input level: Raw time axis

Method. Inspect Δti=ti−ti−1 against the declared sample interval, including duplicates, reversals and clock resets. For bursts, verify start/end times, time zone and the relationship between sample count and duration. For arrays, check relative clock alignment.

Interpretation and limitations. A small timing bias can change phase and inferred direction without noticeably changing a single-channel spectrum. Sorting a corrupted series does not explain why its time axis failed.

06 · Gaps, completeness and longest outage

Input level: Samples and bursts

Method. Create an expected sample grid from the documented sampling plan. Count missing samples, separate isolated losses from consecutive runs, and report both missing fraction and longest gap in seconds. Compare usable duration with the analysis requirements.

Interpretation and limitations. Equal missing percentages can have very different effects: ten isolated losses and one long outage are not equivalent. An absent value remains missing, never zero. A failed derived product can coexist with many individually usable samples.

07 · Hard range and contextual range

Input level: Calibrated channels and bulk values

Method. Use separate limits for sensor capability and environmental plausibility. Check units and offsets before applying them. Mark impossible or saturated values distinctly from unusual values that merely exceed a site history.

Interpretation and limitations. A storm can exceed a historical envelope while remaining physically real. Ranges need a documented basis and version. Neither a universal maximum wave height nor a single height-to-depth ratio is appropriate to all sensors and settings.

08 · Isolated spikes and robust outliers

Input level: Raw channel values

Method. Compare a point with its local neighbourhood and with a robust distribution estimate. A simple illustrative residual is ei=|xi−(xi−1+xi+1)/2|. Inspect adjacent points, raw diagnostics and waveform continuity before classifying a large residual.

Interpretation and limitations. This centred test needs the next sample and is delayed by one sample. A spike can contaminate its neighbours’ residuals. Large crests, breaking and rapid physical change can also trigger a detector; clipping all statistical outliers biases extremes downward.

09 · Flat lines and repeated values

Input level: Raw samples and successive products

Method. Search for runs whose maximum minus minimum is within a tolerance tied to resolution. State the run length in samples and seconds. Compare the apparent freeze with other channels and telemetry counters.

Interpretation and limitations. Quantized low-energy data or rounded peak periods can legitimately repeat. The raw-sample illustration here is distinct from the QARTOD long-term bulk-parameter flat-line test. A repeated zero should be investigated rather than exempted automatically.

10 · Clipping and quantization

Input level: Instrument output

Method. Look for repeated occupancy at upper/lower rails, long horizontal plateaus at crests or troughs, and unusually coarse increments. Compare observed rails with electronic and configured ranges. Inspect the histogram as well as the waveform.

Interpretation and limitations. Clipping often reduces variance and creates harmonics while leaving every recorded sample inside a nominal range. Increasing a software range limit cannot recover the lost crest information.

11 · Mean shifts, drift and segment changes

Input level: Within a burst and across deployments

Method. Divide a sufficiently long record into comparable segments and inspect means, variances and their changes. Use a two-sided difference, |μj+1−μj|, with a documented limit. Check service times, depth changes, timing resets and low-frequency content.

Interpretation and limitations. Segments must span enough waves to avoid confusing phase or grouping with an offset. A genuine long wave or water-level change may alter the mean; removing it is a scientific processing decision, not an automatic QC repair.

12 · Acceleration and impulse checks

Input level: Buoys with appropriate raw channels

Method. Compare acceleration with calibrated sensor limits after establishing axes and gravity treatment. Use recorded diagnostics to distinguish impacts, attitude problems and saturation. Record whether the channel contains gravity or has already been compensated.

Interpretation and limitations. A finite difference of sampled elevation is a noisy proxy, not an interchangeable accelerometer observation. Manufacturer limits and filtering affect what an acceleration threshold means.

13 · Acoustic amplitude, correlation and noise

Input level: Acoustic instruments only

Method. Review signal strength, echo correlation and acoustic noise for each contributing beam/bin and surface-tracking return. Confirm whether values are linear power, amplitude, counts or decibels before computing a ratio or applying a threshold.

Interpretation and limitations. SNR in dB is not formed by dividing two dB values. Low correlation or poor tracking can invalidate a wave estimate even when its final height looks plausible. Instrument-specific diagnostics must follow the relevant manufacturer definitions.

14 · Pressure and surface tracking plausibility

Input level: Pressure or acoustic surface return

Method. Compare mean level and variability with depth, atmospheric referencing, sensor elevation and plausible water-level excursions. Check for lost returns, bottom/structure echoes, emergence and sudden changes of reference distance.

Interpretation and limitations. Do not assume pressure is hydrostatic at every wave frequency. The frequency response and correction limits in Chapter 28 still apply after these basic checks.

15 · Wave-channel velocity diagnostics

Input level: Acoustic wave-estimation channels

Method. For instruments deriving wave information from orbital-velocity channels, inspect each selected beam/bin for range errors and anomalous mean or variability, using the documented coordinate frame and valid-sample mask. Count surviving samples after diagnostic rejection.

Interpretation and limitations. These are instrument-channel checks for wave estimation, not a treatment of wave-driven currents. Do not apply one instrument’s velocity thresholds to a different measurement geometry or convert beam statistics to earth axes without the required transformation.

16 · Sensor response and usable band

Input level: Before inversion or spectral integration

Method. Apply the documented transfer function only over the supported band. Inspect the inverse gain and noise contribution as functions of frequency. Track depth and response changes, and report which frequencies were excluded.

Interpretation and limitations. A correction can amplify noise while producing a smooth-looking spectrum. Unsupported low-frequency buoy response and weak high-frequency pressure response require different investigations; the same cutoff cannot be assumed for every instrument.

17 · Aliasing and sample-rate consistency

Input level: Acquisition and processing

Method. Compare sample rate, instrument response and anti-alias filtering with the highest retained frequency. Check that any downsampling was preceded by suitable filtering. Report acquisition rate separately from output interval.

Interpretation and limitations. A sampled sinusoid at 0.8 Hz and another at 0.2 Hz can coincide at a 1 Hz sample rate. Filtering the already aliased record cannot decide which frequency was originally present.

18 · Spectral validity and variance closure

Input level: Each spectrum

Method. Check finite values, ordered frequencies, positive bin widths and nonnegative variance density within a stated numerical tolerance. Verify one-sided normalization, window-energy correction and the correspondence between integrated density and the processed record variance.

Interpretation and limitations. With a rectangular, demeaned complete record, Parseval provides an exact numerical check. A windowed periodogram closes against its window-weighted quantity, not automatically the unwindowed variance. Frequency density and angular-frequency density require a Jacobian.

19 · Low/high-frequency contamination and noise floor

Input level: Spectral shape

Method. Inspect energy outside the validated instrument band, unusually large low-frequency contributions, narrow electrical/mechanical peaks and high-frequency plateaus. Compare changes against instrument diagnostics and the unfiltered record.

Interpretation and limitations. A low-frequency peak can be a real seiche or infragravity process. Classify it as outside a sea-swell product when appropriate, rather than declaring that the physical observation is false. Negative noise-subtracted bins need explicit handling, not silent clipping.

20 · Stationarity and segmentation

Input level: Spectral-analysis windows

Method. Compare segment means, variances and spectral shapes before pooling a record. Record storms or transitions that make a stationary sea-state approximation weak. Choose duration to balance enough cycles against changing conditions.

Interpretation and limitations. Shorter segments reduce frequency resolution; longer ones can mix different states. Nonstationarity is often a limitation of the summary product rather than a malfunctioning sensor.

21 · Moments and derived-parameter consistency

Input level: Bulk wave products

Method. Recompute mn=ΣfiⁿSiΔfi over the declared valid band. Check Hm0=4√m0, Tm01=m0/m1 and Tm02=√(m0/m2) where denominators are positive. Identify Tp from the stated peak method, and keep H1/3 distinct from Hm0.

Interpretation and limitations. Exclude unsupported products rather than divide by zero in calm or empty-band records. A fitted peak and the largest discrete spectral bin can give different Tp values; neither algorithm should be left undocumented.

22 · Rate of change and plausible evolution

Input level: Successive sea-state products

Method. Compute changes using the actual interval: |Xj−Xj−1|/Δt. Compare with limits developed for that variable and location, and inspect storms, rapidly changing winds, telemetry gaps and processing changes.

Interpretation and limitations. Peak period and peak direction can jump when two valid spectral peaks exchange dominance. Do not apply a smoothness rule for integrated wave height blindly to these discrete peak descriptors.

23 · Climatological and robust context

Input level: Long-term records

Method. Compare observations with seasonally and directionally relevant historical distributions, using a sufficiently representative baseline. Inspect changes in instrument, depth and processing before merging records. Track flags during known storms separately.

Interpretation and limitations. An extreme outside previous experience is not automatically wrong. A contaminated baseline can normalize faults, while a narrow fair-weather baseline can reject the very events the observing programme needs.

24 · Neighbour, duplicate and model comparisons

Input level: Independent observations

Method. Match timestamps, averaging periods, frequency bands and direction conventions. Prefer collocated independent instruments for direct comparisons; allow for propagation, sheltering and depth transformation when comparing sites. Compare residuals and uncertainty, not only correlation.

Interpretation and limitations. A model is supporting evidence, not truth. Two sensors can share the same bias, and nearby sites may observe different seas. Disagreement should normally trigger investigation before automatic rejection.

25 · Directional check factors and channel consistency

Input level: Directional instruments

Method. Use the provider’s definition of the check factor, channel scaling and depth dependence. Inspect its behaviour across the valid band and over time; compare deviations with fouling, mooring changes and sensor health.

Interpretation and limitations. Do not combine unscaled heave spectra in m²/Hz with raw pitch/roll spectra in rad²/Hz and expect a universal dimensionless ratio. A single copied formula without response and channel definitions can be dimensionally invalid.

26 · Directional range, wraparound and convention

Input level: Directions and directional products

Method. Validate the encoded angle before normalizing it. Resolve true/magnetic reference and from/toward meaning. Use the signed circular difference atan2[sin(θ2−θ1),cos(θ2−θ1)] rather than ordinary subtraction.

Interpretation and limitations. 359° to 1° is a two-degree change, not 358°. A numerical wrap can hide an invalid raw value, so record the original. Mean direction becomes poorly defined when directional concentration is small.

27 · Coherence and spectral-matrix feasibility

Input level: Multichannel spectra

Method. For consistent spectral estimates, inspect coherence |Sxy|²/(SxxSyy) and Hermitian symmetry, and check that the spectral matrix is positive semidefinite within numerical tolerance. Confirm channel synchronization and compatible averaging.

Interpretation and limitations. Coherence should lie between zero and one for valid estimates with positive denominators. Near-zero energy makes ratios unstable. Testing each channel separately cannot detect every inconsistent cross-spectral combination.

28 · Directional coefficients and variance conservation

Input level: Frequency–direction spectra

Method. Check coefficient conventions and dimensions, directional normalization and frequency/direction marginals. If a1,b1,a2,b2 are normalized trigonometric moments, r1=√(a1²+b1²) and r2=√(a2²+b2²) cannot exceed one. Integrate angular density in radians.

Interpretation and limitations. These bounds are necessary, not sufficient for joint realizability. A truncated Fourier reconstruction may become negative despite admissible coefficients. Distinguish reconstruction artefacts from corrupted coefficients; do not silently delete negative lobes and change variance.

29 · Per-test flags and product aggregation

Input level: Each processing level

Method. Store results for each test plus an explicitly defined aggregate. Keep sample, burst, bin and product flags separate. Preserve the real-time record and add a versioned reviewed result. Flag a missing value as missing, not failed or passed.

Interpretation and limitations. Flag numbers are categories, not a quality score. Taking their arithmetic mean is meaningless. A good aggregate must not conceal an unevaluated required test or a missing input; document the precedence policy.

30 · Repairs, interpolation and reprocessing

Input level: Reviewed derived copies

Method. Retain original values and mark every replacement with method, bounds, rationale and software version. Compare spectra and moments before and after processing. Withhold products when gaps or corruption exceed the validated repair capability.

Interpretation and limitations. Interpolation invents estimates, not observations. It often suppresses short-period variance; zero-filling adds discontinuities. A repaired copy must never overwrite the evidence used to make the decision.

31 · Uncertainty, false alarms and missed faults

Input level: Acceptance validation

Method. Use controlled faults and independently reviewed real events to measure detection and false-alarm performance. For a known labelled set, report both missed faults and falsely flagged valid data, with the definition of each denominator. Evaluate uncertainty in retained parameters.

Interpretation and limitations. Synthetic faults only cover the failures injected. Passing this chapter’s tests does not validate an operational QC system, and aggressive screening can produce a deceptively quiet but biased archive.

32 · Release, audit trail and metadata

Input level: Distribution and archive

Method. Publish units, coordinates, time convention, valid band, calibration, sampling, window/averaging, thresholds, flag meanings, processing version and relevant uncertainty. Keep raw, provisional and reviewed products distinguishable and link flags to their data variables.

Interpretation and limitations. Metadata formats support exchange; they do not certify scientific quality. Provide a reason for an unavailable product and retain review decisions so another analyst can reproduce the result.

Two valid behaviours that can look like faults.

D · Direction crosses north
Clockwise, propagation toward. Blue: first direction; coral: second. The short arc and signed result use circular arithmetic.

E · Two spectral peaks exchange dominance
Two equal-width, idealized spectral bins keep total variance fixed. Moving energy between them can make the selected peak period jump.

The compass demonstrates arithmetic, not retrieval uncertainty. The peak example represents two discrete bins, not a fitted ocean spectrum. At an exact tie, this lesson reports both peak periods instead of pretending that one is uniquely defined.

Flags, publication and a reproducible decision.

1 · PassTest criterion satisfied2 · Not evaluatedRequired evidence absent3 · SuspectReview needed4 · FailCritical criterion failed9 · MissingNo observation

Keep a per-test result even when a summary flag is supplied. Separate flag categories from uncertainty estimates and from an analyst's final acceptance decision. Record why a test was skipped, whether a product was withheld, and which replacements were introduced. Never average flag codes or interpret a higher numerical flag as a larger measurement error.

Minimum reproducibility record for a QC decision
RecordInformation to preserve
Observation identityStation, sensor serial, deployment, location, depth/datum, sample times, time zone, units and source-file identifier
Measurement chainCalibration, response function, sampling/anti-alias settings, orientation, firmware and diagnostic channels
AnalysisBand limits, detrending, window, segment length/overlap, averaging, spectral normalization and parameter definitions
Test configurationTest name/version, input level, units, thresholds, tolerance, inequality boundary and not-applicable reason
Result and actionPer-test and aggregate flags, suspect/fail/missing fractions, longest gap, withheld parameters and uncertainty
Review historyOriginal values, replacement mask, method, reviewer, rationale, timestamps and versioned outputs
An illustrative machine-readable flag description

For a netCDF variable called wave_height, an ancillary_variables attribute can name its QC variables. An enumerated QC variable can declare flag_values = 1, 2, 3, 4, 9 and corresponding flag_meanings = "pass not_evaluated suspect fail missing". Attribute types must match the selected format/convention rules. Store thresholds, software version and individual test variables as well; a five-value list alone does not explain how quality was assessed.

A CSV export can carry the same scientific information through explicitly named columns and a versioned metadata file. Do not assume a spreadsheet colour will survive export. The definitions and processing history must travel with the values.

Before releasing a dataset, answer four questions.

  1. What was actually measured? Establish the quantity, geometry, calibration, time axis and usable band.
  2. Which evidence supports each decision? Inspect raw values, diagnostics, spectra and relevant independent observations.
  3. What information was changed or lost? Identify missing, rejected, interpolated, filtered and unsupported parts of the record.
  4. Can another analyst reproduce the release? Supply configuration, code version, flags, uncertainty and review history.