A river can be on time while its gauge is not.
An upstream gauge peaks at 14:20 and a downstream gauge peaks at 14:38. Subtracting gives 18 minutes. Before you call that a river travel time, check what the two timestamps mean and whether either clock was wrong. A few minutes of error can matter more than the change you are trying to explain.
This guide follows the difference from two clocks to a corrected comparison, then shows why corrected peak timing is still not a measurement of how long a parcel of water travelled. The worked values are fictional. You can follow them with a clock sketch and a pencil.
Check three things in order: when each sample was taken, how each clock compares with independent time evidence, and whether the two points on the curves represent the same physical feature.
Hydrology subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Route 1 · Chapters 1–3: Read the record
Separate the water level from the time label
Find the source of each timestamp and calculate how clock errors alter a difference.
Route 2 · Chapters 4–6: Correct the timing without guessing
Work a gradual drift between two independent checks
Work a drift correction, identify a reset and keep sampling limits visible.
Route 3 · Chapters 7–9: Interpret and report
Decide whether the two crests represent the same event feature
Separate the gauge result from the river process, build a review and transfer the method.
Full chapter index · Independent worked case · Questions and practical checks
Chapter index
Read the record
Separate the water level from the time label
Find out which clock supplied the time
Work the sign of the correction before adjusting a series
Correct the timing without guessing
Work a gradual drift between two independent checks
Do not smooth a reset into a drift curve
Keep the sampling interval in the answer
CHAPTER 1 OF 9 · Read the record
1. Separate the water level from the time label
River stage is the height of the water surface relative to a stated reference at a site. A series of stage readings shows how that height changed. Each point needs a value and a time. A well-calibrated stage sensor can measure height correctly while its logger attaches the wrong time. In that case the shape can still look convincing, especially if the clock is only a few minutes out.
Write one line of data on paper: “2.08 metres at 14:20.” Circle the height and underline the time. A datum error changes how you interpret the circled part. A clock error changes how you interpret the underlined part. In some investigations both matter, but testing one does not verify the other. A new stage calibration does not synchronise the clock, and a correct clock does not repair a moved sensor.
Suppose a downstream gauge reports a crest before the upstream gauge. That could reflect real local inputs or a different feature, but it could also follow from inconsistent timestamps. Do not repair the series just to put the events in the expected order. Treat the unexpected result as a question: which parts can be checked independently of the assumed river behaviour?
The same caution applies when a lag looks reasonable. Two timing errors can hide within a plausible number. If you intend to compare event timing, the time reference and its uncertainty belong in the evidence, not in a footnote after the interpretation. A plausible sequence is a useful check, but it is not independent proof of a clock correction.
CHAPTER 2 OF 9 · Read the record
2. Find out which clock supplied the time
A reading can be taken at one moment, saved at another and transmitted later. The time a data service received a message is not necessarily the time the river was measured. A connection outage can delay several readings and deliver them together. If you use arrival time for every reading in the batch, the apparent crest can move without any change in the physical event.
A timestamp can also represent a boundary of an interval rather than an instantaneous observation. A fifteen-minute mean labelled 14:30 might cover the preceding quarter hour, the following one, or a window centred on the label. A maximum over an interval tells you the peak occurred somewhere in that interval if the input is adequate; it does not necessarily give the instant the maximum occurred. Record the timestamp convention before aligning stations.
Check the time scale and time zone. A local display and a UTC export can refer to the same instant with different clock labels. A one-hour discontinuity in a region that changes civil time could reflect a conversion problem rather than a logger oscillator suddenly speeding up. Use unambiguous time representations and keep any conversion rule. Do not diagnose drift from labels that have not been put on the same basis.
The practical check is a short chain: sensor or logger clock, stored observation timestamp, transfer timestamp, published display. Find which one the analysis uses. If a file reconstructs times from a start time and sample interval, record that too. Errors in a starting label and in a sampling interval affect the series differently. Treat time as a measured part of the record.
CHAPTER 3 OF 9 · Read the record
3. Work the sign of the correction before adjusting a series
Define clock error as the displayed logger time minus the reference time. A fast clock has positive error. If it displays 14:20 at a true 14:17, its error is +3 minutes and you subtract three minutes to recover the reference time. A slow clock that displays 14:20 at a true 14:22 has error −2 minutes; subtracting that error adds two minutes. Decide the sign convention and write it down once.
For two gauges, let the apparent lag be the downstream label minus the upstream label. The true lag for those chosen samples equals the apparent lag minus the downstream clock error plus the upstream clock error. This works because each sample is corrected at its own station. It is easy to reverse the sign if you remember only that a clock was “four minutes different.” Use a tiny two-line calculation instead.
In a fictional example, the upstream crest is labelled 14:20 and its logger is two minutes fast. The downstream crest is labelled 14:38 and its logger is five minutes fast. Corrected times are 14:18 and 14:33. The separation of those samples is 15 minutes, not 18. Because the downstream label was further ahead, the apparent lag was too large by three minutes.
Change the downstream clock to five minutes slow, keeping both labels the same. Now its corrected time is 14:43 and the corrected separation is 25 minutes. The direction of the correction matters. If two clocks have the same offset at the moments in question, the offset cancels in their difference, but both timestamps can still be wrong for comparison with rain, another river or an external incident log.
CHAPTER 4 OF 9 · Correct the timing without guessing
4. Work a gradual drift between two independent checks
In this fictional case, an operator checks the same downstream logger against a reliable reference at 08:00 and again at 16:00. At the first reference instant the logger is two minutes fast; at the second it is six minutes fast. The reference-time interval is eight hours. The observed change in error is four minutes, or half a minute per hour on average. That average is not proof of a uniform drift.
Suppose the service log shows no reboot, time adjustment or other interruption, and independent evidence makes a roughly linear change plausible for this example. At reference time 12:00, halfway between the checks, the estimated error is four minutes. A reading labelled 12:04 then corresponds to roughly 12:00 under this model. This is a model based on known reference checks, not a shift chosen to make two hydrographs line up.
If you need to solve a less tidy example, keep the time variable straight. The interpolated error is defined against reference time. You can express logger time as reference time plus estimated error and solve for the reference time, or map between the paired logger and reference check times using a consistent affine transformation. Do not mix elapsed time on one clock with a scale estimated for the other without checking the effect.
Retain the raw labels and the paired checks. Record the model, the interval it applies to and the evidence for treating drift as gradual. USGS guidance for processing streamflow records discusses time corrections for field clock errors, including prorating suitable errors across a record period. It does not make a gradual correction appropriate across an independently documented reset or an unexplained discontinuity.
CHAPTER 5 OF 9 · Correct the timing without guessing
5. Do not smooth a reset into a drift curve
Imagine a second logger starts two minutes fast and ends exactly on reference time. A linear correction between the two checks looks tempting. Now read its maintenance log: at midday the clock was reset. Before that moment the offset was close to two minutes, and after it the clock was close to reference time. A line drawn across the whole interval would apply a fictional one-minute error near the reset and distort readings on both sides.
Split the record only where the evidence supports a boundary. Clock checks, restart logs, duplicate time labels, sample sequence numbers and telemetry messages can help identify a jump. A gap by itself does not prove a reset. A repeated label might come from a time-zone conversion. Look for a pattern across independent evidence, and keep ambiguous parts of the sequence explicitly uncertain.
If a reset moved the clock backward, two observations can share a displayed time even though they occurred at different instants. Do not silently drop one to make a tidy time index. Use the original acquisition order, counters or reference events to decide whether both can be placed reliably. If the order or time cannot be recovered, mark that interval as unsuitable for a precise lag estimate.
A useful check is to compare the model with a reference observation it did not use. If a mid-period independent clock check disagrees substantially with the straight-line correction, do not add a smooth curve merely to fit all available points. Ask whether there was a jump, a measurement uncertainty or another clock in the chain. A better fit is not always a better explanation.
CHAPTER 6 OF 9 · Correct the timing without guessing
6. Keep the sampling interval in the answer
Correcting a timestamp does not create samples between readings. Suppose one gauge records at five-minute intervals and another at fifteen-minute intervals. The highest observed value at each gauge is the maximum of its samples. The physical peak may have occurred between them. If a narrow crest was missed entirely, the highest sampled time may be a poor marker even after the clocks are corrected.
In a simple case with a smooth, single crest and reliable neighbouring values, you may be able to put a useful window around when it occurred. Without such assumptions, the sample interval alone is not a universal error bar. A flat plateau can provide several equally plausible peak times. Multiple crests can make the choice of which feature to compare more important than the timestamp correction.
For a teaching estimate, suppose independent evidence places an upstream crest between 14:15 and 14:20, and the matched downstream crest between 14:30 and 14:40, with all times already on the same basis. The possible separation from these bounds runs from 14:30 minus 14:20, or 10 minutes, to 14:40 minus 14:15, or 25 minutes. This is an interval based on the stated bounds, not a statistical confidence interval.
Add clock uncertainty separately if it has not already been included. If the upstream window still has a possible common offset of one minute in either direction and the downstream window two, a conservative bound in this simple case can widen by up to three minutes on each side. If the errors are shared or correlated, the bound or uncertainty calculation may differ. Name the assumptions rather than implying every uncertainty combines the same way.
CHAPTER 7 OF 9 · Interpret and report
7. Decide whether the two crests represent the same event feature
Two corrected peaks do not necessarily represent one feature travelling downstream. Between stations, a tributary may add water, a control may alter flow, or storage may reshape the hydrograph. One crest can broaden, split or be replaced by a larger crest from another input. Before calling the peak difference a propagation measure, examine the wider event, the locations and any important intervening changes.
Peak timing is also not the speed of a water parcel. A stage disturbance can move differently from the water itself. A dye tracer and a river-stage crest are different observables. Even if both are measured accurately, they answer different questions. A trustworthy report might describe a lag between matched stage features without translating it into a water travel time at all.
A curve-matching method can help identify candidate lags, but it cannot by itself distinguish a clock offset from a real propagation delay. Both can shift a feature along the time axis. Do not use the same assumption twice: first shift the series to produce the lag you expect, then cite the shifted series as evidence that the lag is correct. Establish clock corrections with information not derived from the desired river comparison.
A practical test is to compare more than the peak. Check the rising limb, shoulder and recession, and look for known rainfall or tributary changes. A shift that aligns one point but leaves other features incompatible may signal that the curves were reshaped or matched incorrectly. Do not demand that every part of a river response align under one time shift; real hydrology need not behave that way.
CHAPTER 8 OF 9 · Interpret and report
8. Write a review that another reader can reproduce
Preserve the original stage series, original time labels and source. Add a corrected series with a clear record of which timestamps changed, why and under which assumptions. Keep station identifiers, time zones, sampling intervals and the logger checks. If a source record is provisional or later revised, name the version used so a reader can explain any differences they find.
A useful review table has one row per gauge and event: raw feature time, independent clock evidence, applied correction, sampling or feature window, and residual uncertainty. Put the river interpretation in a separate field. This stops a numerical correction from being mistaken for the conclusion. It also makes it possible to revise one gauge’s correction without rewriting the whole account.
For the worked example a short statement could say: “The labels were 18 minutes apart. Independent clock checks support a 15-minute separation between the selected samples after correction. Actual crest timing is less precise because the samples are discrete. The result describes the selected stage features, not water-particle travel time.” If you cannot establish independent clock evidence, say that the apparent lag remains provisional.
Do one check that could prove you wrong. Ask a second reader to reproduce the corrected labels from the clock evidence without showing them the lag you expect. Then ask whether the two features should be matched from the event evidence. Separating those questions reduces the chance that an appealing interpretation will drive the time correction.
CHAPTER 9 OF 9 · Interpret and report
9. Transfer: compare a rainfall logger and a road-water sensor
A fictional neighbourhood study compares a rainfall burst with water ponding in a road sensor record. The rain logger reports a fifteen-minute accumulation ending at 15:15. The road sensor records an instantaneous depth reading at 15:18. Can you say ponding began three minutes after the rain began? No. The rainfall label marks the end of an interval, and the first observed ponding may be later than its actual beginning.
Now suppose the road sensor had a documented two-minute fast clock. The corrected depth reading is at 15:16. That helps align the records, but it still does not identify which minute inside the rain interval started the response. Preserve the time support of both measurements. What transfers from the river case is the distinction between a time label and the event it represents.
Change one fact: a rain series now reports an instantaneous rate every minute, with independently verified time. The timing can be more informative, but a local gauge may not represent rainfall at the road. The physical comparison has a new limit even when time is well known. A better clock does not fix a spatial mismatch.
For the final transfer check, write the two strongest claims you could make: one about the observations and one about the process that may connect them. If the process claim is more precise than the observations can support, revise it. This is a practical way to keep measurements useful without treating them as evidence for more than they can show.
Independent worked case and evidence checks
A fictional upstream logger records a stage feature at 10:12. Independent evidence puts its clock error between +1 and +2 minutes. The corrected sample time is therefore between 10:10 and 10:11. A downstream logger labels a similar feature 10:25; its clock error is between −2 and 0 minutes, so the corrected sample lies between 10:25 and 10:27.
The raw labels differ by 13 minutes. The possible difference between those corrected sample times is 14 to 17 minutes, taking the earliest downstream time minus the latest upstream time for the lower bound, and the latest downstream time minus the earliest upstream time for the upper bound. This bound accounts for the stated clock evidence only. It says nothing yet about where the continuous crest fell between the samples.
Now a telemetry log shows that the downstream label came from server receipt, not measurement time. The calculation no longer supports a measurement-time lag, because the known offset belonged to the logger clock and was applied to a different timestamp. This is not a smaller uncertainty on the same calculation. It is a mismatch between the correction and the measured quantity. Recover the logger’s actual record or leave the lag unresolved.
Finally, suppose a true measurement label is recovered and the two sampled features can be placed in time, but a large tributary rose between the gauges. The clock calculation may still be valid. The claim that one feature propagated between sites is now less secure. A good report does not throw away a correct timing result just because a physical interpretation is uncertain. It keeps the two levels of evidence apart.
Try the same reasoning with a one-day record. At 06:00 reference time, the logger reads 06:03. At 18:00 reference time, it reads 18:09. If linear drift is justified between those paired checks, a reference time of 12:00 corresponds to a logger reading of 12:06. Now suppose a maintenance image shows the logger displaying 12:02 at 12:00. The new observation is four minutes away from the prediction. The simple model no longer fits the available time evidence at the precision you planned to use.
There are at least two next checks worth making. First, verify that the image time comes from a trustworthy source, not from a phone whose clock was also out. Second, look for an adjustment or pause between checks. If you cannot resolve the disagreement, avoid publishing a minute-specific crest time from that interval. A broader interval might be defensible, or you may have to omit the comparison. Which choice is appropriate depends on the question and the evidence, not on how attractive the final graph looks.
For a separate sensitivity check, compare the same two hydrographs with and without the defensible clock correction. Record how far the selected sample lag moves. If the correction is larger than the difference you were trying to detect between two storms, do not claim that one storm travelled faster until you can account for timing uncertainty at both stations. This is especially important when comparing studies that used different data versions, logger intervals or time conventions.
Leave a traceable label on every output. A raw plot is suitable for showing what the logger recorded. A corrected plot is suitable for comparisons supported by the time checks. A plot with a range of plausible corrections can help a reader see what the available record cannot resolve. None of these versions needs to hide the others.
Worked comparison
| Evidence | What it can support | What to avoid |
|---|---|---|
| Two reference clock checks and stable operation | A bounded drift correction under a stated model | Extrapolating through a reset without evidence |
| Sampled peak times on one time basis | Separation between selected samples | Calling it exact continuous crest timing |
| A crest pair and consistent event evidence | A qualified stage-feature comparison | Calling it water-particle travel time |
| Only transmission receipt times | When messages arrived | Assuming the river was measured at that instant |
Questions and practical checks
Is clock drift the same as a fixed offset?
No. A fixed offset is constant over a defined interval. Drift changes the offset over time. A reset may instead introduce a jump.
Can I align two curves by their peaks to estimate clock error?
That can suggest a difference, but it cannot independently distinguish clock error from a physical lag. Use separate clock evidence.
Do two equally fast clocks cancel?
Their equal errors cancel in their difference at the relevant times, but the absolute timestamps can still be wrong.
Does a crest lag tell me how fast water moved?
No. A stage feature and the motion of individual water parcels are different quantities.
Should all clock errors be corrected linearly?
No. A linear model needs a defensible interval and evidence against resets or other changes. Preserve uncertainty when evidence is insufficient.
Should I overwrite original times?
No. Preserve the raw records and document the corrected version, reference and assumptions.
Continue with the connected subjects
Check river observations and references
Gauge datum shifts: a different error that changes the height record
Pressure-based water levels: check the sensor and its reference
Interpret physical timing carefully
Flood-wave celerity and water-particle velocity: distinguish the two motions
Bank storage and flood waves: why hydrographs change shape
Dye-tracer travel times: another measurement with a different physical meaning
Sources and scope
USGS: Processing and Publication of Surface-Water Data, including clock corrections for unit values
USGS: Stage Measurement at Gaging Stations
USGS: quality assurance for stage records, including checks on timing
The numerical cases in this guide are original fictional teaching examples, not measurements from a real site. This is an educational explanation, not a field safety instruction, engineering approval or an official certification. For a real project, use qualified practitioners, current product documentation and site-specific evidence. Sources were checked on 5 November 2026.