Missing water on a spreadsheet is not automatically groundwater leakage
A lake receives more estimated water than its measured outflow and storage increase account for. The difference is tempting to call seepage. But rainfall, evaporation, streamflow and storage observations all contain uncertainty, and some errors may be shared rather than independent.
This guide explains water-budget closure as an inference problem. Original calculations show how sign conventions, covariance, timing, shared references and forced reconciliation can change the interpretation of a residual. The aim is to distinguish an accounting discrepancy from evidence for a particular physical pathway.
The existing Water Balance Modelling guide develops boundaries and basic storage calculations. This article adds the missing uncertainty and identifiability treatment. All examples are fictional educational cases, not water-allocation estimates, leakage certifications or instructions for managing a real reservoir.
Hydrology subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Define the residual · Chapters 1–2
Write one signed balance for one defined system
Separate a residual from a directly observed flux
Quantify uncertainty and dependence · Chapters 3–5
Propagate independent random uncertainty as a stated special case
Include covariance with the signs of the balance
Separate persistent bias from repeatable scatter
Test time, scale and missing pathways · Chapters 6–8
Use compatible storage endpoints and their reference dependence
Align time windows before diagnosing a physical discrepancy
Test spatial support and omitted transfers explicitly
Avoid false closure · Chapters 9–10
Recognise identifiability: one residual can fit several explanations
Do not confuse forced closure with independent validation
Practise and report · Chapters 11–12
Practise covariance, timing and a residual-flux hypothesis
Report the discrepancy, uncertainty and next discriminating observation
CHAPTER 1 OF 12
1. Write one signed balance for one defined system
For a selected control volume and interval, change in water storage equals total inflow minus total outflow. Internal transfers do not appear as external flows unless the boundary definition makes them cross it. The first uncertainty problem is often not a noisy sensor but an unclear accounting boundary.
Use a simple lake example with precipitation volume P, stream inflow I, evaporation volume E, stream outflow O and observed storage change ΔS. Define residual R = P + I − E − O − ΔS. Positive R means the listed net input exceeds the observed storage increase. It does not by itself name the missing process.
Suppose P = 100, I = 500, E = 80, O = 450 and ΔS = 40, all in thousand cubic metres. Residual is 30 thousand m³. That could be consistent with an omitted outward flux, an overestimated input, an underestimated output or storage error, among other possibilities.
Keep all terms in compatible units and periods. Rainfall depth over a lake must be converted using an appropriate area; discharge must be integrated over time; storage must refer to the same endpoints. A balance built from incompatible quantities can produce a residual that is entirely bookkeeping rather than hydrology.
CHAPTER 2 OF 12
2. Separate a residual from a directly observed flux
A residual is the mathematical difference among observed or estimated terms. A physical flux is water crossing a boundary. The two can be related only through assumptions about omitted components and measurement error.
If every listed term were exact and the only omitted pathway were groundwater outflow G, the defined positive residual would equal G. Real budgets rarely satisfy those conditions completely. The residual usually combines unmeasured pathways with errors and approximations in the listed terms.
A positive discrepancy can support investigating leakage without establishing its magnitude. A negative discrepancy can support investigating an omitted input, but it can also arise from errors in evaporation, storage or discharge. Preserve alternative explanations rather than assigning a physical label based solely on sign.
USGS water-budget studies explicitly treat residuals as containing errors and unquantified components. That distinction makes the result more useful: the residual is a diagnostic observation about the whole accounting system. It becomes a particular flux estimate only after the relevant alternatives have been constrained.
CHAPTER 3 OF 12
3. Propagate independent random uncertainty as a stated special case
If component errors are independent and the quoted standard uncertainties are compatible, the variance of a signed sum is the sum of their variances. Negative coefficients do not make an independent variance contribution negative because each coefficient is squared.
For the teaching residual R = P + I − E − O − ΔS, let standard uncertainties be 10, 20, 8, 18 and 5 thousand m³. The independent variance sum is 100 + 400 + 64 + 324 + 25 = 913. Standard uncertainty of R is √913 ≈ 30.2 thousand m³.
Nominal residual 30 thousand m³ is therefore similar to one standard uncertainty under those assumptions. It should not be presented as a precisely identified 30-thousand-m³ leakage observation. An approximate two-sided 95% interval would require additional distribution assumptions and is not created merely by writing a plus–minus sign.
Distinguish standard uncertainty, a confidence interval, a maximum plausible error and a study-specific assigned percentage. They are not interchangeable inputs to the same formula without interpretation. A percentage described as an error allowance may not represent a standard deviation. Retain the meaning and source of every uncertainty estimate.
CHAPTER 4 OF 12
4. Include covariance with the signs of the balance
Errors can be dependent because terms share instruments, models, area estimates, calibration or weather inputs. For a residual expressed as R = aᵀx, variance is aᵀΣa, where a contains signed coefficients and Σ is the covariance matrix. The signs determine whether a covariance increases or decreases residual variance.
For two quantities X and Y, Var(X − Y) = Var(X) + Var(Y) − 2Cov(X,Y). Positive covariance can reduce uncertainty in their difference. For X + Y, the covariance term is positive. Independence is one possible model, not the default truth for every budget.
Take inflow and outflow standard uncertainties 20 and 18 thousand m³ with error correlation 0.8. Their covariance is 0.8 × 20 × 18 = 288 in squared thousand-m³ units. The variance of I − O becomes 400 + 324 − 576 = 148, giving standard uncertainty about 12.2 thousand m³ instead of 26.9 under independence.
This reduction is justified only for a genuinely shared error structure. Nearby gauges do not automatically have positively correlated errors, and correlation between their true flows is not the same as correlation between measurement errors. Use calibration and measurement-chain evidence to support dependence assumptions.
CHAPTER 5 OF 12
5. Separate persistent bias from repeatable scatter
A sensor can give closely repeated readings while sharing a persistent offset. Repetition helps estimate some random variation but does not remove a calibration bias. A budget accumulated over months can therefore be sensitive to a small consistent error.
An unrecognised discharge bias of +0.10 m³/s accumulated over thirty days corresponds to 259,200 m³. That is not a large instantaneous number in some rivers, but it can dominate a small lake-budget residual. The relevant consequence depends on integration time and the size of the inferred missing flux.
Shared multiplicative error can behave differently in a difference or ratio. If inflow and outflow are both overestimated by the same 5%, their difference is also scaled by 5%; the absolute errors partly cancel but do not vanish unless the underlying values or quantity make that appropriate.
Do not assume every term has the same error model. Rainfall spatial representation, evaporation modelling, stage measurement and discharge rating can have different systematic and random components. An uncertainty budget should identify the physical origin of each important error rather than assign one convenient percentage to all terms.
CHAPTER 6 OF 12
6. Use compatible storage endpoints and their reference dependence
Storage change is S₂ − S₁, often derived from water levels and a stage–storage relation. Its uncertainty depends on both endpoint measurements and the geometry model. The errors in S₁ and S₂ may be correlated because they share the same datum or capacity curve.
A constant additive storage offset cancels exactly in S₂ − S₁. A changed datum between endpoints does not. A multiplicative capacity error generally scales the change rather than cancelling. Identify which kind of shared error is being discussed before claiming that common errors disappear.
For a fixed-area teaching basin of 100,000 m², a 0.02 m level difference corresponds to 2,000 m³. If the endpoint datum shifts by 0.03 m between visits, apparent storage change includes 3,000 m³ of reference error. That can exceed the true change being studied.
Real area changes with water level, so stage–storage geometry may need integration. A changing sediment bed can also alter the relation over longer periods. Keep the geometry version and reference checks with the storage series. A final ΔS column without its measurement chain makes later uncertainty review much harder.
CHAPTER 7 OF 12
7. Align time windows before diagnosing a physical discrepancy
A daily inflow total, a midnight-to-midnight outflow total and a storage change measured at noon do not necessarily describe the same interval. During rapid filling or release, that mismatch can create a substantial residual even when every individual observation is valid.
Suppose a teaching basin gains 10,000 m³ during a six-hour morning pulse. If inflow accounting includes the pulse while the storage endpoint is taken before it, the balance can show an apparent missing 10,000 m³. The discrepancy is a timing mismatch, not evidence that the water vanished.
Check time zones, daylight conventions where relevant, logger clock drift, aggregation boundaries and averaging windows. A timestamp label may refer to the start, end or centre of an interval. Converting all labels to one displayed time zone is not enough if those interval definitions remain different.
Temporal aggregation can hide compensating residuals. A positive residual during one event and a negative residual later may cancel in an annual total. Annual closure therefore does not establish that event timing or pathways were modelled correctly. Match the diagnostic scale to the decision instead of relying only on the longest-period balance.
CHAPTER 8 OF 12
8. Test spatial support and omitted transfers explicitly
A rainfall gauge measures at its location, while a lake or catchment budget needs an area-representative input. An evaporation estimate may represent a model grid or a particular exposure. A seepage measurement may cover only a small bed patch. Spatial support changes at every scaling step.
For a fictional lake area 0.5 km², a 10 mm rainfall error corresponds to 5,000 m³. If the same gauge is used to estimate catchment inflow in a model, rainfall error can affect more than one budget term, creating dependence rather than two independent errors.
Human transfers also matter. Pumping, diversions, return flows and operational releases can cross the control boundary. Counting a transfer as both an internal movement and an external input can create artificial closure or imbalance. Draw each pathway before assigning it to a column.
Use alternative boundary and support scenarios to locate consequential assumptions. If a residual changes sign when a plausible ungauged tributary estimate is included, the missing pathway is not uniquely identified. The most useful next measurement may concern the unobserved boundary rather than finer precision at the best-monitored gauge.
CHAPTER 9 OF 12
9. Recognise identifiability: one residual can fit several explanations
Identifiability asks whether the available observations distinguish the proposed explanation from alternatives. One residual equation usually cannot uniquely estimate several unknown fluxes and biases. More decimal places do not add independent equations.
Let R = G + bP − bE in a simplified interpretation, where G is an omitted outward flux, bP an input overestimate and bE an evaporation overestimate under the chosen signs. A residual of 30 can arise from G = 30 with no biases, from bP = 30 with no G, or from many combinations. The same observed balance supports several stories.
Independent information can reduce ambiguity: direct seepage observations, additional gauges, evaporation evidence, tracer constraints or verified transfer records. Each has its own scale and uncertainty. A local observation should not be promoted into a whole-basin constraint without a representation model.
A good conclusion can be bounded rather than unique. The budget may rule out a very large unmeasured transfer while remaining unable to distinguish small transfers from measurement error. Failure to identify a small flux is not proof that it is zero. State the detection or inference limit in the quantity relevant to the question.
CHAPTER 10 OF 12
10. Do not confuse forced closure with independent validation
A model can be made to close by adjusting an uncertain term or assigning the residual to one pathway. This may be a transparent reconciliation procedure, but it is not independent evidence that the adjusted pathway is physically correct.
Suppose measured terms give R = +30 thousand m³. Increasing evaporation by 30 makes the equation close. So would adding groundwater outflow of 30 or reducing inflow by 30. The zero residual after adjustment does not choose among those explanations.
Constrained reconciliation can distribute adjustments using an uncertainty model and conservation. Its output depends on the assumed error structure and any fixed terms. If a poorly known term is declared exact, the procedure must place the discrepancy elsewhere, potentially creating a misleading appearance of certainty.
Retain original and reconciled values separately, explain the constraints and test sensitivity to plausible uncertainty choices. Validate physical interpretations with evidence not used to enforce closure. A balanced account is necessary for conservation, but a balance achieved by construction cannot validate itself.
CHAPTER 11 OF 12
11. Practise covariance, timing and a residual-flux hypothesis
Return to R = 30 thousand m³ with independent standard uncertainty about 30.2. Now assume only the inflow and outflow errors have correlation 0.8 with covariance 288. Replacing their independent contribution by the correlated one reduces total variance from 913 to 337, giving standard uncertainty about 18.4 thousand m³.
The nominal residual is unchanged. What changed is the uncertainty model. A smaller uncertainty is not automatically more credible; the assumed covariance needs evidence. If the correlation were −0.8 instead, variance would increase to 1,489 and standard uncertainty to about 38.6.
Next discover a timing mismatch accounting for +12 thousand m³ of apparent residual. Correcting that justified bookkeeping issue reduces the nominal residual to 18. It does not remove uncertainty in the remaining terms or prove that the remaining 18 is groundwater outflow.
Finally obtain an independent whole-system constraint suggesting an outward groundwater flux between 5 and 15 thousand m³ under a clearly defined scenario. Compare it with the budget rather than forcing agreement. Remaining differences may reflect other errors or omissions. Keep the independent constraint independent so disagreement remains informative.
CHAPTER 12 OF 12
12. Report the discrepancy, uncertainty and next discriminating observation
A reproducible budget provides the boundary, interval, units, sign convention, component values, methods, uncertainty meanings, covariance assumptions and storage reference. Preserve raw observations, derived totals and any reconciled alternatives.
Report the residual before attaching a process name. Then state which omitted pathways are plausible, what the uncertainty allows and which independent observations would distinguish them. A useful statement may be that the budget does not resolve a small exchange rather than that the exchange is absent.
Prioritise the next measurement by consequence. If a shared stage-reference shift controls the residual, repeating the rainfall measurement may do little. If ungauged inflow dominates, refining outflow weighing precision is not the best repair. The uncertainty budget guides a specific next action.
The practical standard is to honour conservation without pretending that every imbalance has one obvious cause. Define the accounting, propagate uncertainty, test dependence and preserve alternatives. Water-budget closure becomes evidence-led reasoning when the reader can see not only the residual but also why a proposed physical explanation is or is not identifiable.
Independent worked case and evidence checks
A fictional monthly lake budget uses thousand cubic metres throughout. Precipitation is 120, stream inflow 800, evaporation 100, stream outflow 760 and storage increase 35. With R = P + I − E − O − ΔS, the residual is 25. A first report calls this 25 thousand m³ of seepage loss. The arithmetic is correct; the physical label has not yet been justified.
Assigned standard uncertainties for this exercise are 12, 40, 15, 38 and 8. Under independence, residual variance is 144 + 1,600 + 225 + 1,444 + 64 = 3,477, giving standard uncertainty about 59.0 thousand m³. The residual is much smaller than that scale. It does not uniquely establish an outward groundwater flux of 25.
Suppose inflow and outflow errors share a calibration component with correlation 0.75. Covariance is 0.75 × 40 × 38 = 1,140. Because the two terms enter with opposite signs, variance becomes 3,477 − 2,280 = 1,197 and standard uncertainty about 34.6. The discrepancy remains unresolved, though the uncertainty is smaller under this justified dependence scenario.
Now suppose precipitation and evaporation errors also have positive covariance 90 in the same squared units. Their opposite signs reduce variance by another 180, giving 1,017 and standard uncertainty about 31.9. This is a mathematical scenario, not a claim that rainfall and evaporation measurement errors always have that dependence. Actual error covariance must be supported by the measurement system.
A timing review finds that one outflow total ended six hours earlier than the storage endpoint. Including the omitted outflow volume adds 18 to O and reduces R from 25 to 7. This correction addresses a known interval mismatch. It is not a calibration adjustment chosen to make the budget look better.
A separate operational record identifies an unlisted return flow of 10 into the lake. Including it as an input raises residual from 7 to 17. The direction of change follows the signed boundary equation. If the same return had already been included inside a modelled inflow estimate, adding it again would double-count it. Boundary definitions must be checked before inserting another column.
A direct seepage study then provides local positive and negative fluxes over a limited shore zone. It supports investigating groundwater exchange but does not measure the whole lake’s net outward flux. Scaling it to the whole bed requires an area representation and temporal model. Calling it independent does not remove that support mismatch.
For an identifiability exercise, express the remaining residual 17 as G + bI − bE, with all values in thousand m³. One explanation is G = 17 and zero biases. Another is G = 5, bI = 12 and bE = 0. A third is G = 25, bI = 0 and bE = 8. All satisfy the same residual equation. Additional observations, not algebra alone, are needed to distinguish them.
For a long-term bias exercise, hold an outflow measurement bias at +0.02 m³/s for thirty days. The accumulated error is 51,840 m³, or 51.84 in the budget units. This exceeds the original residual 25 despite the small instantaneous bias. A dense time series can therefore be highly repeatable and still dominate a monthly closure calculation through shared systematic error.
Finally reconcile the budget by assigning all remaining residual to evaporation. The adjusted equation closes exactly, but the interpretation remains conditional on that assignment. A useful report retains the unadjusted residual, the adjustment, the uncertainty model and the independent evidence that could validate or reject it. Closure is an accounting constraint; physical attribution is a separate inferential achievement.
A further exercise uses a shared lake-area estimate. Precipitation and evaporation depths are 0.10 and 0.08 m over an assumed area 1 million m². Their volumes are 100,000 and 80,000 m³, giving net atmospheric input 20,000 m³. If area is actually 10% larger, both volumes rise by 10% and their difference becomes 22,000 m³. The shared area error partly cancels in absolute terms but still scales the net result. Treating those two volume errors as independent misrepresents the measurement chain.
Next consider two storage endpoints each uncertain because of one shared additive volume reference. The common offset cancels from ΔS, but a stage-dependent geometry error need not cancel equally at the two levels. A capacity curve can be too steep even while its arbitrary zero is harmless. When prioritising a new survey, determine whether uncertainty concerns the reference offset, the curve shape or a changed bed. These are different repairs.
A covariance model must also be mathematically coherent. Pairwise correlation values cannot always be chosen independently for three or more errors. The full covariance matrix must produce non-negative variance for every possible signed combination. If a proposed model yields a negative residual variance, the correct response is to repair the dependence assumptions, not to report zero uncertainty by truncating the impossible result. A physically motivated shared-error construction can be easier to audit than a collection of guessed correlations.
Finally distinguish an uncertainty comparison from a proof of absence. A residual smaller than its uncertainty scale means the budget may not resolve a proposed flux under that model. It does not establish that all omitted pathways are zero, nor does it prove that the uncertainty model is correct. Report the range of physical conclusions supported by the evidence and identify which independent observation would change that range. This preserves the diagnostic value of an imperfect budget without turning uncertainty into either certainty or dismissal.
For a final independent check, hide the residual column and ask a second reader to reconstruct it from the signed terms, then hide the process labels and ask which physical explanations remain possible. The first task checks arithmetic and bookkeeping. The second checks identifiability. A report is stronger when it passes both tests, because a correctly calculated difference can still carry an unjustified physical name.
Worked comparison
| Term | Value in thousand m³ | Coefficient in R | Teaching standard uncertainty |
|---|---|---|---|
| Precipitation | 120 | +1 | 12 |
| Inflow | 800 | +1 | 40 |
| Evaporation | 100 | −1 | 15 |
| Outflow | 760 | −1 | 38 |
| Storage increase | 35 | −1 | 8 |
| Residual | 25 | — | 59.0 under independence |
Questions and practical checks
Is a positive residual automatically leakage?
No. It can contain omitted outward flows, input overestimation, output underestimation, storage errors and timing or boundary problems.
Why does covariance sometimes reduce uncertainty?
Shared errors in terms with opposite balance signs can partly cancel. The effect reverses for other sign structures or negative covariance.
Can repeated measurements eliminate bias?
Not automatically. Repetition helps with some random variation but does not remove a shared calibration or reference error.
Does exact closure validate a model?
Not if closure was imposed by adjusting terms. Independent observations are needed to validate the physical interpretation.
Does an unresolved small residual prove no groundwater exchange?
No. Exchange can exist below the budget’s inference limit or be masked by other components. Report what can and cannot be distinguished.
Continue with the connected subjects
Build the accounting before interpreting the residual
Water Balance Modelling: boundaries, stores, time steps and basic calculations
Gage datum shifts: shared references can change an inferred storage or flow history
Floodplain storage: capacity, occupied volume and boundary flows are distinct
Constrain uncertain components independently
Seepage meters: local flux is not automatically a whole-basin exchange estimate
Eddy-covariance evapotranspiration: footprint, energy closure and gaps
Index-velocity discharge: timing, calibration and accumulated signed volumes
Compare the related measurement and accounting problem
Sources and scope
USGS: uncertainty in lake water-budget components and interpretation of the residual
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 29 October 2026.