The Gold Standard of Hydrology | Recession Constants, Time Steps and the Storage Behind a Falling Flow

Full chapter index

After rain ends, a stream may continue flowing because water drains from catchment storage. The discharge can decline over time as the available storage and hydraulic gradients change. A recession model approximates part of that decline. It does not observe every subsurface pathway directly.

Begin by defining the series. Discharge Q is a rate, commonly measured in m³/s. Time may be recorded in hours or days. A hydrograph gives a sequence of discharge values or interval averages, depending on the data product. That distinction matters when converting the curve into volume.

A declining record is not automatically a period of unforced groundwater drainage. Rain may continue elsewhere in the catchment, a reservoir may reduce its release, pumping may alter groundwater, or evapotranspiration may vary through the day. Identify whether those processes are negligible, measured or explicitly represented before treating the segment as a simple recession.

The same visible curve may also contain more than one contribution. Rapid drainage from shallow storage can dominate early, while slower groundwater drainage becomes more important later. A single fitted constant may summarise the chosen interval without becoming a fixed catchment property at every flow level.

The useful first question is therefore specific: over this selected interval, under these conditions, does a proportional decline provide an adequate description? That is narrower and more testable than asking whether the entire catchment has one true recession constant. It also makes clear why the time window and data screening belong in the final explanation.

Contents · Next chapter

Let Q0 = 12 m³/s at time zero. Suppose each day retains a fraction k = 0.8 of the previous day’s discharge. After one day, Q1 = 12 × 0.8 = 9.6 m³/s. After two days, Q2 = 12 × 0.8² = 7.68 m³/s. The general daily expression is Qn = Q0kⁿ.

The absolute daily drop becomes smaller. The first drop is 2.4 m³/s, the next is 1.92 m³/s, and the next is 1.536 m³/s. Equal proportional declines do not produce equal absolute declines. A straight downward line on ordinary axes represents a different model.

In this retention-factor convention, a k nearer one represents a slower proportional recession over the stated interval. Compare k = 0.9 with k = 0.8 over one day: the former retains 90%, while the latter retains 80%. This interpretation cannot be transferred blindly to a decay-rate parameter λ, for which a larger positive value means faster decline.

The parameter name alone is insufficient because different publications and software may use different conventions. HEC-HMS, for example, defines its recession k using a one-day ratio. Check the definition, reference interval and equation before comparing a reported value with another study.

A dimensional check helps: k is dimensionless, but its meaning is tied to an interval. The exponent n counts those intervals. If time t is in days, Q(t) = Q0(0.8)^t uses the daily factor. Inserting a number of hours directly into the same exponent would silently make each hour behave like an entire day.

Contents · Previous chapter · Next chapter

The same exponential curve can be written Q(t) = Q0 exp(−λt), with t measured in days. Equating the one-day retention factor gives exp(−λ) = 0.8, so λ = −ln(0.8) ≈ 0.22314355 day⁻¹. The reciprocal time scale is τ = 1/λ ≈ 4.48142 days.

The time scale τ is the time needed for the flow to decline to 1/e of its starting value in this model. It is not the half-life. The half-life is t½ = ln(2)/λ ≈ 3.10628 days. At that time the model predicts Q = 6 m³/s, half the starting discharge of 12.

At day four, Q = 12 × 0.8⁴ = 4.9152 m³/s. Substituting t = 4 into the continuous expression gives the same value. These are two parameterisations of the same curve under the stated conversion, not two independent models with different physical predictions.

Natural logarithms are required in λ = −ln(k)/Δt. If base-10 logarithms are used to plot the data, the slope has a different numerical conversion. The decay rate is the negative natural-log slope per unit time. Mixing the two logarithm bases changes the inferred time scale.

The continuous form is convenient for interpolation and integration. It also makes the units explicit: λt must be dimensionless. If t is converted from days to seconds, λ must become λ/86,400 s⁻¹. Preserving the product λt preserves the curve. This is a useful independent check whenever an implementation changes its computational time step.

Predict a threshold-crossing time

Under the exact teaching model, when does discharge fall from 12 to 3 m³/s? Solve 3 = 12 exp(−λt), giving t = ln(12/3)/λ. With λ = 0.22314355 day⁻¹, t ≈ 6.21257 days. This is twice the half-life because reducing to one quarter requires two successive halvings.

This calculation is a conditional crossing time, not a forecast independent of future inputs. A new rainfall pulse, a release change or a failure of the exponential approximation can shift the crossing. State the threshold and the conditions under which the time is predicted.

Sensitivity is easy to see by changing the daily factor. With k = 0.75, the time to one quarter is ln(4)/[−ln(0.75)] ≈ 4.81884 days. With k = 0.85 it is about 8.53005 days. A seemingly modest parameter difference produces several days of change in the extrapolated crossing time.

That spread should not be called a confidence interval unless it came from an appropriate uncertainty analysis. Here it is a scenario comparison showing how the chosen parameter affects the result. Keeping scenario ranges distinct from statistically estimated intervals makes the prediction easier to assess.

Contents · Previous chapter · Next chapter

For two positive observations Q1 and Q2 separated by Δt days, the implied decay rate is λ = ln(Q1/Q2)/Δt. The equivalent daily retention factor is kday = exp(−λ), or directly (Q2/Q1)^(1/Δt). This formula makes the interval correction explicit.

Suppose flow declines from 12 to 7.68 m³/s over two days. The raw ratio is 0.64. That is a two-day retention factor, not a daily factor. The daily factor is √0.64 = 0.8. Reporting 0.64 as the daily k would exaggerate the rate of recession.

With several observations, plot ln Q against time. An exact exponential gives a straight line, with intercept ln Q0 and slope −λ. For noisy records, a fitted line can estimate the average log-space slope over the selected period. Report the period, fitting method and residual pattern rather than only the resulting k.

Fitting in log space and fitting Q directly minimise different discrepancies. Log-space fitting emphasises proportional errors more evenly across flow magnitudes; ordinary least squares on Q gives greater numerical influence to large absolute errors. Neither procedure automatically solves observation error, serial dependence or a poor physical model.

Check consecutive ratios as a diagnostic. In the exact daily example, 9.6/12, 7.68/9.6 and 6.144/7.68 all equal 0.8. A systematic increase or decrease in those ratios suggests changing behaviour. Random measurement noise may also disturb them, so the pattern should be considered alongside data quality and process evidence, not treated as an automatic mechanism detector.

Check sensitivity to one noisy endpoint

Suppose the true two-day endpoint in the teaching example is 7.68 m³/s, but a measurement reports 8.0. The endpoint-based daily factor becomes √(8/12) ≈ 0.81650 rather than 0.8. The fitted recession appears slower, even though the change arose solely in the observation.

If the endpoint is instead reported as 7.4, the daily factor becomes √(7.4/12) ≈ 0.78528. Using a longer record can reduce dependence on one endpoint, but it may introduce a new problem if the record spans changing processes. More timestamps help only when the model and observation treatment remain appropriate.

Compare errors in terms of both absolute and relative magnitude. An uncertainty of 0.3 m³/s is a much larger fraction of a 1 m³/s flow than a 12 m³/s flow. Log-space residuals are sensitive to that relative scale. A constant-variance error model on Q and a constant-variance model on ln Q encode different assumptions about measurement and discrepancy.

Do not infer physical precision from a fitted value such as k = 0.81649658. That many digits mainly reflect arithmetic. Report precision that is consistent with the data and explain how sensitive the inferred time scale is to plausible observation changes.

Contents · Previous chapter · Next chapter

For the daily factor 0.8, the hourly retention factor is khour = 0.8^(1/24) ≈ 0.990745. It is close to one because each interval is much shorter. Applying that factor 24 times gives khour²⁴ = 0.8, recovering the original daily decline.

The six-hour factor is 0.8^(1/4) ≈ 0.945742, and the two-day factor is 0.8² = 0.64. A larger numerical retention factor does not necessarily mean a slower physical recession if the comparison uses different time intervals. Compare daily-equivalent factors or decay rates in a common unit.

A frequent shortcut is to divide the daily 20% decline by 24 and subtract that percentage each hour. That produces 1 − 0.2/24 ≈ 0.991667 as an hourly factor. Compounding it over 24 hours does not give exactly 0.8. The shortcut replaces exponential compounding with a different approximation.

An even larger error is to use k = 0.8 at every hourly step. After 24 steps, the model retains 0.8²⁴, approximately 0.00472 of its original discharge. Starting from 12 m³/s, that leaves about 0.0567 m³/s after one day rather than the intended 9.6 m³/s.

For a spreadsheet or model audit, store the parameter’s reference interval beside its value. Compute the per-step factor as kref^(Δt/Δtref). Changing Δt then changes the computational factor while preserving the physical recession. A successful check reproduces the same discharge at common timestamps under hourly, six-hourly and daily calculations.

Contents · Previous chapter · Next chapter

Discharge is a rate, so a recession volume requires integration over time. With Q(t) = Q0 exp(−λt) and t in days, the volume between day zero and day T is V = 86,400 × Q0[1 − exp(−λT)]/λ. The factor 86,400 converts days into seconds to match m³/s.

Using Q0 = 12 m³/s, kday = 0.8 and T = 4 days gives V = 86,400 × 12 × (1 − 0.8⁴)/0.22314355, approximately 2.743 million m³. This is the volume discharged during those four days under the exact exponential teaching model.

Do not add 12 + 9.6 + 7.68 + 6.144 + 4.9152 and label the result cubic metres. Those values are instantaneous rates at timestamps. A numerical approximation must assign a time width and a rule for representing the curve between observations.

For daily trapezoids, the approximate volume is 86,400 × [(12 + 4.9152)/2 + 9.6 + 7.68 + 6.144] = 2,754,570.24 m³. It is slightly greater than the exact integral because an exponential decline is convex and the straight chords lie above the curve between endpoints.

The difference here is a numerical approximation error under a known curve. In field data, uncertainty about the curve itself can be more important. If the observations are already daily mean discharges, each mean can be multiplied by its interval duration and summed without interpreting it as an instantaneous endpoint. Know which data product is being integrated before choosing the method.

Contents · Previous chapter · Next chapter

A simple linear reservoir model sets Q = aS, where S is active storage volume and a is a positive outflow coefficient. With no inflow and no other losses, conservation gives dS/dt = −Q = −aS. The solution is S(t) = S0 exp(−at), and Q follows the same exponential decline.

Use consistent units. If Q is m³/s and S is m³, a is s⁻¹. For the teaching decay rate in day⁻¹, a = λ/86,400 s⁻¹. The inferred active storage at time zero is S0 = Q0/a = Q0 × 86,400/λ, approximately 4.646 million m³.

This equals the integral of the modelled discharge from zero to infinite time. After day four, the remaining modelled outflow volume is Q4 × 86,400/λ, approximately 1.903 million m³. The four-day outflow plus that remaining amount equals the initial modelled active storage, providing a conservation check.

The word active matters. This is the storage represented by the linear model and its assumed outlet relation. It is not automatically the entire catchment groundwater volume, all water in an aquifer, or a safely extractable supply. Some stored water may not drain to the stream on this time scale or through this outlet.

If evaporation, abstraction or additional inflows operate, dS/dt includes those terms. The observed recession can then differ from the unforced linear-reservoir solution. A fitted exponential may still be a useful empirical approximation, but its integrated tail should not be presented as a measured aquifer reserve without the additional evidence required by that interpretation.

Add an inflow and watch the asymptote change

For a linear store with constant inflow I and outflow Q = aS, conservation becomes dS/dt = I − aS. Expressing the same model in discharge gives dQ/dt = a(I − Q), provided I is a flow in units consistent with Q and a. Its solution is Q(t) = I + (Q0 − I) exp(−at).

Take I = 2 m³/s and a corresponding to the daily factor 0.8. Starting at Q0 = 12 m³/s, the prediction is Q(t) = 2 + 10(0.8)^t. At day one it is 10, at day two 8.4, and at day four 6.096 m³/s. The excess above 2 decays exponentially; the total discharge does not follow the original zero-asymptote equation.

The consecutive ratios of total Q move closer to one as flow approaches the sustained inflow. Fitting a zero-asymptote exponential over progressively later periods can therefore suggest a slower apparent recession even though a is fixed. This is another reason that curvature or changing ratios do not uniquely identify a changing outlet property.

The constant-inflow construction is an explanatory model, not a claim that a real catchment has an unchanging external supply. It demonstrates how an omitted forcing term changes interpretation. Evidence about inflow or sustained contributions is needed before replacing one model with the other.

Contents · Previous chapter · Next chapter

Construct a second fictional catchment response: Q(t) = 8(0.5)^t + 4(0.9)^t m³/s, with t in days. One component declines quickly and the other slowly. Total discharge begins at 12 m³/s, but there is no single daily retention factor that remains constant for the sum.

At days zero through four, total discharges are 12, 7.6, 5.24, 3.916 and 3.1244 m³/s. Consecutive daily ratios are approximately 0.6333, 0.6895, 0.7473 and 0.7979. The apparent recession slows as the fast component becomes less important.

Each component has a constant proportional decline, yet their mixture does not. A straight line fitted through ln Q over the first two days would give a different summary slope from a line fitted over later days. The difference can arise from changing component weights without requiring either component’s own parameter to change.

This does not prove that a curved recession contains exactly two physical reservoirs. Many combinations of processes and measurement effects could produce curvature. The example establishes a possibility and a limitation: one changing apparent constant does not uniquely identify one changing physical property.

Use this changed-context check to evaluate the single-store model. Inspect whether the residuals are systematically curved in log space, whether late-time ratios stabilise, and whether external evidence supports different pathways. Additional complexity is useful when it resolves a repeatable discrepancy and is identifiable from the available data, not simply because a more elaborate curve can be drawn through every point.

Why an integrated tail is especially assumption-sensitive

The infinite tail of the zero-inflow exponential is finite because the curve approaches zero fast enough. In the constant-inflow example, total discharge approaches 2 m³/s, so its integral over infinite time diverges. The excess-flow component above that sustained input still has a finite exponential integral.

Those two mathematical results are not inconsistent. They refer to different systems and different quantities. Integrating total discharge in a system with continuing inflow counts water arriving in the future; it cannot all represent water initially stored at time zero.

When someone interprets an integrated recession as stored water, ask whether the integral counts only drainage from the initial store or also continuing inputs. The balance equation supplies the answer. A fitted curve alone cannot establish that distinction, even if it follows the measured discharge closely over a few days.

Contents · Previous chapter · Next chapter

Keep a transparent calculation sheet

A useful calculation sheet contains columns for timestamp, elapsed time, observed discharge, modelled discharge, residual and any screening note. Put the reference k interval, decay-rate units and time conversion beside the parameters. A reader should be able to reproduce one predicted value without guessing whether the spreadsheet exponent counts hours, days or row numbers.

For unequal time gaps, use the actual elapsed time. A missing day should not make the next observed value behave as though only one model step had passed. In the original example, day zero at 12 and day three at 6.144 are consistent with kday = 0.8. Treating those as adjacent daily records would instead assign the three-day ratio 0.512 to a one-day decline.

Keep screened observations in an audit copy even if they are excluded from fitting. A flagged rainfall pulse or release change helps explain why the simple model was not applied there. Removing the record entirely makes it harder to distinguish a principled segment choice from selection of only the points that fit well.

Finally, calculate one prediction by two equivalent routes. At day two, use both 12 × 0.8² and 12 × exp(−0.22314355 × 2). Then compare a one-day calculation using 24 hourly factors. Agreement checks parameter conversion and implementation; disagreement reveals a unit or exponent error before it becomes a misleading physical conclusion.

Start by checking timestamps, time zone, units and whether values are instantaneous or averaged. Mark rainfall periods, missing records, known release changes and abstraction if those data are available. A smooth-looking line with inconsistent timestamps is not a sound basis for estimating a physical time scale.

Choose recession segments using a rule stated before comparing constants. If one catchment is fitted only during late low flows and another during early event drainage, the estimates answer different questions. Document the selected intervals and test whether reasonable changes in segment boundaries materially alter the conclusion.

Estimate parameters with an explicit error model or fitting criterion, then inspect residuals and held-out segments. A good fit to one selected recession does not establish performance across seasons or flow states. Repeating the analysis across suitable segments can reveal whether a single summary is stable enough for the intended use.

For students, separate four learning tasks: proportional decline, logarithmic conversion, time-step conversion and integration. An error in one task should be repaired there. If the learner can calculate k but labels a rate sum as volume, additional logarithm practice will not address the missing dimensional reasoning.

A useful independent explanation includes the equation, parameter convention, units, selected period, predicted discharge at a new timestamp and a limitation. The learner should also be able to name an observation that would challenge the model, such as a new rainfall pulse or systematic curvature on a log-discharge plot.

Contents · Previous chapter · Next chapter

Contents · Previous chapter · Next chapter

Does k = 0.8 mean 80% of groundwater remains?

It means the modelled discharge retains 80% over the defined interval. Only under additional assumptions, such as the unforced linear storage relation, does the represented active storage share that proportional decline. It is not automatically a fraction of all groundwater in the catchment.

Is a smaller recession constant always faster?

In the retention-factor convention, yes over the same reference interval. In the positive decay-rate convention λ, larger values mean faster decline. Read the equation and units before interpreting the word constant.

Can the exponential curve reach zero?

Mathematically it approaches zero asymptotically while remaining positive at finite time. A real stream may cease flowing because the model’s assumptions and storage-outflow relation no longer apply. A finite observed drying time therefore should not be forced into an exact exponential interpretation.

Why not fit every falling section together?

Different sections may contain different inputs, storage states, regulation or measurement conditions. Combining them can hide a systematic difference. A defensible combined model requires an explanation of why the selected sections should share parameters.

Does a high fitted accuracy prove a groundwater mechanism?

No. A curve fit assesses compatibility with the recorded series under a criterion. Mechanistic attribution needs supporting evidence about inputs, storage, pathways and outlet behaviour. The two-store example shows how different internal constructions can produce similar external patterns over a limited window.

Contents · Previous chapter · Next chapter

The US Army Corps of Engineers’ HEC-HMS Recession Model documentation specifies the daily-ratio convention and explains the exponential representation. Its parameter definition is the reference for the convention used in this guide; the numerical examples here are original.

For research context, see the US Geological Survey’s Basic concepts for the linear model of ground water level recession and New method for calculating a mathematical expression for streamflow recession. Their scope helps distinguish empirical recession descriptions from stronger statements about storage and recharge.

Before accepting a result, ask which k convention is used, what its interval is, whether Q values are instantaneous or averaged, which inputs were screened, how the segment was chosen, and whether the integration used compatible time units. Then ask whether the inferred quantity is discharge, volume, modelled active storage or an independently measured groundwater reserve.

Continue with hydrograph separation to examine what can be inferred from a river hydrograph, and flow-duration curves to see why another useful summary loses chronological information. Return to the Hydrology subject guide for the complete route.

Contents · Previous chapter · Return to the Hydrology guide