Session 14 · Special topic: river routing

Routing I: how a flood moves down a river

Session 8 delayed each day's runoff with one curve, the unit hydrograph, and for the Leaf River that was enough. This special topic asks what that curve stands for. Follow a flood down a real river and it turns out to be a wave that travels and flattens. Add up the travel times of a whole river network and the unit hydrograph appears by itself. That tells you exactly when one curve is enough, and when it is not. The worked case of all three routing sessions is the Rhine above Düsseldorf.

Time about 40 minutesNeeds Session 8; Session 10 for the scores
Watch this session first: a 7-minute narrated video with English subtitles. The page below goes further and lets you try each step yourself.

You will be able to

Where Session 8 stopped

The unit hydrograph of Session 8 is one list of weights for the whole Leaf River, 1,924 km²: 17% of a day's runoff passes the gauge the same day, 35% the next day, and so on. It never asks where in the catchment the runoff was made. It cannot: dCREST gives the catchment one rainfall number per day, so there is no "where".

Now take a basin times larger: the Rhine at Düsseldorf, 147,680 km², from the Alps to the lowlands. The river network used here has 2,242 reaches, each about 11 km long, and the longest path to the gauge measures 1,026 km. Rain in the Alps enters the river a thousand kilometres from Düsseldorf; rain near Cologne enters it about forty kilometres upstream. The two cannot share one curve. To see what replaces the curve, we have to follow the water down the river.

Part of the basinkm²share
The Rhine above Düsseldorf as a river network. Every line is one reach; thicker lines carry more water. The colours mark five parts of the basin used later in this session: four tributary areas, each closed by a gauge, and the rest along the Middle and Lower Rhine. Areas in the table are summed from the reaches of the network, which comes within 0.2% of the gauge's official catchment area.
The Rhine set-up of Sessions 14–16

The numbers come from an experiment with a global CREST set-up. Runoff is made on 1,996 cells of 0.1°, by a water balance that was trained on other river systems, so the Rhine is out of sample. The river network is MERIT-Hydro at 5 arc-minutes. Each reach gets its wave speed from the river's width, its bed slope and Manning's friction formula at mean flow. Nothing in the routing is fitted to the Rhine's gauges. This session uses only the network and its travel times; Session 16 sends real floods through it.

A wave, not the water

How long does a flood take from Basel to Düsseldorf? The natural guess is distance divided by the speed of the water. That guess is too slow, because what travels down a river in flood is not a body of water. It is a wave.

Put extra water into the upper end of a channel. The river is now deeper there, and deeper water runs faster, because a smaller share of it rubs against the bed. The fast, deep water runs into the slower, shallower water ahead of it and raises it. That water, now deeper, speeds up and raises the water ahead of it. The high-water mark is handed on downstream faster than any single drop travels, as a wave runs around a stadium faster than any spectator could. The difference is that in a river the water moves too, only more slowly than the wave.

Try itThe crest overtakes the water

A sketch of 300 km of river seen from the side, heights exaggerated. The dots are floats: they drift with the water, at the speed Manning's formula gives for the depth they are in. The red float starts under the crest. The crest moves at five-thirds of the water's speed at normal depth and leaves every float behind.

How much faster is the wave? Its speed is called the celerity, \(c\). It answers this question: if the river holds a little more water at some place, how much more flow does that make? For a wide river, Manning's formula gives the water speed \(v\) from the depth \(h\), the bed slope \(S\) and a roughness number \(n\):

\[ v = \frac{1}{n}\, h^{2/3}\, S^{1/2}, \qquad q = v\,h = \frac{1}{n}\, h^{5/3}\, S^{1/2}, \qquad c = \frac{dq}{dh} = \frac{5}{3}\, v \]

\(q\): flow per metre of river width (m²/s). \(n\) is about 0.035 for a natural river bed. Because flow grows faster than depth (power 5/3), a small rise in depth adds more flow than the water's own speed would suggest. This is the kinematic wave celerity (Lighthill and Whitham, 1955; Mizukami et al., 2016).

So a flood crest travels at five-thirds of the water's speed. A river 4 m deep on a slope of 0.0003 flows at about 1.2 m/s, and its flood waves move at about 2.1 m/s. A handy conversion: 1 m/s is 86.4 km per day. Models of large rivers work with wave speeds of 1 to 3 m/s, which is roughly 90 to 260 km per day (Lohmann et al., 1996; Mizukami et al., 2016). In the Rhine set-up the median celerity is 2.1 m/s on the main stem, km per day, and 1.1 m/s over all reaches, because small streams are shallow and slow.

A real flood shows the scale. In June 2013 the crest of a Rhine flood passed Basel on 1 June, Maxau near Karlsruhe on 2 June, Cologne on 4 June, and Lobith at the Dutch border on the morning of 7 June: about a week for the length of the river (Hochwasserzentralen am Rhein, 2013). The dates on the lower river also depend on when the Neckar, the Main and the Moselle deliver their own peaks, which is where this session is heading.

Delayed and flattened

Two things happen to a flood wave on its way through a reach of river.

It is delayed. The delay is the length of the reach divided by the celerity, \(L / c\). Five hundred kilometres at 2 m/s take 250,000 seconds, a little under three days.

It flattens. A reach is also a store. While the river rises, part of the inflow stays in the reach and fills it, so less leaves at the lower end than enters at the upper end. When the inflow falls again the reach drains, and more leaves than enters. The peak at the lower end is therefore lower and the wave longer. No water is lost: the area under the curve stays the same. How strongly a river flattens its waves is measured by one number, the diffusivity \(D\) (m²/s). It is large for big rivers on gentle slopes, \(D = q / (2S)\); routing models of large basins use values from a few hundred to a few thousand m²/s (Lohmann et al., 1996).

Try itOne reach: delay and flattening

A one-day flood wave enters the upper end of a reach; the second curve is the flow at the lower end. Lengthen the reach: the wave arrives later and flatter. Raise the celerity: it arrives earlier and has less time to flatten. Raise the diffusivity: the centre of the wave arrives at the same time, but the peak is lower and wider.

The exact curve

With a constant celerity and diffusivity, the flow \(Q\) along a reach obeys the diffusion-wave equation: the wave moves downstream at speed \(c\) and spreads at rate \(D\).

\[ \frac{\partial Q}{\partial t} + c\,\frac{\partial Q}{\partial x} = D\,\frac{\partial^2 Q}{\partial x^2} \]

Release a short pulse at the upper end. The share of it that passes a point \(L\) downstream per unit time, at time \(t\) after the release, is (Hayami, 1951)

\[ f(t) = \frac{L}{\sqrt{4\pi D\,t^{3}}}\; \exp\!\left[-\frac{(L - c\,t)^2}{4 D\,t}\right], \qquad \text{mean} = \frac{L}{c}, \qquad \text{variance} = \frac{2 D L}{c^{3}} \]

Statisticians know this curve as the inverse-Gaussian distribution; here it is a distribution of travel times. The outflow of the reach is the inflow convolved with \(f\), the same operation as in Session 8. Note what the two formulas on the right say: the delay grows with \(L\), the spread only with \(\sqrt{L}\). On a long river the delay dominates.

From one reach to a network

A river network is thousands of reaches joined end to end. Each has a length and a celerity, and so a travel time of its own; in the Rhine set-up the median is 2.6 hours. Add the travel times of all reaches on the way down and every place in the basin gets one number, \(\tau\): its travel time to the outlet. For the Rhine at Düsseldorf it averages 4.4 days over the basin's area and reaches 8.4 days in the farthest headwaters of the Aare.

Now let one millimetre of runoff appear everywhere at the same moment, and leave flattening aside. After \(\tau\) days the gauge sees the water of exactly those places whose travel time is \(\tau\). So count, for each travel time, the area that has it. That count is the time–area diagram, and divided by the total area it says which share of the runoff passes the gauge at which time. It is a unit hydrograph, built from a map.

DataTravel times on the Rhine, and the unit hydrograph they make
Share of the basin area by travel time (steps of a quarter day)
The same per day, and the response of the full router

Move the slider: reaches whose water is passing Düsseldorf light up, reaches whose water has already passed turn grey, and the histogram fills. Lower chart: bars, the share arriving on each day from travel times alone, with the runoff made evenly through day 0; line, the unit response of the full router, which also flattens the wave in every reach.

Two things are worth reading off. The histogram is not a smooth hump: its shape is the shape of the basin, and each bulge is a large tributary area coming in. And in the lower chart the bars and the line nearly coincide. In a network, most of the flattening at the outlet comes from the many different path lengths; the flattening inside each reach adds only a little.

So when is a unit hydrograph enough?

dCREST v1.0 gives a catchment one rainfall number per day and one parameter set. Every part of the catchment therefore makes the same runoff \(r_t\). Send that runoff down the network: the flow at the gauge is the sum over all places \(i\), each with its area \(a_i\) and each delayed by its own travel time \(\tau_i\). Because \(r\) is the same everywhere, the places can be grouped by travel time, and the sum collapses:

\[ Q_t \;=\; \sum_{i} a_i\, r_{t-\tau_i} \;=\; \sum_{d} w_d\, r_{t-d}, \qquad w_d = \text{share of the area with travel time } d \]

The right-hand side is the convolution of Session 8. The step needs the wave speeds not to depend on the flow itself, which is the case for every method in this session (a linear router).

So with one runoff number per catchment, routing through the river network is a convolution with one fixed curve. A detailed network can do nothing that a unit hydrograph cannot; the gamma curve of dCREST is a two-parameter stand-in for the network's histogram.

This was tested on nine CAMELS catchments of 461 to 25,524 km². The released water balance was frozen, and its runoff was routed along every flow path of the 1-km flow-direction grid that the EF5 flood model uses, as a diffusion wave. Three routing numbers were fitted per catchment: a celerity, a diffusivity and a hillslope velocity. The table lists, for each catchment, its area, the longest channel path to the gauge and the number of grid cells, and then the test NSE (Session 10) for three ways of routing the same runoff: none at all, the released gamma unit hydrograph, and the network.

USGS gaugearea, km²channel, kmgrid cellsno routingunit hydrographnetwork
07315200461344590.3760.4390.439
06440200665466920.0310.2380.238
11482500717777730.7610.7900.789
015520001,126751,2030.7410.7410.730
016675001,212711,2450.6750.6900.684
011440001,786721,9610.7410.7450.745
133370003,0501383,5480.6720.6780.678
068465003,4282194,527−0.056−0.056−0.056
0645200025,52451029,1420.3530.4090.420
median0.6720.6780.678

The median test NSE is 0.678 with the unit hydrograph and 0.678 with the network; the largest difference in any catchment is 0.011. Up to 29,142 grid cells of river network bought nothing, exactly as the equation says. In most of the nine, the curve fitted through the network and the gamma curve come out as nearly the same curve.

The column without any routing teaches a second lesson. At a daily time step, several of these catchments hardly notice whether their runoff is routed at all, because the water reaches the gauge within the day; the median barely moves (0.672). Where the response takes longer than a day, routing matters: gauge 06440200 goes from 0.03 to 0.24, and the largest of the nine, gauge 06452000, from 0.35 to 0.41.

When it is not

The collapse of the sum needed one thing: the same runoff everywhere. On the Rhine that fails with every storm, because a rain band covers one part of the basin and misses another. The table gives the unit response at Düsseldorf to one millimetre of runoff made in one day in each of the five parts of the basin, from the full router.

Runoff made inmean delay (days)arrives on days 0–2largest share on day

Water from the Middle and Lower Rhine area needs days on average, and % of it has passed within three days. Water from the Alpine Rhine and the Aare needs days, and % of it arrives in the first three. A single unit hydrograph for the whole basin averages over all of this (mean delay days) and cannot tell a storm on the lower basin from a storm on the Alps.

Try itWhere the rain falls
The same volume of runoff:
Extra flow at Düsseldorf (m³/s) against days since the start of the runoff day

Set the runoff made in one day in each part of the basin. Coloured lines: what each part contributes at Düsseldorf. Dark line: their sum. The routing is linear, so the responses simply add. Dashed: what one unit hydrograph for the whole basin gives for the same total volume. The three buttons put the same volume of water in different places.

Press the three buttons in turn. The dashed curve does not move: a single unit hydrograph sees only the total. The dark curve does. Runoff on the lower basin raises the river at Düsseldorf within a day and keeps it high for four days. The same volume on the Alps does nothing for three days and then arrives as a later, sharper peak. (The router has channels only. In reality Lake Constance holds back part of the Alpine water; Session 16 comes back to that.)

Three more situations break the single curve, and all three appear in Session 16:

In Python

Build a unit hydrograph for the Rhine from nothing but the travel time and the area of each reach, compare it with the response of the full router, and route three days of runoff with it.

Things to try: replace "whole basin" by "Moselle" or "Alpine Rhine and Aare"; halve every travel time (R.tau_d / 2) and see the peak rise; or route with R.uh, the gamma unit hydrograph that was trained together with the water balance, and notice how much slower it is. Session 16 explains why.

Check your understanding

The water of a wide river flows at 1.5 m/s. How fast does a flood crest travel, and how far does it get in a day?

\(c = \tfrac53 \times 1.5 = 2.5\) m/s. One m/s is 86.4 km per day, so the crest covers about 216 km in a day, while the water itself covers about 130 km.

A reach is 200 km long and its celerity is 2 m/s. How long is the delay? What changes if the diffusivity doubles?

\(L/c\) = 100,000 s, about 28 hours. Doubling \(D\) leaves that mean delay unchanged; the variance \(2DL/c^3\) doubles, so the wave arrives lower and wider, by a factor \(\sqrt2\) in spread.

Why is the time–area diagram of a basin a unit hydrograph?

If runoff appears everywhere at once, the water passing the gauge after \(\tau\) days comes from the places whose travel time is \(\tau\). The area with each travel time, as a share of the whole, is therefore the share of the runoff arriving at each time: the response to one unit of runoff.

dCREST v1.0 uses one rainfall number per catchment. Would routing its runoff through a detailed river network improve the flow at the outlet?

No. With the same runoff everywhere, linear routing through any network equals one convolution with the network's travel-time histogram, and the gamma unit hydrograph can take nearly that shape. In nine CAMELS catchments the test NSE changed by at most 0.011.

The same volume of runoff is made once on the Middle and Lower Rhine and once on the Alpine Rhine and Aare. How do the two floods at Düsseldorf differ, and what does a single unit hydrograph say?

From the lower basin the water starts arriving within a day (mean delay days); from the Alps almost nothing arrives in the first three days (mean delay days). A single unit hydrograph gives the same curve for both, because it knows only the total volume.

What to remember

Next: Session 15 opens the toolbox. All routing methods are shortcuts through the same two equations, and they differ in what they neglect, what they need and what they cost to compute.

References

Hayami, S.: On the propagation of flood waves, Bulletin of the Disaster Prevention Research Institute, Kyoto University, 1, 1951, hdl.handle.net/2433/123641.

Hochwasserzentralen am Rhein: Das Rheinhochwasser Juni 2013, report of 5 November 2013, hochwasser.rlp.de.

Lighthill, M. J. and Whitham, G. B.: On kinematic waves I. Flood movement in long rivers, Proceedings of the Royal Society of London A, 229, 281–316, doi:10.1098/rspa.1955.0088, 1955.

Lohmann, D., Nolte-Holube, R., and Raschke, E.: A large-scale horizontal routing model to be coupled to land surface parametrization schemes, Tellus A, 48, 708–721, doi:10.3402/tellusa.v48i5.12200, 1996.

Mizukami, N., Clark, M. P., Sampson, K., Nijssen, B., Mao, Y., McMillan, H., Viger, R. J., Markstrom, S. L., Hay, L. E., Woods, R., Arnold, J. R., and Brekke, L. D.: mizuRoute version 1: a river network routing tool for a continental domain water resources applications, Geoscientific Model Development, 9, 2223–2238, doi:10.5194/gmd-9-2223-2016, 2016.