Soil I: the infiltration curve
This is the most important session of the course. The soil step decides how much of each storm is stored and how much runs off toward the river, and it does so with one elegant idea: the soil's capacity to hold water is not the same everywhere in the catchment. You will see why two storms of the same size at the Leaf River produced a small rise in one case and the largest flood in 34 years in the other.
You will be able to
- explain saturation-excess runoff and why it depends on how wet the soil already is;
- read the variable infiltration capacity (VIC) curve and say what its parameters wm and b do;
- compute, for a given storm and soil state, how much water infiltrates and how much runs off;
- tell the soil step's "wet" and "dry" days apart, and say what the impervious fraction im does.
Wet days and dry days
The soil step first compares two numbers that arrive from the canopy: the water reaching the ground, \(L\), and the evaporative demand still unmet, \(D - E_c\).
- If \(L > D - E_c\), it is a wet day. The demand is met from the incoming water, and the rest, \(L - (D - E_c)\), is water to be shared between the soil and runoff. This session is about wet days.
- Otherwise it is a dry day. All the incoming water evaporates and the soil (and then the aquifer) must supply the rest of the demand. Session 6 covers dry days.
At the Leaf River with its calibrated parameters, only 14% of days in 1989–1999 are wet in this sense: rain heavy enough to exceed what the canopy and the air take back.
On a wet day, a fraction im of the excess falls on impervious ground (rock, roads, roofs, water surfaces) and becomes overland runoff directly, \(P_i\). The rest, \(P_s\), reaches pervious soil:
For the Leaf River im = 0.00034, so \(P_i\) is negligible and \(P_s\) is essentially all the excess.
Runoff from saturated ground
Imagine the catchment's soil as a sponge. Rain soaks in until the sponge is full; after that, any extra rain cannot enter and runs off over the surface or through the topmost layer toward streams. This is saturation-excess runoff, and it is the mechanism CREST uses. Its key property is that the runoff from a storm depends on how full the sponge already was.
A single sponge would behave like a switch: no runoff at all until the soil is full, then all of it. Real catchments are smoother than that, because the soil is not uniform. Valley bottoms near streams are often already saturated, thin soils on slopes fill quickly, and deep soils on ridges can absorb large storms. Even a small storm produces some runoff from the parts that are already full, and the area that is full grows as the storm goes on.
The variable infiltration capacity curve
CREST takes this idea from the VIC model of Liang et al. (1994). Imagine sorting every point of the catchment by how much water its soil can hold, its point capacity, from smallest to largest. The VIC curve gives the capacity at each fraction of the area:
\(f\): fraction of the catchment area whose capacity is at most \(i\). \(W_m\): the largest point capacity anywhere.
Two parameters shape it. wm is the average capacity over the catchment, in mm. b sets how uneven the capacities are. A small b makes nearly every point hold close to wm: the catchment behaves like one big bucket. A large b means many points with small capacity and a few with very large capacity: runoff starts early, even in small storms.
The water already in the soil fills the curve from below up to a level \(A\), the "water level" on the capacity axis. Points with capacity below \(A\) are full; points with larger capacity hold \(A\) millimetres. The area under the filled region is the soil water \(\mathit{SM}\):
A storm delivering \(P_s\) raises the level from \(A\) to \(A + P_s\). Wherever the curve lies below the new level, the points are full and the extra water runs off; elsewhere it is stored. Integrating over the area gives the water stored, the infiltration \(I\), and the runoff \(R\):
If the new level reaches the top of the curve, \(A + P_s \ge W_m\), the whole soil saturates: it stores only what it was missing, \(I = \mathit{wm} - \mathit{SM}\), and the rest runs off.
Two storms at the Leaf River
The two presets above are real days from the model's run with the Leaf River's parameters (wm = 235.2 mm, b = 0.397, so \(W_m\) = 328.6 mm).
| Quantity | 27 Aug 1992 | 16 Feb 1990 |
|---|---|---|
| rain that day, P | 127.6 | 118.8 |
| water reaching the soil, \(P_s\) | 123.1 | 116.3 |
| soil water before, SM (share of wm) | 93.3 (40%) | 231.8 (99%) |
| infiltration \(I\) | 93.7 | 3.4 (saturated) |
| runoff \(R\) | 29.5 | 112.8 |
| runoff share \(R / P_s\) | 24% | 97% |
| highest observed river flow of the event | 7.0 mm/day (29 Aug) | 53.9 mm/day (17 Feb) |
Both days brought more than 115 mm of rain. In late August 1992 the soil had been dried by a summer of evaporation and was 40% full; it absorbed 94 mm and only about a quarter ran off. In February 1990 the soil was practically full after a wet winter, so almost everything ran off. The next day the Leaf River reached its highest flow in water years 1981–2014. The same rain made a modest rise in one case and a record flood in the other, which is the reason soil moisture is the centre of every rainfall–runoff model.
The model gets the February 1990 flood's timing right but its peak too low: 29.4 mm/day against 53.9 observed on 17 February. The soil is not the bottleneck: it turned 97% of the water reaching it into runoff, and the day's runoff entering the routing, 68.5 mm, was more than the observed peak. Two things lowered it. The routing (Session 8) spreads each day's runoff over several days: the unit hydrograph delivers at most 35% of it on any one day. (Releasing the fast reservoir in a single day, qc = 1, would raise the peak to 37.2 mm/day; removing the loss, lf = 0, to 33.0.) And water is missing: from 14 to 25 February the model's river carried 153 mm against 206 mm observed. From 9 February to 10 March the gauge recorded 288 mm of flow, as much as the 289 mm of rain in the gridded data, which no model that evaporates water can reproduce. Peaks of large floods are where daily lumped models are weakest, for reasons in both the model and the data.
How runoff grows with storm size
The VIC curve makes runoff a nonlinear function of both the storm and the soil state. The chart shows runoff \(R\) against the water entering the soil \(P_s\), for four soil wetness levels, using the current wm and b from the widget above.
precipSoil = (precip - adjPET) * (1.0 - IM) # P_s
precipImperv = precip - adjPET - precipSoil # P_i
interflowExcess_w = torch.clamp(SM0 - WM, min=0.0) # X, soil above capacity
SM_c = torch.minimum(SM0, WM)
sm_lt = SM_c < WM
Wmaxm = WM * (1.0 + B) # W_m
A = Wmaxm * (1.0 - _safe_pow(torch.clamp(1.0 - SM_c / WM, min=0.0), 1.0 / (1.0 + B)))
saturated = (precipSoil + A) >= Wmaxm
R_sat = torch.clamp(precipSoil - (WM - SM_c), min=0.0)
b1 = torch.clamp(1.0 - A / Wmaxm, min=0.0)
b2 = torch.clamp(1.0 - (A + precipSoil) / Wmaxm, min=0.0)
infiltration = torch.minimum(WM * (_safe_pow(b1, 1.0 + B) - _safe_pow(b2, 1.0 + B)), precipSoil)
R_inf = torch.clamp(precipSoil - infiltration, min=0.0)
R_lt = torch.where(saturated, R_sat, R_inf) # saturated or not
R_w = torch.where(sm_lt, R_lt, precipSoil) # soil already full: all runs off
The names differ from the equations (adjPET is the demand left after the canopy, WM is wm), but the steps are the same. Three details matter for differentiability. The branches are computed for every catchment and chosen with torch.where, so thousands of catchments run in one pass. _safe_pow adds \(10^{-9}\) to the base of each power, because the derivative of \(x^{1/(1+b)}\) is infinite at \(x = 0\). And the minimum with \(P_s\) guards against rounding errors.
In Python
The VIC arithmetic for the two Leaf River storms. Change SM or Ps to explore, or set b to 0.01 to turn the catchment into a single bucket.
Check your understanding
What happens to the VIC curve when b is very small? And very large?
Small b: \(W_m \approx \mathit{wm}\) and almost every point has capacity close to wm, so the soil behaves like one bucket: no runoff until it is full. Large b: capacities are very uneven, many points fill quickly, and runoff starts even in small storms.
Why did the 118 mm storm of 16 February 1990 produce far more runoff than the 128 mm storm of 27 August 1992?
Because of the soil state. In February 1990 the soil was nearly full, so the water level was already near the top of the curve and almost all new water ran off. In August 1992 the soil was 40% full and could store most of the storm.
On a "dry day" in the soil step's sense, can there still be rain?
Yes. A day is dry when the water reaching the ground does not exceed the remaining evaporative demand. A light shower on a hot day is a dry day: all of it evaporates.
What is the impervious fraction im for, and why is it nearly zero at the Leaf River?
It sends a share of every wet day's excess straight to overland flow, as from rock, roads or open water. The Leaf River is a forested rural basin, and calibration found almost no need for such a direct path.
What to remember
- CREST produces runoff where the soil is already full: saturation excess.
- The VIC curve spreads capacities over the catchment: wm is the mean capacity, b the unevenness, \(W_m = \mathit{wm}(1+b)\) the maximum.
- The soil state sets a water level \(A\); a storm raises it by \(P_s\); the area above the curve becomes runoff \(R\), the rest infiltrates.
- The same storm can produce very different runoff depending on how wet the soil already is.