Session 2 · Foundations

What a conceptual model is

The last session ended with a hidden quantity: how wet the catchment is. A conceptual model represents that wetness as water held in a few imaginary stores, and moves water between them with simple rules. This session builds the smallest such model, shows why it is not enough, and then lays out the full map of dCREST that the next six sessions will fill in.

Time about 30 minutesNeeds Session 1
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

Three ways to model a river

There are broadly three ways to turn weather into river flow.

dCREST is a conceptual model. It is also lumped (the whole catchment is treated as a single point, with one value of each store) and daily (it advances one day at a time). Its soil water balance comes from CREST (Wang et al., 2011), the model behind the EF5 forecasting system; its groundwater store follows CRESTPhys, a CREST extension that exists only in a fork of EF5, and its snow, canopy and routing steps are its own. What makes it unusual is that it is written so that it can be differentiated, which Sessions 11 and 12 will explain.

The vocabulary

Five words describe every model in this course.

WordMeaningExamples in dCREST
Forcingthe inputs that drive the model, given day by dayP, T, PET
Statethe amount of water in a store at the end of a day: the model's memorysnowpack SWE, soil water SM, groundwater G
Fluxwater moving from one place to another during a daymelt M, infiltration I, baseflow Qb
Parametera fixed number that shapes a rule; one value per catchment, constant in timesoil capacity wm, melt factor ddf
Outputwhat we compare with observationsriver flow Q

States are measured in millimetres of water and fluxes in millimetres per day. Because the model steps one day at a time, a flux of 3 mm/day moves 3 mm of water during that step.

One store, one rule: the mass balance

Every store obeys the same bookkeeping. The water in it at the end of today equals the water at the end of yesterday, plus what came in today, minus what went out:

\[ S_t \;=\; S_{t-1} \;+\; \text{inflow}_t \;-\; \text{outflow}_t \]

Writing a model means choosing the rules for the outflows. A common choice is the linear reservoir: each day the store releases a fixed fraction \(k\) of its contents, \(Q_t = k\,S\). A store with \(k = 0.1\) per day releases 10% of its water each day, so with no new input its outflow halves about every 7 days.

The smallest useful model: one bucket

Here is a one-store model of the Leaf River. Each day it does three things, in this order:

  1. add the day's precipitation to the bucket: \(S \leftarrow S + P\);
  2. evaporate \(E = \min(\alpha \cdot PET,\; S)\) and subtract it, where \(\alpha\) scales the atmosphere's demand and the minimum stops the bucket going negative;
  3. release \(Q = k\,S\) to the river and subtract it.

It has two parameters, \(k\) and \(\alpha\). Try to make the blue line follow the black one. The score shown is the Nash–Sutcliffe efficiency (NSE): 1 is a perfect match, 0 means no better than predicting the average flow every day. Session 10 explains it properly.

Try itA one-bucket model of the Leaf River

Water years 1990–1991. The model starts from 50 mm in the bucket on 1 October 1987 and runs for two years before what is shown, so the starting value hardly matters.

Whatever you choose, something is wrong. A large \(k\) produces sharp peaks but the river empties in days and falls to nothing between storms. A small \(k\) keeps the river flowing but smears every flood into a low bump. The real river does both at once: it has a fast response and a slow one. And with a single store, summer rain fills the same bucket that feeds winter floods, so the model cannot reproduce the strong seasonal contrast of Session 1.

Key idea

A useful model needs several stores that respond at different speeds, and rules that make the response depend on how wet the catchment already is. That is exactly what dCREST adds: a soil store that decides how much rain runs off, a fast path and a slow path, and evaporation from several places.

The map of dCREST

dCREST has five stores and moves water between them with a fixed sequence of rules, once per day. The diagram is the map for the rest of the course; each of the next six sessions zooms into one part of it.

dCREST's stores and fluxes. Water enters at the top as precipitation and passes through snow, canopy and soil. The soil sends it along three paths: overland flow through a fast reservoir, interflow, and drainage to a groundwater store (the aquifer). Evaporation leaves from the canopy, the soil and the aquifer. The three outflows are summed, reduced by a loss fraction and spread over the following days by a unit hydrograph to give river flow. The italic names are the parameters that act at each place.

Each day the model goes down the diagram in a fixed order:

  1. Snow (Session 3): decide whether precipitation is rain or snow; melt some snow if it is warm.
  2. Canopy (Session 4): leaves catch some water; the atmosphere's demand evaporates from the canopy first.
  3. Soil (Sessions 5 and 6): split the water reaching the soil into water stored and runoff; split runoff into overland flow and interflow; on dry days, evaporate from the soil.
  4. Aquifer (Session 7): drain the soil into groundwater, spill any excess, release baseflow; on dry days the aquifer can also feed evaporation.
  5. Fast reservoir and losses (Session 8): delay overland flow; remove a small fraction as losses.
  6. Unit hydrograph (Session 8): spread each day's runoff over the following days to reach the gauge.

Because the steps run in order, what one store does today depends on what happened upstream earlier the same day, and on its own state from yesterday. The states carry the memory from day to day.

The sixteen parameters

These are the numbers that tailor dCREST to a catchment. Each has an allowed range. The last column shows the values found for the Leaf River by calibration (Session 11); they are the values used everywhere in this course unless a widget lets you change them.

ParameterWhat it controlsUnitsRangeLeaf RiverSession
A word of caution

Parameters have physical-sounding names, but after calibration they are effective values: whatever makes this simple structure reproduce the river best. For example, the Leaf River's calibrated canopy can hold 32 mm of water, far more than real leaves hold. Session 4 explains why calibration chose that, and Session 13 discusses how far such values can be interpreted physically.

Lumped, daily, and one catchment at a time (almost)

The original CREST is distributed: it runs the same rules in every cell of a grid covering a region and routes water from cell to cell. dCREST is lumped: one set of stores per catchment, no routing between cells. That makes it cheap enough to run thousands of times, which calibration needs.

dCREST also runs many catchments side by side. Its inputs are tables with one column per catchment, of shape (days × catchments), and each column is processed independently. On a graphics card, 665 catchments take about as long as one. In this course we use one column: the Leaf River.

In Python

The one-bucket model is ten lines. Change k and alpha and run again. Can you find values that give an NSE above 0.6?

Check your understanding

Is the soil water SM a state, a flux or a parameter? And the soil capacity wm?

SM is a state: it changes every day and carries memory. wm is a parameter: one fixed number for the catchment that shapes how the soil behaves.

A linear reservoir holds 80 mm and releases 5% per day. With no inflow, how much is released on the first day, and how much remains after the second?

Day 1: \(0.05 \times 80 = 4\) mm, leaving 76 mm. Day 2: 3.8 mm, leaving 72.2 mm.

Why can a single bucket not reproduce both sharp floods and steady flow in dry weather?

One store has one release rate. A fast rate empties the store quickly (sharp floods, but no flow between storms); a slow rate keeps flow going but flattens floods. The river needs a fast path and a slow path at the same time.

In what order does dCREST process a day?

Snow, canopy, soil, aquifer, fast reservoir and losses; then, after all days are done, the unit hydrograph spreads each day's runoff over the following days.

What to remember