Reference

Glossary

The symbols and terms used in the course. Variables carry their units, the name used in the package source (src/dcrest/kernels and model.py) and the session that introduces them. All water amounts are depths over the catchment, in millimetres.

Variables

Forcing, states and fluxes of one day, in the order the model computes them.

SymbolMeaningUnitsIn the codeSession

Parameters

The sixteen calibrated parameters, their allowed ranges and the released values for the Leaf River (USGS 02472000).

ParameterMeaningUnitsRangeLeaf RiverSession

Scores

ScoreDefinitionPerfectSession
error \(e_t\)\(Q_{\text{sim},t} - Q_{\text{obs},t}\), model minus observation on day \(t\)010
biasmean error; the volume ratio \(\beta\) expresses it as a ratio010
MAEmean absolute error, \(\overline{|e_t|}\), in mm/day010
RMSEroot-mean-square error, \(\sqrt{\overline{e_t^2}}\), in mm/day010
standard deviation \(\sigma\)the typical size of a series' departures from its mean–10
correlation \(r\) (Pearson)how closely two series rise and fall together, from −1 to 1110
rank correlation \(\rho\) (Spearman)correlation of the ranks: whether two lists put items in the same order113
NSENash–Sutcliffe efficiency, \(1 - \sum e_t^2 / \sum (Q_{\text{obs},t} - \overline{Q}_{\text{obs}})^2\); 0 means no better than predicting the observed mean every day110
KGEKling–Gupta efficiency, \(1 - \sqrt{(r-1)^2 + (\alpha-1)^2 + (\beta-1)^2}\)110
\(r\)correlation between simulated and observed flow (timing)110
\(\alpha\)variability ratio, \(\sigma_{\text{sim}} / \sigma_{\text{obs}}\) (size of the swings)110
\(\beta\)bias ratio, \(\mu_{\text{sim}} / \mu_{\text{obs}}\) (volume)110
log-flow NSENSE computed on \(\log(Q + 0.01)\); weighs low flows as much as high ones110
losswhat calibration minimises; here 1 − NSE on the training period011

Terms

TermMeaningSession
catchmentall the land that drains to one point on a river; also watershed, drainage basin1
stream gaugea station that records the river's water level and converts it to discharge1
dischargeriver flow; volume per second, or here a depth per day over the catchment1
water year1 October to 30 September, named after the year in which it ends1
water balance\(P = ET + Q + \Delta S\): water in equals water out plus the change in storage1
runoff ratioriver flow divided by precipitation over a period1
forcingthe daily inputs that drive the model: precipitation, temperature, potential evaporation1
conceptual modela model built from storages (buckets) and rules for moving water between them2
lumpedone value of each store for the whole catchment, no spatial detail2
state, flux, parameterwhat the model stores from day to day; what moves during a day; what is fixed and calibrated2
degree-day methodsnowmelt proportional to how far the temperature is above a threshold3
interceptionrain caught by leaves and evaporated before reaching the ground4
infiltration-capacity curvethe distribution of soil storage capacity across the catchment (VIC curve) that decides how much water infiltrates and how much runs off5
saturation excessrunoff produced because the soil is full, not because rain falls faster than it can soak in5
interflowwater moving sideways through the soil to the river6
baseflowthe slow contribution of groundwater that keeps a river flowing between storms7
unit hydrographhow runoff produced on one day is spread over the following days at the gauge8
convolutionadding up the delayed contributions of all past days' runoff with the unit-hydrograph weights8
spin-upyears simulated before the scored period so that the initial states no longer matter9
training and test periodsthe nine water years used to choose the parameters (1 October 1999 to 30 September 2008) and the decade used only to judge them (1 October 1989 to 30 September 1999)10
flow duration curvedaily flows sorted from highest to lowest, against the share of days they are exceeded10
baselinea simple prediction that needs no model, such as the average flow or the seasonal cycle10
calibrationchoosing parameter values so that the simulation matches observations on the training period11
effective parametera value that makes the model behave like the whole catchment; not a measurable property2, 11
loss landscapethe score as a function of the parameters11
equifinalitydifferent parameter sets fitting equally well (Beven, 2006)11
SCE-UAshuffled complex evolution, a population-based search that needs only scores (Duan et al., 1992)11
gradient descent, Adamstepping downhill on the loss along its gradient; Adam adds momentum and per-parameter step sizes (Kingma and Ba, 2015)11
learning rate \(\eta\)the step length of gradient descent11
momentuma running average of recent gradients that lets the search roll through small bumps11
complex, shufflein SCE-UA, a sub-group of the population that evolves on its own; and the regular pooling and re-dealing of all points among the complexes11
\(\theta\), \(u\)a parameter in its physical units, and the unbounded number the optimiser actually moves (Session 11, logistic transform)11
logistic transform\(\theta = \theta_{\min} + (\theta_{\max} - \theta_{\min})/(1 + e^{-u})\); keeps a parameter inside its range while the optimiser moves \(u\) freely11
gradientthe list of partial derivatives of a score with respect to each parameter (or input)12
automatic differentiationcomputing exact derivatives of a program by applying the chain rule to each recorded operation; reverse mode runs backwards from the score12
finite differenceapproximating a derivative by re-running the model with a slightly changed input12
dead (inert) parametera parameter whose gradient is exactly zero because it acts only through an inactive branch12
kinka point where a quantity's slope changes abruptly, such as where a cap starts to bind; the left and right slopes differ there12
batchingrunning many catchments at once as columns of one array, so one model run and one backward pass serve them all12
GPUgraphics processing unit: a processor that does the same arithmetic on many numbers at once12
checkpointingstoring intermediate values only every so often and recomputing the rest during the backward pass, to save memory12
CAMELSCatchment Attributes and Meteorology for Large-sample Studies: 671 little-disturbed US catchments with forcing, flow and attributes13
Daymet, Maurertwo gridded daily weather datasets built from station records (Thornton et al., 1997; Maurer et al., 2002), averaged over each catchment in CAMELS1, 10
SAC-SMA, Snow-17the Sacramento soil moisture accounting model, a conceptual model used operationally by the US National Weather Service, and its companion snow model10
HBVa widely used conceptual model from Sweden, with snow, soil and two groundwater stores10
VICthe Variable Infiltration Capacity land-surface model, whose infiltration curve dCREST uses (Session 5)10
mHM, FUSEthe mesoscale Hydrologic Model; and the Framework for Understanding Structural Errors, which assembles many conceptual models from interchangeable parts10
LSTMlong short-term memory, a type of neural network for sequences; here one network learns all catchments at once10
aridity indexmean potential evaporation divided by mean precipitation; above 1 is dry13
regional modelone model (here, the LSTM neural network) trained on many catchments at once13
routingmoving the runoff a model makes on the land to a point on the river: when it arrives, and how spread out8, 14
reacha stretch of river treated as one unit by a routing method; in the Rhine case about 10 km long14
river networkthe reaches of a basin and the order in which they flow into each other, down to the outlet14
flood wavea rise in flow that travels down a river; it moves faster than the water itself14
celerity \(c\)the speed of a flood wave; about 5/3 of the water's velocity in a wide channel14
attenuationthe lowering and spreading of a flood crest as it travels, caused by water stored along the way14
travel timethe time a wave needs from a reach to the outlet: the sum of length ÷ celerity along its path14
time–area diagramthe share of a basin's area by travel time to the outlet; with the same runoff everywhere it is the basin's unit hydrograph14
lumped, distributeda lumped model gives the whole catchment one rainfall value and one set of stores (dCREST v1.0); a distributed model computes runoff in many cells or sub-basins and routes it between them14
Saint-Venant equationsconservation of mass and of momentum for flow in a channel; every routing method is a simplification of them15
dynamic, diffusive and kinematic wavethe Saint-Venant equations in full; without the inertia terms; and without the water-surface slope as well (waves then only move downstream and do not flatten)15
local inertialthe Saint-Venant equations without the advection term; used by floodplain models such as CaMa-Flood15
Nash cascadea chain of equal linear reservoirs; its response is the gamma curve of dCREST's unit hydrograph15
Muskingum methodrouting a reach with storage \(S = K[X I + (1 - X) O]\): one line of arithmetic per reach and step; \(K\) is the travel time, \(X\) sets the flattening15
Muskingum–Cungethe Muskingum method with \(K\) and \(X\) computed from reach length, celerity, width and slope, so that it behaves like a diffusive wave (Cunge, 1969)15
Courant number\(c\,\Delta t/\Delta x\): how many grid cells or reaches a wave crosses in one time step; explicit schemes need it to stay near or below 115
time step, sub-stepthe interval at which the router updates the flow; a daily model may take many routing sub-steps per day to stay stable15
topological levelthe position of a reach in the upstream-to-downstream order; reaches of the same level can be computed at the same time15
floodplainthe land beside a river that floods at high water; it stores water and delays and caps the flood crest16
regulationthe change of a river's flow by dams and reservoirs, which hold back floods and release water later16