Get Started with airGR

Introduction

airGR is a package that brings into the R software the hydrological modelling tools used and developed at the Catchment Hydrology Research Group at Irstea (France), including the GR rainfall-runoff models and a snowmelt and accumulation model, CemaNeige. Each model core is coded in Fortran to ensure low computational time. The other package functions (i.e. mainly the calibration algorithm and the efficiency criteria calculation) are coded in R.

The airGR package has been designed to fulfill two major requirements: to facilitate the use by non-expert users and to allow flexibility regarding the addition of external criteria, models or calibration algorithms. The names of the functions and their arguments were chosen to this end. airGR also contains basics plotting facilities.

Six hydrological models and one snowmelt and accumulation model are implemented in airGR. The snow model can be used alone or together with the daily hydrological models.

The models can be called within airGR using the following functions:

The GRP forecasting model and the Otamin predictive uncertainty tool are not available in airGR.

In this vignette, we show how to prepare and run a calibration and a simulation with airGR hydrological models.

Loading data

In the following example, we use a data sample contained in the package. For real applications, the user has to import its data into R and to prepare it with an adequate data.frame format as described below.

First, it is necessary to load the airGR package:

library(airGR)

Below is presented an example of a data.frame of daily hydrometeorological observations time series for a fictional catchment included in the airGR package that contains:

data(L0123001)
summary(BasinObs)
##      DatesR                 P                T                 E        
##  Min.   :1984-01-01   Min.   : 0.000   Min.   :-18.700   Min.   :0.000  
##  1st Qu.:1991-04-02   1st Qu.: 0.000   1st Qu.:  4.100   1st Qu.:0.600  
##  Median :1998-07-02   Median : 0.300   Median :  9.100   Median :1.400  
##  Mean   :1998-07-02   Mean   : 2.915   Mean   :  9.147   Mean   :1.764  
##  3rd Qu.:2005-10-01   3rd Qu.: 3.600   3rd Qu.: 14.500   3rd Qu.:2.900  
##  Max.   :2012-12-31   Max.   :66.800   Max.   : 28.400   Max.   :5.500  
##                                                                         
##       Qls             Qmm         
##  Min.   :   70   Min.   : 0.0168  
##  1st Qu.: 1640   1st Qu.: 0.3936  
##  Median : 4050   Median : 0.9720  
##  Mean   : 6130   Mean   : 1.4713  
##  3rd Qu.: 7850   3rd Qu.: 1.8840  
##  Max.   :99500   Max.   :23.8800  
##  NA's   :772     NA's   :772

The usual functions (e.g. read.table()) can be used to load real-case data sets.

Preparation of functions inputs

To run a model, the functions of the airGR package (e.g. the models, calibration and criteria calculation functions) require data and options with specific formats.

To facilitate the use of the package, there are several functions dedicated to the creation of these objects:

InputsModel object

To run a GR hydrological model or CemaNeige, the user has to prepare the input data with the CreateInputsModel() function. As arguments, this function needs the function name corresponding to the model the user wants to run, a vector of dates, a vector of precipitation and a vector of potential evapotranspiration.

In the example below, we already have the potential evapotranspiration. If the user does not have these data, it is possible to compute it with the Oudin’s formula with the PEdaily_Oudin() function (this function only needs Julian days, daily average air temperature and latitude).

Missing values (NA) of precipitation (or potential evapotranspiration) are not allowed.

InputsModel <- CreateInputsModel(FUN_MOD = RunModel_GR4J, DatesR = BasinObs$DatesR,
                                 Precip = BasinObs$P, PotEvap = BasinObs$E)
str(InputsModel)
## List of 3
##  $ DatesR : POSIXlt[1:10593], format: "1984-01-01" "1984-01-02" ...
##  $ Precip : num [1:10593] 4.1 15.9 0.8 0 0 0 0 0 2.9 0 ...
##  $ PotEvap: num [1:10593] 0.2 0.2 0.3 0.3 0.1 0.3 0.4 0.4 0.5 0.5 ...
##  - attr(*, "class")= chr [1:3] "InputsModel" "daily" "GR"

RunOptions object

The CreateRunOptions() function allows to prepare the options required to the RunModel*() functions, which are the actual models functions.

The user must at least define the following arguments:

To select a period for which the user wants to run the model, select the corresponding indexes for different time periods (not the POSIXt dates), as follows:

Ind_Run <- seq(which(format(BasinObs$DatesR, format = "%d/%m/%Y")=="01/01/1990"), 
               which(format(BasinObs$DatesR, format = "%d/%m/%Y")=="31/12/1999"))
str(Ind_Run)
##  int [1:3652] 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 ...

The initialization of hydrological models is of the utmost importance. Indeed, an inaccurate initialization causes poor quality discharge simulations during the earliest stages of the running period. For example, in the GR models, by default, the production and the routing store levels store level are respectively set to 30 % and 50 % of their capacity, which may be far from their ideal value. Two solutions are offered to accurately initialize the GR models in airGR: manually predefining the initial states (e.g. from a previous run) or running the models during a warm up period before the actual running period. It is generally advised to set up this warm up period to be equal or superior to one year.

As a consequence, it is possible to define in CreateRunOptions() the following arguments:

RunOptions <- CreateRunOptions(FUN_MOD = RunModel_GR4J,
                               InputsModel = InputsModel, IndPeriod_Run = Ind_Run,
                               IniStates = NULL, IniResLevels = NULL, IndPeriod_WarmUp = NULL)
## Warning in CreateRunOptions(FUN_MOD = RunModel_GR4J, InputsModel = InputsModel, :     Model warm up period not defined -> default configuration used 
##      The year preceding the run period is used
str(RunOptions)
## List of 6
##  $ IndPeriod_WarmUp: int [1:365] 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 ...
##  $ IndPeriod_Run   : int [1:3652] 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 ...
##  $ IniStates       : num [1:67] 0 0 0 0 0 0 0 0 0 0 ...
##  $ IniResLevels    : num [1:3] 0.3 0.5 NA
##  $ Outputs_Cal     : chr "Qsim"
##  $ Outputs_Sim     : chr [1:20] "DatesR" "PotEvap" "Precip" "Prod" ...
##  - attr(*, "class")= chr [1:3] "RunOptions" "GR" "daily"

The CreateRunOptions() function returns warnings if the default initialization options are used:

InputsCrit object

The CreateInputsCrit() function allows to prepare the input in order to calculate a criterion. It is possible to define the following arguments:

Missing values (NA) are allowed for observed discharge.

InputsCrit <- CreateInputsCrit(FUN_CRIT = ErrorCrit_NSE, InputsModel = InputsModel, 
                               RunOptions = RunOptions, Qobs = BasinObs$Qmm[Ind_Run])
str(InputsCrit)
## List of 4
##  $ BoolCrit: logi [1:3652] TRUE TRUE TRUE TRUE TRUE TRUE ...
##  $ Qobs    : num [1:3652] 1.99 1.8 2.86 2.4 3.31 ...
##  $ transfo : chr ""
##  $ epsilon : NULL
##  - attr(*, "class")= chr "InputsCrit"

CalibOptions object

Before using the automatic calibration tool, the user needs to prepare the calibration options with the CreateCalibOptions() function. For that, it is necessary to define the following arguments:

CalibOptions <- CreateCalibOptions(FUN_MOD = RunModel_GR4J, FUN_CALIB = Calibration_Michel)
str(CalibOptions)
## List of 3
##  $ FixedParam       : logi [1:4] NA NA NA NA
##  $ SearchRanges     : num [1:2, 1:4] 4.59e-05 2.18e+04 -1.09e+04 1.09e+04 4.59e-05 ...
##  $ StartParamDistrib: num [1:3, 1:4] 169.017 247.151 432.681 -2.376 -0.649 ...
##  - attr(*, "class")= chr [1:3] "CalibOptions" "GR4J" "HBAN"

Criteria

The evaluation of the quality of a simulation is estimated through the calculation of criteria. These criteria can be used both as objective-functions during the calibration of the model, or as a measure for evaluating its performance on a control period.

The package offers the possibility to use different criteria:

It is also possible to create user-defined criteria. For doing that, it is only necessary to define the function in R following the same syntax as the criteria functions included in airGR.

Calibration

The objective of the calibration algorithm is to identify the model parameters: by comparing the model outputs with observed data, this algorithm determines the combination of parameters that represents the best the behavior of the watershed.

In the airGR package, a function called Calibration_Michel() is implemented. This functions allows running a calibration with the package models. The calibration algorithm optimizes the error criterion selected as objective-function. This algorithm works in two steps:

  1. a screening of the parameters space is performed using either a rough predefined grid or a user-defined list of parameter sets
  2. a simple steepest descent local search algorithm is performed from the best set of parameters found at the first step
OutputsCalib <- Calibration_Michel(InputsModel = InputsModel, RunOptions = RunOptions,
                                   InputsCrit = InputsCrit, CalibOptions = CalibOptions,
                                   FUN_MOD = RunModel_GR4J, FUN_CRIT = ErrorCrit_NSE)
## Grid-Screening in progress (0% 20% 40% 60% 80% 100%)
##   Screening completed (81 runs)
##       Param =  247.151 ,   -0.020 ,   83.096 ,    2.384
##       Crit NSE[Q]       = 0.7688
## Steepest-descent local search in progress
##   Calibration completed (21 iterations, 234 runs)
##       Param =  257.238 ,    1.012 ,   88.235 ,    2.208
##       Crit NSE[Q]       = 0.7988
Param <- OutputsCalib$ParamFinalR
Param
## [1] 257.237556   1.012237  88.234673   2.207958

The Calibration_Michel() function is the only one implemented in the airGR package to calibrate the model, but the user can implement its own calibration function. Two vignettes explain how it can be done (2.1 Plugging in new calibration and 2.2 MCMC parameter estimation).

The Calibration_Michel() function returns a vector with the parameters of the chosen model, which means that the number of values can differ depending on the model that is used. It is possible to use the Calibration_Michel() function with user-implemented hydrological models.

Control

This step assesses the predictive capacity of the model. Control is defined as the estimation of the accuracy of the model on data sets that are not used in its construction, and in particular its calibration. The classical way to perform a control is to keep data from a period separated from the calibration period. If possible, this control period should correspond to climatic situations that differ from those of the calibration period in order to better point out the qualities and weaknesses of the model. This exercise is necessary for assessing the robustness of the model, that is to say its ability to keep stable performances outside of the calibration conditions.

Performing a model control with airGR is similar to running a simulation (see below), followed by the computation of one or several performance criteria.

Simulation

Simulation run

To run a model, the user has to use the RunModel*() functions (InputsModel, RunOptions and parameters). All the data needed have already been prepared in the previous steps defined in this guide.

OutputsModel <- RunModel_GR4J(InputsModel = InputsModel, RunOptions = RunOptions, Param = Param)
str(OutputsModel)
## List of 20
##  $ DatesR  : POSIXlt[1:3652], format: "1990-01-01" "1990-01-02" ...
##  $ PotEvap : num [1:3652] 0.3 0.4 0.4 0.3 0.1 0.1 0.1 0.2 0.2 0.3 ...
##  $ Precip  : num [1:3652] 0 9.3 3.2 7.3 0 0 0 0 0.1 0.2 ...
##  $ Prod    : num [1:3652] 196 199 199 201 200 ...
##  $ Pn      : num [1:3652] 0 8.9 2.8 7 0 0 0 0 0 0 ...
##  $ Ps      : num [1:3652] 0 3.65 1.12 2.75 0 ...
##  $ AE      : num [1:3652] 0.2833 0.4 0.4 0.3 0.0952 ...
##  $ Perc    : num [1:3652] 0.645 0.696 0.703 0.74 0.725 ...
##  $ PR      : num [1:3652] 0.645 5.946 2.383 4.992 0.725 ...
##  $ Q9      : num [1:3652] 1.78 1.52 3.86 3.17 3.45 ...
##  $ Q1      : num [1:3652] 0.2 0.195 0.271 0.387 0.365 ...
##  $ Rout    : num [1:3652] 53.9 53.6 55.3 56.1 56.9 ...
##  $ Exch    : num [1:3652] 0.181 0.18 0.177 0.197 0.207 ...
##  $ AExch1  : num [1:3652] 0.181 0.18 0.177 0.197 0.207 ...
##  $ AExch2  : num [1:3652] 0.181 0.18 0.177 0.197 0.207 ...
##  $ AExch   : num [1:3652] 0.362 0.36 0.353 0.393 0.414 ...
##  $ QR      : num [1:3652] 2.05 1.99 2.36 2.55 2.78 ...
##  $ QD      : num [1:3652] 0.381 0.375 0.447 0.584 0.572 ...
##  $ Qsim    : num [1:3652] 2.43 2.37 2.8 3.14 3.35 ...
##  $ StateEnd:List of 3
##   ..$ Store          :List of 3
##   .. ..$ Prod: num 189
##   .. ..$ Rout: num 48.9
##   .. ..$ Exp : num NA
##   ..$ UH             :List of 2
##   .. ..$ UH1: num [1:20] 0.514 0.54 0.148 0 0 ...
##   .. ..$ UH2: num [1:40] 0.056306 0.057176 0.042253 0.012187 0.000578 ...
##   ..$ CemaNeigeLayers:List of 2
##   .. ..$ G  : num NA
##   .. ..$ eTG: num NA
##   ..- attr(*, "class")= chr [1:3] "IniStates" "GR" "daily"
##  - attr(*, "class")= chr [1:3] "OutputsModel" "daily" "GR"

Results preview

Although it is possible for the user to design its own graphics from the outputs of the RunModel*() functions, the airGR package offers the possibility to make use of the plot.OutputsModel() function (or plot() with an OutputsModel object). This function returns a dashboard of results including various graphs (depending on the model used):

plot(OutputsModel, Qobs = BasinObs$Qmm[Ind_Run])

Moreover, if the CemaNeige model is used, the air temperature and the simulated snowpack water equivalent time series are plotted.

Efficiency criterion

To evaluate the efficiency of the model, it is possible to use the same criterion as defined at the calibration step or to use another criterion.

OutputsCrit <- ErrorCrit_NSE(InputsCrit = InputsCrit, OutputsModel = OutputsModel)
## Crit. NSE[Q] = 0.7988
OutputsCrit <- ErrorCrit_KGE(InputsCrit = InputsCrit, OutputsModel = OutputsModel)
## Crit. KGE[Q] = 0.7854
##  SubCrit. KGE[Q] cor(sim, obs, "pearson") = 0.8985 
##  SubCrit. KGE[Q] sd(sim)/sd(obs)          = 0.8161 
##  SubCrit. KGE[Q] mean(sim)/mean(obs)      = 1.0437

References

Le Moine, Nicolas. 2008. “Le Bassin Versant de Surface Vu Par Le Souterrain : Une Voie d’amélioration Des Performances et Du Réalisme Des Modèles Pluie-Débit ?” PhD thesis, Université Pierre et Marie Curie, Paris 6. http://webgr.irstea.fr/wp-content/uploads/2012/07/2008-LE_MOINE-THESE.pdf.

Mathevet, Thibault. 2005. “Quels Modèles Pluie-Débit Globaux Au Pas de Temps Horaire ? Développements Empiriques et Comparaison de Modèles Sur Un Large échantillon de Bassins Versants.” PhD thesis, Paris: ENGREF. http://webgr.irstea.fr/wp-content/uploads/2012/07/2005-MATHEVET-THESE.pdf.

Mouelhi, Safouane. 2003. “Vers Une Chaîne Cohérente de Modèles Pluie-Débit Conceptuels Globaux Aux Pas de Temps Pluriannuel, Annuel, Mensuel et Journalier.” PhD thesis, Paris, ENGREF. http://webgr.irstea.fr/wp-content/uploads/2012/07/2003-MOUELHI-THESE.pdf.

Mouelhi, Safouane, Claude Michel, Charles Perrin, and Vazken Andréassian. 2006a. “Stepwise Development of a Two-Parameter Monthly Water Balance Model.” Journal of Hydrology 318 (1-4): 200–214. doi:10.1016/j.jhydrol.2005.06.014.

———. 2006b. “Linking Stream Flow to Rainfall at the Annual Time Step: The Manabe Bucket Model Revisited.” Journal of Hydrology 328 (1-2): 283–96. doi:10.1016/j.jhydrol.2005.12.022.

Perrin, Charles, Claude Michel, and Vazken Andréassian. 2003. “Improvement of a Parsimonious Model for Streamflow Simulation.” Journal of Hydrology 279 (1-4): 275–89. doi:10.1016/S0022-1694(03)00225-7.

Pushpalatha, Raji, Charles Perrin, Nicolas Le Moine, Thibault Mathevet, and Vazken Andréassian. 2011. “A Downward Structural Sensitivity Analysis of Hydrological Models to Improve Low-Flow Simulation.” Journal of Hydrology 411 (1–2): 66–76. doi:10.1016/j.jhydrol.2011.09.034.

Valéry, Audrey, Vazken Andréassian, and Charles Perrin. 2014. “’As Simple as Possible but Not Simpler’: What Is Useful in a Temperature-Based Snow-Accounting Routine? Part 2 - Sensitivity Analysis of the Cemaneige Snow Accounting Routine on 380 Catchments.” Journal of Hydrology, no. 517(0): 1176–87. doi:10.1016/j.jhydrol.2014.04.058.