a non-stationary spatial weather generator for statistical
TRANSCRIPT
A non-stationary spatial weather generator for
statistical modelling of daily precipitation
Pradeebane VAITTINADA AYAR and Juliette BLANCHET
International Workshop on Stochastic Weather Generators for HydrologicalApplications
Berlin, Germany – 20th September 2017
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Outline
1 Context
2 Overview of the model
3 Parameter estimation
4 First evaluations
5 Conclusions and Perspectives
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 2/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Outline
1 Context
2 Overview of the model
3 Parameter estimation
4 First evaluations
5 Conclusions and Perspectives
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 3/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Context
SCHADEX : method for flood determination and dam dimensioning at EDF
Need for a realistic daily rainfall simulator to provide inputs for SCHADEX
Currently available : Regional MEWP [cf. Evin et al., 2016] univariate model
Aim to build a generic rainfall simulator (working over different catchements) :
accounting for spatial dependencies (daily time scale and one km2 spatialresolution)well performing for high precipitation quantiles at catchement scale
First choice : to model rain occurrence and intensity at thesame time only from observations.
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 3/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Context
SCHADEX : method for flood determination and dam dimensioning at EDF
Need for a realistic daily rainfall simulator to provide inputs for SCHADEX
Currently available : Regional MEWP [cf. Evin et al., 2016] univariate model
Aim to build a generic rainfall simulator (working over different catchements) :
accounting for spatial dependencies (daily time scale and one km2 spatialresolution)well performing for high precipitation quantiles at catchement scale
First choice : to model rain occurrence and intensity at thesame time only from observations.
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 3/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Context
SCHADEX : method for flood determination and dam dimensioning at EDF
Need for a realistic daily rainfall simulator to provide inputs for SCHADEX
Currently available : Regional MEWP [cf. Evin et al., 2016] univariate model
Aim to build a generic rainfall simulator (working over different catchements) :
accounting for spatial dependencies (daily time scale and one km2 spatialresolution)well performing for high precipitation quantiles at catchement scale
First choice : to model rain occurrence and intensity at thesame time only from observations.
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 3/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
(a) Ardeche catchement at Sauze
Ardèche at Sauze basin
Longitude (km)
Latit
ude
(km
)
48 212 376 541 705 869 1033 1198
1620
1771
1922
2073
2224
2375
2526
2677
(b) Stations locations and altitudes
0
500
1000
1500
2000
Elevation (m)
●
●●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
CHASSERADES
SAINTE−MARGUERITE−LAFIGERE
CUBIERES
LANGOGNE
BERZEME
SABLIERESALBA−LA−ROMAINE
BOURG−SAINT−ANDEOL
LOUBARESSEMIRABEL
PRIVAS
SAINT−ETIENNE−DE−LUGDARES
SAINT−MONTAN
BAGNOLS−LES−BAINS
CHATEAUNEUF−DE−RANDON
FLORAC
LE−PONT−DE−MONTVERT
SAINT−SAUVEUR−DE−GINESTOUX
VILLEFORT
LAPALUD
VALS−LES−BAINSCHAUDEYRAC
LE−BLEYMARD
Ardèche at Sauze
X (km) − Lambert II extended
Y (
km)
− L
ambe
rt II
ext
ende
d
●
●
●●
●
●●
●●
●●
●
●
●
●
●
●
●
●
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
ST−PIERREVILLE
MAYRESMONTPEZAT−SOUBEYROL
ANTRAIGUES−AIZAC
AUBENAS
LABLACHÈRE
JOYEUSE
VALLON−PONT−D'ARC
MALONS
SENECHAS
ST−MAURICE−DE−VENTALON
MAS−DE−LA−BARQUE
BESSEGES
STE−EULALIE
USCALADES−ET−RIEUTORD
MAZAN
LAC−D'ISSARLÈS
ISSANLAS−MEZEYRAC
MASMEJEAN/−BASTIDE−PUYLAURENT
695 709 722 736 749 763 776 790
1910
1921
1933
1944
1956
1967
1979
1990
FIGURE 1: Area : 2260 km2 – 42 Stations – Altitudes : from 47 to 1425 mAt least 20 years long time series
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 4/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Outline
1 Context
2 Overview of the model
3 Parameter estimation
4 First evaluations
5 Conclusions and Perspectives
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 5/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Model framework
∀x ∈ D,Y (x) a r.v. characterising the precipitation over a given domain. For a givenday, the rainfield is equal to a realisation of {Y (x)}x∈D
What is {Y (x)}x∈D ?The rainfield {Y (x)}x∈D is obtained from a model based on a transformedlatent censored gaussian field quite widely used in the literature [e.g., Vischelet al., 2009; Rasmussen, 2013; Allard and Bourotte, 2015].
Gaussian field : fully defined by its mean vector and covariance structure.
Model stepsMarginal model
Spatial model
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 5/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Model framework
∀x ∈ D,Y (x) a r.v. characterising the precipitation over a given domain. For a givenday, the rainfield is equal to a realisation of {Y (x)}x∈D
What is {Y (x)}x∈D ?The rainfield {Y (x)}x∈D is obtained from a model based on a transformedlatent censored gaussian field quite widely used in the literature [e.g., Vischelet al., 2009; Rasmussen, 2013; Allard and Bourotte, 2015].
Gaussian field : fully defined by its mean vector and covariance structure.
Model stepsMarginal model
Spatial model
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 5/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Model framework
∀x ∈ D,Y (x) a r.v. characterising the precipitation over a given domain. For a givenday, the rainfield is equal to a realisation of {Y (x)}x∈D
What is {Y (x)}x∈D ?The rainfield {Y (x)}x∈D is obtained from a model based on a transformedlatent censored gaussian field quite widely used in the literature [e.g., Vischelet al., 2009; Rasmussen, 2013; Allard and Bourotte, 2015].Gaussian field : fully defined by its mean vector and covariance structure.
Model stepsMarginal model
Spatial model
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 5/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Model framework
∀x ∈ D,Y (x) a r.v. characterising the precipitation over a given domain. For a givenday, the rainfield is equal to a realisation of {Y (x)}x∈D
What is {Y (x)}x∈D ?The rainfield {Y (x)}x∈D is obtained from a model based on a transformedlatent censored gaussian field quite widely used in the literature [e.g., Vischelet al., 2009; Rasmussen, 2013; Allard and Bourotte, 2015].Gaussian field : fully defined by its mean vector and covariance structure.
Model stepsMarginal model
Spatial model
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 5/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Marginal model
Extended Generalized Pareto of Type II (EGPII)Ü Model the full range of rainfall intensities without any threshold selection.Naveau et al. [2016] : P(Y 6 y) = G
(Hξ( yσ
))
where
Hξ, the Generalized Pareto Distribution
G(v) = vκ
Courtesy : Naveau et al. [2016] ξ = 0.5
0 1 2 3 4 5 6 7
0.0
0.2
0.4
0.6
0.8
1.0
x
Den
sity
f(x)
ξ=0.5
κ = 1κ = 2κ = 5Gamma
0 2 4 6 8 10
050
100
150
200
U = −ln(−ln(F))
prec
ipita
tion
(mm
/day
)
κ = 1κ = 2κ = 5Gamma
ξ=0.510 20 50 100 1000 10000
Temps de retour T (annees)
0 1 2 3 4 5 6 7
0.0
0.2
0.4
0.6
0.8
1.0
x
Den
sity
f(x)
ξ=0.5
κ = 1κ = 2κ = 5Gamma
0 2 4 6 8 10
050
100
150
200
U = −ln(−ln(F))
prec
ipita
tion
(mm
/day
)
κ = 1κ = 2κ = 5Gamma
ξ=0.510 20 50 100 1000 10000
Temps de retour T (annees)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 6/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Marginal model
Extended Generalized Pareto of Type II (EGPII)Ü Model the full range of rainfall intensities without any threshold selection.Naveau et al. [2016] : P(Y 6 y) = G
(Hξ( yσ
))where
Hξ, the Generalized Pareto Distribution
G(v) = vκ
Courtesy : Naveau et al. [2016] ξ = 0.5
0 1 2 3 4 5 6 7
0.0
0.2
0.4
0.6
0.8
1.0
x
Den
sity
f(x)
ξ=0.5
κ = 1κ = 2κ = 5Gamma
0 2 4 6 8 10
050
100
150
200
U = −ln(−ln(F))
prec
ipita
tion
(mm
/day
)
κ = 1κ = 2κ = 5Gamma
ξ=0.510 20 50 100 1000 10000
Temps de retour T (annees)
0 1 2 3 4 5 6 7
0.0
0.2
0.4
0.6
0.8
1.0
x
Den
sity
f(x)
ξ=0.5
κ = 1κ = 2κ = 5Gamma
0 2 4 6 8 10
050
100
150
200
U = −ln(−ln(F))pr
ecip
itatio
n (m
m/d
ay)
κ = 1κ = 2κ = 5Gamma
ξ=0.510 20 50 100 1000 10000
Temps de retour T (annees)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 6/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Marginal model
Extended Generalized Pareto of Type II (EGPII)Ü Model the full range of rainfall intensities without any threshold selection.Naveau et al. [2016] : P(Y 6 y) = G
(Hξ( yσ
))where
Hξ, the Generalized Pareto Distribution
G(v) = vκ
Constraints
Low values driven by κ, large values driven by ξ
Weilbull type lower tail behavior (bounded orshort-tailed)
Frechet type upper tail behavior (unbouded orheavy-tailed)
Courtesy : Naveau et al. [2016] ξ = 0.5
0 1 2 3 4 5 6 7
0.0
0.2
0.4
0.6
0.8
1.0
x
Den
sity
f(x)
ξ=0.5
κ = 1κ = 2κ = 5Gamma
0 2 4 6 8 10
050
100
150
200
U = −ln(−ln(F))
prec
ipita
tion
(mm
/day
)
κ = 1κ = 2κ = 5Gamma
ξ=0.510 20 50 100 1000 10000
Temps de retour T (annees)
0 1 2 3 4 5 6 7
0.0
0.2
0.4
0.6
0.8
1.0
x
Den
sity
f(x)
ξ=0.5
κ = 1κ = 2κ = 5Gamma
0 2 4 6 8 10
050
100
150
200
U = −ln(−ln(F))pr
ecip
itatio
n (m
m/d
ay)
κ = 1κ = 2κ = 5Gamma
ξ=0.510 20 50 100 1000 10000
Temps de retour T (annees)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 6/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Marginal model
Extended GP-II : without threshold
F (y ;σ, ξ) =
(
1 −(
1 + ξyσ
)−1/ξ)κ
if ξ 6= 0,(1 − exp
{−yσ
})κ, if ξ = 0.
ξ is set to 0 Þ powered exponential used as marginaldistribution.
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 7/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Marginal model
Extended GP-II : without threshold
F (y ;σ, ξ) =
(
1 −(
1 + ξyσ
)−1/ξ)κ
if ξ 6= 0,(1 − exp
{−yσ
})κ, if ξ = 0.
ξ > 0 Ü too large return levels for long return periods e.g. 1000 to 10000 years(standard return period in dam construction)
ξ is set to 0 Þ powered exponential used as marginaldistribution.
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 7/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Marginal model
Extended GP-II : without threshold
F (y ;σ, ξ) =
(
1 −(
1 + ξyσ
)−1/ξ)κ
if ξ 6= 0,(1 − exp
{−yσ
})κ, if ξ = 0.
ξ > 0 Ü too large return levels for long return periods e.g. 1000 to 10000 years(standard return period in dam construction)
ξ is set to 0 Þ powered exponential used as marginaldistribution.
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 7/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Spatial dependance
Let the following transformation :H(x) = Φ−1(F (Y (x))), where Φ is the CDF of the normal N(0, 1).
Þ Since Y (x) > 0, H(x) > T (x) ≡ Φ−1(FY(x)(0))
Þ Then H(x) = max(T (x),G(x)),
G(x) the latent Gaussian process GP(0, 1,C(h))with marginal N(0, 1), C(h) the spatial covariance function.
ProcedureEstimate C(h) from the H-transformed data,
Simulate G and deduce H,
Retrieve Y by transforming back with Y (x) = F−1Y(x)(Φ(H(x))).
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 8/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Spatial dependance
Let the following transformation :H(x) = Φ−1(F (Y (x))), where Φ is the CDF of the normal N(0, 1).
Þ Since Y (x) > 0, H(x) > T (x) ≡ Φ−1(FY(x)(0))
Þ Then H(x) = max(T (x),G(x)),
G(x) the latent Gaussian process GP(0, 1,C(h))with marginal N(0, 1), C(h) the spatial covariance function.
ProcedureEstimate C(h) from the H-transformed data,
Simulate G and deduce H,
Retrieve Y by transforming back with Y (x) = F−1Y(x)(Φ(H(x))).
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 8/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance function
CS(h) = σ2exp(−hλ
), the stationary exponential covariance (isotropic) function
Geometrical anisotropy widely used in the literature
h(xi , xj) =√
(xi − xj)TΣ−1(xi − xj)
with Σ a covariance matrix, Σ−1 = MT M :
M =
(cosψ sinψ
−b sinψ b cosψ
)
with b the elongation coefficient, ψ ∈[−π2 ,
π2
]
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 9/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance function
CS(h) = σ2exp(−hλ
), the stationary exponential covariance (isotropic) function
Geometrical anisotropy widely used in the literature
h(xi , xj) =√
(xi − xj)TΣ−1(xi − xj)
with Σ a covariance matrix, Σ−1 = MT M :
M =
(cosψ sinψ
−b sinψ b cosψ
)
with b the elongation coefficient, ψ ∈[−π2 ,
π2
]
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 9/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance function
CS(h) = σ2exp(−hλ
), the stationary exponential covariance (isotropic) function
Geometrical anisotropy widely used in the literature
h(xi , xj) =√
(xi − xj)TΣ−1(xi − xj)
with Σ a covariance matrix, Σ−1 = MT M :
M =
(cosψ sinψ
−b sinψ b cosψ
)
with b the elongation coefficient, ψ ∈[−π2 ,
π2
]How to include topographical information in the
covariance ?
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 9/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance function
CS(h) = σ2exp(−hλ
), the stationary exponential covariance (isotropic) function
Geometrical anisotropy widely used in the literature
h(xi , xj) =√
(xi − xj)TΣ−1(xi − xj)
with Σ a covariance matrix, Σ−1 = MT M :
M =
(cosψ sinψ
−b sinψ b cosψ
)
with b the elongation coefficient, ψ ∈[−π2 ,
π2
]Non-stationarity introduced by Higdon et al. [1999] : kernel
convolution in anisotropic function.
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 9/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Spatial non-stationary covariance function
Generalized by Paciorek and Schervish [2006]
One solution for spatial non-stationarity :
CNS(xi , xj) = |Σi |14 |Σj |
14
∣∣∣∣Σi + Σj
2
∣∣∣∣− 12
exp
−
√(xi − xj)T
(Σi + Σj
2
)−1
(xi − xj)
Σi ,Σj : 2× 2 covariance matrix (local kernels, ∀i, j Σi = Σj ) Ü classical anisotropy∀xi ∈ D, Σ−1
i = MTi Mi :
Mi =1λi
(cosψi sinψi
−bi sinψi bi cosψi
)=
(Ai cosψi Ai sinψi
−Bi sinψi Bi cosψi
)
with Ai , Bi (> 0) et ψi as functions of the covariates (lon, lat, elev).
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 10/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Spatial non-stationary covariance function
Generalized by Paciorek and Schervish [2006]
One solution for spatial non-stationarity :
CNS(xi , xj) = |Σi |14 |Σj |
14
∣∣∣∣Σi + Σj
2
∣∣∣∣− 12
exp
−
√(xi − xj)T
(Σi + Σj
2
)−1
(xi − xj)
Σi ,Σj : 2× 2 covariance matrix (local kernels, ∀i, j Σi = Σj ) Ü classical anisotropy
∀xi ∈ D, Σ−1i = MT
i Mi :
Mi =1λi
(cosψi sinψi
−bi sinψi bi cosψi
)=
(Ai cosψi Ai sinψi
−Bi sinψi Bi cosψi
)
with Ai , Bi (> 0) et ψi as functions of the covariates (lon, lat, elev).
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 10/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Spatial non-stationary covariance function
Generalized by Paciorek and Schervish [2006]
One solution for spatial non-stationarity :
CNS(xi , xj) = |Σi |14 |Σj |
14
∣∣∣∣Σi + Σj
2
∣∣∣∣− 12
exp
−
√(xi − xj)T
(Σi + Σj
2
)−1
(xi − xj)
Σi ,Σj : 2× 2 covariance matrix (local kernels, ∀i, j Σi = Σj ) Ü classical anisotropy∀xi ∈ D, Σ−1
i = MTi Mi :
Mi =1λi
(cosψi sinψi
−bi sinψi bi cosψi
)=
(Ai cosψi Ai sinψi
−Bi sinψi Bi cosψi
)
with Ai , Bi (> 0) et ψi as functions of the covariates (lon, lat, elev).
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 10/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Outline
1 Context
2 Overview of the model
3 Parameter estimation
4 First evaluations
5 Conclusions and Perspectives
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 11/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
The estimation procedure is performed over six sub-set based on :
a rainy days classification in K = 3 weather types [WT,Garavaglia et al., 2010],
S = 2 seasons defined as low risk (LO dec. to aug.) and high risk (HI sep. to nov.)seasons.
Marginal distribution parameters : maximum likelihood,
Covariance function parameters : censored maximum likelihood [Pesonenet al., 2015].
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 11/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
The estimation procedure is performed over six sub-set based on :
a rainy days classification in K = 3 weather types [WT,Garavaglia et al., 2010],
S = 2 seasons defined as low risk (LO dec. to aug.) and high risk (HI sep. to nov.)seasons.
Marginal distribution parameters : maximum likelihood,
Covariance function parameters : censored maximum likelihood [Pesonenet al., 2015].
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 11/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance estimation and censored likelihood
gd(xi) not observed : only gd(xi) > Td(xi).
Let Zd = {xi |gd(xi)is censored} (i.e. stations without rainfall the day d),
Let Od = {xi |gd(xi)is observed}
β the parameters of C(h)
The censored likelihood given the {gd(xi)}xi∈Zd is Lc(β) =∏
d
Lcd(β) with :
1 Lcd(β) = fGP(gd(x1), . . . , gd(xN);β), if it rains at all stations
2 Lcd(β) = fGP({gd(xi)}xi∈Od ;β)
×P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd(xi)}xi∈Od ;β),if not
fGP are GP densities with covariance function C
P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd ,m(xi)}xi∈Od ;β) is a conditional GP CDF
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 12/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance estimation and censored likelihood
gd(xi) not observed : only gd(xi) > Td(xi).
Let Zd = {xi |gd(xi)is censored} (i.e. stations without rainfall the day d),
Let Od = {xi |gd(xi)is observed}
β the parameters of C(h)
The censored likelihood given the {gd(xi)}xi∈Zd is Lc(β) =∏
d
Lcd(β) with :
1 Lcd(β) = fGP(gd(x1), . . . , gd(xN);β), if it rains at all stations
2 Lcd(β) = fGP({gd(xi)}xi∈Od ;β)
×P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd(xi)}xi∈Od ;β),if not
fGP are GP densities with covariance function C
P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd ,m(xi)}xi∈Od ;β) is a conditional GP CDF
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 12/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance estimation and censored likelihood
gd(xi) not observed : only gd(xi) > Td(xi).
Let Zd = {xi |gd(xi)is censored} (i.e. stations without rainfall the day d),
Let Od = {xi |gd(xi)is observed}
β the parameters of C(h)
The censored likelihood given the {gd(xi)}xi∈Zd is Lc(β) =∏
d
Lcd(β) with :
1 Lcd(β) = fGP(gd(x1), . . . , gd(xN);β), if it rains at all stations
2 Lcd(β) = fGP({gd(xi)}xi∈Od ;β)
×P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd(xi)}xi∈Od ;β),if not
fGP are GP densities with covariance function C
P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd ,m(xi)}xi∈Od ;β) is a conditional GP CDF
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 12/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance estimation and censored likelihood
gd(xi) not observed : only gd(xi) > Td(xi).
Let Zd = {xi |gd(xi)is censored} (i.e. stations without rainfall the day d),
Let Od = {xi |gd(xi)is observed}
β the parameters of C(h)
The censored likelihood given the {gd(xi)}xi∈Zd is Lc(β) =∏
d
Lcd(β) with :
1 Lcd(β) = fGP(gd(x1), . . . , gd(xN);β), if it rains at all stations
2 Lcd(β) = fGP({gd(xi)}xi∈Od ;β)
×P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd(xi)}xi∈Od ;β),if not
fGP are GP densities with covariance function C
P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd ,m(xi)}xi∈Od ;β) is a conditional GP CDF
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 12/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance estimation and censored likelihood
gd(xi) not observed : only gd(xi) > Td(xi).
Let Zd = {xi |gd(xi)is censored} (i.e. stations without rainfall the day d),
Let Od = {xi |gd(xi)is observed}
β the parameters of C(h)
The censored likelihood given the {gd(xi)}xi∈Zd is Lc(β) =∏
d
Lcd(β) with :
1 Lcd(β) = fGP(gd(x1), . . . , gd(xN);β), if it rains at all stations
2 Lcd(β) = fGP({gd(xi)}xi∈Od ;β)
×P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd(xi)}xi∈Od ;β),if not
fGP are GP densities with covariance function C
P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd ,m(xi)}xi∈Od ;β) is a conditional GP CDF
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 12/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Covariance estimation and censored likelihood
gd(xi) not observed : only gd(xi) > Td(xi).
Let Zd = {xi |gd(xi)is censored} (i.e. stations without rainfall the day d),
Let Od = {xi |gd(xi)is observed}
β the parameters of C(h)
The censored likelihood given the {gd(xi)}xi∈Zd is Lc(β) =∏
d
Lcd(β) with :
1 Lcd(β) = fGP(gd(x1), . . . , gd(xN);β), if it rains at all stations
2 Lcd(β) = fGP({gd(xi)}xi∈Od ;β)
×P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd(xi)}xi∈Od ;β),if not
fGP are GP densities with covariance function C
P({Gd(xi) 6 Td(xi)}xi∈Zd |{Gd(xi) = gd ,m(xi)}xi∈Od ;β) is a conditional GP CDF
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 12/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Outline
1 Context
2 Overview of the model
3 Parameter estimation
4 First evaluations
5 Conclusions and Perspectives
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 13/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
First evaluations
Calibration over the whole period with available data (1948-2013),
100 simulations over the period,
The simulations are only conditioned by the season and weather type the daybelongs to.
Evaluation of the marginal and spatial dependance properties. In particular, theadded-value of NS-covariance is explored. To this end three different types ofexponential covariance function are tested :
isotropic (ISO)anisotropic (AN)non-stationary (NSAN)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 13/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
First evaluations
Calibration over the whole period with available data (1948-2013),
100 simulations over the period,
The simulations are only conditioned by the season and weather type the daybelongs to.
Evaluation of the marginal and spatial dependance properties. In particular, theadded-value of NS-covariance is explored. To this end three different types ofexponential covariance function are tested :
isotropic (ISO)anisotropic (AN)non-stationary (NSAN)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 13/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Mean rainfall & occurrence frequency (LO WT2)
●●
●
●
●●
●
●
●●●
●
●
● ●●
●●
●● ●
●
●●
●
●●● ●
●
● ●
●●
● ●
●
●
4 6 8 10 12
Mean
NSAN● ●
●
●
●●●
●
●
●●
●●●
●
● ●
●
●●
● ●●
10 15 20 25 30
Mean>0
●● ●
●
● ●
●
●
●●● ●
● ●
●●
●●
●●●
●●
●
●●
●
●
●●
● ●●
● ●
●●
●
●
●
●●
●
●●
●
●
●●
30 35 40 45 50 55
P1
0
200
400
600
800
1000
1200
1400
Elevation (m)
●
●
●●
● ●
●●●
● ●
●
4 6 8 10 12
AN●● ●
●
●
●
●
● ●
●
●
●●
10 15 20 25 30
●●
●●
● ●●
●
●
●
●
●
● ●●●● ●
●●
●●
●
● ●
●● ●
● ●●
●●
●
●
●
●● ● ●
●●
●
30 35 40 45 50 55
●●●●
●
●
●●
●
●
●
4 6 8 10 12
ISOPR (mm/day)
●
●
●
●
●
● ●
●
● ●●
●● ●
● ●
●
●●●
●●
●
10 15 20 25 30
PR (mm/day)●
●
●
● ●
●● ●
●
●
●
●●
●●
●
●● ●●
● ●
●
● ●●
●
●
●
●
● ● ●●●
●
30 35 40 45 50 55
%
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 14/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Interannual variabilityLO WT2
020
040
060
080
010
00
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: NSAN
MeanQ5−Q95OBS
RMSE = 146.66 mmCOR = 0.49
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: AN
MeanQ5−Q95OBS
RMSE = 147.44 mmCOR = 0.47
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: ISO
MeanQ5−Q95OBS
RMSE = 146.08 mmCOR = 0.48
020
040
060
080
010
00
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: NSAN
MeanQ5−Q95OBS
RMSE = 146.66 mmCOR = 0.49
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: AN
MeanQ5−Q95OBS
RMSE = 147.44 mmCOR = 0.47
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: ISO
MeanQ5−Q95OBS
RMSE = 146.08 mmCOR = 0.48
020
040
060
080
010
00
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: NSAN
MeanQ5−Q95OBS
RMSE = 146.66 mmCOR = 0.49
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: AN
MeanQ5−Q95OBS
RMSE = 147.44 mmCOR = 0.47
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
LO GT2 cov mod: ISO
MeanQ5−Q95OBS
RMSE = 146.08 mmCOR = 0.48
NSAN
AN
ISO
HI WT2
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: NSAN
MeanQ5−Q95OBS
RMSE = 138.21 mmCOR = 0.45
020
040
060
080
0Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: AN
MeanQ5−Q95OBS
RMSE = 137.45 mmCOR = 0.45
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: ISO
MeanQ5−Q95OBS
RMSE = 137.51 mmCOR = 0.45
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: NSAN
MeanQ5−Q95OBS
RMSE = 138.21 mmCOR = 0.45
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: AN
MeanQ5−Q95OBS
RMSE = 137.45 mmCOR = 0.45
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: ISO
MeanQ5−Q95OBS
RMSE = 137.51 mmCOR = 0.45
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: NSAN
MeanQ5−Q95OBS
RMSE = 138.21 mmCOR = 0.45
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: AN
MeanQ5−Q95OBS
RMSE = 137.45 mmCOR = 0.45
020
040
060
080
0
Years
PR
(m
m/y
ear)
1950 1956 1962 1968 1974 1980 1986 1992 1998 2004 2010
HI GT2 cov mod: ISO
MeanQ5−Q95OBS
RMSE = 137.51 mmCOR = 0.45
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 15/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Log-odds ratio
The log-odds ratio can be seen as analogous to spatial correlation of continuousvariable for binary variables [Charles et al., 1999]
TABLE 1: Contingency table for two stations xi, xj
NR(xi) R(xi)
NR(xj) n00 n01
R(xj) n10 n11
The log-odds ratio is given by :
LOR = log(
n00n11
n01n10
)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 16/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Log-odds ratioLO WT2
−
−
−
−−
− −
−
−−
−
−
−−−−−
−
−−−−
−−
−−−−−−−−−−−−− −−−
−
−−−
−
− −
−
−−−
−−− −
−
−
−
−−−
−−−−−−−−
−
−− −−−−−−−−
−−−
−− −
−−−
−
−
−−−
−
−−
−−
−
−
−−−− − −− −−−−−
− −−−−−−−
− −
−−−
−
−
−−−− −−
−−
−−
−−−−−−−−
−−−−−
−− −−−
−−
−−
−
−−
−−
−−− −−− −
−
−−
−−−−−
−−−−−−−−−− −
−
−−−
−−−−−
−
−
−−
−−−−−−−
−−−−
−−−−−−
−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−−−
−−−−−− −−−
−
−
−−−−−−
−
−−−
−−−−−−−
−−−−−−
−−−−−−−−
−
−−−− −−
−
−−−−−
−−−−−−
−−
−−−−
−−−
−−−−
−−
−−−−−− −
−
−−−
−−−−
−−−−−−−−
−−−−−−−−
−−−−−− −
−
−−
−−−−−−−−−−−− −−−−−−−−
−−−−−−
−−−−
−
−−−−− −−−−−−
−−−
−−−
−
−−−−−−
−
−−−
−−−−−−−−
−−−−−
−−−−−−
−−−−−
−−
−−
−−−−−−−−
−−−−−−−
−−−−−−−
−−−
−−
−−−−−−−−
−−−−
−−−−−−−
−
−−
−−−
−
−−−−−− −−
−−−−−−−−−−
−−
−
−−−
−
−−−−−−−−−−
−−
−−−−−−−
−−−
−−
−−−−−−−−−−−−−
−−−−−−−
−−−
−−−−− −−
−
−−
−−
−− −−−−−
−−−
−−−− −−−−
−−−−−
−−−−−− −−
−−−−−
−−−−−−− −−−−−−−
−
−−−−−−−−
−−−−
−−−−−−−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
−−−− −−−
−−−−−
−−−−
−−−−− −−−−
−−−−
−− −− −−−−−−−
−−−−−− −−−−−−−
−−−−−−−−−−
−−−−−
−−−−−−−−−−
−− −−−−
−−−−−−−−−
− −−−−−−−
−
−−−−−−
−−−−−− − −−−−−
−−−
−−
−−−−−−−−
−−
0 2 4 6 8
02
46
8LO GT2 cov mod: NSAN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−
−−
−
−
−−−−−−
−−
−−
−−−−−−−−
−− −−−−− −−−
−−
−−
−
− −
−
−−−
−−− −
−−
−
−−−−
−−−−−−−
−
−− −−
−−−−−−−−
−−
− −
−−−
−
−
−−−
−−−
−−−
−
−−−− − −− −−−−−
− −−−−−−−
− −
−−−
−
−
−−−−
−−−
−
−−
−−−−−−−−
−−−− −−−
−−−−
−
−−−
−−
−−
−−− −−−
−−
−−
−−−−−−−−−−−−−−− −
−
−−−
−−−− −−
−
−−−−−
−−−−
−−−−−−−−−
−−
−−
−
−
−−−−−−
−−
−−−−− − −
− −−−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−−
−−−−−−−
−−−−− −−−−−−−
−−
−
−−−− −−
−
−−−−−
−−−−−−
−−−−− −−−−
−−−−
−−
−−−−−− −
−
−−−−−−
−−−−
−−−− −−−−−−−−−
−−−−−−−
−
−−
−
−−−−−−−−−−− −
−−−−−−−−−−−−−
−−−−
−
−−−−− −
−−−−− −−−
−−−
−
−−−−−−
−−
−−−−−−−−−−
−−−− −−−−−−−
−−−−−
−−
−−
−−−−−−−−
−−−−−−−−−−−
−−
−
−
−−
−−
−−−−−−−−
−−−−−−−−−−−
−
−−
−−−
−
−−−−−− −−
−−−−−
−−−−−−−−
−−−
−
−−−−−−−−−−
−−
−−−−−−
−−−−
−−
−−−−−−−−
−−−−− −−−−−−
−−−−
−−−−−−−
−
−− −−
−− −−−−
−−
−−−−−− −−−
−−−−− −
−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−− −−−−−−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
−−−− −−−
−−
−−−−
−−−−−−−− −
−−−−−−−
−− −− −−−−−−−
−−−−−− −−−−−−−−−−
−−−−−−−−−−
−−
−−−−−−−−−−−
− −−−−
−−−−
−−−−−− −−
−−−−−
−
−−−−−−
−−−−−− − −−−−−
−−−
−−
−−−−
−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: AN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−−
−
− −
−−−
−
−
−−−−−−
−−
−−
−−−−−−−−
−− −−−−−−−−
−−
−−
−
− −
−−−−
−−− −
−−
−
− −−−−
−−−−−−
−
−− −−
−−−−−−−−
−−
− −−
−−
−−
−−−
−−−
−−−
−
−−−− − −− −−−−−
−−−−−−−
−− −
−−−
−
−−
−
−−−
−−
−
−−
−−−−−−−−
−−−− −−−
−−−−
−
−−−
−−
−−
−−− −−−−−
−−
−−−−−−−−−
−−−−−− −
−
−−
−−−−− −−
−
−−−−
−
−−−
−−
−−−−−−−−−
−
−−
−
−
−−−−−−
−−
−−−−− − −
− −−−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−− −−−−
−−−−
−−−− −−−−−−−−−
−
−−−− −−
−−
−−−−
−−−−−−
−−−−− −−−−
−−−−
−−
−−−−−− −
−
−−−−−−
−−−−
−−−− −−−−−−−−−
−−−−−−−
−
−−
−
−−−−−−−−−−− −
−−−−−−−−−−−−−
−−−−
−
−−−−− −
−−−−− −−−
−−−
−
−−−−−−
−−
−−−−−−−−−−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−−−−
−−−−−−−−−−−
−−
−
−−−
−−
−−−−−−−−
−−−−−−−−−−− −
−−−−−
−
−−−−−− −−
−−−−−
−−−−−−−−
−−−
−
−−−−−−−−
−−
−−−−−−−−
−−−−
−−
−−−−−−−−
−−−−− −−−−−−
−−−−
−−−−−−−
−
−− −−
−− −−−−
−−
−− −−−− −−−−
−−−− −−−−−−− −−−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−− −−−−−
−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
− −−− −−−
−−
−−−−−
−−−−−−− −
−−−−−−−
−− −− −
−−−−−−−−−−−− −
−−−−−−−−−
−−−−−−−−−−−−−−−−−−−−−− −
− −−−−
−−−−
−−−−−− −−
−−−−−
−−
−−−−−−
−−−−− − −−−−−
−−
−−−
−−−−
−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: ISO
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−
−
−
−−
− −
−
−−
−
−
−−−−−
−
−−−−
−−
−−−−−−−−−−−−− −−−
−
−−−
−
− −
−
−−−
−−− −
−
−
−
−−−
−−−−−−−−
−
−− −−−−−−−−
−−−
−− −
−−−
−
−
−−−
−
−−
−−
−
−
−−−− − −− −−−−−
− −−−−−−−
− −
−−−
−
−
−−−− −−
−−
−−
−−−−−−−−
−−−−−
−− −−−
−−
−−
−
−−
−−
−−− −−− −
−
−−
−−−−−
−−−−−−−−−− −
−
−−−
−−−−−
−
−
−−
−−−−−−−
−−−−
−−−−−−
−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−−−
−−−−−− −−−
−
−
−−−−−−
−
−−−
−−−−−−−
−−−−−−
−−−−−−−−
−
−−−− −−
−
−−−−−
−−−−−−
−−
−−−−
−−−
−−−−
−−
−−−−−− −
−
−−−
−−−−
−−−−−−−−
−−−−−−−−
−−−−−− −
−
−−
−−−−−−−−−−−− −−−−−−−−
−−−−−−
−−−−
−
−−−−− −−−−−−
−−−
−−−
−
−−−−−−
−
−−−
−−−−−−−−
−−−−−
−−−−−−
−−−−−
−−
−−
−−−−−−−−
−−−−−−−
−−−−−−−
−−−
−−
−−−−−−−−
−−−−
−−−−−−−
−
−−
−−−
−
−−−−−− −−
−−−−−−−−−−
−−
−
−−−
−
−−−−−−−−−−
−−
−−−−−−−
−−−
−−
−−−−−−−−−−−−−
−−−−−−−
−−−
−−−−− −−
−
−−
−−
−− −−−−−
−−−
−−−− −−−−
−−−−−
−−−−−− −−
−−−−−
−−−−−−− −−−−−−−
−
−−−−−−−−
−−−−
−−−−−−−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
−−−− −−−
−−−−−
−−−−
−−−−− −−−−
−−−−
−− −− −−−−−−−
−−−−−− −−−−−−−
−−−−−−−−−−
−−−−−
−−−−−−−−−−
−− −−−−
−−−−−−−−−
− −−−−−−−
−
−−−−−−
−−−−−− − −−−−−
−−−
−−
−−−−−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: NSAN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−
−−
−
−
−−−−−−
−−
−−
−−−−−−−−
−− −−−−− −−−
−−
−−
−
− −
−
−−−
−−− −
−−
−
−−−−
−−−−−−−
−
−− −−
−−−−−−−−
−−
− −
−−−
−
−
−−−
−−−
−−−
−
−−−− − −− −−−−−
− −−−−−−−
− −
−−−
−
−
−−−−
−−−
−
−−
−−−−−−−−
−−−− −−−
−−−−
−
−−−
−−
−−
−−− −−−
−−
−−
−−−−−−−−−−−−−−− −
−
−−−
−−−− −−
−
−−−−−
−−−−
−−−−−−−−−
−−
−−
−
−
−−−−−−
−−
−−−−− − −
− −−−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−−
−−−−−−−
−−−−− −−−−−−−
−−
−
−−−− −−
−
−−−−−
−−−−−−
−−−−− −−−−
−−−−
−−
−−−−−− −
−
−−−−−−
−−−−
−−−− −−−−−−−−−
−−−−−−−
−
−−
−
−−−−−−−−−−− −
−−−−−−−−−−−−−
−−−−
−
−−−−− −
−−−−− −−−
−−−
−
−−−−−−
−−
−−−−−−−−−−
−−−− −−−−−−−
−−−−−
−−
−−
−−−−−−−−
−−−−−−−−−−−
−−
−
−
−−
−−
−−−−−−−−
−−−−−−−−−−−
−
−−
−−−
−
−−−−−− −−
−−−−−
−−−−−−−−
−−−
−
−−−−−−−−−−
−−
−−−−−−
−−−−
−−
−−−−−−−−
−−−−− −−−−−−
−−−−
−−−−−−−
−
−− −−
−− −−−−
−−
−−−−−− −−−
−−−−− −
−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−− −−−−−−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
−−−− −−−
−−
−−−−
−−−−−−−− −
−−−−−−−
−− −− −−−−−−−
−−−−−− −−−−−−−−−−
−−−−−−−−−−
−−
−−−−−−−−−−−
− −−−−
−−−−
−−−−−− −−
−−−−−
−
−−−−−−
−−−−−− − −−−−−
−−−
−−
−−−−
−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: AN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−−
−
− −
−−−
−
−
−−−−−−
−−
−−
−−−−−−−−
−− −−−−−−−−
−−
−−
−
− −
−−−−
−−− −
−−
−
− −−−−
−−−−−−
−
−− −−
−−−−−−−−
−−
− −−
−−
−−
−−−
−−−
−−−
−
−−−− − −− −−−−−
−−−−−−−
−− −
−−−
−
−−
−
−−−
−−
−
−−
−−−−−−−−
−−−− −−−
−−−−
−
−−−
−−
−−
−−− −−−−−
−−
−−−−−−−−−
−−−−−− −
−
−−
−−−−− −−
−
−−−−
−
−−−
−−
−−−−−−−−−
−
−−
−
−
−−−−−−
−−
−−−−− − −
− −−−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−− −−−−
−−−−
−−−− −−−−−−−−−
−
−−−− −−
−−
−−−−
−−−−−−
−−−−− −−−−
−−−−
−−
−−−−−− −
−
−−−−−−
−−−−
−−−− −−−−−−−−−
−−−−−−−
−
−−
−
−−−−−−−−−−− −
−−−−−−−−−−−−−
−−−−
−
−−−−− −
−−−−− −−−
−−−
−
−−−−−−
−−
−−−−−−−−−−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−−−−
−−−−−−−−−−−
−−
−
−−−
−−
−−−−−−−−
−−−−−−−−−−− −
−−−−−
−
−−−−−− −−
−−−−−
−−−−−−−−
−−−
−
−−−−−−−−
−−
−−−−−−−−
−−−−
−−
−−−−−−−−
−−−−− −−−−−−
−−−−
−−−−−−−
−
−− −−
−− −−−−
−−
−− −−−− −−−−
−−−− −−−−−−− −−−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−− −−−−−
−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
− −−− −−−
−−
−−−−−
−−−−−−− −
−−−−−−−
−− −− −
−−−−−−−−−−−− −
−−−−−−−−−
−−−−−−−−−−−−−−−−−−−−−− −
− −−−−
−−−−
−−−−−− −−
−−−−−
−−
−−−−−−
−−−−− − −−−−−
−−
−−−
−−−−
−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: ISO
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−
−
−
−−
− −
−
−−
−
−
−−−−−
−
−−−−
−−
−−−−−−−−−−−−− −−−
−
−−−
−
− −
−
−−−
−−− −
−
−
−
−−−
−−−−−−−−
−
−− −−−−−−−−
−−−
−− −
−−−
−
−
−−−
−
−−
−−
−
−
−−−− − −− −−−−−
− −−−−−−−
− −
−−−
−
−
−−−− −−
−−
−−
−−−−−−−−
−−−−−
−− −−−
−−
−−
−
−−
−−
−−− −−− −
−
−−
−−−−−
−−−−−−−−−− −
−
−−−
−−−−−
−
−
−−
−−−−−−−
−−−−
−−−−−−
−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−−−
−−−−−− −−−
−
−
−−−−−−
−
−−−
−−−−−−−
−−−−−−
−−−−−−−−
−
−−−− −−
−
−−−−−
−−−−−−
−−
−−−−
−−−
−−−−
−−
−−−−−− −
−
−−−
−−−−
−−−−−−−−
−−−−−−−−
−−−−−− −
−
−−
−−−−−−−−−−−− −−−−−−−−
−−−−−−
−−−−
−
−−−−− −−−−−−
−−−
−−−
−
−−−−−−
−
−−−
−−−−−−−−
−−−−−
−−−−−−
−−−−−
−−
−−
−−−−−−−−
−−−−−−−
−−−−−−−
−−−
−−
−−−−−−−−
−−−−
−−−−−−−
−
−−
−−−
−
−−−−−− −−
−−−−−−−−−−
−−
−
−−−
−
−−−−−−−−−−
−−
−−−−−−−
−−−
−−
−−−−−−−−−−−−−
−−−−−−−
−−−
−−−−− −−
−
−−
−−
−− −−−−−
−−−
−−−− −−−−
−−−−−
−−−−−− −−
−−−−−
−−−−−−− −−−−−−−
−
−−−−−−−−
−−−−
−−−−−−−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
−−−− −−−
−−−−−
−−−−
−−−−− −−−−
−−−−
−− −− −−−−−−−
−−−−−− −−−−−−−
−−−−−−−−−−
−−−−−
−−−−−−−−−−
−− −−−−
−−−−−−−−−
− −−−−−−−
−
−−−−−−
−−−−−− − −−−−−
−−−
−−
−−−−−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: NSAN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−
−−
−
−
−−−−−−
−−
−−
−−−−−−−−
−− −−−−− −−−
−−
−−
−
− −
−
−−−
−−− −
−−
−
−−−−
−−−−−−−
−
−− −−
−−−−−−−−
−−
− −
−−−
−
−
−−−
−−−
−−−
−
−−−− − −− −−−−−
− −−−−−−−
− −
−−−
−
−
−−−−
−−−
−
−−
−−−−−−−−
−−−− −−−
−−−−
−
−−−
−−
−−
−−− −−−
−−
−−
−−−−−−−−−−−−−−− −
−
−−−
−−−− −−
−
−−−−−
−−−−
−−−−−−−−−
−−
−−
−
−
−−−−−−
−−
−−−−− − −
− −−−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−−
−−−−−−−
−−−−− −−−−−−−
−−
−
−−−− −−
−
−−−−−
−−−−−−
−−−−− −−−−
−−−−
−−
−−−−−− −
−
−−−−−−
−−−−
−−−− −−−−−−−−−
−−−−−−−
−
−−
−
−−−−−−−−−−− −
−−−−−−−−−−−−−
−−−−
−
−−−−− −
−−−−− −−−
−−−
−
−−−−−−
−−
−−−−−−−−−−
−−−− −−−−−−−
−−−−−
−−
−−
−−−−−−−−
−−−−−−−−−−−
−−
−
−
−−
−−
−−−−−−−−
−−−−−−−−−−−
−
−−
−−−
−
−−−−−− −−
−−−−−
−−−−−−−−
−−−
−
−−−−−−−−−−
−−
−−−−−−
−−−−
−−
−−−−−−−−
−−−−− −−−−−−
−−−−
−−−−−−−
−
−− −−
−− −−−−
−−
−−−−−− −−−
−−−−− −
−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−− −−−−−−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
−−−− −−−
−−
−−−−
−−−−−−−− −
−−−−−−−
−− −− −−−−−−−
−−−−−− −−−−−−−−−−
−−−−−−−−−−
−−
−−−−−−−−−−−
− −−−−
−−−−
−−−−−− −−
−−−−−
−
−−−−−−
−−−−−− − −−−−−
−−−
−−
−−−−
−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: AN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−−
−
− −
−−−
−
−
−−−−−−
−−
−−
−−−−−−−−
−− −−−−−−−−
−−
−−
−
− −
−−−−
−−− −
−−
−
− −−−−
−−−−−−
−
−− −−
−−−−−−−−
−−
− −−
−−
−−
−−−
−−−
−−−
−
−−−− − −− −−−−−
−−−−−−−
−− −
−−−
−
−−
−
−−−
−−
−
−−
−−−−−−−−
−−−− −−−
−−−−
−
−−−
−−
−−
−−− −−−−−
−−
−−−−−−−−−
−−−−−− −
−
−−
−−−−− −−
−
−−−−
−
−−−
−−
−−−−−−−−−
−
−−
−
−
−−−−−−
−−
−−−−− − −
− −−−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−− −−−−
−−−−
−−−− −−−−−−−−−
−
−−−− −−
−−
−−−−
−−−−−−
−−−−− −−−−
−−−−
−−
−−−−−− −
−
−−−−−−
−−−−
−−−− −−−−−−−−−
−−−−−−−
−
−−
−
−−−−−−−−−−− −
−−−−−−−−−−−−−
−−−−
−
−−−−− −
−−−−− −−−
−−−
−
−−−−−−
−−
−−−−−−−−−−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−−−−
−−−−−−−−−−−
−−
−
−−−
−−
−−−−−−−−
−−−−−−−−−−− −
−−−−−
−
−−−−−− −−
−−−−−
−−−−−−−−
−−−
−
−−−−−−−−
−−
−−−−−−−−
−−−−
−−
−−−−−−−−
−−−−− −−−−−−
−−−−
−−−−−−−
−
−− −−
−− −−−−
−−
−− −−−− −−−−
−−−− −−−−−−− −−−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−− −−−−−
−
−−−−−−−−
−−−−−−−−−−−
−−−
−−−−−−−− −−−−−−−
−−−− −
− −−− −−−
−−
−−−−−
−−−−−−− −
−−−−−−−
−− −− −
−−−−−−−−−−−− −
−−−−−−−−−
−−−−−−−−−−−−−−−−−−−−−− −
− −−−−
−−−−
−−−−−− −−
−−−−−
−−
−−−−−−
−−−−− − −−−−−
−−
−−−
−−−−
−−−−
−−
0 2 4 6 8
02
46
8
LO GT2 cov mod: ISO
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
NSAN
AN
ISO
HI WT2
−
−−
−−
− −
−
−−
−
−
−−−
−−
−
−−
−−
−−
−−−− −−
−−−−
−−−−−−
−
−−−
−
− −
−
−−−
−−−−
−−
−
−−−−−
−−− −− −
−
−−−−
−−−−−−−−
−−
− −
−
−−
−
−
−− −
−
−−
−−
−
−
−−−− −− −
−−−−−
−−−−−−
−−
− −
−−−
−
−
−−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−
−−
−−
−
−−
−
−
−−− −−−
−−
−−
−−−
− −−−− −−−−−−− −
−
−−−
−−−−
−−
−
−
−−−−−−
−
−−−−−−−−−
−
−
−−
−
−
−−−−−−
−−
−−−−− − −− −
−
−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−−
−−−− −−−
−−−−− −
−− −−−−−−
−
−−−− −−
−
−−−−−
−
−−−− −
−
−−−−
−−− −
−−−−
−−
−−− −−
− −−
−−
−−−−
−−−−
−−−− −−−−−−− −−
−−−−−− −
−
−−
−
−−− −− −−−−−− −−−−−−− −
−−− −−
−
−−−−
−
−−−−− −
−−−
−−−−−−−−
−
−−−−−−
−
−−−
−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−−
−−−−−− −−
−−−−−−
−−−−−−−−
−
−−
−−
−−−−−− −−
−−−−
−−−−− −−
−−−
−−
−−
−−−−−− −−
−−−−−−−−−−−
−−
−−
−−
−−−−−− −
−−−
−−
−−−−−−−
−−−
−−
− −−−−− −−−−−−−−
−−−−−−
−−−−
−−−−− −
−
−−
−−
−−−−−−−
−−−
−−−− −−−−
−−−− −−−−−−−−−
−−−−−
−−−−−
−−−−−−−−−
−
−−−−−−−−
−−−−
−−−−−−−
−−−−−− −−
−−−−
−−−−− −
−−−
−−
− −−−−−− −−−−−−−
−−−−
−
−−−−
−
−−−
−−−−
−− −−
−−−− −−−−−−−
−−− −
−−−− −−−−
−−−−−−−− −−−−
−−−
−−−−−−−− −−−
−−−−−−−−−−−−
−−−−−
−−− −−−−−− −−−
−−−−−
−
−−−−−−−
−−−−−−−−−−−
−−−
−−
−−−−−− −
−−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: NSAN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−
−−
−
−
−−−
−−−
−−
−−
−−
−−−− −−
−−−−
−−−−−−
−
−−−
−
− −
−
−−
−
−−−−
−−
−
−−−−
−−
−− −− −
−
−−−−−
−−−−−−
−−
−
− −
−−−
−
−
−− −
−−
−
−−
−
−
−−−− −− −−
−−−−
−−−−−−−−
− −
−
−−
−
−
−−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−−
−
−−
−
−−
−−
−−− −−−
−−
−−
−−−− −
−−− −−−−−−− −
−
−−−
−−−− −−
−
−−−−−
−−−
−−−−−−−−− −
−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−− −−−−−−− −−−
−
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−− −−−−−−
−
−−−− −−
−
−−−−−
−
− −−− −−
−−−− −−−−
−−−−
−−
−−− −−− −
−
−−
−−−−
−−−−−−−− −−−−−−− −−
−−−−−− −
−
−−
−
−−− −− −−−−−− −
−−−−−− −−−− −− −
−−−
−−
−−−−− −
−−−−
−−−−−
−−
−
−−−−−−
−−
−−−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−− −−
−−−−
−−−−−−−−−−
−
−−
−−
−−−−−− −−
−−−−−−−−− −−
−−−
−−−−
−−−−−− −−
−−−−−−−−−−−−−
−−−−
−−−−−− −−
−−
−−
−−−−−−
−−−−
−−
−−−−−− −−
−−−−−−−−−−−−
−−−
−−−−−
− −−
−− −
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−−−−
−−−−
−−−−−− −−
−−−−
−−−−− −−
−−−−
− −−−−−− −−−−−−−
−−−−
−−−−−−
−−
−
−−−−
−− −− −−−− −
−−−−−−−−− −−−
−− −−−−−−−−−−
−− −−−−−−−
−−−−−−−− −−−
−−−−−−−−−−−−
−−−−−
−−
− −
−−−−−−−−
−−−−−
−
−−−−−−
−−−−−−
−−−−−−
−−−
−−
−−−−−−
−−
−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: AN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−−−
−
−
−−− −−
−
− −
−−
−−−−−− −
−−−−
−−−−−−−
−−
−−
−
− −
−−−−
−−−− −−
−
− −−−−
−−− −− −
−
−−−−−
−−−−−−
−−
−− −
−−−
−−
−− −
−−
−
−−−
−
−−−− −− −−
−−−−
−−−−−−
−−
− −
−−−
−
−−
−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−−
−
−−−
−−
−−
−−− −−−−
−
−−
−−−− −
−−− −−−−−−− −
−
−−−
−−−− −−−
−−−−
−−−
−−
−−−−−−−− −−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−− −−−−−−− −−
−−
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−− −−−−−−
−
−−−− −−
−−
−−−−
−
− −−− −−
−−−− −−−−
−−−−
−−
−−− −−− −
−
−−−−−
−−
−−−−−−− −−−−−−− −−
−−−−−−−−
−−
−
−−− −
− −−−−−− −−−−−−− −
−−− −− −−
−−
−−
−−−−− −
−−−−−
−−−−
−−
−
−−−−−−
−−
−−−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−− −−
−−−
−
−−−−−−−−−
−
−
−−
−−
−−−−−− −−
−−−−−−−−− −− −
−−−−
−−
−−−−−− −−
−−−−−−−−−−−−−
−−−−
−−−−−
− −−−−
−−
−−−−−−
−−−−
−−
−−−−−− −−
−−−−−−−−−−−− −
−−
−−−−−
− −−
−− −
−
−−−−−−
−
−−− −−−− −−−
−−−−− −−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−−−−
−−−−
−−−−−− −−
−−−−
−−−−− −−
−−−−
− −−−−−− −−−−−−−
−−−−
− −−−−−−−
−
−−−−
−− −− −−−− −
−−−−−−−−− −−−
−− −−−−−−−−−−
−− −−−−−−−
−−−−−−−− −−− −−−−−−−−−−−−
−−−−−
−−
− −
−−−−−−−−
−−−−−
−−
−−−−−−
−−−−− −−−−−−
−−
−−−
−−−−−−
−−
−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: ISO
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−
−−
−−
− −
−
−−
−
−
−−−
−−
−
−−
−−
−−
−−−− −−
−−−−
−−−−−−
−
−−−
−
− −
−
−−−
−−−−
−−
−
−−−−−
−−− −− −
−
−−−−
−−−−−−−−
−−
− −
−
−−
−
−
−− −
−
−−
−−
−
−
−−−− −− −
−−−−−
−−−−−−
−−
− −
−−−
−
−
−−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−
−−
−−
−
−−
−
−
−−− −−−
−−
−−
−−−
− −−−− −−−−−−− −
−
−−−
−−−−
−−
−
−
−−−−−−
−
−−−−−−−−−
−
−
−−
−
−
−−−−−−
−−
−−−−− − −− −
−
−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−−
−−−− −−−
−−−−− −
−− −−−−−−
−
−−−− −−
−
−−−−−
−
−−−− −
−
−−−−
−−− −
−−−−
−−
−−− −−
− −−
−−
−−−−
−−−−
−−−− −−−−−−− −−
−−−−−− −
−
−−
−
−−− −− −−−−−− −−−−−−− −
−−− −−
−
−−−−
−
−−−−− −
−−−
−−−−−−−−
−
−−−−−−
−
−−−
−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−−
−−−−−− −−
−−−−−−
−−−−−−−−
−
−−
−−
−−−−−− −−
−−−−
−−−−− −−
−−−
−−
−−
−−−−−− −−
−−−−−−−−−−−
−−
−−
−−
−−−−−− −
−−−
−−
−−−−−−−
−−−
−−
− −−−−− −−−−−−−−
−−−−−−
−−−−
−−−−− −
−
−−
−−
−−−−−−−
−−−
−−−− −−−−
−−−− −−−−−−−−−
−−−−−
−−−−−
−−−−−−−−−
−
−−−−−−−−
−−−−
−−−−−−−
−−−−−− −−
−−−−
−−−−− −
−−−
−−
− −−−−−− −−−−−−−
−−−−
−
−−−−
−
−−−
−−−−
−− −−
−−−− −−−−−−−
−−− −
−−−− −−−−
−−−−−−−− −−−−
−−−
−−−−−−−− −−−
−−−−−−−−−−−−
−−−−−
−−− −−−−−− −−−
−−−−−
−
−−−−−−−
−−−−−−−−−−−
−−−
−−
−−−−−− −
−−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: NSAN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−
−−
−
−
−−−
−−−
−−
−−
−−
−−−− −−
−−−−
−−−−−−
−
−−−
−
− −
−
−−
−
−−−−
−−
−
−−−−
−−
−− −− −
−
−−−−−
−−−−−−
−−
−
− −
−−−
−
−
−− −
−−
−
−−
−
−
−−−− −− −−
−−−−
−−−−−−−−
− −
−
−−
−
−
−−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−−
−
−−
−
−−
−−
−−− −−−
−−
−−
−−−− −
−−− −−−−−−− −
−
−−−
−−−− −−
−
−−−−−
−−−
−−−−−−−−− −
−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−− −−−−−−− −−−
−
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−− −−−−−−
−
−−−− −−
−
−−−−−
−
− −−− −−
−−−− −−−−
−−−−
−−
−−− −−− −
−
−−
−−−−
−−−−−−−− −−−−−−− −−
−−−−−− −
−
−−
−
−−− −− −−−−−− −
−−−−−− −−−− −− −
−−−
−−
−−−−− −
−−−−
−−−−−
−−
−
−−−−−−
−−
−−−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−− −−
−−−−
−−−−−−−−−−
−
−−
−−
−−−−−− −−
−−−−−−−−− −−
−−−
−−−−
−−−−−− −−
−−−−−−−−−−−−−
−−−−
−−−−−− −−
−−
−−
−−−−−−
−−−−
−−
−−−−−− −−
−−−−−−−−−−−−
−−−
−−−−−
− −−
−− −
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−−−−
−−−−
−−−−−− −−
−−−−
−−−−− −−
−−−−
− −−−−−− −−−−−−−
−−−−
−−−−−−
−−
−
−−−−
−− −− −−−− −
−−−−−−−−− −−−
−− −−−−−−−−−−
−− −−−−−−−
−−−−−−−− −−−
−−−−−−−−−−−−
−−−−−
−−
− −
−−−−−−−−
−−−−−
−
−−−−−−
−−−−−−
−−−−−−
−−−
−−
−−−−−−
−−
−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: AN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−−−
−
−
−−− −−
−
− −
−−
−−−−−− −
−−−−
−−−−−−−
−−
−−
−
− −
−−−−
−−−− −−
−
− −−−−
−−− −− −
−
−−−−−
−−−−−−
−−
−− −
−−−
−−
−− −
−−
−
−−−
−
−−−− −− −−
−−−−
−−−−−−
−−
− −
−−−
−
−−
−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−−
−
−−−
−−
−−
−−− −−−−
−
−−
−−−− −
−−− −−−−−−− −
−
−−−
−−−− −−−
−−−−
−−−
−−
−−−−−−−− −−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−− −−−−−−− −−
−−
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−− −−−−−−
−
−−−− −−
−−
−−−−
−
− −−− −−
−−−− −−−−
−−−−
−−
−−− −−− −
−
−−−−−
−−
−−−−−−− −−−−−−− −−
−−−−−−−−
−−
−
−−− −
− −−−−−− −−−−−−− −
−−− −− −−
−−
−−
−−−−− −
−−−−−
−−−−
−−
−
−−−−−−
−−
−−−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−− −−
−−−
−
−−−−−−−−−
−
−
−−
−−
−−−−−− −−
−−−−−−−−− −− −
−−−−
−−
−−−−−− −−
−−−−−−−−−−−−−
−−−−
−−−−−
− −−−−
−−
−−−−−−
−−−−
−−
−−−−−− −−
−−−−−−−−−−−− −
−−
−−−−−
− −−
−− −
−
−−−−−−
−
−−− −−−− −−−
−−−−− −−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−−−−
−−−−
−−−−−− −−
−−−−
−−−−− −−
−−−−
− −−−−−− −−−−−−−
−−−−
− −−−−−−−
−
−−−−
−− −− −−−− −
−−−−−−−−− −−−
−− −−−−−−−−−−
−− −−−−−−−
−−−−−−−− −−− −−−−−−−−−−−−
−−−−−
−−
− −
−−−−−−−−
−−−−−
−−
−−−−−−
−−−−− −−−−−−
−−
−−−
−−−−−−
−−
−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: ISO
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−
−−
−−
− −
−
−−
−
−
−−−
−−
−
−−
−−
−−
−−−− −−
−−−−
−−−−−−
−
−−−
−
− −
−
−−−
−−−−
−−
−
−−−−−
−−− −− −
−
−−−−
−−−−−−−−
−−
− −
−
−−
−
−
−− −
−
−−
−−
−
−
−−−− −− −
−−−−−
−−−−−−
−−
− −
−−−
−
−
−−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−
−−
−−
−
−−
−
−
−−− −−−
−−
−−
−−−
− −−−− −−−−−−− −
−
−−−
−−−−
−−
−
−
−−−−−−
−
−−−−−−−−−
−
−
−−
−
−
−−−−−−
−−
−−−−− − −− −
−
−−−− −−−−−−− −
−−−
−
−−−−−−
−
−−−
−−−− −−−
−−−−− −
−− −−−−−−
−
−−−− −−
−
−−−−−
−
−−−− −
−
−−−−
−−− −
−−−−
−−
−−− −−
− −−
−−
−−−−
−−−−
−−−− −−−−−−− −−
−−−−−− −
−
−−
−
−−− −− −−−−−− −−−−−−− −
−−− −−
−
−−−−
−
−−−−− −
−−−
−−−−−−−−
−
−−−−−−
−
−−−
−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−−
−−−−−− −−
−−−−−−
−−−−−−−−
−
−−
−−
−−−−−− −−
−−−−
−−−−− −−
−−−
−−
−−
−−−−−− −−
−−−−−−−−−−−
−−
−−
−−
−−−−−− −
−−−
−−
−−−−−−−
−−−
−−
− −−−−− −−−−−−−−
−−−−−−
−−−−
−−−−− −
−
−−
−−
−−−−−−−
−−−
−−−− −−−−
−−−− −−−−−−−−−
−−−−−
−−−−−
−−−−−−−−−
−
−−−−−−−−
−−−−
−−−−−−−
−−−−−− −−
−−−−
−−−−− −
−−−
−−
− −−−−−− −−−−−−−
−−−−
−
−−−−
−
−−−
−−−−
−− −−
−−−− −−−−−−−
−−− −
−−−− −−−−
−−−−−−−− −−−−
−−−
−−−−−−−− −−−
−−−−−−−−−−−−
−−−−−
−−− −−−−−− −−−
−−−−−
−
−−−−−−−
−−−−−−−−−−−
−−−
−−
−−−−−− −
−−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: NSAN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−
−−
−
−
−−−
−−−
−−
−−
−−
−−−− −−
−−−−
−−−−−−
−
−−−
−
− −
−
−−
−
−−−−
−−
−
−−−−
−−
−− −− −
−
−−−−−
−−−−−−
−−
−
− −
−−−
−
−
−− −
−−
−
−−
−
−
−−−− −− −−
−−−−
−−−−−−−−
− −
−
−−
−
−
−−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−−
−
−−
−
−−
−−
−−− −−−
−−
−−
−−−− −
−−− −−−−−−− −
−
−−−
−−−− −−
−
−−−−−
−−−
−−−−−−−−− −
−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−− −−−−−−− −−−
−
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−− −−−−−−
−
−−−− −−
−
−−−−−
−
− −−− −−
−−−− −−−−
−−−−
−−
−−− −−− −
−
−−
−−−−
−−−−−−−− −−−−−−− −−
−−−−−− −
−
−−
−
−−− −− −−−−−− −
−−−−−− −−−− −− −
−−−
−−
−−−−− −
−−−−
−−−−−
−−
−
−−−−−−
−−
−−−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−− −−
−−−−
−−−−−−−−−−
−
−−
−−
−−−−−− −−
−−−−−−−−− −−
−−−
−−−−
−−−−−− −−
−−−−−−−−−−−−−
−−−−
−−−−−− −−
−−
−−
−−−−−−
−−−−
−−
−−−−−− −−
−−−−−−−−−−−−
−−−
−−−−−
− −−
−− −
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−−−−
−−−−
−−−−−− −−
−−−−
−−−−− −−
−−−−
− −−−−−− −−−−−−−
−−−−
−−−−−−
−−
−
−−−−
−− −− −−−− −
−−−−−−−−− −−−
−− −−−−−−−−−−
−− −−−−−−−
−−−−−−−− −−−
−−−−−−−−−−−−
−−−−−
−−
− −
−−−−−−−−
−−−−−
−
−−−−−−
−−−−−−
−−−−−−
−−−
−−
−−−−−−
−−
−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: AN
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
−−
−
−
−
− −
−−−
−
−
−−− −−
−
− −
−−
−−−−−− −
−−−−
−−−−−−−
−−
−−
−
− −
−−−−
−−−− −−
−
− −−−−
−−− −− −
−
−−−−−
−−−−−−
−−
−− −
−−−
−−
−− −
−−
−
−−−
−
−−−− −− −−
−−−−
−−−−−−
−−
− −
−−−
−
−−
−−−−−
−−
−−
−−−−−− −−
−−−−−
−−−−−−
−
−−−
−−
−−
−−− −−−−
−
−−
−−−− −
−−− −−−−−−− −
−
−−−
−−−− −−−
−−−−
−−−
−−
−−−−−−−− −−
−−
−
−
−−−−−−
−−
−−−−− − −− −−
−−−− −−−−−−− −−
−−
−
−−−−−−
−
−−−
−−−− −−−−
−−−− −−− −−−−−−
−
−−−− −−
−−
−−−−
−
− −−− −−
−−−− −−−−
−−−−
−−
−−− −−− −
−
−−−−−
−−
−−−−−−− −−−−−−− −−
−−−−−−−−
−−
−
−−− −
− −−−−−− −−−−−−− −
−−− −− −−
−−
−−
−−−−− −
−−−−−
−−−−
−−
−
−−−−−−
−−
−−−−−− −− −−
−−−− −−−−−−−
−−−−−
−−
−
−
−−−−−− −−
−−−
−
−−−−−−−−−
−
−
−−
−−
−−−−−− −−
−−−−−−−−− −− −
−−−−
−−
−−−−−− −−
−−−−−−−−−−−−−
−−−−
−−−−−
− −−−−
−−
−−−−−−
−−−−
−−
−−−−−− −−
−−−−−−−−−−−− −
−−
−−−−−
− −−
−− −
−
−−−−−−
−
−−− −−−− −−−
−−−−− −−−−−−− −−
−−−−−
−−−−−−− −−−−−−−−
−−−−−−−−
−−−−−−−
−−−−
−−−−−− −−
−−−−
−−−−− −−
−−−−
− −−−−−− −−−−−−−
−−−−
− −−−−−−−
−
−−−−
−− −− −−−− −
−−−−−−−−− −−−
−− −−−−−−−−−−
−− −−−−−−−
−−−−−−−− −−− −−−−−−−−−−−−
−−−−−
−−
− −
−−−−−−−−
−−−−−
−−
−−−−−−
−−−−− −−−−−−
−−
−−−
−−−−−−
−−
−−
0 2 4 6 8
02
46
8
HI GT2 cov mod: ISO
Observation log−odds ratio
Sim
ulat
ion
log−
odds
rat
io
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 16/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Correlogram (zero included)LO WT2
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: NSAN
MeanQ5−Q95OBS
GT2 RMSE = 0.1GT2 COR = 0.98
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: AN
MeanQ5−Q95OBS
GT2 RMSE = 0.11GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: ISO
MeanQ5−Q95OBS
GT2 RMSE = 0.09GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: NSAN
MeanQ5−Q95OBS
GT2 RMSE = 0.1GT2 COR = 0.98
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: AN
MeanQ5−Q95OBS
GT2 RMSE = 0.11GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: ISO
MeanQ5−Q95OBS
GT2 RMSE = 0.09GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: NSAN
MeanQ5−Q95OBS
GT2 RMSE = 0.1GT2 COR = 0.98
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: AN
MeanQ5−Q95OBS
GT2 RMSE = 0.11GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
LO GT2 cov mod: ISO
MeanQ5−Q95OBS
GT2 RMSE = 0.09GT2 COR = 0.99
NSAN
AN
ISO
HI WT2
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: NSAN
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.98
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: AN
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: ISO
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: NSAN
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.98
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: AN
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: ISO
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: NSAN
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.98
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: AN
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.99
0.0
0.2
0.4
0.6
0.8
1.0
Distance (km)
Cor
rela
tion
0 6 12 19 27 35 42 50 58 65 73 81 88
HI GT2 cov mod: ISO
MeanQ5−Q95OBS
GT2 RMSE = 0.04GT2 COR = 0.99
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 17/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Outline
1 Context
2 Overview of the model
3 Parameter estimation
4 First evaluations
5 Conclusions and Perspectives
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 18/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Conclusions & Perspectives
Good agreement in terms of marginal statistics with an underestimation of therain occurrence probability at stations
Underestimation of the spatial dependency
NS-covariance does not really improve the results in terms of marginal statisticsand spatial dependency (Is it worth it ?)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 18/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Conclusions & Perspectives
Good agreement in terms of marginal statistics with an underestimation of therain occurrence probability at stations
Underestimation of the spatial dependency
NS-covariance does not really improve the results in terms of marginal statisticsand spatial dependency (Is it worth it ?)
Another way to characterize spatial dependency,
Add cross-validation procedure,
Remove days with too many missing values in the dataset,
3-days cumulated amounts are really important for flood determination : Þadd temporal correlation (AR process)
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 18/18
Context Overview of the model Parameter estimation First evaluations Conclusions and Perspectives
Conclusions & Perspectives
LO WT2
HI WT2
P. Vaittinada Ayar & J. Blanchet SWGEN-Hydro – Berlin, Sept. 20, 2017 18/18
Thank you for yourattention
Reference
Allard, Denis and Bourotte, Marc [2015]. “Disaggregating daily precipitations into hourly values with a transformed censored latentGaussian process”. In: Stochastic Environmental Research and Risk Assessment. Vol. 29. no. 2, pp. 453–462.
Charles, Stephen P., Bates, Bryson C., and Hughes, James P. [1999]. “A spatiotemporal model for downscaling precipitationoccurrence and amounts”. In: Journal of Geophysical Research : Atmospheres. Vol. 104. no. D24, pp. 31657–31669.
Evin, G., Blanchet, J., Paquet, E., Garavaglia, F., and Penot, D. [2016]. “A regional model for extreme rainfall based on weatherpatterns subsampling”. In: Journal of Hydrology. Vol. 541, Part B, pp. 1185 –1198.
Garavaglia, F., Gailhard, J., Paquet, E., Lang, M., Garcon, R., and Bernardara, P. [2010]. “Introducing a rainfall compound distributionmodel based on weather patterns sub-sampling”. In: Hydrology and Earth System Sciences. Vol. 14. no. 6, pp. 951–964.
Higdon, Dave, Swall, J, and Kern, J [1999]. “Non-stationary spatial modeling”. In: Bayesian statistics. Vol. 6. no. 1, pp. 761–768.
Naveau, Philippe, Huser, Raphael, Ribereau, Pierre, and Hannart, Alexis [2016]. “Modeling jointly low, moderate, and heavy rainfallintensities without a threshold selection”. In: Water Resources Research. Vol. 52. no. 4, pp. 2753–2769.
Paciorek, Christopher J. and Schervish, Mark J. [2006]. “Spatial modelling using a new class of nonstationary covariance functions”.In: Environmetrics. Vol. 17. no. 5, pp. 483–506.
Pesonen, Maiju, Pesonen, Henri, and Nevalainen, Jaakko [2015]. “Covariance matrix estimation for left-censored data”. In:Computational Statistics & Data Analysis. Vol. 92, pp. 13 –25.
Rasmussen, P. F. [2013]. “Multisite precipitation generation using a latent autoregressive model”. In: Water Resources Research.Vol. 49. no. 4, pp. 1845–1857.
Vischel, Theo, Lebel, Thierry, Massuel, Sylvain, and Cappelaere, Bernard [2009]. “Conditional simulation schemes of rain fields andtheir application to rainfall–runoff modeling studies in the Sahel”. In: Journal of Hydrology. Vol. 375. no. 1–2. Surface processesand water cycle in West Africa, studied from the AMMA-CATCH observing system, pp. 273 –286.