bias correction with cdf-transform · qq-mapping !cdf observations (1980 2016) = cdf observations...
TRANSCRIPT
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Bias correction with CDF-Transform
Komlan A. KPOGO-NUWOKLO, Henning W. RUST, Uwe ULBRICH, ChristosVAGENAS and Edmund MEREDITH
Institut fur Meteorologie, Freie Universitat Berlin, Berlin, Germany
Berlin-Workshop on Bias Correction in Climate Studies
Berlin-Dahlem, 4-6 October 2016
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 1 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
• Systematic biases in the Regional Climate Models
6.8 7.0 7.2 7.4 7.6
50.9
51.0
51.1
51.2
51.3
51.4
Observation
lon
lat
4.0
4.5
5.0
5.5
6.0
6.5
6.8 7.0 7.2 7.4 7.6
50.9
51.0
51.1
51.2
51.3
51.4
Model
lon
lat
4.0
4.5
5.0
5.5
6.0
6.5
Figure: Mean of daily precipitation (mm/day) at Wupper catchment (case of July)
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 2 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
• Difference in the distribution
0 10 20 30 40 50 60
0.0
0.2
0.4
0.6
0.8
1.0
lon= 7.09 lat= 51.18
Precipitation [mm/day]
CD
F
ObsModel
Need to correct the whole distribution and not only the mean
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 3 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Existing bias correction methods
Mean-based approaches
Linear scaling [Lenderink et al., 2007]
Local intensity scaling [Schmidli et al., 2006]
Variance scaling [Leander and Buishand, 2007]
Power transformation (for precipitation) [Chen et al., 2013]
Distribution-based approaches
Quantile-mapping [Panofsky and Brier, 1968]
CDF-Transform [Michelangeli et al., 2009]
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 4 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Existing bias correction methods
Mean-based approaches
Linear scaling [Lenderink et al., 2007]
Local intensity scaling [Schmidli et al., 2006]
Variance scaling [Leander and Buishand, 2007]
Power transformation (for precipitation) [Chen et al., 2013]
Distribution-based approaches
Quantile-mapping [Panofsky and Brier, 1968]
CDF-Transform [Michelangeli et al., 2009]
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 4 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Existing bias correction methods
Mean-based approaches
Linear scaling [Lenderink et al., 2007]
Local intensity scaling [Schmidli et al., 2006]
Variance scaling [Leander and Buishand, 2007]
Power transformation (for precipitation) [Chen et al., 2013]
Distribution-based approaches
Quantile-mapping [Panofsky and Brier, 1968]
CDF-Transform [Michelangeli et al., 2009]
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 4 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Table of contents
1 Quantile-mapping
2 CDF-Transform
3 Application
4 Ongoing work
5 Summary
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 5 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Quantile-mapping
• Fo → CDF of observations
• Fm → CDF of model output
Fo(xo) = Fm(xm) =⇒ xbc = F−1o [Fm(xm)]
0 1 2 3 4 5 6 7 8 9 100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
CDF
Observation
Model
xbc
xm
Quantile-mapping can be used with both empirical and parametric CDF
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 6 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Limit of the Quantile-mapping
Future simulation
Observations do not cover the same time period as the model output
Example:
• Model output→ 1980− 2026 (historical+future)
• Observations→ 1980− 2016 (historical)
QQ-mapping→ CDF observations (1980− 2016) = CDF observations (1980− 2026)
Quantile mapping does not take into account any information on the distribution of thefuture modelled dataset
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 7 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Limit of the Quantile-mapping
Future simulation
Observations do not cover the same time period as the model output
Example:
• Model output→ 1980− 2026 (historical+future)
• Observations→ 1980− 2016 (historical)
QQ-mapping→ CDF observations (1980− 2016) = CDF observations (1980− 2026)
Quantile mapping does not take into account any information on the distribution of thefuture modelled dataset
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 7 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Limit of the Quantile-mapping
Future simulation
Observations do not cover the same time period as the model output
Example:
• Model output→ 1980− 2026 (historical+future)
• Observations→ 1980− 2016 (historical)
QQ-mapping→ CDF observations (1980− 2016) = CDF observations (1980− 2026)
Quantile mapping does not take into account any information on the distribution of thefuture modelled dataset
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 7 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Limit of the Quantile-mapping
Future simulation
Observations do not cover the same time period as the model output
Example:
• Model output→ 1980− 2026 (historical+future)
• Observations→ 1980− 2016 (historical)
QQ-mapping→ CDF observations (1980− 2016) = CDF observations (1980− 2026)
Quantile mapping does not take into account any information on the distribution of thefuture modelled dataset
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 7 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
CDF-Transform
Historical period Future periodObservation Fo,h(x) Fo,f (x)Model Fm,h(x) Fm,f (x)
• If it is possible to estimate Fo,f , then future model output can be corrected throughquantile-mapping:
xbc = F−1o,f [Fm,f (x)]
How can we approximate Fo,f ?
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 8 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
CDF-Transform
Historical period Future periodObservation Fo,h(x) Fo,f (x)Model Fm,h(x) Fm,f (x)
• If it is possible to estimate Fo,f , then future model output can be corrected throughquantile-mapping:
xbc = F−1o,f [Fm,f (x)]
How can we approximate Fo,f ?
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 8 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
CDF-Transform
Historical period Future periodObservation Fo,h(x) Fo,f (x)Model Fm,h(x) Fm,f (x)
• If it is possible to estimate Fo,f , then future model output can be corrected throughquantile-mapping:
xbc = F−1o,f [Fm,f (x)]
How can we approximate Fo,f ?
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 8 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
A solution to approximate Fo,f
Historical period Future periodObservation Fo,h(x) Fo,f (x)Model Fm,h(x) Fm,f (x)
• T : [0, 1] −→ [0, 1],
T (Fm,h(x)) = Fo,h(x) (1)
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 9 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
A solution to approximate Fo,f
Historical period Future periodObservation Fo,h(x) Fo,f (x)Model Fm,h(x) Fm,f (x)
• T : [0, 1] −→ [0, 1],
T (Fm,h(x)) = Fo,h(x) (2)
T (Fm,f (x)) = Fo,f (x) (3)
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 10 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
A solution to approximate Fo,f
Historical period Future periodObservation Fo,h(x) Fo,f (x)Model Fm,h(x) Fm,f (x)
• T : [0, 1] −→ [0, 1],
T (Fm,h(x)) = Fo,h(x) (4)
T (Fm,f (x)) = Fo,f (x) (5)
T is modelled by replacing x by F−1m,h(u),
T (u) = Fo,h(F−1m,h(u)) (6)
ans thus,
Fo,f (x) = Fo,h(F−1m,h(Fm,f (x))) (7)
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 11 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Application
Wupper catchment (Germany)
Model output from COSMO-CLM (1979-2015)
WFDEI used as reference data (1979-2013)
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 12 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
QQplot of daily precipitation (mm/day): January (left) - July (right)
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●
●●●●●●
●●●
● ●●● ●●●
● ●●
●
●
●
0 10 20 30 40 50
010
2030
4050
WFDEI
CO
SM
O−
CLM
lon= 7.46 lat= 51.11
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●
●●●●●●●●●●
●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●
●●●●
●●
●●
●
●●●
●●
●
●
●
0 10 20 30 40 50
010
2030
4050
WFDEI
CO
SM
O−
CLM
●
●
Model without correctionModel with CDFt correction
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●
●●●●●●
●●●●
●●●●●●●
●●●●●
●●●●●●●●
●● ●●●
● ●
●
●●
●
0 10 20 30 40
010
2030
40
WFDEI
CO
SM
O−
CLM
lon= 7.46 lat= 51.11
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●
●●●●●●●●●●●●●●
●●●●●●●●●●●●●●●
●●●●●●●●●
●●●●●●●●
●●●●●●●●
●●●●●
●●●●
●●●●●
●
●
●
●
●
●
●
●
0 10 20 30 40
010
2030
40
WFDEI
CO
SM
O−
CLM
●
●
Model without correctionModel with CDFt correction
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 13 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Monthly mean (left) and standard deviation (right) of daily precipitation
lon= 7.46 lat= 51.11
Mea
n of
pre
cipi
tatio
n [m
m/d
ay]
23
45
67
8
●
●
●
●
●
●●
●
●
● ●
●
J F M A M J J A S O N D
●
●
●
●
●
● ●
●
●
●●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
WFDEICOSMO−CLM without correctionCOSMO−CLM with CDFt correction
lon= 7.46 lat= 51.11
Std
of p
reci
pita
tion
[mm
/day
]
24
68
10
J F M A M J J A S O N D
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
WFDEICOSMO−CLM without correctionCOSMO−CLM with CDFt correction
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 14 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Deviation in the mean [Mean(reference)−Mean(model)] for January and July
6.8 7.0 7.2 7.4 7.6
50.9
51.1
51.3
Model without correction
lon
lat
−1.0
−0.5
0.0
0.5
1.0jan
6.8 7.0 7.2 7.4 7.6
50.9
51.1
51.3
Model with CDFt correction
lon
lat
−1.0
−0.5
0.0
0.5
1.0jan
6.8 7.0 7.2 7.4 7.6
50.9
51.1
51.3
Model without correction
lon
lat
−2
−1
0
1
2jul
6.8 7.0 7.2 7.4 7.6
50.9
51.1
51.3
Model with CDFt correction
lon
lat
−2
−1
0
1
2jul
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 15 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Bias correction for decadal climate predictions with CDF-t
• Goal: Bias correction of daily MiKlip decadal predictions ( spatial resolution = 0.11◦ )
Historical (1979-2014) Future (2015-2004)MiKlip F (M11)
h F (M11)f
WATCH F (W11)h F (W11)
f
CDF-t =⇒ F (W11)f (x) = F (W11)
h
{F
(M11)−1h
[F (M11)
f (x)]}
• Problem: MiKlip historical simulations are only available for spatial resolution of 0.44◦
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 16 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Bias correction for decadal climate predictions with CDF-t
• Goal: Bias correction of daily MiKlip decadal predictions ( spatial resolution = 0.11◦ )
Historical (1979-2014) Future (2015-2004)MiKlip F (M11)
h F (M11)f
WATCH F (W11)h F (W11)
f
CDF-t =⇒ F (W11)f (x) = F (W11)
h
{F
(M11)−1h
[F (M11)
f (x)]}
• Problem: MiKlip historical simulations are only available for spatial resolution of 0.44◦
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 16 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Use CDF-t as statistical downscaling method
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 17 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Summary
Compared to the quantile-mapping, CDF-t has advantage to take into account the changein the distribution for the future period
Satisfactory results are obtained with CDF-t
CDF-t can be used for statistical downscaling
Need of reliable multivariate and spatial bias correction methods
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 18 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
References I
Chen, J., Brissette, F. P., Chaumont, D., and Braun, M. (2013).Finding appropriate bias correction methods in downscaling precipitation for hydrologic impact studies over north america.Water Resources Research, 49(7):4187–4205.
Leander, R. and Buishand, T. A. (2007).Resampling of regional climate model output for the simulation of extreme river flows.Journal of Hydrology, 332(3):487–496.
Lenderink, G., Buishand, A., and Deursen, W. v. (2007).Estimates of future discharges of the river rhine using two scenario methodologies: direct versus delta approach.Hydrology and Earth System Sciences, 11(3):1145–1159.
Michelangeli, P.-A., Vrac, M., and Loukos, H. (2009).Probabilistic downscaling approaches: Application to wind cumulative distribution functions.Geophysical Research Letters, 36(11).
Panofsky, H. A. and Brier, G. W. (1968).Some applications of statistics to meteorology.University Park : Penn. State University, College of Earth and Mineral Sciences.
Schmidli, J., Frei, C., and Vidale, P. L. (2006).Downscaling from gcm precipitation: a benchmark for dynamical and statistical downscaling methods.International journal of climatology, 26(5):679–689.
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 19 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Acknowledgements
The BINGO project has received funding from the European Union’sHorizon 2020 Research and Innovation programme, under the GrantAgreement number 641739.
THANK YOU FOR YOUR ATTENTION
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 20 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
Future simulation CDFs off the range of the historical ones
Fo,f (x) = Fo,h(F−1m,h(Fm,f (x)))
−2 0 2 4 6 8 10
0.0
0.2
0.4
0.6
0.8
1.0
x
CD
F(x
)
Fo,hFm,hFm,fFo,f without shiftFo,f with shift
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 21 / 22
Introduction Quantile-mapping CDF-Transform Application Ongoing Summary
K. A. KPOGO-NUWOKLO et al. Bias correction with CDF-t 22 / 22