probabilistic downscaling to detect regional present and
TRANSCRIPT
Probabilistic downscaling to detect regional present andfuture climate hazards
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson
University of Warwick, ECMWF, University of Bristol
28th April 2020
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 1 / 55
Funding
University of Warwick
Alan Turing Institute
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 2 / 55
Introduction
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 3 / 55
Introduction
Figure: Precipitation
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 4 / 55
Introduction
We have computer simulations of the weather/climate, they are calledmodel fields.
Temperature
Precipitable water content
Humidity
Geopotential
Wind speed and velocity
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 5 / 55
Introduction
We have computer simulations of the weather/climate, they are calledmodel fields.
Temperature
Precipitable water content
Humidity
Geopotential
Wind speed and velocity
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 5 / 55
Introduction
Figure: Air temperature
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 6 / 55
Introduction
(a) Model fields (b) Precipitation
Figure: Sample from data
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 7 / 55
Introduction
40 years of data (1979 - 2019)
Observed precipitation (∼ 10 km)
Simulated model fields (∼ 80 km)
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 8 / 55
Introduction
(a) Model fields (b) Precipitation
Figure: Sample from data
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 9 / 55
Introduction
Can we use the model fields on a coarse grid to forecast the precipitationon the fine grid? This task is known as downscaling.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 10 / 55
Introduction
1989-01 1989-03 1989-05 1989-07 1989-09 1989-11 1990-01Time
0
5
10
15
20
25
30
35ra
infa
ll (m
m)
Figure: Forecast (orange) and observed (blue)
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 11 / 55
Introduction
1.2± 1.5 mm
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 12 / 55
Statistical model
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 13 / 55
Statistical model
Occurrence of precipitation
Quantity of precipitation
Autocorrelation of precipitation
Affected by the model fields
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 14 / 55
Statistical model
Occurrence of precipitation
Quantity of precipitation
Autocorrelation of precipitation
Affected by the model fields
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 14 / 55
Statistical model
Occurrence of precipitation
Quantity of precipitation
Autocorrelation of precipitation
Affected by the model fields
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 14 / 55
Statistical model
Occurrence of precipitation
Quantity of precipitation
Autocorrelation of precipitation
Affected by the model fields
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 14 / 55
Statistical model
0 5 10 15 20lag (day)
0.0
0.2
0.4
0.6
0.8
1.0au
toco
rrela
tion
London: autocorrelation of rain
Figure: Autocorrelation of precipitation
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 15 / 55
Statistical model
We are dealing with a zero-inflated random variable.
Sometimes, theprecipitation is exactly 0 mm, sometimes it is a positive number.
We introduce the compound-Poisson.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 16 / 55
Statistical model
We are dealing with a zero-inflated random variable. Sometimes, theprecipitation is exactly 0 mm,
sometimes it is a positive number.
We introduce the compound-Poisson.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 16 / 55
Statistical model
We are dealing with a zero-inflated random variable. Sometimes, theprecipitation is exactly 0 mm, sometimes it is a positive number.
We introduce the compound-Poisson.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 16 / 55
Statistical model
We are dealing with a zero-inflated random variable. Sometimes, theprecipitation is exactly 0 mm, sometimes it is a positive number.
We introduce the compound-Poisson.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 16 / 55
Statistical model
Zt = number of times it rains at day t
R(i)t = amount of precipitation in a rain event at day t
Yt = amount of precipitation at day t
Yt |Zt =Zt∑i=1
R(i)t
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 17 / 55
Statistical model
Zt = number of times it rains at day t
R(i)t = amount of precipitation in a rain event at day t
Yt = amount of precipitation at day t
Yt |Zt =Zt∑i=1
R(i)t
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 17 / 55
Statistical model
Zt = number of times it rains at day t
R(i)t = amount of precipitation in a rain event at day t
Yt = amount of precipitation at day t
Yt |Zt =Zt∑i=1
R(i)t
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 17 / 55
Statistical model
Zt = number of times it rains at day t
R(i)t = amount of precipitation in a rain event at day t
Yt = amount of precipitation at day t
Yt |Zt =Zt∑i=1
R(i)t
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 17 / 55
Statistical model
Zt ∼ Poisson(λt)
R(i)t ∼ Gamma(1/ωt , 1/(ωtµt))
Yt |Zt =Zt∑i=1
R(i)t ∼ Gamma(Zt/ωt , 1/(ωtµt))
Yt ∼ Compound-Poisson
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 18 / 55
Statistical model
Zt ∼ Poisson(λt)
R(i)t ∼ Gamma(1/ωt , 1/(ωtµt))
Yt |Zt =Zt∑i=1
R(i)t ∼ Gamma(Zt/ωt , 1/(ωtµt))
Yt ∼ Compound-Poisson
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 18 / 55
Statistical model
Zt ∼ Poisson(λt)
R(i)t ∼ Gamma(1/ωt , 1/(ωtµt))
Yt |Zt =Zt∑i=1
R(i)t ∼ Gamma(Zt/ωt , 1/(ωtµt))
Yt ∼ Compound-Poisson
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 18 / 55
Statistical model
Zt ∼ Poisson(λt)
R(i)t ∼ Gamma(1/ωt , 1/(ωtµt))
Yt |Zt =Zt∑i=1
R(i)t ∼ Gamma(Zt/ωt , 1/(ωtµt))
Yt ∼ Compound-Poisson
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 18 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ) + θλ
Zt−1 − λt−1√λt−1
+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ) + θλ
Zt−1 − λt−1√λt−1
+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ) + θλ
Zt−1 − λt−1√λt−1
+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt
+ φλ(lnλt−1 − kλ) + θλZt−1 − λt−1√
λt−1+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ)
+ θλZt−1 − λt−1√
λt−1+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ) + θλ
Zt−1 − λt−1√λt−1
+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ) + θλ
Zt−1 − λt−1√λt−1
+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ) + θλ
Zt−1 − λt−1√λt−1
+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
We make λt , µt , and ωt a function of the model fields xt.
We make Zt and Yt |Zt undergo an autoregressive and movingaverage process.
For example
λt = exp
[βλxt + φλ(lnλt−1 − kλ) + θλ
Zt−1 − λt−1√λt−1
+ kλ
]
µt = exp
[βµxt + φµ(lnµt−1 − kµ) + θµ
yt−1 − Zt−1µt−1
µt−1√
Zt−1ωt−1+ kµ
]
ωt = exp [βωxt + kω]
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 19 / 55
Statistical model
Figure: Graphical model
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 20 / 55
Statistical model
1980 1982 1984 1986 1988 1990time
0
10
20
30
40
50
60
70
rain
fall
(mm
)
(a) Time series
0 10 20 30 40 50 60 70rainfall (mm)
0
500
1000
1500
2000
2500
3000
3500
4000
cum
ulat
ive
frequ
ency
(b) Cumulative
Figure: Simulated data
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 21 / 55
Statistical model
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20lag (day)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
auto
corre
latio
n
Figure: Simulated data autocorrelation ARMA(5,5)
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 22 / 55
Markov chain Monte Carlo
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 23 / 55
Markov chains Monte Carlo
This model is wrong so it is important to quantify how wrong themodel is.
The problem of modelling fitting can be done using Bayesianinference to get uncertainity quantification.
This was done using Markov chains Monte Carlo (MCMC).
We use Monte Carlo which uses random numbers to solve a modelfitting problem. Flucations in solutions, (samples), reflect theuncertainity.
We use a Markov chain to draw dependent samples in such a waythey converge to the right answer.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 24 / 55
Markov chains Monte Carlo
This model is wrong so it is important to quantify how wrong themodel is.
The problem of modelling fitting can be done using Bayesianinference to get uncertainity quantification.
This was done using Markov chains Monte Carlo (MCMC).
We use Monte Carlo which uses random numbers to solve a modelfitting problem. Flucations in solutions, (samples), reflect theuncertainity.
We use a Markov chain to draw dependent samples in such a waythey converge to the right answer.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 24 / 55
Markov chains Monte Carlo
This model is wrong so it is important to quantify how wrong themodel is.
The problem of modelling fitting can be done using Bayesianinference to get uncertainity quantification.
This was done using Markov chains Monte Carlo (MCMC).
We use Monte Carlo which uses random numbers to solve a modelfitting problem. Flucations in solutions, (samples), reflect theuncertainity.
We use a Markov chain to draw dependent samples in such a waythey converge to the right answer.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 24 / 55
Markov chains Monte Carlo
This model is wrong so it is important to quantify how wrong themodel is.
The problem of modelling fitting can be done using Bayesianinference to get uncertainity quantification.
This was done using Markov chains Monte Carlo (MCMC).
We use Monte Carlo which uses random numbers to solve a modelfitting problem. Flucations in solutions, (samples), reflect theuncertainity.
We use a Markov chain to draw dependent samples in such a waythey converge to the right answer.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 24 / 55
Markov chains Monte Carlo
This model is wrong so it is important to quantify how wrong themodel is.
The problem of modelling fitting can be done using Bayesianinference to get uncertainity quantification.
This was done using Markov chains Monte Carlo (MCMC).
We use Monte Carlo which uses random numbers to solve a modelfitting problem. Flucations in solutions, (samples), reflect theuncertainity.
We use a Markov chain to draw dependent samples in such a waythey converge to the right answer.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 24 / 55
Markov chains Monte Carlo
In Bayesian inference, we are interested in studying the posteriordistribution
posterior ∝ likelihood× prior
π(β, z1:T |x1:T , y1:T )︸ ︷︷ ︸posterior
∝ p(y1:T , z1:T |x1:T , β)︸ ︷︷ ︸likelihood
× π(β)︸︷︷︸prior
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 25 / 55
Markov chains Monte Carlo
In Bayesian inference, we are interested in studying the posteriordistribution
posterior ∝ likelihood× prior
π(β, z1:T |x1:T , y1:T )︸ ︷︷ ︸posterior
∝ p(y1:T , z1:T |x1:T , β)︸ ︷︷ ︸likelihood
× π(β)︸︷︷︸prior
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 25 / 55
Markov chains Monte Carlo
p(data|parameters)︸ ︷︷ ︸likelihood
p(parameters|data)︸ ︷︷ ︸posterior
∝ p(data|parameters)︸ ︷︷ ︸likelihood
× p(parameters)︸ ︷︷ ︸prior
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 26 / 55
Markov chains Monte Carlo
p(data|parameters)︸ ︷︷ ︸likelihood
p(parameters|data)︸ ︷︷ ︸posterior
∝ p(data|parameters)︸ ︷︷ ︸likelihood
× p(parameters)︸ ︷︷ ︸prior
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 26 / 55
Markov chains Monte Carlo
p(data|parameters)︸ ︷︷ ︸likelihood
p(parameters|data)︸ ︷︷ ︸posterior
∝ p(data|parameters)︸ ︷︷ ︸likelihood
× p(parameters)︸ ︷︷ ︸prior
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 26 / 55
Markov chains Monte Carlo
But the prior is subjective!
1980 1982 1984 1986 1988 1990time
0
5
10
15
20
25
30
35
rain
fall
(mm
)
(a) Sensible
1980 1982 1984 1986 1988 1990time
0
50
100
150
200
250
300
350
rain
fall
(mm
)
(b) Not sensible
Figure: Simulated data
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 27 / 55
Markov chains Monte Carlo
But the prior is subjective!
1980 1982 1984 1986 1988 1990time
0
5
10
15
20
25
30
35
rain
fall
(mm
)
(a) Sensible
1980 1982 1984 1986 1988 1990time
0
50
100
150
200
250
300
350
rain
fall
(mm
)
(b) Not sensible
Figure: Simulated data
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 27 / 55
Markov chains Monte Carlo
We use Gibbs sampling to split the posterior into two parts
Sample Zt |β,Z1:T\t , y1:T , x1:T
Sample β|y1:T ,Z1:T , x1:T
We use slice sampling and elliptical slice sampling.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 28 / 55
Markov chains Monte Carlo
We use Gibbs sampling to split the posterior into two parts
Sample Zt |β,Z1:T\t , y1:T , x1:T
Sample β|y1:T ,Z1:T , x1:T
We use slice sampling and elliptical slice sampling.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 28 / 55
Markov chains Monte Carlo
We use Gibbs sampling to split the posterior into two parts
Sample Zt |β,Z1:T\t , y1:T , x1:T
Sample β|y1:T ,Z1:T , x1:T
We use slice sampling and elliptical slice sampling.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 28 / 55
Markov chains Monte Carlo
We use slice sampling to sample Zt |β,Z1:T\t , y1:T , x1:T .
yt = 0⇒ Zt = 0
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 29 / 55
Markov chains Monte Carlo
We use slice sampling to sample Zt |β,Z1:T\t , y1:T , x1:T .
yt = 0⇒ Zt = 0
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 29 / 55
Markov chains Monte Carlo
Slice sampling:
s ∼ Uniform(0, π(z(i)))
z(i+1) ∼ Uniform(z : π(z) > s)
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 30 / 55
Markov chains Monte Carlo
We use elliptical slice sampling to sample β.
It can sample a posterior witha Normal prior. For the technical:
Chain at β(i)
ν ∼ prior
s ∼ Uniform(0, posterior(β(i)))
θ ∼ Uniform(0, 2π)
β(i+1) = β(i) cos θ + ν sin θ if posterior(β(i+1)) > s, else try another θwith a smaller support containing zero
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 31 / 55
Markov chains Monte Carlo
We use elliptical slice sampling to sample β. It can sample a posterior witha Normal prior.
For the technical:
Chain at β(i)
ν ∼ prior
s ∼ Uniform(0, posterior(β(i)))
θ ∼ Uniform(0, 2π)
β(i+1) = β(i) cos θ + ν sin θ if posterior(β(i+1)) > s, else try another θwith a smaller support containing zero
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 31 / 55
Markov chains Monte Carlo
We use elliptical slice sampling to sample β. It can sample a posterior witha Normal prior. For the technical:
Chain at β(i)
ν ∼ prior
s ∼ Uniform(0, posterior(β(i)))
θ ∼ Uniform(0, 2π)
β(i+1) = β(i) cos θ + ν sin θ if posterior(β(i+1)) > s, else try another θwith a smaller support containing zero
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 31 / 55
Markov chains Monte Carlo
Even more technical:
Initial: [θmin, θmax] = [θ − 2π, θ]
If: θ < 0, θmin = θ
Else: θmax = θ
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 32 / 55
Markov chains Monte Carlo
0 2000 4000 6000 8000 10000Sample number
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0Po
isson
Rate
_geo
pote
ntia
l
Figure: Example of a chain
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 33 / 55
Forecasting
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 34 / 55
Forecasting
Simulate the future using different samples from the posterior (MCMCchain). Any variation reflect the uncertainity.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 35 / 55
Forecasting
1989-01 1989-03 1989-05 1989-07 1989-09 1989-11 1990-01Time
0
5
10
15
20
25
30
35ra
infa
ll (m
m)
Figure: Forecast (orange) and observed (blue)
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 36 / 55
Forecasting
Figure: Probability precipitation > 15 mm
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 37 / 55
Forecasting
1989-01 1989-03 1989-05 1989-07 1989-09 1989-11 1990-01time
0.0
0.1
0.2
0.3
0.4
0.5
0.6fo
reca
sted
pro
babi
lity
of >
15
mm
of r
ain
Figure: Probability precipitation > 15 mm
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 38 / 55
Forecasting
0.0 0.2 0.4 0.6 0.8 1.0false positive rate
0.0
0.2
0.4
0.6
0.8
1.0tru
e po
sitiv
e ra
te
0 mm5 mm10 mm15 mm
Figure: ROC curve
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 39 / 55
Downscaling
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 40 / 55
Downscaling
We interpolate the model fields for now
(a) Coarse grid (b) Fine grid
Figure: Interpolation
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 41 / 55
Downscaling
We impose a prior on β where neighbouring locations with similartopography have similar values
β1β2...βN
∼ N(0, τ−1K
)
where [K]i ,j = exp[−ν
2(wi − wj)
2]
and wi is the topography
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 42 / 55
Downscaling
We impose a prior on β where neighbouring locations with similartopography have similar values
β1β2...βN
∼ N(0, τ−1K
)
where [K]i ,j = exp[−ν
2(wi − wj)
2]
and wi is the topography
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 42 / 55
Downscaling
1_poisson_rate
0.25
0.50
0.75
1.00
1.25
1.50
1.75
2.00
2.25
Figure: Simulation from prior
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 43 / 55
Downscaling
0 2000 4000 6000 8000 10000sample number
0.7
0.6
0.5
0.4
0_0_
Poiss
onRa
te_c
onst
Figure: MCMC chains
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 44 / 55
Downscaling
Figure: Forecast at Wales
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 45 / 55
Downscaling
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 46 / 55
Downscaling
Instead of interpolating the model fields, we can learn the model fields onthe coarse grid. We use a Gaussian process.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 47 / 55
Downscaling
x ′i ,t = model fields at time t, location i on coarse grid
xj ,t = model fields at time t, location j on fine gridx ′1,tx ′2,t
...x ′M,t
∼ N(0, τ−1K)
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 48 / 55
Downscaling
We can sample the model fields on the fine grid given:
model fields on the coarse grid
precipitation & Z on the fine grid
x1,tx2,t
...xN,t
∣∣∣∣∣∣∣∣∣
x ′1,tx ′2,t
...x ′M,t
, {Y1:T} , {Z1:T}
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 49 / 55
Downscaling
We can sample the model fields on the fine grid given:
model fields on the coarse grid
precipitation & Z on the fine gridx1,tx2,t
...xN,t
∣∣∣∣∣∣∣∣∣
x ′1,tx ′2,t
...x ′M,t
, {Y1:T} , {Z1:T}
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 49 / 55
Conclusion
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 50 / 55
Conclusion
We used the model fields on a coarse grid to forecast the precipitationon the fine grid.
We have quantified uncertainity.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 51 / 55
Conclusion
We used the model fields on a coarse grid to forecast the precipitationon the fine grid.
We have quantified uncertainity.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 51 / 55
Conclusion
1989-01 1989-03 1989-05 1989-07 1989-09 1989-11 1990-01Time
0
5
10
15
20
25
30
35ra
infa
ll (m
m)
Figure: Forecast (orange) and observed (blue)
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 52 / 55
Conclusion
The task of downscaling is computational expensive.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 53 / 55
Conclusion
To transistion from weather forecasting to climate prediction, we needto forecast many years into the future.
Sherman Lo, Ritabrata Dutta, Peter Dueben, Peter Watson (University of Warwick, ECMWF, University of Bristol)Probabilistic downscaling to detect regional present and future climate hazards28th April 2020 54 / 55