Nathalie Saint-Geours
PhD thesis defended on November 29, 2012
Supervised by: Christian Lavergne Jean-Stéphane Bailly – Frédéric Grelot
Introduction
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Spatially distributed models
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 2 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Spatially distributed models
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models
spatially distributed outputs
numerical model inputs
outputs
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Spatially distributed models
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models
numerical model inputs
outputs
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 2 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Spatially distributed models
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models spatially distributed inputs
numerical model
outputs
other inputs
spatial inputs +
u1=8
u3=5 u4=2.3 u2=0.1
u5=-3
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models spatially distributed inputs spatially distributed outputs
numerical model
other inputs
spatial inputs +
u1=8
u3=5 u4=2.3 u2=0.1
u5=-3
spatial output
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 2 / 55
Spatially distributed models
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models spatially distributed inputs spatially distributed outputs
numerical model
other inputs
spatial inputs +
u1=8
u3=5 u4=2.3 u2=0.1
u5=-3
spatial output
spatial scale issues !
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models spatially distributed inputs spatially distributed outputs
numerical model
other inputs
spatial inputs +
u1=8
u3=5 u4=2.3 u2=0.1
u5=-3
spatial output
spatial scale issues ! point output
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 2 / 55
Spatially distributed models
Numerical models
= blackbox
to describe, understand, support decision making, forecast
Spatially distributed models spatially distributed inputs spatially distributed outputs
numerical model
other inputs spatial inputs +
u1=8
u3=5 u4=2.3 u2=0.1
u5=-3
spatial output
spatial scale issues ! point output
aggregated output
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Uncertainty in spatial modelling
All models are wrong, some are useful (G. Box)
Many sources of uncertainty
lack of knowledge natural variability measurement errors model assumptions. . .
Spatial uncertainty
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 3 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Uncertainty in spatial modelling
All models are wrong, some are useful (G. Box)
Many sources of uncertainty
lack of knowledge natural variability measurement errors model assumptions. . .
Spatial uncertainty
spatial dependence spatial scale issues
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Uncertainty in spatial modelling
All models are wrong, some are useful (G. Box)
Many sources of uncertainty
lack of knowledge natural variability measurement errors model assumptions. . .
Spatial uncertainty
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 3 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Uncertainty in spatial modelling
All models are wrong, some are useful (G. Box)
Many sources of uncertainty
lack of knowledge natural variability measurement errors model assumptions. . .
Spatial uncertainty
spatial structure of uncertainty
spatial scale issues
z p(z)
z p(z)
spatial structure of uncertainty
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Uncertainty in spatial modelling
All models are wrong, some are useful (G. Box)
Many sources of uncertainty
lack of knowledge natural variability measurement errors model assumptions. . .
Spatial uncertainty
spatial structure of uncertainty spatial dependence
z p(z)
z p(z)
spatial dependence
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 3 / 55
Uncertainty in spatial modelling
All models are wrong, some are useful (G. Box)
Many sources of uncertainty
lack of knowledge natural variability measurement errors model assumptions. . .
Spatial uncertainty
spatial structure of uncertainty spatial dependence
spatial scale issues
z p(z)
z p(z)
z p(z)
z p(z)
scale issues
Many sources of uncertainty
lack of knowledge natural variability measurement errors model assumptions. . .
Spatial uncertainty
spatial structure of uncertainty spatial dependence
spatial scale issues
what impact on the uncertainty of model outputs?
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 3 / 55
Sensitivity analysis
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Model
Inputs Output
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Model
Inputs Output
Modelling uncertainty on model inputs
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 4 / 55
Sensitivity analysis
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Model
Inputs Output
Modelling uncertainty on model inputs
Propagating uncertainty
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Model
Inputs Output
Modelling uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 4 / 55
Sensitivity analysis
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Model
Inputs Output
Modelling uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Ranking sources of uncertainty (sensitivity indices)
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Modelling uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Ranking sources of uncertainty (sensitivity indices) Sensitivity analysis is not model validation.
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 4 / 55
Sensitivity analysis
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Ranking sources of scalar uncertainty
(sensitivity indices)
Resulting uncertainty on model output
Resulting scalar uncertainty
on model output Propagating
scalar uncertainty
Modelling uncertainty on scalar inputs
for scalar models only
How the uncertainty of a model output can be apportioned to different sources of uncertainty in the model inputs
Ranking sources of spatial uncertainty
(sensitivity indices)
Resulting uncertainty on model output
Resulting spatial uncertainty
on model output Propagating
spatial uncertainty
Modelling uncertainty on spatial inputs
what about spatial models?
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 4 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Research questions
investigate the use of sensitivity analysis
for spatial models
VB-GSA: Variance-Based Global Sensitivity Analysis (Sobol’ 1991)
low CPU models
2 research questions
1 how to compute sensitivity indices forspatial inputs?
2 how to account forscale issueswith VB-GSA?
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Research questions
Research goal
investigate the use of sensitivity analysis
for spatial models
(Sobol’ 1991) low CPU models
2 research questions
1 how to compute sensitivity indices forspatial inputs?
2 how to account forscale issueswith VB-GSA?
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 5 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Research questions
Research goal
investigate the use of sensitivity analysis
for spatial models
Scope
VB-GSA: Variance-Based Global Sensitivity Analysis (Sobol’ 1991)
low CPU models
questions 2 how to account forscale issueswith VB-GSA?
Research goal
investigate the use of sensitivity analysis
for spatial models
Scope
VB-GSA: Variance-Based Global Sensitivity Analysis (Sobol’ 1991)
low CPU models
2 research questions
1 how to compute sensitivity indices forspatial inputs?
2 how to account forscale issueswith VB-GSA?
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 5 / 55
An inductive approach
One case-study
flood risk model on the Orb river floodplain
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 6 / 55
An inductive approach
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis
methodological problems &
observations
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 6 / 55
An inductive approach
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis
methodological problems &
observations
Theoretical framework
to explain observations to elaborate methods
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis
methodological problems &
observations
Theoretical framework
to explain observations to elaborate methods new properties
of sensitivity indices
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 6 / 55
An inductive approach
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis
methodological problems &
observations
Theoretical framework
to explain observations to elaborate methods new properties
of sensitivity indices
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 6 / 55
An inductive approach
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis Inputs ModelOutput Modelling
uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Estimating sensitivity indices
preliminary step:
delimit the model under study
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis Inputs ModelOutput Modelling
uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Estimating sensitivity indices
preliminary step:
delimit the model under study
but: initial model not fully suitable to our needs
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 6 / 55
An inductive approach
One case-study
flood risk model on the Orb river floodplain
first tests of sensitivity
analysis Inputs ModelOutput Modelling
uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Estimating sensitivity indices
preliminary step:
delimit the model under study
but: initial model not fully suitable to our needs better describe the model
(not presented)
investigate the use of sensitivity analysis for spatial models
VB-GSA: Variance-Based Global Sensitivity Analysis low CPU models
2 research questions
1 how to deal withspatial inputsin VB-GSA?
2 how to account forscale issueswith VB-GSA?
Applied goals
design a flood risk model on the Orb floodplain
perform sensitivity analysis
Focus on
maps of flood damages at different scales
uncertainty in spatial data
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 7 / 55
Research goal
investigate the use of sensitivity analysis for spatial models
Scope
VB-GSA: Variance-Based Global Sensitivity Analysis low CPU models
2 research questions
1 how to deal withspatial inputsin VB-GSA?
2 how to account forscale issueswith VB-GSA?
Applied goals
design a flood risk model on the Orb floodplain
perform sensitivity analysis
Focus on
maps of flood damages at different scales
uncertainty in spatial data
investigate the use of sensitivity analysis for spatial models
VB-GSA: Variance-Based Global Sensitivity Analysis low CPU models
2 research questions
1 how to deal withspatial inputsin VB-GSA?
2 how to account forscale issueswith VB-GSA?
Applied goals
design a flood risk model on the Orb floodplain
perform sensitivity analysis
Focus on
maps of flood damages at different scales
uncertainty in spatial data
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 7 / 55
Outline of the presentation
2 The Orb case-study
3 Variance-based sensitivity indices for spatial inputs
4 Spatial scale issues
The Orb case-study
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 9 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
The Orb floodplain (Hérault)
(required by project funders - EU Floods Directive)
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
The Orb floodplain (Hérault)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 10 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
The Orb floodplain (Hérault)
(required by project funders - EU Floods Directive)
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
The Orb floodplain (Hérault)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 10 / 55
The Orb floodplain (Hérault)
−→ assessment of flood damage reduction
(required by project funders - EU Floods Directive)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 11 / 55
Terrain elevation
Terrain elevation Hazard map (water depths)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 11 / 55
Terrain elevation Elements at risk
(landuse map) Hazard map (water depths)
Terrain elevation Elements at risk
(landuse map) Hazard map (water depths)
Depth-damage functions
+
water depth
€
corn vineyard house
damage cm
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 11 / 55
Terrain elevation Elements at risk
(landuse map) Hazard map (water depths)
Depth-damage functions
+
water depth
€
corn vineyard house
damage cm
numerical model
Terrain elevation Elements at risk
(landuse map) Hazard map (water depths)
Depth-damage functions
+
water depth
€
corn vineyard house
damage cm
numerical model
Expected damages (for one flood event)
€
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 11 / 55
Terrain elevation Elements at risk
(landuse map)
Depth-damage functions
+
water depth
€
corn vineyard house
damage cm
numerical model
Expected damages (for one flood event)
€
Hazard maps
Terrain elevation Elements at risk
(landuse map)
Depth-damage functions
+
water depth
€
corn vineyard house
damage cm
numerical model
Expected damages (for one flood event)
€
Hazard maps
Frequencies of flood events
30 years 50 years 100 years
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 11 / 55
Terrain elevation Elements at risk
(landuse map)
Depth-damage functions
+
water depth
€
corn vineyard house
damage cm
numerical model Hazard maps
Frequencies of flood events
30 years 50 years 100 years
Expected annual damages Total : 12.2 M€ /year
numerical model
Expected annual damages Total : 12.2 M€ /year Terrain
elevation Elements at risk
(landuse map)
Depth-damage functions
+
numerical model
Hazard maps Frequencies of
flood events
Depth-damage functions
water depth
€
damage cm
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 11 / 55
Terrain elevation Elements at risk
(landuse map)
Depth-damage functions
+
numerical model
Hazard maps Frequencies of
flood events
Before After
Depth-damage functions
water depth
€
damage cm
12.2M€ /y 5.7M€ /y
After Before
Terrain elevation Elements at risk
(landuse map)
Depth-damage functions
+
numerical model
Hazard maps Frequencies of
flood events
Depth-damage functions
water depth
€
damage cm
12.2M€ /y 5.7M€ /y
After Before
Annual avoided damages (ΔAAD)
6.5 M€/y
- =
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 11 / 55
Orb case-study: model output
Map ofexpected annual avoidedflood damages
What spatialsupport (=aggregated area) for model output?
Orb river
Mediterranean sea ΔAAD [k€/year]
4 km N
1 0
< -6 -6 to -2 -2 to -0.1
-0.1 t o 0.1
4 to 6 > 8 0.5 to 2
What spatialsupport (=aggregated area) for model output?
Orb river
Mediterranean sea ΔAAD [k€/year]
4 km N
1 0
< -6 -6 to -2 -2 to -0.1
-0.1 t o 0.1
4 to 6 > 8 0.5 to 2
house of Mr Poujol 1.73 k€ /year
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 12 / 55
Orb case-study: model output
Map ofexpected annual avoidedflood damages
What spatialsupport (=aggregated area) for model output?
Orb river
Mediterranean sea ΔAAD [k€/year]
4 km N
1 0
< -6 -6 to -2 -2 to -0.1
-0.1 t o 0.1
4 to 6 > 8 0.5 to 2
221 k€ /year Sérignan district
What spatialsupport (=aggregated area) for model output?
Orb river
Mediterranean sea ΔAAD [k€/year]
4 km N
1 0
< -6 -6 to -2 -2 to -0.1
-0.1 t o 0.1
4 to 6 > 8 0.5 to 2
6523 k€ /year total floodplain
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 12 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Main characteristics of the Orb case-study
A spatially distributed model
∀ν ⊂R2, Yν = Z
x∈ν
Y(x)dx
apoint-basedmodel:
∀x∈R2, Y(x) =Floc(U,Z(x))
adeterministicmodel
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Main characteristics of the Orb case-study
A spatially distributed model
a point-basedmodel:
∀x∈R2, Y(x) =Floc(U,Z(x))
a deterministicmodel
numerical model
other inputs Z1(x)
Z2(x)
Z3(x) U1 ... Uk
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 13 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Main characteristics of the Orb case-study
A spatially distributed model a spatially additivemodel:
∀ν ⊂R2, Yν = Z
x∈ν
Y(x)dx
∀x∈R2, Y(x) =Floc(U,Z(x))
a deterministicmodel
numerical model
other inputs Z1(x)
Z2(x)
Z3(x) U1 ... Uk
aggregated output
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Main characteristics of the Orb case-study
A spatially distributed model a spatially additivemodel:
∀ν ⊂R2, Yν = Z
x∈ν
Y(x)dx
a point-basedmodel:
∀x∈R2, Y(x) =Floc(U,Z(x))
numerical model
other inputs Z1(x)
Z2(x)
Z3(x) U1 ... Uk
aggregated output
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 13 / 55
Main characteristics of the Orb case-study
A spatially distributed model a spatially additivemodel:
∀ν ⊂R2, Yν = Z
x∈ν
Y(x)dx
a point-basedmodel:
∀x∈R2, Y(x) =Floc(U,Z(x))
a deterministicmodel
numerical model
other inputs Z1(x)
Z2(x)
Z3(x) U1 ... Uk
aggregated output
1st research question
Variance-based sensitivity indices for spatial inputs
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 14 / 55
Problem
Ranking sources of scalar uncertainty
(sensitivity indices)
Resulting uncertainty on model output
Resulting scalar uncertainty
on model output Propagating
scalar uncertainty
Modelling uncertainty on scalar inputs
for scalar models only
Ranking sources of spatial uncertainty
(sensitivity indices)
Resulting uncertainty on model output
Resulting spatial uncertainty
on model output Propagating
spatial uncertainty
Modelling uncertainty on spatial inputs
what about spatial models?
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 15 / 55
Problem
Model
Inputs Output
Modelling uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Ranking sources of uncertainty (sensitivity indices)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 16 / 55
Modelling uncertainty on spatial inputs
scalar input U
scalar input U
probability density function
u
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 16 / 55
Modelling uncertainty on spatial inputs
scalar input U
probability density function
u p(u)
random sample u(1) = 12.3 u(2) = 4.5 u(3) = 8.9 u(4) = 7.4 u(5) = 48.5 u(6) = 23 u(7) = 8.5 u(8) = 10.5
scalar input U
probability density function
u
random sample u(2) = 4.5 u(3) = 8.9 u(4) = 7.4 u(5) = 48.5 u(6) = 23 u(7) = 8.5 u(8) = 10.5
spatial input Z(x)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 16 / 55
Modelling uncertainty on spatial inputs
scalar input U
probability density function
u p(u)
random sample u(1) = 12.3 u(2) = 4.5 u(3) = 8.9 u(4) = 7.4 u(5) = 48.5 u(6) = 23 u(7) = 8.5 u(8) = 10.5
spatial input Z(x)
stochastic process
2D
scalar input U
probability density function
u
random sample u(2) = 4.5 u(3) = 8.9 u(4) = 7.4 u(5) = 48.5 u(6) = 23 u(7) = 8.5 u(8) = 10.5
spatial input Z(x)
stochastic process
2D
n random realisations
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 16 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Modelling uncertainty on spatial inputs
grid of 5m resolution from aerial photography
measure & interpolation errors 2D Gaussian error field variogram model (estimated)
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Modelling uncertainty on spatial inputs
Terrain elevation (DEM)
grid of 5m resolution from aerial photography
2D Gaussian error field variogram model (estimated)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 17 / 55
Modelling uncertainty on spatial inputs
Terrain elevation (DEM)
grid of 5m resolution from aerial photography
Modelling uncertainty
measure & interpolation errors 2D Gaussian error field variogram model (estimated)
range a variogram model
h η
nugget σ²
Terrain elevation (DEM)
grid of 5m resolution from aerial photography
Modelling uncertainty
measure & interpolation errors 2D Gaussian error field variogram model (estimated)
range a variogram model
h η
nugget σ²
n=100 random realisations of DEM geostatistical
algorithm
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 17 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Modelling uncertainty on spatial inputs
Vector data
from remote sensing + field surveys
classification errors confusion matrix
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Modelling uncertainty on spatial inputs
Landuse map
Vector data
from remote sensing + field surveys
confusion matrix
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 18 / 55
Modelling uncertainty on spatial inputs
Landuse map
Vector data
from remote sensing + field surveys
Modelling uncertainty
classification errors confusion matrix
confusion matrix vineyard corn corn
vineyard 0.82 0.18 0.18 0.82
Landuse map
Vector data
from remote sensing + field surveys
Modelling uncertainty
classification errors confusion matrix
confusion matrix vineyard corn corn
vineyard 0.82 0.18 0.18 0.82
n=1000 random realisations of landuse map Monte Carlo
simulation
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 18 / 55
Propagating uncertainty
Model
Inputs Output
Modelling uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Ranking sources of uncertainty (sensitivity indices)
Model
Inputs Output
Modelling uncertainty on model inputs
Propagating uncertainty
Resulting uncertainty on model output
Resulting uncertainty on model output
Ranking sources of uncertainty (sensitivity indices)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 19 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Propagating uncertainty
What uncertainty should be propagated?
uncertainty on the set of random realisations (fixedP2D) additional uncertainty on stochastic processP2D (second-level)
stochastic process
2D
n random realisations of spatial input Z(x)
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Propagating uncertainty
What uncertainty should be propagated?
uncertainty on some map attributes
stochastic process
2D
n random realisations of spatial input Z(x)
map attributes (scalar descriptors)
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 20 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Propagating uncertainty
What uncertainty should be propagated?
uncertainty on some map attributes
uncertainty on the set of random realisations (fixedP2D)
stochastic process
2D
n random realisations of spatial input Z(x)
map attributes (scalar descriptors)
What uncertainty should be propagated?
uncertainty on some map attributes
uncertainty on the set of random realisations (fixedP2D) additional uncertainty on stochastic processP2D (second-level)
stochastic process
2D
n random realisations of spatial input Z(x)
map attributes (scalar descriptors)
vineyard corn corn vineyard 0.82 0.18
0.18 0.82
a h
η σ²
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 20 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Propagating uncertainty
Benchmark of methods
7 methods reviewed
CPU cost? / many spatial inputs? / anyP2D? / meta-modelling? main idea: spatial input Z(x)→somescalar parameters
stochastic process
2D
n random realisations of spatial input Z(x)
map attributes (scalar descriptors)
vineyard corn corn vineyard 0.82 0.18
0.18 0.82
a h
η σ²
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Propagating uncertainty
Benchmark of methods 7 methods reviewed
Macro Parameter Dimension Reduction Map Labelling Joint Meta-model Trigger Input
Map Attributes Second
Level
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 21 / 55
Introduction The Orb case-study Spatial inputs Scale issues General conclusion
Propagating uncertainty
Benchmark of methods
7 methods reviewed +comparisonon analytical test cases:
CPU cost? / many spatial inputs? / anyP2D? / meta-modelling?
Macro Parameter Dimension Reduction Map Labelling Joint Meta-model Trigger Input
Map Attributes Second
Level
Benchmark of methods
7 methods reviewed +comparisonon analytical test cases:
CPU cost? / many spatial inputs? / anyP2D? / meta-modelling?
main idea: spatial input Z(x)→somescalar parameters
Macro Parameter Dimension Reduction Map Labelling Joint Meta-model Trigger Input
Map Attributes Second
Level
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 21 / 55
Propagating uncertainty
Benchmark of methods
7 methods reviewed +comparisonon analytical test cases:
CPU cost? / many spatial inputs? / anyP2D? / meta-modelling?
main idea: spatial input Z(x)→somescalar parameters
Macro Parameter Dimension Reduction Map Labelling Joint Meta-model Trigger Input
Map Attributes Second
Level
Nov. 29th 2012 Sensitivity analysis of spatially distributed models 22 / 55