Robust Rescaling Methods for Integrated
Water, Food,
Energy Security Management under
Uncertainty:
downscaling from GLOBIOM model
Tatiana Ermolieva
Yuri Ermoliev, Petr Havlik, Aline Mosnier, Michael Obersteiner
Motivations for developing
downscaling of land use and
land-cover changes
2
Global models (multiregional) cannot produce results at fine resolutions
Local models oversee global drivers and tendencies. Conversely, aggregate global models miss local implications, drivers, processes, uncertainties
New estimation problem; fine resolution information/data recovery from aggregate scales using all available information
Although GIS provides detailed geographical information, the social, economic,
environmental data and drivers ussually exist on aggregate level, e.g., national, regional
Local – global interdependencies
Approach: fussion (integration) of Global Biosphere Management (GLOBIOM)
partial equilibrium land use planning model and probabilistic (in general, non-Bayesian) downscaling to “project” GLOBIOM results to finer resolutions
At IIASA – assessment and downscaling of GLOBIOM SSP
Natural Forests
Managed
Forests
Forest
Cropland
Pastures
Other
natural land
GLOBIOM is a land use planning model
producing projections of land use and l.c.
changes:
Recursive dynamics:
land use change
transmitted from one
period to another
GLOBIOM is a multi-sectoral model
4 Crops for Food Sawn wood Pulp Fuel wood Sugarcane Corn Cassava Soybeans Wheat etc…LAND USE PRIMARY
PRODUCTS PROCESS FINAL PRODUCTS
Wood
Refineries
Biofuel Biogas
Heat Saw and Pulp
mills Cattle Sheep and Goats Pigs Poultry Slaughterhouse Lamb Beef Pork Chicken Eggs
Deterministic optimization problem: to achieve
the highest consumption level at the lowest cost
under constraints on land availability,
environmental pollution, etc.
Stochastic optimization problem - yields/weather,
supply, prod. costs, demand elasticities are in
general stochastic;
to achieve highest consumption at lowest cost
under food security constraints (represented by
Combination of biophysical and
economic models
Quantity Price Supply Demand EquilibriumLand productivity depends on
geography (climate, soil,
The economic model:
Main model outputs
Land use change
Production
Consumption
Prices
Trade flows
Water use
GHG emissions
Main exogenous drivers
Population growth
GDP growth
Technological change
Bio-energy demand (POLES
team)
Diet patterns (FAO, 2006)
6
Recursive dynamic: solution computed every 10-year period and
transmitted to the next period
30 regions
A global model with the possibility to
zoom into one region and …
30 Regions are
interconnected
through international
8
Land uses and land use
changes
allowed in GLOBIOM
Crop land to Planted forest Natural forest to Crop land
Natural forest to Grass land Grass land to Crop land
Grass land to Planted forest Natural land to Crop land
Natural land to Planted forest Natural land to Grass land
Crop land to Natural land Planted forest to Natural land
Grass land to Natural land
Land use
transformations
Land
uses
Crop land
Natural forest
Grass land Planted forest
Natural land
i
Land use type
Region
r
tTime
t irA
Land in land use
t ijr
A
Land converted from i to j in region r and time t
j t jir j t ijr t ir t irA
A
A
A
1At regional level
Land use change
downscaling:
GLOBIOM model
t ijlr
z
Find shares of aggregate area change transformed from i to j in l that
t ilr l t jir t jilr l t ijr t ijlr t ilr
z
A
z
A
A
A
1
l r m i t ilrL
A
=
1
n j m i l t ijlr t ijlr t ijlrz
q
z
1 1 1)
/
ln(
1
l t ijlrq
Downscaling problem
t ijr j t ijr t ijlrA
A
z
1
l t ijlrz
i.e.Cross-entropy: Maximize Kulback-Leibler information distance
Maximum Entropy and Minimax
Likelihood
Maximal Likelihood: let random value has N independent observations:
)
1
(
, …, . The problem is to estimate the
(N
)
true probability distribution:
j
pj ob Pr by maximizing the probability of observing the sample:
r j j p 1 1 subject to , ,p
j
0
N N j j
1 or maximizing . solution is .The log-likelihood function is the sample mean approximation of the
In downscaling, information is given not by observations but by various
constraints connecting the distribution with characteristics of observ. variables
Let P be a set of all possible distributions satisfying constraints:
Proposition: If , then .
x
(
x
1,...,
x
r)
r j j r j j p x ln p x ln x max
(
x
1,...,
x
r)
x
r j j j p p Ep 1 * ln
r j j N k j p k ob 1 1 ) ( Pr N pNj
j /
r j j j r j i p p i 1 1 ln ln Examples of priors
Priors for downscaling aggregate land use changes are estimated
based on comparative profitability of land use activities
Crop land to Planted
forest
1 t Pltlr t PltlrA
y
s t slr t slr t jr t lrP
y
A
1 t r t r t r t lr t lr T T T T min max min t lrT
Yield of planted forest
Crop land production value
Normalized transportation time
Average time to closest market
l t lr t r l Plt t r l Plt t lr t lr t r l Plt t r l Plt t lr t Plt CropL r lA
y
A
y
q
1 1 , , , , 1 1 1 , , , , 1 _ , ,)
(
)
(
Grass land to Planted
forest
l t r l Grass t r l Grass t Pltlr t Pltlr t lr t r l Grass t r l Grass t Pltlr t Pltlr t lr t Plt GrassL r lA
y
A
y
A
y
A
y
q
1 1 , , , , 1 1 1 1 , , , , 1 1 _ , ,)
(
)
(
1 t Pltlr t PltlrA
y
Yield of planted forest
1 t Pltlr t Pltlr
A
y
s t slr t slr t jr t lrP
y
A
1 t r t r t r t lr t lr T T T T min max min t lrT
Yield of planted forest
Crop land production value
Normalized transportation time
Forest land to Crop
land
s t slr t slr t jr t lr P y A 1Crop land production value
1 , , , , t Forest r l t Forest r l
A
y
Forest yield
Timber price
l t r l t Forest r l t Forest r l t Forest r l t r l t Forest r l t Forest r l t Forest r l t CropL ForestL r lP
A
y
P
A
y
q
, 1 , , 1 , , , , , 1 , , 1 , , , , _ , , t Forest r lP
, ,Dynamics of forest and planted forest land, 2010 – 2100, in 1000 ha
20 Optimal land use transition
shares: “Crop to grass” and “crop to other natural land”, 2010–2100
Robust downscaling
j i s ij ij ij s ij ij ij sz
z
q
z
z
q
z
z
F
,ln(
/
)
ln(
/
(
))
max
)
(
s i j s ij ij ij sz
z
q
z
F
,ln(
/
)
max
)
(
S s 1
s1
H
1
/
2
A set of priors defines alternative feasible distributions Q
qs,s 1: S
22
ij s ij ij S s s H ij ij ij S s i j s ij ij ij s H q z z z q z z z F( ) max ln( / ) ln min ( ln ) 1 1 , ) ln ( min 1 s ij S s s ij ij H
z q
ks
k
ij s ij ij s z q c : ln
S s s s H c 1 min K k 1,
K k k k u u 1 max
K k 1
ksuk cs u (u1,....,uk)Minimize F(z) with respect to z. F(z) is non-convex in : (z,
)
S s 1
ks
s
k
S s 1
ks
s
k Denote Dual problem
ij s ij ij kksuk z lnq s 1:S zZ
ij s ij ij S s s H ij ij ij S s i j s ij ij ij s H q z z z q z z z F( ) max ln( / ) ln min ( ln ) 1 1 , Minimization of F(z) (z,
)
k k k ij zij lnzij
uis equivalent to minimization w.r.t. (z,u) of
s.t.
24
Robust
Alternativ
Robust
Alternativ
e
26
Robust
Alternativ
Robust
Alternativ
e
28
30