DISAT contribution: Development of a methodology for probabilistic assessments of climate change impacts on typical Mediterranean agric. crops (e.g. durum.

DISATDISAT contribution: contribution:

Development of a methodology for Development of a methodology for probabilistic assessments of climate change probabilistic assessments of climate change

impacts on typical Mediterranean agric. impacts on typical Mediterranean agric. crops (e.g. durum wheat)crops (e.g. durum wheat)

R. Ferrise, M. Moriondo and M. BindiR. Ferrise, M. Moriondo and M. Bindi

ENSEMBLES WORK PACKAGE 6.2 MEETINGENSEMBLES WORK PACKAGE 6.2 MEETING

HELSINKI, 26-27 APRIL 2007HELSINKI, 26-27 APRIL 2007

Structure of the presentationStructure of the presentation

1.1. What are the main objectives of What are the main objectives of our study?our study?

2.2. What have we achieved so far?What have we achieved so far?

3.3. What are we planning to do in the What are we planning to do in the next 6 months ?next 6 months ?

4.4. What are our main questions What are our main questions requiring discussion in this requiring discussion in this meeting?meeting?

1. What are the main objectives 1. What are the main objectives of our study?of our study?

– Select/Test impact models to simulate different Select/Test impact models to simulate different Mediterranean ecosystem tasks:Mediterranean ecosystem tasks:

• Forestry - damageForestry - damage due to forest fire due to forest fire • Agriculture - lossesAgriculture - losses due to water and heat stresses due to water and heat stresses

– Apply the selected models Apply the selected models to estimate a range of to estimate a range of impacts using the probabilistic representation impacts using the probabilistic representation provided by RT1provided by RT1

• Previous monthsPrevious months– Impact model selection and calibration for typical Mediterranean crops Impact model selection and calibration for typical Mediterranean crops

(e.g. olive, grapevine and durum wheat) and forest fire risk assessments(e.g. olive, grapevine and durum wheat) and forest fire risk assessments

– Collection of data required for the calibration and testing of impact Collection of data required for the calibration and testing of impact models and as reference data for model inputmodels and as reference data for model input

•Last 6 monthsLast 6 months2.12.1 Develop a simple statistical modelDevelop a simple statistical model that that

emulate process-based crop yield models (e.g. SIRIUS-emulate process-based crop yield models (e.g. SIRIUS-Quality durum wheat model) and can be used in Quality durum wheat model) and can be used in probabilistic climate change assessmentsprobabilistic climate change assessments

2.22.2 Create yield response surfacesCreate yield response surfaces altering the altering the baseline climatebaseline climate

2.32.3 Define critical thresholds of impactsDefine critical thresholds of impacts using using yield cumulative distribution from the last 30-years (e.g. yield cumulative distribution from the last 30-years (e.g. yield thresholds)yield thresholds)

2.42.4 Obtain preliminary estimates of risk Obtain preliminary estimates of risk probabilities probabilities overlapping response surfaces and joint overlapping response surfaces and joint distribution of T and P changesdistribution of T and P changes

2. What have we achieved so 2. What have we achieved so far?far?

2.1 Statistical model2.1 Statistical model to emulate to emulate process-based crop yield process-based crop yield models models

• 9 representative sites over the Mediterranean 9 representative sites over the Mediterranean Basin were selected to perform a Basin were selected to perform a scenario scenario sensitivity analysissensitivity analysis

• Crop yield model SIRIUS simulations Crop yield model SIRIUS simulations were were carried out for each scenario with different soils carried out for each scenario with different soils and N-ratesand N-rates

• The outputs of the model were used The outputs of the model were used to train a to train a neural network back-propagation modelneural network back-propagation model

• The Artificial Neural Network (ANN) was tested The Artificial Neural Network (ANN) was tested using the using the One-Leave-Out Cross ValidationOne-Leave-Out Cross Validation

2.1 Statistical mode to emulate process-based crop yield 2.1 Statistical mode to emulate process-based crop yield

modelsmodels

• Sites:Sites: 9 grid points (50 Km side), 9 grid points (50 Km side),

representatives of the Mediterranean representatives of the Mediterranean

Basin climatic variabilityBasin climatic variability

• Baseline Climate: Baseline Climate: 30 years (1975-30 years (1975-

2005) of daily Temperature (min and 2005) of daily Temperature (min and

max), Rainfall and Global Radiation max), Rainfall and Global Radiation

(from MARS JRC archive)(from MARS JRC archive)

• Temperature changes:Temperature changes: from 0°C from 0°C

to 8°C with 2°C stepto 8°C with 2°C step

• Precipitation changes:Precipitation changes: from -40% from -40%

to +20% with 20% stepto +20% with 20% step

• COCO22 Scenarios: Scenarios: from 350 ppm to from 350 ppm to

650 ppm with 100 ppm step650 ppm with 100 ppm step

Scenario Sensitivity Analysis:Scenario Sensitivity Analysis:

Soil types used for SIRIUS simulationsSoil types used for SIRIUS simulations

Low Medium High

Nitrogen Availability (Kg ha-1)

110 170 270

Nitrogen levels used for SIRIUS Nitrogen levels used for SIRIUS

simulationssimulations

Sandy Sandy-loam Loamy

Sand % 93 72 5

Clay% 3.6 25.4 17.5

Soil Water Capacity (mm)

53 115 215

SIRIUS simulations:SIRIUS simulations:– For each of the 9 grid cells For each of the 9 grid cells SIRIUS was run forSIRIUS was run for the combination the combination

of the of the different climatic scenariosdifferent climatic scenarios with with 3 different soils3 different soils and and

3 levels of Nitrogen3 levels of Nitrogen fertilization fertilization

– Sowing DateSowing Date was set using a was set using a

climatic criterion: at least 5 climatic criterion: at least 5

consecutive days with mean consecutive days with mean

Temperature < 14°C and Temperature < 14°C and

Rainfall < 2mm, starting from Rainfall < 2mm, starting from

October 1October 1stst and not later than and not later than

February 14February 14thth

– Nitrogen FertilizationNitrogen Fertilization was was

split in three times: 1/4 at split in three times: 1/4 at

sowing, 1/4 at tillering and 2/4 sowing, 1/4 at tillering and 2/4

at jointingat jointing

2.1 Statistical model to emulate process-based crop yield 2.1 Statistical model to emulate process-based crop yield

modelsmodels

• Training the Artificial Neural Network:Training the Artificial Neural Network:

– A A neural network back-propagation modelneural network back-propagation model was trained for emulating the was trained for emulating the SIRIUS outputs:SIRIUS outputs:

Network layersNetwork layers: 3: 3

Input nodesInput nodes: 5 (variables: CO: 5 (variables: CO22, SWC, N level, T(AMJ), Prec.(AMJ)), SWC, N level, T(AMJ), Prec.(AMJ))

Hidden layer nodesHidden layer nodes: 20: 20

OutputOutput: 1: 1

ANN model structureANN model structureTesting resultsTesting results

2.1 Statistical model to emulate process-based crop yield 2.1 Statistical model to emulate process-based crop yield

modelsmodels

• Leave-One-Out Cross Validation Test:Leave-One-Out Cross Validation Test:

Pearson’s correlation coefficients between ANN and SIRIUS estimates of crop yields for Pearson’s correlation coefficients between ANN and SIRIUS estimates of crop yields for

each of the 9 grid cells with all climate scenarios, 3 soils and 3 N-rates.each of the 9 grid cells with all climate scenarios, 3 soils and 3 N-rates.

Grid CellIdentification Number

Latitude Longitude AltitudePearson’s correlation

Coefficient

30090 36.64 27.31 45 0.90

34058 40.12 9.29 684 0.94

34080 39.29 22.11 305 0.95

37043 41.12 0.34 312 0.96

37063 41.42 12.30 2 0.95

43031 42.93 -7.40 590 0.92

43046 43.93 1.81 210 0.95

44056 44.62 8.05 348 0.93

45061 45.05 11.23 10 0.96

2.1 Statistical model to emulate process-based crop yield 2.1 Statistical model to emulate process-based crop yield

modelsmodels

2.2 Create yield response surfaces2.2 Create yield response surfaces

• SIRIUS and ANN model yield response surfaces were SIRIUS and ANN model yield response surfaces were

compared for a study area in France compared for a study area in France ((43.6 N, 5.0 E)43.6 N, 5.0 E)::

– Yield Response Surfaces were estimated altering the 30-years Yield Response Surfaces were estimated altering the 30-years

baseline climate (from MARS-JRC archive):baseline climate (from MARS-JRC archive):

50 km

Temperature changes:Temperature changes: from 0°C to from 0°C to

+8°C+8°C

Precipitation changes:Precipitation changes: from -40% to from -40% to

+20%+20%

COCO22 concentration scenarios: concentration scenarios: 350 ppm 350 ppm

and 550 ppmand 550 ppm

Soil Water Content:Soil Water Content: 115 mm 115 mm

Nitrogen Fertilization:Nitrogen Fertilization: 170 Kg N ha-1 170 Kg N ha-1

2.2 Create yield response surfaces2.2 Create yield response surfaces

• The ANN reproduced quite well the SIRIUS crop The ANN reproduced quite well the SIRIUS crop yields in the different scenariosyields in the different scenarios

France

y = 0.9369x + 0.0207

R2 = 0.9459

2

3

4

5

6

7

8

2 3 4 5 6 7 8

SIRIUS Outputs (Mg ha-1)

AN

N O

utp

uts

(M

g h

a-1)

SIRIUS and ANN estimated response surfaces for a SIRIUS and ANN estimated response surfaces for a

grid box in Southern France, for two COgrid box in Southern France, for two CO22 scenarios scenarios

Comparison between ANN and SIRIUS estimates Comparison between ANN and SIRIUS estimates

of crop yields used to draw the response surfacesof crop yields used to draw the response surfaces

2.3 Define critical thresholds2.3 Define critical thresholds

• Critical threshold of impact was obtained:Critical threshold of impact was obtained:– Calculating the cumulative probability of selected parameter (in Calculating the cumulative probability of selected parameter (in

this case yield)this case yield)

– Selecting, as threshold, the values that correspond to the 20% of Selecting, as threshold, the values that correspond to the 20% of probabilityprobability

0

20

40

60

80

100

120

0 2 4 6 8 10 12Yield (Mg ha-1)

Cu

mu

lati

ve

dis

trib

uti

on

(%

)

Cumulative distribution of yield in a pilot study areaCumulative distribution of yield in a pilot study area

5.35 Mg ha-1

2.4 Estimating risk probability2.4 Estimating risk probability

• The The trained ANNtrained ANN was applied was applied to estimate to estimate response surfacesresponse surfaces in a study area (Tuscany in a study area (Tuscany 50x 50 km grid cells)50x 50 km grid cells)

• The The statistical software “R”statistical software “R” was adopted was adopted to to calculatecalculate a polynomial regression model a polynomial regression model based on response surfacesbased on response surfaces

• The The regression model was applied to regression model was applied to calculate yieldcalculate yield using data from perturbed using data from perturbed physics experiment of Hadley Centre physics experiment of Hadley Centre for future for future scenariosscenarios

• The The perturbed yields were compared with perturbed yields were compared with yield thresholdyield threshold to define risk probability to define risk probability

2.4 Estimating risk probability2.4 Estimating risk probability

• Study Area: TuscanyStudy Area: Tuscany

Grid Cell Id. Number Latitude Longitude Altitude Yield Threshold (Mg ha-1)

39061 42.4 11.2 14 4.70

40061 42.8 11.1 129 4.86

40060 42.8 10.5 16 4.81

41062 43.2 11.7 350 6.17

41061 43.3 11.2 324 6.08

41060 43.3 10.5 107 5.88

42062 43.7 11.8 600 6.02

42061 43.7 11.2 229 5.72

42060 43.7 10.5 78 5.79

Sites: Sites: 9 grid cells (50 Km side)9 grid cells (50 Km side)

Baseline climate:Baseline climate: from MARS from MARS

JRC ArchiveJRC Archive

COCO22 concentration scenario: concentration scenario:

a1ba1b

Soil properties:Soil properties: from the Eusoils from the Eusoils

databasedatabase

Nitrogen level:Nitrogen level: 170 Kg ha 170 Kg ha-1-1

2.4 Estimating risk probabilities2.4 Estimating risk probabilities

• Risk probability Risk probability overlapping response surfaces and joint overlapping response surfaces and joint distribution of T an P changes distribution of T an P changes for a grid box in Tuscany for a grid box in Tuscany (2000-2020)(2000-2020)

Coor.:Coor.: 43.7N, 11.2E 43.7N, 11.2E COCO22 Scenario: Scenario: a1b a1b SWC:SWC: 115 mm 115 mm N level:N level: 170 Kg N ha 170 Kg N ha-1-1

• Risk probability Risk probability overlapping response surfaces and joint overlapping response surfaces and joint distribution of T an P changes for a grid box in Tuscany distribution of T an P changes for a grid box in Tuscany (2020-2040)(2020-2040)

Coor.:Coor.: 43.7N, 11.2E 43.7N, 11.2E

COCO22 Scenario: Scenario: a1b a1b

SWC:SWC: 115 mm 115 mm N level:N level: 170 Kg N ha 170 Kg N ha-1-1

2.4 Estimating risk probabilities2.4 Estimating risk probabilities

• Risk probability Risk probability overlapping response surfaces and joint overlapping response surfaces and joint distribution of T an P changes for a grid box in Tuscany distribution of T an P changes for a grid box in Tuscany (2040-2060)(2040-2060)

Coor.:Coor.: 43.7N, 11.2E 43.7N, 11.2E

COCO22 Scenario: Scenario: a1b a1b

SWC:SWC: 115 mm 115 mm N level:N level: 170 Kg N ha 170 Kg N ha-1-1

2.4 Estimating risk probabilities2.4 Estimating risk probabilities

• Risk probability Risk probability overlapping response surfaces and joint overlapping response surfaces and joint distribution of T an P changes for a grid box in Tuscany distribution of T an P changes for a grid box in Tuscany (2060-2080)(2060-2080)

Coor.:Coor.: 43.7N, 11.2E 43.7N, 11.2E

COCO22 Scenario: Scenario: a1b a1b

SWC:SWC: 115 mm 115 mm N level:N level: 170 Kg N ha 170 Kg N ha-1-1

2.4 Estimating risk probabilities2.4 Estimating risk probabilities

• Risk probability Risk probability overlapping response surfaces and overlapping response surfaces and joint distribution of T an P changes for a grid box in joint distribution of T an P changes for a grid box in Tuscany (2080-2100)Tuscany (2080-2100)

Coor.:Coor.: 43.7N, 11.2E 43.7N, 11.2E COCO22 Scenario: Scenario: a1b a1b SWC:SWC: 115 mm 115 mm N level:N level: 170 Kg N ha 170 Kg N ha-1-1

2.4 Estimating risk probabilities2.4 Estimating risk probabilities

• Change in risk probability in the 9 grid Change in risk probability in the 9 grid cells in the next decadescells in the next decades

-10

-5

0

5

10

15

20

Change in r

isk p

robabilit

y

(%)

2010

2020

2030

2040

2050

2060

2070

2080

2090

Periods

39061 40060 40061

41060 41061 41062

42060 42061 42062

2.4 Estimating risk probabilities2.4 Estimating risk probabilities

– Next 6 months (months 33-38)Next 6 months (months 33-38)•Move on other agric. Crops (grapevine and olive) and Move on other agric. Crops (grapevine and olive) and

forest fire risks to:forest fire risks to:

– Development simple statistical models Development simple statistical models that emulates that emulates process-based crop yield models process-based crop yield models

– Create preliminary yield response surfacesCreate preliminary yield response surfaces altering the altering the baseline climatebaseline climate

– Define critical thresholds of impactsDefine critical thresholds of impacts using yield using yield cumulative distribution from the last 30-yearscumulative distribution from the last 30-years

– Obtain preliminary estimates Obtain preliminary estimates of risk probabilities of risk probabilities overlapping response surfaces and joint distribution of T an overlapping response surfaces and joint distribution of T an P changesP changes

3. What are we planning to do 3. What are we planning to do in the next 6 months ?in the next 6 months ?

4. What are our main questions 4. What are our main questions requiring discussion in this requiring discussion in this

meeting?meeting?• To get information from:

– Chris about overall progress in ENSEMBLES (i.e. new developments, ongoing activities and future plans)

– Clare (from RT2B) about the latest status of RT2B on the provision of climate model outputs or their derivatives for use in impact assessment (i.e. by WP 6.2)

• To discuss with:– Chris, Clare and Glen about:

• the various methods of applying probabilistic climate projections in impact studies,

• the format and delivery of climate information for use in impact assessments in ENSEMBLES