alexis berg ( locean-lsce ), benjamin sultan ( locean ), nathalie de noblet ( lsce )

27
What are the dominant features of rainfall leading to realistic large-scale yield prediction over West Africa ? Alexis Berg (LOCEAN-LSCE), Benjamin Sultan (LOCEAN), Nathalie de Noblet (LSCE) Group Meeting VARCLIM 11/12/09

Upload: wynn

Post on 15-Jan-2016

33 views

Category:

Documents


0 download

DESCRIPTION

What are the dominant features of rainfall leading to realistic large-scale yield prediction over West Africa ?. Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE ). Group Meeting VARCLIM 11/12/09. Context & objectives. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

What are the dominant features of rainfall leading to realistic large-scale yield

prediction over West Africa ?

Alexis Berg (LOCEAN-LSCE), Benjamin Sultan (LOCEAN), Nathalie de Noblet (LSCE)

Group Meeting VARCLIM 11/12/09

Page 2: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

• Climate impacts on crop production (variability, mean).

Particularly true in the Tropics (subsistence farming, low levels of managements, high climate variability)

• Linking climate models and crop models impact assessment (seasonal time-scale, climate change)

• Many sources of error: climate model, crop model, combination of both.

• GCM biases: in particular, rainfall (key variable for crop simulation…)

What consequence on the performance of yield prediction ?

Context & objectives

Case study on West Africa, focus on rainfall.

Model rainfall progressively corrected towards observations: how does the model “skill” respond (=ability to simulate observed yield variability) ?

Page 3: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

What is a crop model ?

Weather, CO2, radiation (+ nutrient stress)

Carbon assimilation(Monteith, Farqhuar…)

• Plot-scale: homogenous conditions, ’one’ plant

(allometric rules…)

= f(stage)

Leaves (LAI)

Roots

Stems

Grains

Sowing Vegetative stage Reproductive stage

Grain filling Maturation/dessication Harvest

Biomass

Stages=f(T° sum)

LAI

Page 4: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

What is a DGVM ?

Global Climate ModelGlobal Climate Model

Atmosphere

Ocean

Sea-iceVegetation

DGVMClimate, CO2Surface fluxes (LE, H,

CO2), albedo, roughness

Atmosphere model(1 grid cell)

Ex:ORCHIDEE

: croplands = grasslands. But they ARE different…

My work : to include a more realistic representation of tropical croplands in ORCHIDEE.

Page 5: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

• Climate impacts on crop production (variability, mean).

Particularly true in the Tropics (subsistence farming, low levels of managements, high climate variability)

• Linking climate models and crop models impact assessment (seasonal time-scale, climate change)

• Many sources of error: climate model, crop model, combination of both.

• GCM biases: in particular, rainfall (key variable for crop simulation…)

What consequence on the performance of yield prediction ?

Context & objectives

Case study on West Africa, with a large-scale crop model ORCHIDE-mil, focus on rainfall:

Model rainfall progressively corrected towards observations: how does the model “skill” respond (=ability to simulate observed yield variability) ?

Page 6: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )
Page 7: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Experimental setup

Yield simulations over West Africa, with a range of different forcing datasets where rainfall is increasingly realistically represented, from “model rain” to observations - using NCEP, CRU and IRD data.

- NCEP: interpolated at 1°x1°, 6h (Ngo Duc et al)

- CRU: 1°x1°, monthly

- IRD: 1°x1°, daily

Obs. – but CRU and IRD amounts are different.

“Model”

Page 8: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

- “Raw” NCEP

- NCEP with corrected annual cumulative rainfall (CRU or IRD data)

- NCEP with corrected monthly cumulative rainfall (CRU or IRD data)

- “monthly-permuted” IRD daily events (CRU or IRD annual amounts)

- IRD daily events (CRU or IRD annual amounts)

“Model”

Model + cumulative rainfall interannual variability

Model + cum. interannual variability + monthly cycle

Model + cum. interannual variability + monthly cycle + frequency

Model + cum.interannual variability + monthly cycle + frequency + real chronology of rainfall events

5 levels of realism:

Experimental setup

Yield simulations over West Africa, with a range of different forcing datasets where rainfall is increasingly realistically represented, from “model rain” to observations - using NCEP, CRU and IRD data.

Obs.

“Model”- NCEP: interpolated at 1°x1°, 6h (Ngo Duc et al)

- CRU: 1°x1°, monthly

- IRD: 1°x1°, daily

Re

alis

m

Page 9: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Average (1961-1990) time-lat. rainfall in NCEP and IRD

Page 10: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Exemple of seasonal rainfall over one pixel, one year

• NCEP overestimates rainfall frequency (« drizzle rains »)

• First rains occur too late in NCEP

Pdf of rainfall events in IRD and NCEP Difference in simulated sowing dates between IRD and NCEP

(blank = missing data)

Page 11: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Mali

Niger

Burkina-Faso

Senegal

NCEPCRUIRD

Annual rainfall (1968-1990) over different countries in NCEP, CRU and IRD

• CRU and IRD annual amounts are well correlated, but CRU rainfall is more abundant

• NCEP rainfall tends to be too small, and not correlated with observations

a)

Page 12: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Model skill

Simulated yields are aggregated at national scale (pixels are averaged).

We are only interested in interannual variability: all time series are detrended.

Model skill: correlation between observed (FAO) and simulated national yields over 1968-1990.

Page 13: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Model skill

Simulated yields are aggregated at national scale (pixels are averaged).

We are only interested in interannual variability: all time series are detrended.

Model skill: correlation between observed (FAO) and simulated national yields over 1968-1990.

Page 14: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of cumulative rainfall variability

Page 15: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of cumulative rainfall variability

Sudano-Sahelian West Africa: water-limited environment, rainfed crops

Observed yields are strongly correlated with observed annual rainfall – not with NCEP rainfall (since NCEP and observed rainfall are not well

correlated)

NCEP rainfall Observed rainfall (IRD/CRU)

Mali 0.21 0.41 / 0.32

Niger 0.18 0.58 / 0.61

Burkina-Faso 0.15 0.47 / 0.46

Senegal 0.26 0.72 / 0.68

Annual rainfall is the first “climate signal” in yield data.

Page 16: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of cumulative rainfall variability

Accordingly, simulated yields are (~always) significantly correlated with annual rainfall in input.

Correlations over 1968-1990 between simulated yields and annual rainfall. Dotted line shows the 5% significance level. Black bars are simulations with IRD annual rainfall, grey bars the ones with CRU annual rainfall

Page 17: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

… As a consequence from these two relationships (in observations and in the model), yields simulated with NCEP can not be expected to correlate well with observations

In other words, one can not simulate yield variability without the right cumulative rainfall variability in input.

Page 18: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of daily rainfall distribution

Page 19: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of daily rainfall distribution

More realistic representation of daily rainfall temporal characteristics (frequency, intensity) higher rain/yield correlations in the model…

Page 20: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of daily rainfall distribution

Rainfall/yield correlations are a first order measure of how water-limited crop productivity is in the model.

Yield/rainfall correlation

Page 21: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Rainfall with a proper frequency acts as a stronger constrain on crop productivity than “drizzle” rainfall.

Positive bias in simulated plant productivity caused by drizzle rainfall: small and frequent rain events reduce water stress, increasing the plant’s ability to assimilate carbon. Well-known bias in crop modelling: using large-scale climate model outputs as forcing tends to artificially increase crop production (e.g., Baron et al., 2005).

<

Page 22: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

‘Drizzle bias’ also undermines the model skill, as it weakens the correlation between input rainfall and simulated yield.

Observed Yields Simulated yields

Annual Rainfall

Stronger rainfall/sim correlations result in an increase of obs./sim correlations.

Page 23: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of intraseasonal distribution

At the scale considered here, information on the chronology of rainfall – whether monthly or daily - does not add to the model skill.

Page 24: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Counterintuitive:

intraseasonal distribution of rainfall (in particular, dry spells) has a significant impact on crop yield (Winkel et al., 1997)

+ the model (daily time step) is able to capture sub-seasonal effects

= model score should improve ?

Page 25: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Effect of intraseasonal distribution

Frequency distribution of the relative differences in yields between FREQ and OBS simulations. Calculations are done at the pixel scale (empty bars) and at the country scale (full grey bars) - all pixels (or countries) and all years considered.

Page 26: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

This suggests that intraseasonal distribution variability does not show a spatial consistency large enough to impact simulated yields aggregated on a wider scale.

Around each pixel, the area within which intraseasonal rainfall events are significantly correlated is no larger than a few pixels (1.27 on average).

Similarly, interannual variations in sowing dates are not spatially correlated beyond a few pixels

Area (pixel) of >0.5 inter-pixel correlation of 4-day dry spell occurrence, over JJAS 68-90

Page 27: Alexis Berg ( LOCEAN-LSCE ), Benjamin Sultan ( LOCEAN ), Nathalie de Noblet ( LSCE )

Conclusions

• The two essential rainfall features for the model to skilfully simulate large-scale yield variability are cumulative annual rainfall variability and rainfall temporal characteristics (frequency/intensity).

• At this scale, having the right chronology of rain events does not increase the model score.

Resolution-dependant ? Region-dependant ?

These results give indications on the characteristics of rainfall that climate models should ideally be able to simulate (or that should be bais-corrected/downscaled…) if seasonal climate forecasts are to be used to drive crop simulations

• The increase in model score here, as reanalysis rainfall is progressively corrected, suggests that improvements in GCM simulations are likely to translate into more accurate yield predictions.