Crop-Weather Models Observed soybean yields (GA yield trials) vs. seasonal rainfall, temperature, simulated yields
The Challenge Nonlinearity. Crop response to environ- ment nonlinear, non- monotonic. Dynamics. Crops respond not to mean conditions but to dynamic interactions: – Soil water balance – Phenology
Crop Model Concept after Rabbinge, 1993
DSSAT v4.02.0
The Challenge: The Scale Mismatch Problem
Information Pathways predicted crop yields observed climate predictors ?
Information Pathways downscaled dynamic model stochastic generator crop model (observed weather) crop model (hindcast weather) analog years predicted crop yields statistical climate model statistical yield model observed climate predictors categorize
Linking Approaches Classification and analog methods (e.g., ENSO phases) Synthetic daily weather conditioned on forecast: stochastic disaggregation (Corrected) daily climate model output Statistical function of simulated response
Stochastic Disagregation of Monthly Rainfall Data for Crop Simulation Studies Stochastic disaggregation, and deterministic bias correction of GCM outputs for crop simulation studies
Linkage to crop simulation models Seasonal Climate Forecasts Crop simulation models (DSSAT) Crop forecasts <<<GAP>>> Daily Weather Sequence
a) Stochastic disaggregation Monthly rainfall Stochastic disaggregation Crop simulation models (DSSAT) Weather Realizations Crop forecasts GCM ensemble forecasts Stochastic weather generator >> >>
b) Bias correction of daily GCM outputs 24 GCM ensemble members Bias correction of daily outputs Crop simulation models (DSSAT) Weather Realizations Crop forecasts >> >>
Stochastic disaggregation of monthly rainfall amounts
Rainfall amounts and frequency prediction Katumani, Machakos Province, Kenya Skill of the MOS corrected GCM data OND
Structure of a stochastic weather generator u f(u) u<=p c ? x f(x) Generate ppt.=0 p c =p 01 p c =p 11 Wet-day non-ppt. parameters: μ k,1 ; σ k,1 Dry-day non-ppt. parameters: μ k,0 ; σ k,0 Generate today’s non- ppt. variables Generate uniform random number Precipitation sub-modelNon-precipitation sub-model (after Wilks and Wilby, 1999) Generate a non- zero ppt. (Begin next day)INPUT OUTPUT
Precipitation sub-model p 01 =Pr{ppt. on day t | no ppt. on day t-1} p 11 =Pr{ppt. on day t | ppt. on day t-1} f(x)=α/β 1 exp[-x/β 1 ] + (1-α)/β 2 exp[-x/β 2 ] μ= αβ 1 + (1-α)β 2 σ 2 = αβ (1-α)β α(1-α)(β 1 -β 2 ) Max. Likelihood (MLH) Markovian process Mixed- exponential Occurrence model: Intensity model:
Long term rainfall frequency: First lag auto-correlation of occurrence series: π=p 01 /(1+p 01 -p 11 ) r 1 =p 11 -p 01
Temperature and radiation model zAzB z(t)=[A]z(t-1)+[B]ε(t) z k (t)=a k,1 z 1 (t-1)+a k,2 z 2 (t-1)+a k,3 z 3 (t-1)+ b k,1 ε 1 (t)+b k,2 ε 2 (t)+b k,3 ε 3 (t) b k,1 ε 1 (t)+b k,2 ε 2 (t)+b k,3 ε 3 (t) T k (t)= μ k,0 (t)+σ k,0 z k (t); if day t is dry μ k,1 (t)+σ k,1 z k (t); if day t is wet Trivariate 1 st order autoregressive conditional normal model
Decomposing monthly rainfall totals R m =μ x π Dimensional analysis: where: R m - mean monthly rainfall amounts, mm d -1 μ - mean rainfall intensity, mm wd -1 π - rainfall frequency, wd d -1
Conditioning weather generator inputs μ = R m /π we condition α in the intensity model π = R m / μ we condition p 01, p 11 from the frequency and auto-correlation equations …and other higher order statistics
Conditioning weather generator outputs First step: Iterative procedure - by fixing the input parameters of the weather generator using climatological values, generate the best realization using the test criterion |1-R mSim /R m | j <= 5% Second step: Rescale the generated daily rainfall amounts at month j by (R m /R mSim ) j
Applications A.1 Diagnostic case study –Locations: Tifton, GA and Gainesville, FL –Data: A.2 Prediction case study –Location: Katumani, Kenya –Data: MOS corrected GCM outputs (ECHAM4.5) –ONDJF ( )
Crop Model: CERES-Maize in DSSATv3.5 Crop: Maize (McCurdy 84aa) Sowing dates: Apr – Tifton Mar – Gainesville Soils: Tifton loamy sand #25 – Tifton Millhopper Fine Sand – Gainesville Millhopper Fine Sand – Gainesville Soil depth: 170cm; Extr. H 2 O:189.4mm – Tifton 180cm; Extr. H 2 O:160.9mm – Gainesville 180cm; Extr. H 2 O:160.9mm – Gainesville Scenario: Rainfed Condition Simulation period: Simulation Data (Tifton, GA and Gainesville, FL)
Sensitivity of RMSE and correlation of yield Tifton, GAGainesville, FL A.1 Diagnostic Case RmRmRmRm π μ
Gainesville, FL Sensitivity of RMSE and R of rainfall amount, frequency and intensity at month of anthesis (May) RmRmRmRm μ π RmRmRmRm π μ
Gainesville, FL μ π RmRm 1000Realizations Predicted Yields
A.2 Case study: Katumani, Machakos Province, Kenya Skill of the MOS corrected GCM data OND
Simulation Data (Katumani, Machakos Province, Kenya) Crop Model: CERES-Maize Crop: Maize (KATUMANI B) Sowing dates (Nov ) Soil depth :130cm Extr. H 2 O:177.0mm Scenario: Rainfed Simulation period: Sowing strategy: conditional-forced
Sensitivity of RMSE and correlation of yield π1 (Conditioned) R m (Hindcast) π2 (Hindcast) R m +π2
R m (Hindcast) R m + π2 π1 (Conditioned) π2 (Hindcast)
Bias correction of daily GCM outputs (precipitation)
Statement of the problem RmRmRmRm Climatology, Monthly rainfall
RmRmRmRm Variance, Monthly rainfall
π μ Intensity Frequency
Proposed bias correction (a)-correcting frequency (b)-correcting intensity
Application Location: Katumani, Machakos, Kenya Climate model: ECHAM4.5 (Lat. 15S;Long. 35E) Crop Model: CERES-Maize Crop: Maize (KATUMANI B) Sowing dates (Nov ) Soil depth :130cm; Extr. H 2 O:177.0mm Scenario: Rainfed Simulation period: Sowing strategy: conditional-forced
Results RmRmRmRm μ Variance, R m μ Variance, μ
π
Sensitivity of RMSE and correlation of yield
Comparison of yield predictions using disaggregated, MOS- corrected monthly GCM predictions, and bias-corrected daily gridcell GCM simulations
Bias corrected seasonal rainfall (OND) RmRmRmRm μ π
Comparison of MOS corrected and bias corrected seasonal rainfall (OND)
Why are we successful? Is the procedure applicable in every situation? Inter-annual correlation (R) of monthly rainfall
Inter-annual variability of monthly rainfall for November
Extracting Useful Information from Daily GCM Rainfall for Cropping System Modeling
Temporal mismatch… Seasonal Climate Forecasts Cropping system models Yield forecasts, water balance etc. <<<GAP>>> Daily Weather Sequences Cropping system models require daily weather inputs
GCM Rainfall vs. Observed Rainfall Ines and Hansen (2006). Agric. For. Meteorol. Mean amount (mm d -1 ) Intensity (mm wd -1 ) Frequency (wd d -1 ) Obs GCM Source: wikipedia
Weather within Climate Hypothesis Maize Yield (kg ha -1 ) Years Correlation=0.65 “Observed” yield Uncorrected ECHAM4.5 GCM BIAS Machakos Southern Province, Katumani, Kenya Cropping season: Oct-Feb (Maize crop)
Deterministic Bias Correction GCM ensemble members Bias correction of daily outputs Crop simulation models (DSSAT) Weather Realizations Crop forecasts >> >>
Bias Correction of Daily GCM Rainfall (a)-correcting frequency (b)-correcting intensity Ines and Hansen (2006) Hansen et al. (2006) can be varied
BC-GCM Rainfall vs. Observed Rainfall Ines and Hansen (2006). Agric. For. Meteorol. Mean amount (mm d -1 ) Intensity (mm wd -1 ) Frequency (wd d -1 ) Source: wikipedia
Dry spell length (days) Cum. Frequency Yield, kg ha -1 During Anthesis (Nov 15-Dec 31), for 25 years BIAS BC-Obs BIAS Uncorr-Obs
Sample Bias-Corrected (BC) Rainfall (mm) Day of Year (year: 1995) Member 1-corr Observed Cropping season Member 1-uncorr mm BC fails to correct the temporal structure of daily rainfall
Corrected Monthly Rainfall Frequency after BC R=0.06 R=0.74 Observed Mean24Mem Observed Before After Mean24Mem
Combined BC-DisAg Stochastic disaggregation GCM ensemble members Bias correction of daily outputs Crop simulation models (DSSAT) Weather Realizations Crop forecasts >> >>
Simulated Number of Dry days (Nov. 15-Dec. 31) R 2 =0.49 R 2 =0.45 RAWBC BC-DisAg2
PDF of dry spell lengths (days) during anthesis period (Nov. 15-Dec. 31) from a) uncorrected, b) BC only and c) BC-DisAg2 (best trial). Dry spell length distributions (Nov. 15-Dec. 31)
1 – rainfall freq information derived from indv. members 2 – mean rainfall freq information derived from ensem. members year Maize Yield (kg ha -1 ) Performance of the information extracted from daily GCM rainfallMethodR(-)MBE (Mg ha -1 ) d(-)MSE (Mg ha -1 ) 2 MSE R (Mg ha -1 ) 2 MSE S (Mg ha -1 ) 2 Uncorrected BC only BC-DisAg BC-DisAg
Lessons learned… Simultaneous Bias Correction (BC) of GCM rainfall frequency and intensity improves the “weather within climate” information contained in the daily GCM rainfall, however- BC does not correct the temporal structure of daily GCM rainfall… GCM daily rainfall are highly auto-correlated. Combined BC-DisAg improves the temporal structure of daily rainfall hence improved the simulations of dry spell lengths and frequency, thus improving the systematic bias in the simulated yields.
Linear Programming Subject to:
Definition of terms Z = value of overall performance x j = level of activity j c j = increase in Z that would result from each unit increase in level of activity j b i = amount of resource i that is available for allocation to activities j a ij = amount of resource i consumed vy each unit of activity j
Example Max Z = 2x 1 + 3x 2 Subject to: x 1 ≤ 4 2x 2 ≤ 12 3x 1 + 2x 2 ≤ 18 Non-negativity constraint: x 1 ≥ 0; x 2 ≥ 0
Graphical solution x 1 ≤ 4 2x 2 ≤ 12 x1x1 x2x2 3x1 + 2x 2 ≤ 18 FEASIBLE REGION (0,0) (0,6) (2,6) (4,3) Z max = 2x 1 + 3x 2
Non-Linear Programming Subject to:
Graphical solution; linear constraints x1x1 x2x2 FEASIBLE REGION
Graphical solution; non-linear constraint x1x1 x2x2 FEASIBLE REGION
Crop-water management Example: Bata Minor, Bhakra Irrigation System, Kaithal, Haryana, India IRRIGATION SYSTEM Physical properties (soil, water quality, GW depth … ) Management practices (water, crop mgt … ) WEATHER WEATHER EXTERNAL CONSTRAINTS We can explore options in agricultural and water management Need to characterize and understand these complexities INPUT INPUT OUTPUT OUTPUT Yield, water balance, water productivities … NEED to develop a regional model (deterministic-stochastic)
RS-simulation model framework Pink: INVERSE MODELING Red: FORWARD MODELING
STUDY AREA Snapshot of Kaithal Irrigation Circle (Landsat 7ETM+)
Crop-water management optimization model Objective function Subject to water availability Decision variables: Water management Decision variables: Crop management By definition: Soil properties Salinity
Crop-water management Optimization Take the relaxed constraints Where: Fitness function:
Optimized wheat yields Current Optimized Current scenario
Crop-water management options Note: A Rainfall of 91 mm was recorded during the simulation period a In terms to T a /T p (irrigation scheduling criterion) b In terms of emergence dates