1. G2 Crop CIS meeting Ispra, May 14 – 15, 2012 Presented by: Institute of Geodesy and Cartography

2. ISPRA2012-01-25 Utility assessment of BioPAR products for wheat yield forecasting in Europe. Crop yield estimation. Detailed description of methods and comparison of results on MARSOP and BioPar data

3. ISPRA2012-01-25 10400 - Utility Assessment – IGiK contribution The objective of the work is to test the performance of MARS and BioPar indicators for yield forecast on an European window. The purpose is to show and assess their practical use in crop monitoring/yield forecasting. The work is aimed at comparing the differences in yield estimation accuracy, based on the two data sets. Objective

4. ISPRA2012-01-25 European agro-climatic zones Iglesias, A., Garrote, L., Quiroga, S., Moneo, M.: Impacts of climate change in agriculture in Europe. PESETA-Agriculture study. EUR 24107 EN; DOI 10.2791/33218; EC 2009.

5. ISPRA2012-01-25 Another grouping of regions mean ordinal number of the decade in which the annual maximum of NDVI occurred

6. ISPRA2012-01-25 Statistical model Partial Least Squares Regression Partial Least Squares Regression (PLSR) - to choose a few components being linear combinations of explanatory variables X and to perform linear regression of response variable Y on these variables instead of performing regression with use of all X-variables Y - response variable (yield value); X n - explanatory variables (values of vegetation indices); n - sequential number of ten-day period taken into account; d_beg, d_end – number of ten-day period corresponding to the beginning and the end of growing season, respectively (different for different agro-climatic zones); c Nn - function f – coefficients generated by the PLS regression algorithm.

7. ISPRA2012-01-25 Statistical model Partial Least Squares Regression Partial Least Squares Regression (PLSR) - generalization of multiple regression - many (correlated) predictor variables - few observations - to derive orthogonal components using the cross-covariance matrix between the response variable and the explanatory variables - dimension reduction technique similar to Principal Component Regression (PCR)  PCR - the coefficients reflect the covariance structure between the predictor variables X  PLSR – the coefficients reflect the covariance structure between the predictor X and response Y variables

8. ISPRA2012-01-25 Model evaluation One-leave-out One-leave-out cross-validation: - for each year of data the PLS regression model was built with this year excluded - the yield prediction for excluded year was performed - predicted and actual yield values were compared

9. ISPRA2012-01-25 Model evaluation One-leave-out One-leave-out cross-validation: Performances were evaluated in terms of cross-validation mean errors: MPE Mean Percentage Error (MPE) MAPE Mean Absolute Percentage Error (MAPE) RMSE Root Mean Square Error (RMSE) Yield_obs i – actual yield in year i, Yield_pred i –yield prediction made for year i, N – number of observations (years) taken into account

10. ISPRA2012-01-25 Results - cross validation Agro-climatic zones B i o P a r M A R S

11. ISPRA2012-01-25 Results - cross validation maxNDVI B i o P a r M A R S

12. Chosen regions For each european NUTS region WA - wheat area harvested (from Eurostat, mean value of 11 considered years) TA - total arable land area (from arable land mask) 12 Ispra, May 14 – 15, 2012

13. Chosen regions DK0 153.02 Atlantic Central ES41 133.69 Mediterranean North DE2 128.95 Continental North DEE 87.51 Continental North ES24 87.31 Mediterranean North mean 29.69 lowest 0.09 13 Ispra, May 14 – 15, 2012

14. Prediction errors 14 Ispra, May 14 - 15, 2012

15. Prediction errors 15 Ispra, May 14 - 15, 2012

16. Year 2009 yield prognosis 16 Ispra, May 14 - 15, 2012

17. Year 2009 yield prognosis 17 Ispra, May 14 - 15, 2012

18. Year 2009 yield prognosis 18 Ispra, May 14 - 15, 2012

19. Year 2009 yield prognosis 19 Ispra, May 14 - 15, 2012

20. Year 2009 yield prognosis 20 Ispra, May 14 - 15, 2012

21. Year 2009 prediction errors 21 Ispra, May 14 - 15, 2012

22. Year 2009 prediction errors 22 Ispra, May 14 - 15, 2012

23. Year 2009 prediction errors 23 Ispra, May 14 - 15, 2012

24. Year 2009 regression coefficients 24 Ispra, May 14 - 15, 2012

25. Year 2009 regression coefficients 25 Ispra, May 14 - 15, 2012

26. Year 2009 regression coefficients 26 Ispra, May 14 - 15, 2012

27. Models for aggregated data 27 Ispra, May 14 - 15, 2012 A strategy to increase the number of observations by grouping the NUTS As the number of years of yield data is small, the possibility of building PLS Regression models for aggregated data was investigated. Levels of NUTS-2 regions aggregation considered: o agro-climatic zone, o country, o country / agro-climatic zone, o NUTS-1 / agro-climatic zone.

28. Models for aggregated data 28 Ispra, May 14 - 15, 2012 For each NUTS-2 region, yield data was standardized. yield standardized = (yield – mean) / standard deviation Standardized yield values and values of vegetation indices from all NUTS-2 regions constituting one aggregated region were used to build PLS regression model for aggregated region.

29. Models for aggregated data 29 Ispra, May 14 - 15, 2012 Cross-validation The predictive ability of the model for aggregated region was assessed with cross-validation. For each year of the data: The PLS regression model was built on the basis of data that did not contain data for year considered (the standardization procedure for each NUTS-2 region was repeated). For each NUTS-2 region constituting the aggregated region, the prediction of standardized yield for year considered was performed and the destandardized yield value was calculated. This predicted yield value was compared with observed yield. Cross-validation MAPE, MPE, Nash-Sutcliffe coefficient were calculated.

30. Models for aggregated data 30 Ispra, May 14 - 15, 2012 Nash–Sutcliffe model efficiency coefficient

31. Models for aggregated data 31 Ispra, May 14 - 15, 2012 Nash–Sutcliffe efficiencies can range from −∞ to 1. An efficiency of 1 (E = 1) corresponds to a perfect match of modeled discharge to the observed data. An efficiency of 0 (E = 0) indicates that the model predictions are as accurate as the mean of the observed data, whereas an efficiency less than zero (E < 0) occurs when the observed mean is a better predictor than the model or, in other words, when the residual variance (described by the numerator in the expression above), is larger than the data variance (described by the denominator). Essentially, the closer the model efficiency is to 1, the more accurate the model is. NSC = 1 - a perfect match of modeled to the observed data. NSC = 0 - the model predictions are as accurate as the mean of the observed data NSC < 0 - the observed mean is a better predictor than the model The closer the model efficiency is to 1, the more accurate the model is.

32. Aggregation for agro-climatic zones 32 Ispra, May 14 - 15, 2012 Nash–Sutcliffe efficiencies can range from −∞ to 1. An efficiency of 1 (E = 1) corresponds to a perfect match of modeled discharge to the observed data. An efficiency of 0 (E = 0) indicates that the model predictions are as accurate as the mean of the observed data, whereas an efficiency less than zero (E < 0) occurs when the observed mean is a better predictor than the model or, in other words, when the residual variance (described by the numerator in the expression above), is larger than the data variance (described by the denominator). Essentially, the closer the model efficiency is to 1, the more accurate the model is. Number of regions NDVI_MARSfAPAR_MARSNDVI_BioParfAPAR_BioPar MPEMAPENSCMPEMAPENSCMPEMAPENSCMPEMAPENSC Alpine5-0.307.460.04-0.498.05-0.03-1.198.48-0.11-0.948.18-0.05 Atlantic Central48-0.907.01-0.02-0.626.880.01-0.866.97-0.01-0.916.97-0.01 Atlantic North3-0.726.78-1.06-0.486.61-1.030.008.11-1.43-0.188.16-1.28 Atlantic South6-1.3713.35-0.37-1.0412.77-0.18-0.7512.85-0.43-1.3011.87-0.09 Boreal4-2.4110.910.34-2.8912.190.07-5.0317.04-0.61-4.1514.35-0.24 Continental North 30-1.559.840.20-1.6510.380.14-2.6911.87-0.15-2.4811.61-0.13 Continental South 14-2.439.410.56-0.779.530.54-0.5812.860.340.0111.580.42 Mediterranean North 14-2.0012.530.01-1.9512.500.02-1.8612.260.05-1.7912.220.05 Mediterranean South 6-6.0020.470.16-6.1020.590.15-5.8120.420.20-5.8020.250.21

33. Country / agro-climatic zone 33 Ispra, May 14 - 15, 2012 Nash–Sutcliffe efficiencies can range from −∞ to 1. An efficiency of 1 (E = 1) corresponds to a perfect match of modeled discharge to the observed data. An efficiency of 0 (E = 0) indicates that the model predictions are as accurate as the mean of the observed data, whereas an efficiency less than zero (E < 0) occurs when the observed mean is a better predictor than the model or, in other words, when the residual variance (described by the numerator in the expression above), is larger than the data variance (described by the denominator). Essentially, the closer the model efficiency is to 1, the more accurate the model is. CountryAgro-Climatic zone Number of NUTS regions NSC MARSBioPar VCIFCIVCIFCI Austria AT_Alpine40.490.500.410.38 AT_Continental North10.240.330.060.17 BelgiumBE_Atlantic Central8-0.270.03-0.460.16 Germany DE_Atlantic Central50.14-0.020.140.17 DE_Continental North70.150.09-0.22-0.26 DenmarkDK_Atlantic Central1-0.88-1.05-0.54-0.52 Spain ES_Atlantic South1-0.53-1.10-0.46-0.51 ES_Mediterranean North60.640.630.740.76 ES_Mediterranean South10.630.610.390.45 FinlandFI_Boreal30.360.27-1.31-1.12 France FR_Atlantic Central120.600.540.20-0.02 FR_Atlantic South4-0.400.00-0.08-0.20 FR_Mediterranean North1-1.24-0.72-1.04 Hungary HU_Continental North10.690.650.530.61 HU_Continental South60.690.740.630.71 IrelandIE_Atlantic North2-0.91-1.25-2.94-2.63 Italy IT_Alpine1-0.48-0.440.440.36 IT_Mediterranean North110.810.770.86 IT_Mediterranean South50.52 0.45 LithuaniaLT_Continental North10.440.280.470.48 NederlandsNL_Atlantic Central9-0.430.02-0.16-0.03 PolandPL_Continental North160.800.790.770.80 PortugalPT_Atlantic South2-0.46-0.080.02-0.10 RomaniaRO_Continental South3-0.56-0.330.000.15 Sweden SE_Atlantic Central50.610.700.69 SE_Boreal10.070.11-0.19-0.24 SlovakiaSK_Continental North40.840.790.640.82 Great Britain UK_Atlantic Central80.550.50-0.30-0.27 UK_Atlantic North1-0.47-0.76-0.08-0.05 18191716

34. THANK YOU VERY MUCH