Agricultural and Forest Meteorology 209 (2015) 1–10 Contents lists available at ScienceDirect Agricultural and Forest Meteorology j our na l ho me page: www.elsevier.com/locate/agrformet Agrometeorological analysis and prediction of wheat yield at the departmental level in France David Gouache a,* , Anne-Sophie Bouchon a , Elodie Jouanneau b , Xavier Le Bris c a ARVALIS – Institut du Végétal, Station Expérimentale, 91720 Boigneville, France b ARVALIS – Institut du Végétal, Chemin des Bissonnets, 14980 Rots, France c ARVALIS – Institut du Végétal, Station Expérimentale de La Jaillière, 44370 La Chapelle Saint Sauveur, France a r t i c l e i n f o Article history: Received 24 July 2014 Received in revised form 20 January 2015 Accepted 27 April 2015 Keywords: Wheat Yield Model Cross-validation Prediction Agrometeorology a b s t r a c t Predicting annual crop yields is of interest for many agricultural applications. We present a prediction scheme at the departmental level, circa 100 km by 100 km, of winter wheat yields in France, applied for 23 departments, using official yield statistics from 1986 to 2010. Each model is a linear combination of 5–7 variables, selected from an initial pool of over 250 candidates. Candidate variables were generated using a phenological model and a crop water balance model, applied to a representative cropping situation for the department. Variable selection was carried out with forward stepwise regression methods. The variable selection process was cross-validated, so as to select robust variables. Model prediction performance was also evaluated by cross-validation. Satisfactory models were created for 20 departments, with root mean square error of prediction ranging from 0.25 t/ha to 0.39 t/ha. During use, whole season weather data is not available: this is complemented by frequential calculation over the past 20 years of historical weather data. We assessed the impact of time of prediction on model error by hindcasting yields for all 25 years of the dataset. We estimate that predictions can start 20 days after heading on average. We analysed predictive performance in an independent dataset and propose recommendations for use of these models outside their training dataset. The models give new insight as to the climatic factors that are key in determining yield in France. © 2015 Elsevier B.V. All rights reserved. 1. Introduction One of the main objectives of agrometeorological studies is to estimate crop production as a function of weather, soil and crop management parameters (Hoogenboom, 2000). Such studies have applications spanning very large temporal and spatial scales, from within-season adjustment of crop management practices to national and international forecasting of crop yields for political and economical stakeholders in famine early warning systems or commodity trading, from evaluation of strategic evolutions of crop management at the grower’s scale to assessments of climate change impacts (Hoogenboom, 2000; Gommes et al., 2010). Predicting crop yields at international, national or sub-national levels has received attention from researchers for many years, especially for impor- tant staple crops, such as wheat (Hill et al., 1979; Feyerherm and Paulsen, 1981; Walker, 1989; Travaso, 1990; Hébel et al., 1993; Abbreviations: RMSE, root mean square error; RMSEP, root mean square error of prediction. * Corresponding author. Tel.: +33 1 64 99 24 49; fax: +33 1 64 99 30 39. E-mail address: d.gouache@arvalisinstitutduvegetal.fr (D. Gouache). Supit, 1997; Landau et al., 2000; Qian et al., 2009; Gobin 2010; de Wit et al., 2010). Although different approaches have been mobilized to develop crop yield prediction schemes, these can be broadly categorized into two approaches: regression models and process-based models. The first family attempts to link synthetic weather variables calculated over all or part of the growing sea- son to regional yield variations, whereas the second makes use of crop simulation models that describe crop–soil–climate interac- tions with interwoven equations representing the response of crop functions to weather. Despite regular controversy between the two methods (Landau et al., 1998, 1999; Jamieson et al., 1999; Semenov and Shewry, 2011), both types of approaches have shown useful and convincing results. Common across both types of approaches are the difficulties in choosing relevant scale, as yields at aggregated scales depend on yields obtained across a wide variety of fields differing in soil properties, local weather, and crop management (Bakker et al., 2005; Challinor et al., 2003). Background tendencies for yield increases, or decreases, through improved management and cultivars, or reduced investments, respectively, also need to be accounted for, and this is mostly accomplished by detrend- ing observed yields, although the choice of the trend model in and of itself is not straightforward (Supit, 1997; Lobell and Field, http://dx.doi.org/10.1016/j.agrformet.2015.04.027 0168-1923/© 2015 Elsevier B.V. All rights reserved.