Prediction of sunflower grain oil concentration as a function of variety, crop management and environment using statistical models

Abstract

Sunflower (Helianthus annuus L.) raises as a competitive oilseed crop in the current environmentally friendly context. To help targeting adequate management strategies, we explored statistical models as tools to understand and predict sunflower oil concentration. A trials database was built upon experiments carried out on a total of 61 varieties over the 2000–2011 period, grown in different locations in France under contrasting management conditions (nitrogen fertilization, water regime, plant density). 25 literature-based predictors of seed oil concentration were used to build 3 statistical models (multiple linear regression, generalized additive model (GAM), regression tree (RT)) and compared to the reference simple one of Pereyra-Irujo and Aguirrezábal (2007) based on 3 variables. Performance of models was assessed by means of statistical indicators, including root mean squared error of prediction (RMSEP) and model efficiency (EF). GAM-based model performed best (RMSEP=1.95%; EF=0.71) while the simple model led to poor results in our database (RMSEP=3.33%; EF=0.09). We computed hierarchical contribution of predictors in each model by means of R2 and concluded to the leading determination of potential oil concentration (OC), followed by post-flowering canopy functioning indicators (LAD2 and MRUE2), plant nitrogen and water status and high temperatures effect. Diagnosis of error in the 4 statistical models and their domains of applicability are discussed. An improved statistical model (GAM-based) was proposed for sunflower oil prediction on a large panel of genotypes grown in contrasting environments.