Cross-validation of predictive models for deoxynivalenol in wheat at harvest

Camardo Leggieri, M.; Fels-Klerx, H.J. van der; Battilani, P.


To date, several models that predict deoxynivalenol (DON) in wheat at harvest are available. This study aimed to evaluate the performance of two of such models, including a mechanistic model developed in Italy and an empirical model developed in the Netherlands. To this end, field data collected in the periods 2002-2004 and 2009-2011 in Italy, and in the period 2001-2010 in the Netherlands were used. These historical data covered farm observations at 1,306 wheat fields, of which 155 in the Netherlands and 1,151 in Italy. A subset of 10% of the Italian data, derived by random sampling from the total Italian dataset, was used to validate both the Italian and the Dutch model. Additionally, the Italian mechanistic model was validated using the total Dutch dataset. Before validating the Dutch model, it was recalibrated using the remaining 90% of the Italian data. Results showed that predictions of both modelling approaches (mechanistic and empirical) for independent wheat fields were in accordance. Applying a threshold for DON concentration of 1,250 ?g/kg, the mechanistic DON model predicted 90% of the samples correctly. Results for cross-validation of the mechanistic DON model and the recalibrated empirical DON model showed that 93% of the samples were correctly predicted. In general, no more than 6% of underestimates were observed.