Descriptive modelling to predict deoxynivalenol in winter wheat in the Netherlands

Abstract

Predictions of deoxynivalenol (DON) content in wheat at harvest can be useful for decision-making by stakeholders of the wheat feed and food supply chain. The objective of the current research was to develop quantitative predictive models for DON in mature winter wheat in the Netherlands for two specific groups of end-users. One model was developed for use by farmers in underpinning Fusarium spp. disease management, specifically the application of fungicides around wheat flowering (model A). The second model was developed for industry and food safety authorities, and considered the entire wheat cultivation period (model B). Model development was based on observational data collected from 425 fields throughout the Netherlands between 2001 and 2008. For each field, agronomical information, climatic data and DON levels in mature wheat were collected. Using multiple regression analyses, the set of biological relevant variables that provided the highest statistical performance was selected. The two final models include the following variables: region, wheat resistance level, spraying, flowering date, several climatic variables in the different stages of wheat growing, and length of the period between flowering and harvesting (model B only). The percentages of variance accounted for were 64.4% and 65.6% for models A and B, respectively. Model validation showed high correlation between the predicted and observed DON levels. The two models may be applied by various groups of end-users to reduce DON contamination in wheat-derived feed and food products and, ultimately, reduce animal and consumer health risks.

Fusarium species and DON contamination associated with head blight in winter wheat over a 7-year period (2003–2009) in Belgium

“…A comparison between samples collected by hand before harvest and machineharvested samples revealed that the method provided a good estimate of DON concentration at harvest. Mean DON content showed considerable variation between the years as described in other European regions (Isebaert et al 2009;Van Der Felz-Klerx et al 2010). In order to help the industry to limit potential food or feed safety problems, a tool based on weather conditions near anthesis was developed to define the annual risk of DON contamination.…”

Predicting Fusarium Head Blight Epidemics with Boosted Regression Trees

“…What became clearer with the BRT partial-dependence plots (and previously unnoticed) was the large jump in the risk of major FHB epidemics with mean RH ≥ 70% for these three group 7 preanthesis predictors. In fact, most FHB predictive models to date have used RH-type predictors based on RH ≥ 80 to 85% (7,31,39,41,47,54) or RH ≥ 90% (13,23). Notable exceptions are the use of an RH > 75% criterion by Schaafsma and Hooker (48) in toxin prediction in winter wheat and an RH ≥ 70% criterion in toxin prediction in oat (57) and wheat (58) in Europe.…”

“…In one sense, averaging the brt i models represented an intermediary step on the way to the brt u model, which was developed using predictors across all windows. Although BRTs and model averaging were promising, we can surmise from the wider cumulative modeling efforts to date (12,13,25,30,39,47,48,54,58) that no one modeling algorithm is likely to suffice in describing or predicting the highly sporadic FHB epidemics (26). As more algorithms are considered, FHB researchers may do well to consider ensemble prediction methods which combine several algorithms into a single predictive algorithm (46).…”

“…The overall inclusion of R-based predictors by the BRT modeling algorithm contrasts with the lack of this predictor type in logistic regression models we presented earlier (51), where R-type predictors were dropped due to low selection frequency by the algorithms used. The majority of FHB models do, in fact, include rainfall-based predictors (7,12,13,23,25,29,31,39,41,47,54). Nevertheless, the strongest predictors of FHB epidemics, based on the full brt u model, were, in general, from one of three groups: mean RH per 24-h day (group 7), mean T per day (group 16), and the number of hours (24-h day) in which T was 9 to 30°C and RH ≥ 90% simultaneously (group 36).…”

“…In this sense, we do not regard BRTs and logistic regression as mutually exclusive efforts in FHB modeling. Many RH-based predictors are formulated on a 80 to 90% threshold (7,12,13,23,27,31,39,41,47,54), after deliberating the empirical links between weather and FHB epidemics (44). However, our BRT results challenged that threshold assumption for the pre-anthesis period by showing that the risk of FHB epidemics rises most dramatically when mean RH is >70%, a lower threshold than previously considered (48,57, Table 2 of Shah et al [51]) and boosted regression tree (brt prefix; see Figure 3 caption) model test error rates.…”