Update on Fusarium Head Blight Forecasting Erick De

Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden

Brief History of Modeling Effort Years Location years Deployment 1999 -2001 50 2002 -2003 120 Individual states and groups of states 2004 -2014 527 Regional (30 states) Individual states • Primarily logistic regression models • Now exploring Boosted Regression Tree (BRTs)

Boosted Regression Trees • Origins in machine learning community • Fits individual trees in forward, additive manner • New trees focus on cases misclassified by previous trees • Combines many simple predictive trees into single predictive model (1, 000 models)

FHB Data Sets • 527 cases; 70% training, 30% testing • Representing 15 states and 26 years • 350 weather-based predictors – 5, 7, 10, 14 days prior to or post-anthesis – Temp, atmospheric moisture, rain • Binary predictors – Corn residue – Wheat type (winter or spring) – Genetic resistance of variety

Response Variable • Binary representation of FHB epidemics – 1 if FHB severity is >10% – 0 if severity is <10%

Model Performance

Relative Influence Binary Predictors • Corn residue and wheat type low relative influence dropped • Genetic resistance retained

Relative Influence Weather Based Predictors • Pre-anthesis – Mean RH% – Temperature and RH combination • Hours that temp. 9 -30 and RH>90% • Post-anthesis – Mean temperature – Rain – Temperature RH combination

Partial Dependence Plots Variables summarize weather 7 -days prior to anthesis

Partial Dependence Plots Mean RH (%) Mean Temperature C Variables summarize weather 7 -days prior to anthesis

Visualize Interactions VS S MS & MR Mean RH(%)

Potential Value of BRTs? • Helpful tools for variable selection – Removal of corn residue and wheat type – Addition of rain post-anthesis • Insights on relationship between variables and FHB epidemics – RH and temp thresholds • Visualization of interactions – RH and Level of genetic resistance

Reality Check • Prediction accuracy improved over logistic models • Application of models considerably more complex (1, 000 predictive models) • Looking to apply what we have learned in other model frameworks better suited for application

Questions • For more information: – Shah et al 2014, Phytopathology 104: 702 -714