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

14
Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden

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

Page 1: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Update on Fusarium Head BlightForecasting

Erick De Wolf, Denis Shah, Peirce Paul, and Larry

Madden

Page 2: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Brief History of Modeling EffortYears Location years Deployment

1999-2001 50 Individual states

2002-2003 120 Individual states and groups of states

2004-2014 527 Regional (30 states)

• Primarily logistic regression models• Now exploring Boosted Regression Tree (BRTs)

Page 3: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

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)

Page 4: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

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

Page 5: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Response Variable

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

Page 6: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Model Performance

Page 7: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Relative Influence Binary Predictors

• Corn residue and wheat type low relative influence dropped

• Genetic resistance retained

Page 8: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

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

Page 9: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Partial Dependence Plots

Variables summarize weather 7-days prior to anthesis

Page 10: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Partial Dependence Plots

Mean RH (%) Mean Temperature C

Variables summarize weather 7-days prior to anthesis

Page 11: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Visualize Interactions

Mean RH(%)

VS

S

MS & MR

Page 12: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

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

Page 13: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

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

Page 14: Update on Fusarium Head Blight Forecasting Erick De Wolf, Denis Shah, Peirce Paul, and Larry Madden.

Questions

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