Herbicide resistance prediction: a mechanistic model vs a random forest model

Abstract

Herbicide Resistance is a major issue in weed control. Prediction tools can help to detect herbicide resistance development in early stages and enable farmers to take countermeasures. These tools can be simulation models combining population dynamics and genetics or AI methods like random forest. To evaluate and train prediction models the data base used is important. Models using population dynamics and genetics depend on a numerous number of plant physiological traits such as seed dormancy, germination probability or seed production, and knowledge about the genetical inheritance (how many genes are involved). Furthermore, the initial conditions like the distribution of the seedbank or proportion of resistant plants in the population are important. To train an AI based prediction tool a large data set of field history data together with the resistance status of the fields is needed. We show that a random forest model trained with an artificial data set generated by a mechanistic model (HERRMANN, 2000) is able to make predictions of the resistance status for real data sets with acceptable accuracies. Simulated model data are therefore equivalent to field data and can be used to investigate the effect of sampling schemes on the detection of resistance.

Keywords

• als inhibitor
• classification
• herbicide resistance prediction
• population dynamic
• random forest