Marriage between variable selection and prediction methods to model plant disease risk

Abstract

Predicting the risk of a disease in a pathosystem based on a set of climatic variables usually requires handling a high number of input variables, many of which are often irrelevant and/or redundant. Building linear predictive models entails not only dimensionality issues but also the negative impact of multicollinearity. Several feature selection methods have proved to be efficient in both linear and non-linear models, regardless of those issues. However, in a machine learning (ML) context, it is necessary to evaluate these feature selection methods embedded into the model fitting algorithm to obtain the greatest accuracy. The aim of this work was to assess different combinations of variable selection methods with linear and non-linear predictors to fit climate-based models that predict the occurrence of a disease in a pathosystem. Four selection methods were compared: stepwise, which is frequently used in linear models, combined with VIF and p-value statistical criteria (Step+VIF+Pv), and other methods commonly used in ML: filter (F), genetic algorithm (GA), and Boruta (B). The disease risk predictors were constructed with a logistic linear regression model (LR) and the random forest (RF) algorithm, using all the available variables and the subgroups of variables selected by each feature selection method. Data from three pathosystems were processed: two involving Begomovirus –one in common bean (Phaseolus vulgaris L) and the other in soybean (Glycine max)– and the third one involving Mal de Rio Cuarto virus in maize (Zea mays L.). The data sets differed in sample size and number of variables. The accuracy of RF prediction did not vary among feature selection methods. Step+VIF+Pv was used to reduce the model outperformed the other feature selection methods in fitting LR. Our proposal suggests that the appropriate pairing of variable selection and prediction models would improve the modeling of plant disease risk.