Projection Pursuit Nonparametric Regression Used for Predicting Insect Densities from Visual Abundance Classes

Abstract

Two visual abundance methods were used to estimate the density of healthy or mummified aphids, Aphis gossypii Glover, on leaves and whole cucumber plants in greenhouses. Additional explanatory variables (e.g., proportion of infested leaves per plant or number of leaves per plant) were collected at the same time. Precise counts were also made on the same spatial units and fit to visual class data along with the explanatory variables using multiple regression with Poisson errors, logistic regression, and projection pursuit regression. Smaller subsets of data were used to test the robustness of the regression models. Multiple regression with Poisson errors gave the least precise fits in all cases, and the projection pursuit regression gave slightly better predictions than logistic regression. The accuracy of the logistic regression was similar for both the reference and the validation data sets. Visual abundance estimates and the associated explanatory variables are easy to gather, and when substituted into the projection pursuit regression model, they can be used to estimate aphid densities. Such models are useful in metapopulation studies of highly abundant organisms where extensive census data are required.