Assessing Whether Habitat Suitability Models Can Predict Abundance of an Insect Pest

Predicting pest abundance across landscapes is a major challenge in pest management. Habitat suitability models offer one approach to produce spatially explicit forecasts of pest abundance, but they can often fail to effectively capture heteroscedastic pest distributions. While pests are unlikely to occur in poor habitats, a gradient between low and high pest abundance can occur in suitable habitats, and it is unknown which types of habitat suitability models are most effective in dealing with such wedge-shaped relationships between abundance and habitat suitability. Here we used six habitat suitability models to assess whether they could predict the abundance of a major native pest of blueberries, the blueberry maggot (Rhagoletis mendax). We used satellite-derived images to gather environmental data to build each habitat suitability model, and predictions were compared with R. mendax abundance from a 4-year field survey (2009-2012) in New Jersey, USA. We also modeled how habitat suitability predicts various quantiles of abundance with quantile regressions rather than only the mean. Only two models (BioClim and random forest) detected a positive wedge-shaped relationship between observed pest abundance and predicted habitat suitability, while four showed a negative relationship. The random forest model performed best in quantile regressions but was poor in transferability; the BioClim model performed poorly in interpolations but was transferrable across spaces. Given the heteroscedastic nature of pest distributions, habitat suitability models may fail to predict pest density unless they capture the gradient of abundance in suitable areas. Different modeling approaches may also be needed for effective interpolation and transferability.