Testing MaxEnt model performance in a novel geographic region using an intentionally introduced insect

Abstract

Species distribution models (SDM's), such as the popular MaxEnt software, are frequently used to guide conservation programmes, predict the potential distribution of invasive species, forecast the impacts of climate change and develop applied ecological research (e.g., biological control). Many applications of these models require that SDM's be transferred in either space and/or time. However, few studies to date have tested the transferability, and thus usefulness, of MaxEnt models using fully independent data. Moreover, numerous authors have raised concerns over how model complexity (controlled primarily by feature class and regularisation multiplier settings) may affect MaxEnt model transferability. In this paper, we evaluated the usefulness of MaxEnt models when transferred in space using the native Australian insect Dasineura rubiformis Kolesik (Diptera: Cecidomyiidae) which has been intentionally introduced as a biological control agent of the invasive plant Acacia mearnsii De Wild (Fabaceae) in South Africa. MaxEnt's models were developed using native-range records only (Australia) and projected over South Africa to identify the potential climatic suitability for the insect, using a range of model parameter configurations. Model transferability was assessed using independent post-release data from South Africa. Our results demonstrated that MaxEnt scores were positively correlated with increased probabilities of recording D. rubiformis at field sites in South Africa. However, the accuracy of the MaxEnt models and their climatic suitability projections depended on the parameter configurations (i.e. features classes and regularisation multipliers) used to calibrate the models. Taken together, MaxEnt models can be useful in predicting where a species may establish, even when projected into a novel geographic area, but users need to be aware of how model parameter configurations and model extrapolation can influence model outputs and uncertainty. We demonstrate how statistical validation of the D. rubiformis MaxEnt models can be used to develop an environmental management programme towards improving the biological control of A. mearnsii using D. rubiformis by selecting optimal release sites.