Invasive Species Compendium

Detailed coverage of invasive species threatening livelihoods and the environment worldwide

Abstract

Effect of the Maxent model's complexity on the prediction of species potential distributions.

Abstract

Ecological niche modeling (ENM) is widely used in the study of biological invasions and conservation biology. Maxent is the most popular algorithm and is being increasingly used to estimate species' realized and potential distributions. Most modelers use the default Maxent setting to fit niche models, which originated from an earlier study containing 266 species, with the purpose of seeking their realized distributions. However, recent studies have shown that Maxent uses a complex machine learning method. It is sensitive to sampling bias and tends to overfit training data, and is only transferrable at low thresholds. Default settings based on Maxent outputs are sometimes not reliable, making it difficult to interpret. Using Halyomorpha halys and classical modeling approaches (i.e., niche models that were calibrated in native East Asia and transferred to North America), we tested the complexity and performance of the Maxent model under different settings of regulation multipliers and feature combinations, and chose a fine-tuned setting with the lowest complexity. We then compared the response curves and model interpolative and extrapolative validations between models calibrated using default and fine-tuned settings. Our purpose was to explore the effects of the model's complexity on niche model performance in order to improve the development and application of Maxent in China. We argue that selection of environmental variables is crucial for model calibration, which should include ecological relevance and spatial correlation. Reducing sampling bias and delimitating a proper geographic background, together with the comparison of response curves and complexity of Maxent models built under different settings, is very important for fitting a good niche model. In the case of H. halys, the default and fine-tuned settings are different, however the response curve is much smoother in the fine-tuned model, and the omission error is lower in introduced areas when compared to default model, suggesting that the fine-tuned model reflects the response of H. halys to environmental factors more reasonably and precisely predicts the potential distribution.