Assessing the accuracy and stability of variables importance in species distribution models through sample perturbations

Variable importance derived from species distribution models (SDMs) is widely used to identify key environmental drivers of species distributions. However, there is a lack of systematic evaluation regarding the reliability of the variable importance inference and whether the individual occurrence records will have a significant impact. Here, we used virtual species and developed a sample perturbations diagnostic framework. By establishing different scenarios and constructing metrics such as Driver Rank Accuracy (DRA), Change Variable Importance (|∆VI|) etc., we evaluated the accuracy and stability of various SDM models in inferring the variable importance. Furthermore, we also revealed the underlying mechanisms contributing to the instability of inference. The project includes the data and code used, where the code contains drawing and calculation, and the data contains complete virtual species data and intermediate results.