Assessing the accuracy and stability of variable importance in species distribution modelling algorithms through sample perturbations

Variable importance derived from species distribution modelling (SDM) is widely used to identify key environmental drivers of species distributions. However, the reliability of these inferences has rarely been systematically evaluated, particularly with respect to their sensitivity to individual occurrences. Here, we used virtual species and developed a sample perturbations diagnostic framework, in which each occurrence was systematically removed (leave‐one‐out) to assess how small changes in the input data affect variable importance inference. Using this framework, we evaluated and compared ten SDM algorithms across multiple scenarios. We constructed a set of evaluation metrics to quantify both the accuracy of driver identification and the stability of permutation importance estimates, defined as their variability under sample perturbations. We further investigated the sample level mechanisms underlying instability. Our results showed that random forest (RF), gradient boosting machine (GBM), and maximum entropy (Maxent) achieved the best overall performance. RF exhibited the relatively high inferential accuracy and stability, but showed limited ability to differentiate the relative importance of environmental variables. GBM and Maxent achieved slightly lower accuracy than RF, but were more stable and effective for distinguishing the relative contributions of different variables. Instead, several algorithms displayed relatively high accuracy but were sensitive to sample perturbations. Scenario analyses indicated that stability generally increased under conditions of high sample sizes, low geographical sampling bias, and moderate prevalence. Highly influential occurrences were not necessarily associated with extreme environmental conditions. Removal experiments further demonstrated that variable importance inference in several algorithms was sensitive to a relatively small subset of occurrences, highlighting the disproportionate contribution of these occurrences. Overall, this study provides a sample level diagnostic framework for evaluating the reliability of variable importance in SDM, offering practical guidance for improving the interpretability of environmental driver inference in ecological research and conservation applications.