Spatially explicit metrics improve the evaluation of species distribution models facing sampling biases

The proliferation of open repositories offering georeferenced occurrences on biodiversity has boosted the use of species distribution models (SDMs). However, the need of presence-only records from these repositories yields a substantial limitation due to sampling biases, which can introduce uncertainty and skew SDM predictions. Furthermore, most predictions rely only on non-spatial metrics such as the AUC and the TSS to evaluate model performance. These metrics may not adequately account for spatially biased predictions, whereas the use of spatially explicit metrics could be more informative. Here, the effectiveness of both non-spatial and spatially explicit metrics is evaluated in response to predictions affected by sampling biases. Using SDMs, the distribution of 31 fleshy-fruited plants was predicted as a case study with contrasting settings to generate pseudo-absences and sampling bias corrections. Then, the performance of predictions was assessed with two non-spatial and three alternative, spatially explicit metrics. Predictions were affected by substantial sampling biases, particularly from West to East. Significant discrepancies were found between non-spatial and spatially explicit metrics. The non-spatial metrics failed to detect predictions affected by sampling biases, often yielding higher scores for less reliable predictions. In contrast, spatially explicit metrics benefited the implementation of bias corrections. Moreover, the method to generate pseudo-absences was more influential in determining spatial differences between predictions than either the bias correction or the geographical characteristics of input occurrences. Overall, our findings reveal the utility of spatially explicit metrics as a complementary tool to evaluate SDMs affected by sampling biases. • Non-spatial metrics are unable to discriminate SDMs affected by sampling biases. • The method to generate pseudo-absences drives spatial differences among predictions. • Bias corrections have contrasting effects on spatial and non-spatial metrics. • SDMs benefit from the use of alternative, spatially explicit metrics.