Influences of temporal independence of data on modelling species distributions

Modelling species distributions has been widely used to understand present and future potential distributions of species, and can provide adaptation and mitigation information as references for conservation and management under climate change. However, various methods of data splitting to develop and validate functions of the models do not get enough attention, which may mislead the interpretation of predicted results. We used the Taiwanese endemic birds to test the influences of temporal independence of datasets on model performance and prediction. Training and testing data were considered to be independent if they were collected during different survey periods (1993–2004 and 2009–2010). The results indicated no significant differences of six model performance measures (AUC, kappa, TSS, accuracy, sensitivity, and specificity) among the combinations of training and testing datasets. Both species- and grid cell-based assessments differed significantly between predictions by the annual and pooled training data. We also found an average of 85.8% similarity for species presences and absences in different survey periods. The remaining dissimilarity was mostly caused by species observed in the late survey period but not in the early one. The method of data splitting, yielding training and testing data, is critical for resulting model species distributions. Even if similar model performance exists, different methods can lead to different species distributional maps. More attention needs to be given to this issue, especially when amplifying these models to project species distributions in a changing world.