Effectiveness and implications of spatial background restrictions on model performance and predictions: a special reference for Rattus species

Abstract Controlling background data selection in presence-only models is crucial for addressing sampling biases and enhancing model performance. While numerous studies have evaluated the impact of various background data selection techniques across different taxa, research remains limited on how spatially restricted background areas and employing random and biased distribution methods, influence model performance for Rattus species predictions. These species often present challenging collection conditions and low trap success rates, potentially leading to spatial biases in the occurrence records that may affect the accuracy of model predictions. Thus, this study examined methods to assess model accuracy variability for Rattus species by applying spatial background restrictions within the study area. These restrictions were defined by four main criteria: (1) areas within islands with documented species occurrences, (2) areas within the species’ extent of occurrence according to IUCN range maps, (3) defined road distance, and (4) varying buffer areas around recorded species occurrences. To further assess the effects of spatial background restrictions on model performance, we used two methods to distribute the background sampling points: random and biased (bias file) method. Our findings demonstrated that the selection of spatial background restrictions and the distribution methods for background sampling points play a critical role in influencing model performance and the accuracy of predicted habitat suitability for Rattus species. Our findings highlight that defining a specific spatial restriction, such as restricting background selection to within 5 km of a road, improves model performance. However, overly narrow or restrictive buffer sizes, such as the 20 km buffer size used in this study, fail to capture the full environmental variability of the species, which can diminish model accuracy. Furthermore, the method used to distribute background sampling points whether random or biased affects species predictive outcomes. To ensure reliable predictions, we recommend a systematic evaluation of different spatial restriction methods and distribution approaches, along with a thorough analysis of their impacts on model performance. This approach not only reveals how outcomes vary across different modeling scenarios but also provides a strong basis for determining the most reliable predictions. By carefully assessing these factors, researchers can refine and optimize habitat suitability models for Rattus species, ultimately enhancing predictive accuracy and ensuring more consistent and dependable results.