Improving species distribution models by optimizing background points: impacts on current and future climate projections.

Abstract

Species Distribution Models (SDM) are often fit using presence-background data due to the lack of reliable absence records. To calibrate these models, background records are required, yet the optimal number of records and if they should be proportional to study area or the number of occurrences remains uncertain. This study addresses three key questions: (i) how does varying background proportions affect predictive accuracy? (ii) How do background proportions influence future species distribution projections under climate change? and (iii) should the number of background records be determined based on study region size or presence record availability? To investigate these questions, we simulated 280 virtual species distributions worldwide under present and future climate conditions. Model outputs were evaluated against simulated “true” distributions under both present and future scenarios. Results indicate that sampling background records proportional to either presence points or study area yields comparable average performance. Optimal performance occurred with a 0.5–1 ratio of background records to presence points when sampled proportionally to presences, and with approximately 5 % of the study area sampled when proportional to region size. Species prevalence also modulated the optimal presence-background ratio. Increasing the number of background records across suitable and unsuitable areas had contrasting effects for both strategies tested, emphasizing the need to assess model performance separately for both. Notably, background proportions influenced baseline predictions but had minimal impact on future projections, where niche-related variables dominated model performance. These findings offer practical insights for SDM practitioners. Adjusting background sampling strategies enhances current prediction accuracy, while future projections remain robust across different sampling approaches, ensuring more reliable modelling outcomes.