Background selection complexity influences Maxent predictive performance in freshwater systems

Absence data are often lacking for species distribution modeling (SDM) purposes. This necessitates selecting background or pseudo-absence observations that influence SDM performance. Little is understood about how background selection affects SDM prediction in lode systems. Here we test six background selection methods that implement different combinations of three selection filters concerning 1) sampling biases in species occurrence data, 2) geographic restriction to regions accessible to the species modeled, and 3) species occurrence relative to stream size, a key habitat factor. These six methods are used with Maxent to develop binary presence-absence predictions of 71 freshwater mussel distributions in the Midwestern United States. Prediction accuracy was evaluated with a separate validation presence-absence dataset derived from intensive surveys. Pairwise comparisons of background selection methods across species recorded in the validation dataset revealed significant differences relative to the Area Under Curve (AUC), the similarity between the prediction and observation, and the True Skill Statistic (TSS) metrics. The prediction specificity for those species absent in the validation dataset was also significantly affected by the background selection method. Implementing the sampling bias filter increased prediction similarity with validation data, AUC and TSS for species with validation presences, as well as prediction specificity for species without validation presences. Our results provide much needed insight into how background selection influences presence-background SDM performance in lotic systems. These findings can guide how to leverage available data and biological understanding to produce accurate SDM predictions that prioritize research objectives and goals regardless of study system or habitat.