The interplay of various sources of noise on reliability of species distribution models hinges on ecological specialisation.

Digitized species occurrence data provide an unprecedented source of information for ecologists and conservationists. Species distribution model (SDM) has become a popular method to utilise these data for understanding the spatial and temporal distribution of species, and for modelling biodiversity patterns. Our objective is to study the impact of noise in species occurrence data (namely sample size and positional accuracy) on the performance and reliability of SDM, considering the multiplicative impact of SDM algorithms, species specialisation, and grid resolution. We created a set of four 'virtual' species characterized by different specialisation levels. For each of these species, we built the suitable habitat models using five algorithms at two grid resolutions, with varying sample sizes and different levels of positional accuracy. We assessed the performance and reliability of the SDM according to classic model evaluation metrics (Area Under the Curve and True Skill Statistic) and model agreement metrics (Overall Concordance Correlation Coefficient and geographic niche overlap) respectively. Our study revealed that species specialisation had by far the most dominant impact on the SDM. In contrast to previous studies, we found that for widespread species, low sample size and low positional accuracy were acceptable, and useful distribution ranges could be predicted with as few as 10 species occurrences. Range predictions for narrow-ranged species, however, were sensitive to sample size and positional accuracy, such that useful distribution ranges required at least 20 species occurrences. Against expectations, the MAXENT algorithm poorly predicted the distribution of specialist species at low sample size.