Species‐observer link and kernel density estimation of background points allow for sampling bias correction in bird species distribution models

Species distribution models (SDMs), broadly referring to both species distribution and ecological niche modelling frameworks, are widely used to predict habitat suitability. However, their performance can be biased by uneven sampling effort in occurrence data. Building on two existing approaches, we propose a novel method for sampling bias correction, consisting of the estimation of observer kernel densities for individual species and their subsequent weighting according to the relative contribution of individual observers to the total number of focus species presences. This approach, the ‘presence-weighted observer-oriented approach' (PW-OOA), aimed to provide a better estimation of sampling effort, thus further improving SDM prediction performance. Using bird occurrence data from the Czech Republic, we modelled the distributions of 109 species using four approaches to bias correction: spatial thinning of species presences (STSP), target group occurrences background (TGOB), TGOB+ (tuned up by adjusting kernel smoothing bandwidths) and the new PW-OOA method. We compared the results with simple random background sampling. Models were evaluated using independent reference (presence–absence) data. The PW-OOA method outperformed the other approaches, with the greatest improvement detected for species with higher prevalence. However, as internal validation can be misleading with biased occurrences, we recommend TGOB+ as the most robust approach without independent data; with such data, PW-OOA is superior. While no single optimal combination of bandwidth and observers' weights was identified across species, the PW-OOA method provides a flexible framework to account for observer-specific sampling biases. This study demonstrates the crucial importance of considering the behavior of individual observers and sampling intensity smoothing when correcting for sampling bias in SDMs based on unstructured opportunistic occurrence data.