||Data contained in biological atlases are typically not collected using standardized survey techniques that allow for robust estimation of species detection probabilities. Experts realize that there are spatial patterns in detection probabilities, due to for example spatial variation in the accessibility of terrain or numbers of volunteers. This greatly complicates estimation of the environmental (such as climatic) distribution of species if spatial patterns in detection probabilities are correlated with spatial patterns in environmental variables. However, existing methods to analyze biological atlas data commonly assume that there are no such spatial patterns in detection probabilities. We discuss how Bayesian Image Restoration (BIR) techniques can be used to relax this assumption by jointly estimating the environmental distribution and spatially varying detection probabilities of species. The central idea behind BIR is that location-specific detection probabilities can be estimated by relating them to spatial covariates, hypothesized by experts to be related to detection probabilities. We explain the theory underpinning BIR, and how to fit autologistic models (a common model used to estimate species distributions in relation to environmental variables) within a BIR framework using WinBugs (a freely available software package commonly used to implement Bayesian methods). Implementation of BIR is demonstrated by fitting the autologistic models to maps of recorded presences of plant species of the German atlas of vascular plants (FLORKART). We were able to estimate species- and location-specific detection probabilities, and estimated that gradients in species detection probabilities were correlated with environmental variables used in the autologistic models to predict probabilities of species occurrence. We conclude that BIR can be used to investigate the existence of spatial patterns in detection probabilities and to test the reliance of inferences on the distribution of species to assumptions concerning detection probabilities. In the absence of a standardized survey design, such inferences may depend strongly on the particular chosen model, and this is made explicit within the BIR framework. When the true distributions of species are obscured by spatial patterns in detection probabilities, BIR techniques have the potential to provide better distribution estimates than conventional methods, if reliable covariates relating to detection probabilities are available.