Mapping species assemblages at a national scale: Practical lessons from scalable joint Species Distribution Modelling

Scalable joint Species Distribution Models (sjSDMs) are a statistical tool that can estimate the habitat suitability of multiple species simultaneously and their interspecific associations. When used with unbiased presence–absence data and relevant environmental variables, sjSDMs can be applied to both ecological understanding and ecosystem management. We tested two possible applications of this tool: inferring species interactions from co- occurrence patterns and creating distribution maps of species assemblages at a national scale. To do this, we used an opportunistic dataset comprising over 1.5 million georeferenced observations, mostly from citizen science. We included 249 wild bee species and 462 flowering-plant species in the Netherlands, exploring for the first time the potential of this approach for wild bees at such a large scale. To account for the limitations of a presence-only training dataset, we generated pseudo-absence data from well-sampled cells. We compared the association matrix—an output of sjSDMs showing the co-occurrence patterns of the modelled species—to a bee–flower visitation dataset in the Netherlands. Contrary to our expectations, higher values in the association matrix were not correlated with records of those species pairs in the visitation dataset. Moreover, the distribution predictions of the sjSDMs had accuracy metrics similar to those of the same model when ignoring co- occurrence patterns. Our results show that sjSDMs are not inherently more accurate than other SDMs and did not provide relevant ecological insights into the bee–flower community of our study system. In this presentation, we introduce these results and discuss how sampling biases, resolution, and environmental variables are critical to the efficacy of sjSDMs. We also offer a few comments on promising approaches that combine SDMs with other statistical and monitoring tools, such as spatial random-effect models, observational models, and remote sensing.