A trait-informed deep learning model to map species distributions and assemblages of Swiss flora

Biodiversity models extend predictions of species assemblages beyond direct observations and play a key role in applications such as conservation planning. Yet, predicting the spatial distribution of species assemblages remains a persistent challenge in biogeography, as both abiotic constraints and biotic interactions shape community composition through complex interactions. Multi-species distribution models are promising because they can integrate information about species co-occurrence, but they remain largely limited to environmental filtering. Functional trait data, such as morphological, physiological, or phenological characteristics, offer considerable yet underutilized potential to inform models about species responses to the environment and ecological assembly rules. Recent deep learning-based approaches to species distribution modelling provide a promising avenue in this context, as they can model large numbers of species simultaneously and offer a flexible framework for integrating heterogeneous data sources. 

Here, we propose a two-step deep learning approach to incorporate traits into assemblage predictions. In a first step, we model trait-informed habitat suitability by training a deep learning species distribution model using opportunistic occurrence records, high-resolution environmental predictors, and species-level trait data. In a second step, we use vegetation plot data to train a transformer-based model that predicts species assemblages from these habitat suitability score predictions and learned trait representations. By leveraging attention mechanisms, the model captures associations between species, potentially accounting for biotic dependencies and co-occurrence structure. We evaluate our framework on plant assemblages across Switzerland. Our results show that attention-based modelling improves the prediction not only of individual species distributions, but also of assemblage composition across landscapes. From these predictions, we also obtained more accurate estimations of community indices, such as species richness and community-weighted trait statistics. Overall, this study demonstrates how the flexibility of deep learning enables application-driven integration of diverse data sources. It highlights the value of combining large-scale presence-only observational data with a smaller, but high-quality set of vegetation plots. Our approach aims to bridge species distributions modelling and trait-based community ecology, offering scalable and ecologically grounded predictions of biodiversity patterns.