<scp>CAMELIA:</scp> Enhancing species assemblage predictions by integrating structural and functional indices: A case study on plant communities in the French Alps

Abstract Addressing the challenge of predicting biodiversity requires models that fully capture the complex dynamics of species and communities. While species distribution models (SDMs) predict individual species' distributions based on environmental variables, they often fail to reconstruct realistic communities. Macroecological models (MEMs), on the other hand, predict community indices but lack species‐level resolution. To bridge this gap, we introduce CAMELIA (Community Assemblage Modelling macroEcoLogically Informed with AI), a neural network framework integrating species distribution data with community constraints to improve biodiversity predictions. Approach and Methods. CAMELIA uses a multitask neural network to jointly optimize species occurrence probabilities and community indices. It recalibrates species probabilities from stacked SDMs using observed or predicted community indices. We tested CAMELIA on plant communities in the French Alps, using MEM predictions (CAMELIA‐MEM), high‐quality field data (CAMELIA‐OBS) and degraded field data (CAMELIA‐PROXY) for various community indices: species richness, community mean and community standard deviation for plant height, specific leaf area and leaf nitrogen content. CAMELIA substantially improved community predictions over stacked SDMs. With observed indices, it reconstructed communities with functional indices highly correlated with observations (for community mean plant height: R 2 = 0.98 vs. 0.74 with S‐SDMs), reduced the mean errors in predicting species richness by 38%, and increased the Sørensen coefficient by an average of 33%. Even with lower‐quality data, CAMELIA‐MEM outperformed stacked SDMs in community standard deviation ( R 2 +6.9%) and species composition (+10.6% for Sørensen). With noisy community indices (CAMELIA‐PROXY), CAMELIA remained robust, with R 2 values for all indices falling between CAMELIA‐MEM and CAMELIA‐OBS, often closer to the latter (e.g. R 2 for community mean plant height = 0.92), demonstrating resilience to imperfect data. CAMELIA provides a flexible and effective framework for improving community structure predictions. By optimizing species probabilities to align with community constraints, it produces more ecologically realistic assemblages than stacked SDMs. Its adaptability to different data qualities makes it a valuable tool for biodiversity modelling. Future developments should explore the integration of remotely sensed biodiversity indices to further enhance predictions, particularly for conservation applications.