Explainable neural networks for trait-based multispecies distribution modelling—A case study with butterflies and moths

Species response traits mediate environmental effects on species distribution. Traits are used in joint and multispecies distribution models (JSDMs and MSDMs) to enable community-wide shared parameters that characterise niche filtering along environmental gradients. Multispecies machine learning SDMs, however, do not use traits as their inclusion requires an additional taxonomic dimension that is incompatible with their usual tabular inputs. This has confined trait mediation in SDMs to hierarchical Bayesian models. Here we provide a novel artificial neural network (ANN) architecture that solves this dimensionality problem. Our ANN includes species traits (via a time distributed layer) and is therefore able to identify not only species-specific responses to the environment, but also shared responses across the community that are mediated by species traits. Model performance evaluated at the species level not only quantifies the reliability of species predictions, but also their departure from an average response dictated by traits only. We apply our model to two unique long-term spatio-temporal of butterfly and moth datasets collected across the United Kingdom between 1990 and 2019. In addition to species traits, predictors include numerous metrics derived from weather, land-cover and topology data. For butterflies and moths we show convincing model performance for classifying species occupancy. We use SHAP (Shapley Additive exPlanations) to explain the ANN and show how trait-mediated and species-specific responses can be approximated, hence yielding ecological insights on the key drivers of species distribution. We highlight a range of drivers of change that determine occupancy, including wind, temperature as well as habitat type. We demonstrate that a trait-based approach can be encoded as an ANN by using a time distributed layer. This brings ANNs unmatched predictive capabilities to the field of MSDMs, at the same time of lifting their reputed drawback of poor explainability.