Improving the predictability and interpretability of co-occurrence modelling through feature-based joint species distribution ensembles

Background/Question/Methods

Species Distribution Models (SDMs) are vital tools for predicting species occurrences and are used in many practical tasks including conservation and biodiversity management. However, the expanding minefield of SDM methodologies makes it difficult to select the most reliable method for large co-occurrence datasets, particularly when time constraints make designing a bespoke model challenging. To facilitate these practical out-of-sample prediction tasks, we propose a novel method that considers three major questions in community ecology: (1) how can we incorporate multiple functional forms for species associations; (2) how can we understand what characteristics of species co-occurrence data are driving model performance; and (3) how can we optimise community predictions without the need for building complex bespoke models? We propose an ensemble method that uses descriptive features of binary co-occurrence datasets to predict model weightings for a set of candidate SDMs. We demonstrate how this method may be applied through a simple case-study that uses five independent Joint Species Distribution Models (JSDMs) and Stacked Species Distribution Models (SSDMs) to predict out-of-sample observations for a diversity of co-occurrence datasets. Moreover, we introduce a novel SSDM that offers the potential to include multiple functional forms for each species while delivering robust community predictions.

Results/Conclusions

Our case-study highlights two major findings. First, the ability for the feature-based ensemble to offer more robust species co-occurrence predictions compared to individual candidate SDMs, while also providing insights into the data features that impact model performance. By providing interpretable outputs, this method can facilitate both community-level predictions using a combination of multiple SDMs, as well as individual model selection based on predicted performance. Second, this case-study demonstrates the competitiveness of the novel SSDM method for forecasting species co-occurrences, even when using a simple univariate Generalised Linear model (GLM) as the base model prior to stacking. This method provides an avenue for building bespoke models that can account for individual species variation when generating community-level predictions. Given these findings, we conclude that feature-based ensembles can provide ecologists with a useful tool for generating species distribution predictions in a way that is reliable and informative, and may help reduce computational time in the long-term when applying these methods to practical prediction tasks. Moreover, the flexibility of the ensemble and the novel SSDM method both offer exciting prospects for incorporating a diversity of functional forms while prioritising out-of-sample prediction.