Harnessing the power of machine and deep learning for transferring joint species distribution models considering the structure of biotic interactions

The transferability of single or joint species distribution models ((j)SDMs) depends on their ability to predict beyond the observed environmental range and to remain consistent despite shifts in biotic interactions. Transfer accuracy may be improved by recent advances in the application of deep learning that provide greater flexibility and potentially superior predictive accu- racy than traditional approaches. We implemented jSDMs with deep and machine learning algorithms and measured the transfer accuracy from continental to regional areas in communities with different species composition. We ran jSDMs with deep neural networks (DNN), elastic net (EN), and stacked SDMs (sSDM) with random forests (RF). We used 134 689 occurrence records representing 1776 species of six taxonomic groups (beetles, birds, bryophytes, fungi, lichens and plants) from 2387 forest plots in Europe. We employed an agnostic modelling approach that covered most of the environmental con- ditions by including more than 100 satellite-derived variables and 98 climatic variables. The predictive power of the models within the training continental area was evaluated using AUC, whereas the transfer accuracy in the regional area was evalu- ated with the Boyce index calculated with independent presence records. We found that the DNN–jSDMs outperformed other models at continental scale, but model transfer from continental to regional extent was less accurate. We found that the accuracy of regional predictions was higher for taxonomic groups with better representation in the continental data, such as birds, bryophytes and plants. Depending on the algorithm and the taxonomic group, we achieved acceptable (Boyce > 0) to accurate (Boyce > 0.5) transferability for 32–78% of the species. Our findings underscored the need of considering trade-offs among hyperparameter tuning, spatial scales and model complexity. Our findings also suggest that the varying biotic interac- tion structures and, particularly, the different species compositions of the transfer areas, may affect model transferability more than previously considered.

Peer Review, Open Access, PDF available, RBINS Collection(s)

artificial intelligence, biotic interactions, deep neural networks, ecological niche, jSDM, random forest