Multi-scale neural networks enhance species distribution modelling across predictors and taxonomic groups

Ecological processes and patterns are scale-dependent, as reflected in how relationships between species occurrence and the environment vary across scales. In species distribution models (SDMs), spatial scale can strongly affect model predictions and performance. Despite this, the choice of scale is often overlooked. The development of SDMs using deep learning models has enabled spatially structured data, such as patches of environmental raster data, to be considered as model inputs, making the question of scale even more crucial. Here, we evaluate convolutional neural network-based multi-species SDMs considering different scales on a dataset from Switzerland, including more than 8 million observations for 2390 plant and 1006 animal species comprising amphibians, butterflies, beetles, and mammals. We investigate how scale affects model performance and compare single- and multi-scale models. Our results reveal stronger scale effects for remote sensing indices than for bioclimatic and edaphic variables. In single-scale models, we find that plant species perform better with smaller spatial scales, while the opposite is true for animal species. Multi-scale models consistently improve predictive performance and reduce sensitivity to arbitrary scale selection across all modalities and taxonomic groups. The effect of scale is nonetheless smaller than the increase in performance achieved by considering multiple predictor groups. To improve the interpretability of such complex models, we use attribution methods to compute the relative contribution of different scales and predictor groups. Our study demonstrates the importance of accounting for scale and the potential of deep learning methods to integrate ecological complexity with spatial data across scales.