Convolutional neural networks outperform other presence-only species distribution modeling algorithms

Species distribution models (SDMs) are essential tools for ecologists and conservationists because they can identify environmental determinants of species occurrences and predict species distributions. Unfortunately, most SDM algorithms rely on point-based or localized averages of environmental conditions and do not capture spatial heterogeneity and landscape patterns that shape species distributions. However, species-environment relationships are hierarchical, with different environmental factors shaping realized niche at different spatial extents. Using presence-only data, we evaluated whether convolutional neural networks (CNNs), which capture three crucial aspects of spatial context (heterogeneity, pattern, and multiscale relationships), outperform other SDM algorithms in predicting species distributions. We benchmarked CNNs against other widely used algorithms, including Maxent and ensemble models, and modeled 225 species from diverse geographic regions and taxonomic groups. We also assessed the efficiency of data augmentation in mitigating CNNs' sensitivity to limited training data. We found that CNNs consistently outperformed other algorithms. CNNs utilizing augmented data achieved median AUC_ROC of 0.77 and AUC_PRG of 0.78, compared to, for example, 0.74 and 0.61, respectively, for ensemble models. Notably, for rare species with <30 occurrences, CNNs with augmentation maintained high performance (AUC_ROC = 0.75), again exceeding ensemble models (AUC_ROC = 0.68). While CNNs required longer inference times, their model fitting was as fast as for other algorithms. Our results demonstrate CNNs' ability to incorporate multiscale spatial complexity and enhance predictive accuracy, particularly for data-limited species. CNNs have the potential to transform biodiversity modeling and inform conservation by enabling more spatially explicit and ecologically meaningful representations of the realized niche.