Synergistic mechanistic and deep learning approaches for modeling insect thermal responses and geographic distributions
In the authors' words
Aim The Hutchinsonian shortfall, the lack of knowledge about species' abiotic tolerance, limits our understanding of species responses to environmental change and consequently also the outlining of their distribution in space and time. Such environmental tolerance is commonly inferred with two distinct approaches: i) trait performance studies, which use lab experiments to model species' functional responses, often overlooking real-world complexity, and ii) species distribution models, which infer environmental suitability from occurrence data, but offer limited causal insights. In this paper, we argue that these two strategies bring complementary information that can be integrated to estimate species-environment relationships more robustly and accurately map species distributions. Location Global scale Time period 2000-2019 Major taxa studied Given their exceptional diversity and climatic sensitivity, insects provide a powerful system for studying these responses. We selected key insect species representing i) disease vectors: the yellow fever (Aedes aegypti) and Asian tiger (Aedes albopictus) mosquitoes; ii) agricultural pests: the cotton bollworm (Helicoverpa armigera) and the fall armyworm (Spodoptera frugiperda); and iii) biocontrol agents: the green lacewing (Chrysoperla carnea) and the 14-spotted ladybug (Propylea quatuordecimpunctata). Methods We compared predicted insect thermal suitability and distributions derived from two complementary methodological approaches: mechanistic models based on experimentally derived thermal responses of ecological rates, and deep learning-based species distribution models based on occurrence data. We further tested hybrid frameworks integrating mechanistic and correlational information through model ensembling, mechanistic regularization, and mechanistically informed deep learning. We assessed the efficacy of these methods under increasing data scarcity by progressively removing up to 80% of species occurrence records using three strategies: removal from the warmest or coldest parts of species' thermal ranges, removal from the least-sampled locations based on geographic sampling bias, and random removal as a baseline. Results Our findings highlight that occurrence-derived thermal suitability closely reproduces the observed distributions of insects. Compared to the mechanistic estimates, these thermal response curves are consistently shifted towards lower temperatures, potentially reflecting niche underfilling. However, when challenged with progressive data removal, correlative methods fail to extrapolate reliably in data-deficient regimes, where mechanistic knowledge becomes increasingly more crucial. Hybrid approaches in turn leveraged complementary information, improving predictive performance and reducing the sensitivity of environmental suitability estimates to sampling biases. Main conclusions Ultimately, we show how the synergistic integration of experimental ecology with observational biogeography can lead to more accurate and ecologically robust predictions for species distributions and their environmental responses to climatic changes.
Appeared: Wednesday, November 18. bioRxiv. Preprint, not yet peer-reviewed.