Mapping species thermal suitability by integrating mechanistic ecological theory and deep learning

The Hutchinsonian shortfall, i.e. lack of knowledge about the tolerance of species to abiotic conditions, represents one major drawback besetting our understanding of biodiversity and its response to a changing environment. The environmental responses and tolerances of species are 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 (SDM), which infer environmental responses from occurrence data, yet offering limited causal insight. Here, we argue that the described strategies bring complementary information that can be integrated to better estimate species responses and more accurately map species distributions.

We select insect species whose ecology is strongly impacted by climate and are either important for ecosystem health (pollinators and biocontrol species) or disease vectors that threaten public health (Aedes mosquitoes). For each species, we retrieve from published studies the ecological model that estimates the trait-derived probability of occurrence as a function of temperature. Concurrently, we use occurrence records with average temperature data during the species activity season to train a neural network architecture (multi-layer perceptron) and estimate the SDM-based thermal response curves. The performance of the two approaches are compared on an independent test set, together with their respective thermal responses and identified thermal optimal and extremes. Among the two approaches, trait-based ecological models underperform deep learning SDMs in mapping distributions, possibly because the latter have been trained on the similar task of predicting species distribution. Interestingly, the thermal response curve of species occurrence seems consistently overestimated by the ecological model, suggesting that, when thermal functions generated in the lab are applied under real-world conditions, they might need correction with a shift towards lower temperatures. Last, we explore possible hybrid approaches, such as trait-guided deep learning or model ensembling, that can leverage both the mechanistic understanding of species ecology and the power of deep learning SDMs trained on the vast amount of community-based observational data.

Ultimately, we show how integrating experimental ecology with observational biogeography can lead to more accurate and ecologically grounded predictions for species distributions and their environmental responses to climatic changes.