Climate‐informed models benefit hindcasting but present challenges when forecasting species–habitat associations

Abstract

Although species distribution models (SDMs) are commonly used to hindcast finescale population metrics, there remains a paucity of information about how well these models predict future responses to climate. Many conventional SDMs rely on spatiallyexplicit but time-invariant conditions to quantify species distributions and densities. We compared these status quo 'static' models with more climate-informed 'dynamic' SDMs to assess whether the addition of time-varying processes would improve hindcast performance and/or forecast skill. Here, we present two groundfish case studies from the Bering Sea -a high latitude system that has recently undergone considerable warming. We relied on conventional statistics (R 2 , % deviance explained, UBRE or GCV) to evaluate hindcast performance for presence-absence, numerical abundance and biomass of arrowtooth flounder Atheresthes stomias and walleye pollock Gadus chalcogrammus. We then used retrospective skill testing to evaluate near-term forecast skill. Retrospective skill testing enables direct comparisons between forecasts and observations through a process of fitting and forecasting nested submodels within a given time series. We found that the inclusion of time-varying covariates improved hindcasts. However, dynamic models either did not improve or decreased forecast skill relative to static SDMs. This is likely a result of rapidly changing temperatures within the ecosystem, which required models to predict species responses to environmental conditions that were outside the range of observed values. Until additional model development allows for fully dynamic predictions, static model forecasts (or persistence forecasts from dynamic models) may serve as reliable placeholders, especially when anomalous conditions are anticipated. Nonetheless, our findings demonstrate support for the use of retrospective skill testing rather than selecting forecast models a priori based on their ability to quantify species-habitat associations in the past.

Recommendations for quantifying and reducing uncertainty in climate projections of species distributions

“…However, for many mobile species, there may be a point at which historical spatial relationships begin to break down and no longer accurately predict species distributions, thus care should be taken when interpreting projected SDMs that contain these spatial structures and perhaps supplement model evaluations with expert opinion and guidance (e.g., Warren et al, 2020 ). Indeed, Barnes et al ( 2022 ), showed more complex SDMs improved model fit but failed to skillfully forecast species distributions. Furthermore, our results indicated that not capturing the appropriate mechanisms driving species distributions can lead to poor model performance when projecting.…”

“…We primarily attribute this to the lesser flexibility of covariate responses, compared to the nonlinear splines, trees, or neural networks seen in other model types, and potential effects of spatial fields and spatially autocorrelated covariates. We acknowledge that the simulation framework may have unfairly considered GLMMs given that more flexible responses (e.g., splines) can be incorporated into such models and additional model structure can better help to resolve complex ecological processes (Barnes et al, 2022 ; Barnett et al, 2021 ). Our results indicate that the impact of extrapolation on model performance is difficult to predict, and more research is needed on methods for measuring and improving extrapolation, and the trade‐offs between resolving ecological processes and more accurately defining response curves (e.g., Brodie et al, 2020 ).…”

“…For example, species distribution models (SDMs) have become a common tool for resource managers to describe and predict species distributions as a function of various biotic and abiotic factors, thereby providing critical insight into core habitats, range shifts, habitat connectivity, and potential impacts of anthropogenic pressures (Elith & Leathwick, 2009 ; Guisan & Thuiller, 2005 ; Robinson et al, 2017 ). SDMs have been used extensively for predicting and projecting changes into the future (Robinson et al, 2017 ), but not all models perform well when applied to novel conditions (Barnes et al, 2022 ; Muhling et al, 2020 ), which raises concerns about the realism of SDM projections. This is particularly pertinent as novel climate conditions emerge (Smith et al, 2022 ) because SDMs trained on historical conditions typically have less skill at predicting into novel environmental conditions (Muhling et al, 2020 ).…”

An ensemble approach to species distribution modelling reconciles systematic differences in estimates of habitat utilization and range area

“…While the ensemble approach described here is useful for predicting large‐scale patterns across several species, we recommend that future work explores SDMs that can handle dynamic environmental covariates (e.g. Barnes et al., 2022) and the mechanistic links between environmental and biological processes (Thorson et al., 2021), that incorporate life history processes including life stage‐specific habitat needs and that provide more accurate predictions for species with unique distributions (e.g. a strip of habitat within an ecosystem).…”

Species redistribution creates unequal outcomes for multispecies fisheries under projected climate change

“…S1, S4, S7, and S10) (84). Although the models varied in their percent deviance explained relative to an intercept-only null model, note that higher deviance explained for historical data does not always indicate strong predictive performance for SDM projection models under climate change (54,85). We instead quantified the predictive ability of each model for each species by splitting data into a test and training set and using the test set to compute the predictive density.…”