Accounting for dispersal and intraspecific variation in forecasts of species distribution under climate change

Dispersal and local adaptation play an important role in driving species distributions under climate change. Many studies aim to estimate relationships between species occurrences and environmental variables to predict range shift and biodiversity loss, ignoring dispersal and intraspecific variation contributing to complex spatial and temporal dynamics. We accounted for dispersal and intraspecific variation in forecasts of species distribution under climate change with species distribution models (SDMs) using two cold‐adapted, low‐dispersal Platycerus species (Coleoptera: Lucanidae), each with distinct subspecies distributions, as focus species. The results showed that the subspecies‐level model performed significantly better than the species‐level model when considering dispersal constraints in SDMs. Whether or not dispersal or intraspecific variation is accounted for, the predicted species range of Platycerus albisomni is expected to decrease in the future. For Platycerus takakuwai, accounting for dispersal constraints in SDMs indicated that its potential distribution area would increase at the subspecies level under climate change, but decrease at the species level. These divergent results show that SDMs at the subspecies level can detect impacts of climate change that may be overlooked in species‐level models. Therefore, models that consider intraspecific variation and dispersal constraints may provide a more realistic perspective on the impacts of climate change. Because accurate mapping of potential habitats is needed for conservation purposes, demographic studies should include dispersal explicitly and explore how and when intraspecific variation in dispersal affects local population dynamics. This approach could help evaluate species' habitat shifts, thus enabling suitable conservation strategies.


INTRODUCTION
Climate change poses serious threats to biodiversity and ecosystems, as it imposes shifts in geographical distribution, biotic interactions, phenotypic plasticity and adaptation (Benito Garz on et al., 2019;Gárate-Escamilla et al., 2019;Thompson et al., 2013).The ability to track climate changes and persist in suitable habitats is crucial for a species to avoid extinction, as future extinction risk is often associated with range contractions (Borges et al., 2019;Urban, 2015).Thus, accurately predicting the potential distributions of species under climate change is essential for designing conservation strategies to protect climate-vulnerable species (Melo-Merino et al., 2020).Species distribution models (SDMs), also known as ecological niche models or habitat suitability models, are statistical tools that associate species occurrence with environmental variables to predict species range shifts in past or future climates (Elith et al., 2006;Guisan & Thuiller, 2005).SDMs have become an important tool to understand a broad range of ecological questions, from the effects of climate change to challenges for the management of threatened species (Guillera-Arroita et al., 2015;Guisan et al., 2013;Hao et al., 2019).
Under climate change, species must disperse into newly suitable habitats as rapidly as climate shifts across landscapes (Lemes et al., 2022).Urban (2015) showed that dispersal ability significantly affected extinction risk.Many SDM studies integrate dispersal ability into SDMs by considering dispersal within a relevant timeframe, the existence of dispersal barriers, species-specific dispersal rates throughout evolutionary time or the delimitation of accessible areas (Chapman et al., 2019;Mendes et al., 2020;Seaborn et al., 2020;Shipley et al., 2022).Furthermore, intraspecific variation with environmental differences in niche preference provides strong evidence that climate tolerances vary throughout a species' range (DeMarche et al., 2019;Lecocq et al., 2019).Therefore, both dispersal constraints and intraspecific variation can contribute to the complex range dynamics, especially when coupled with the effects of climate change.
More studies carrying out comprehensive analyses incorporating intraspecific variation and dispersal constraints are needed to gain a better understanding of range shifts during periods of intense environmental climate change.
Previous research has examined the effects of dispersal ability and intraspecific variation in isolation.In stag beetle species (Platycerus species), Zhang and Kubota (2021a) showed that dispersal constraints played a key role in species distribution and improved model performance for dispersal-limited Platycerus species.Zhang and Kubota (2021b) demonstrated the importance of intraspecific variation in the environmental space and assessed the predicted suitability differences at the species and subspecies levels, indicating that subspecies-based SDMs were more accurate.Here, we predicted distribution shifts under climatic change in consideration of both factors of two cold-adapted Platycerus species: Platycerus albisomni Kubota et al., 2008 andPlatycerus takakuwai Fujita, 1987.Both species live in cool-temperature deciduous broad-leaved forests and are particularly sensitive to climate change, as well as having limited dispersal ability (Zhang & Kubota, 2021a;Zhu et al., 2022).Future changes in the geographic distribution at the sub-specific level may thus affect the distribution at the species level because of their local adaptation and limited dispersal abilities.We anticipate that our results will help to realistically predict range shifts under climate change and provide a robust basis for management strategies.By consolidating dispersal information and intraspecific variation, we gain greater insight into the overall utility and limitations of SDMs in ecology and conservation biology.
Climate data were generated according to Zhang and Kubota (2021a).Ten variables were selected with a spatial resolution of 2.5 arc-min from the WorldClim database (v.1.4;http://www.worlclim.org; Hijmans et al., 2005) using the "Present" period: mean diurnal range (Bio2), temperature seasonality (Bio4), maximum temperature of the warmest month (Bio5), mean temperature of the wettest quarter (Bio8), mean temperature of the driest quarter (Bio9), mean temperature of the coldest quarter (Bio11), precipitation in the wettest month (Bio13), precipitation seasonality (Bio14), precipitation in the warmest quarter (Bio18) and precipitation in the coldest quarter (Bio19).These variables represent temperature and precipitation measures that are not strongly correlated and are particularly significant in determining species distribution (Zhang & Kubota, 2021a).Future climate conditions were obtained from global climate model data from MIROC-ESM (http://www.worldclim.com/cmip5_2.5m)under the Representative Concentration Pathway 8.5 (RCP 8.5) scenario, which represents the highest predicted level of greenhouse gas emissions in 2070 (IPCC, 2014).Species distribution models were calculated using the maximum entropy algorithm implemented in Maxent (Version v.3.4.1;Phillips et al., 2006Phillips et al., , 2017)), a general-purpose machine learning method that uses presence-only occurrence and background data and is particularly suited to noisy or sparse information (Elith et al., 2006).Maxent has been shown to perform well in comparison to other methods for modelling species distributions, especially when the data are presence only.From the area delimited as our study area, 10,000 background data points were randomly obtained.We ran Maxent with default parameters for 10 replicates.A fivefold cross-validation was set to assess the predictive performance of the Maxent.In each repetition, 75% of the occurrences were selected as training data and 25% for testing.Model performance was tested using the area under the receiver operating curve (AUC; Fielding and Bell, 1997).AUC values range between 0 and 1, with 0.5 representing random discrimination and 1 representing perfect discrimination.However, the AUC values are sensitive to the proportion between the extent of the species' distribution and the extent of the study area (Lobo et al., 2008).Therefore, true skill statistic (TSS; Allouche et al., 2006), Kappa (Cohen, 1960) and Boyce index (Hirzel et al., 2006) were used to compare model preference.AUC, TSS and Kappa were calculated using the default setting of 'biomod2', and Boyce index was calculated using the 'ecospat' package in R (Di Cola et al., 2017;Thuiller et al., 2016).We used Mann-Whitney U tests with Holm-Bonferronicorrected p-values to detect differences among models with dispersal and intraspecific variation.To dynamically explore the consequences of climate change over time, we used dispersal kernels to represent the dispersal constraint of each subspecies.This method describes the probability of dispersal from the end location of a disperser to the source point (Engler et al., 2012).Source points were defined by the mid-range longitude and latitude of the locations of each subspecies (MR scenario, Zhang & Kubota, 2021a).The dispersal constraint factor was implemented with ArcGIS 10.4.1 using the 'Euclidean Distance' function of the 'Spatial Analyst Tools'.Then we converted it into raster layer as ASCII output projection and added it to the SDM as an additional model predictor (Figure S1).
To evaluate the importance of dispersal and intraspecific variation in suitable habitats under climate change, we examined four SDM scenarios: (i) SDMs with no dispersal constraint at the species level and subspecies level in the present period; (ii) SDMs with dispersal constraint at the species level and subspecies level in the present period; (iii) SDMs with no dispersal constraint at the species level and subspecies level in the future period; and (iv) SDMs with dispersal constraint at the species level and subspecies level in the future period.The combined ranges of subspecies models (subspecies level) results were obtained by merging each subspecies model in ArcGIS v.10.4.1 using the merge tool.The area of the suitable habitats at the species and subspecies levels in the present and future climatic scenarios was computed based on the number of grid cells among projected climatic extent where species exhibited moderate to high presence probability (i.e., >0.2) using spatial analyst tools in ArcGIS 10.4.1.

RESULTS
Overall, the models performed well when assessed with the AUC, TSS, Boyce index and Kappa (Table 1).SDMs with dispersal constraints had higher AUC values and Boyce index than those that did not consider dispersal at the species level and subspecies level (Mann-Whitney U test, p < 0.05), whereas TSS and Kappa did not exhibit the same results.When accounting for dispersal constraints, subspecies-level models had higher performance than species-level models (Table 1); dispersal-constrained models had less potential area for both the species and subspecies at present and future scenarios.
The potential distribution of P. albisomni was greater under the species-level model than under the subspecies-level model at present and future conditions when dispersal was considered (Figures 1   and S2).
The potential habitat area was smaller for the species-level models than for the subspecies-level models for P. albisomni in present conditions with no dispersal constraints (Figure S2).For P. takakuwai, the potential habitat area was smaller in the subspecieslevel models than at the species level when not accounting for dispersal constraints in SDMs.When considering dispersal constraint, the potential distribution in subspecies-level models was greater than in species-level models under present conditions (Figure S2).
P. takakuwai was predicted to increase at the subspecies level under climate change, but to decrease at the species level.Moreover, the potential distribution area was twice as large at the subspecies level than at the species level under future condition (Figures 1 and S2).

DISCUSSION
Climate change has profoundly impacted range shifts and intraspecific ecological variation via local adaptations of species (Aitken & Whitlock, 2013;Atkins & Travis, 2010;Thuiller, 2007;Valladares et al., 2014).We evaluated model performance using four metrics by SDM with and without dispersal and intraspecific variation.Based on AUC values and Boyce index results, model performance was better for SDMs that considered dispersal and intraspecific variation than for models that considered only a single factor.However, this was not the case for TSS and Kappa values.The different metrics have varying T A B L E 1 Mean value of area under the receiver operating characteristic curve (AUC) value, true skill statistic (TSS), Boyce index and Kappa for species distribution models of P. albisomni and P. takakuwai with no dispersal constraints and those incorporating dispersal constraints.abilities to capture different aspects of model evaluation, which may explain the observed result.A multimetric approach, with its capacity to unravel the impact of the different factors that affect the score of a single-value statistic, may more accurately identify the skill of a prediction and where its weaknesses are.Interestingly, when accounting for dispersal constraints, the distribution of Platycerus takakuwai was predicted to increase at the subspecies level under climate change, but to decrease at the species level.These divergent results show that SDMs at the subspecies level can detect impacts of climate change that may be hard to detect in species-level models, although the model does not take into account the biotic interactions among subspecies (e.g., competitive exclusion due to common resource use).
Here, we evaluated the interactions between two factors, namely the dispersal and environmental diversity within species, in SDMs instead of each factor in isolation to better understand the consequences of climate change.For most species, dispersal has an important impact on fitness, community composition and patterns of biodiversity (Hessen et al., 2019;Kroiss & Hillerslambers, 2015).Many studies rely on estimates of dispersal at species level to predict range shift and biodiversity loss, ignoring the dispersal variation within species caused by complex spatial and temporal dynamics (Snell et al., 2019).However, this variation could contribute to gene flow and long-distance dispersal and has important consequences for understanding and predicting population dynamics (Binks et al., 2019).
For example, intraspecific variation in dispersal can affect variation in demography via growth and survival rates, and consequently determine population growth (Janzen, 1970;Snell et al., 2019).In addition, this variation in dispersal can impact genotypes and phenotypic traits, affecting local dynamics within and among populations (Castorani et al., 2017).In the future, researchers should pay close attention to demographic studies that include dispersal explicitly and explore how and when intraspecific variation in dispersal affects local population dynamics to better evaluate species' habitat shifts.
The importance of SDMs for conservation management is indisputable (Guisan et al., 2013;Velazco et al., 2020).However, SDM outputs are sensitive to input assumptions at all stages of the modelling process (Muscatello et al., 2021;Watling et al., 2015).One major source of input assumptions is the environmental data used to develop SDMs.These data are often derived from remotely sensed imagery or interpolated from weather station data, and they can be subject to errors and biases.Another source of input assumptions is the choice of modelling algorithm and parameter settings.Dispersal constraint is major source of uncertainty in estimating the potential distributions of the subspecies.The dispersal constraint method developed in this study provides a simple estimation but does not account for the heterogeneity of dispersal among individuals and spatial locations.The model necessarily sacrifices some accuracy in favour of generality and cannot include as many parameters as other models developed for a single species (e.g., Dullinger et al., 2004) or for addressing more theoretical insights.In addition, the interpretation of SDM results comes with several major caveats, including sampling bias, biotic interactions, habitat fragmentation and human impacts.
While subject to these caveats, our results showed that the potential Predicted present and future distributions of Platycerus albisomni and Platycerus takakuwai modelled with Maxent at the subspecies and species levels.The coloured areas represent moderate-to-high probability of presence (>0.2).Species prediction scenarios at the subspecies and species levels without dispersal were cited fromZhang and Kubota (2021b).
Subspecies had different responses to climate change.The potential distribution area of P. t. takakuwai was predicted to increase to twice the present potential area, but the area of P. t. akitai would decrease to one third of the present area under climate change.Overall, our results indicated that the incorporation of dispersal ability and intraspecific variation in SDMs significantly altered their outcome.
change in distribution area would be variable with climate change (Figures1 and S2).Underestimation or overestimation of SDM predictions could result in excessive spending of resources or the neglect of species in areas mistakenly designated as unimportant.Therefore, more accurate mapping of potential habitats is critical for conservation purposes.In the future, to help develop effective conservation strategies, researchers should use SDMs that integrate multiple abiotic and biotic factors, which may be overlooked by traditional SDMs.
and 159 for P. takakuwai (69 for P. t. takakuwai, 82 for P. t. akitai Fujita, 1987 and 8 for P. t. namedai Zhang and Kubota (2021b) to species distribution model without and with dispersal constraints, respectively.Bolded values were calculated in the present study, and the others were cited fromZhang and Kubota (2021b).When accounting for dispersal constraints, subspecies-level models had significantly higher performance than species-level models (Mann-Whitney U test, *p ≤ 0.05; **p ≤ 0.01; ***p ≤ 0.001).