Oscillayers: A dataset for the study of climatic oscillations over Plio‐Pleistocene time‐scales at high spatial‐temporal resolution

Abstract Motivation In order to understand how species evolutionarily responded to Plio‐Pleistocene climate oscillations (e.g. in terms of speciation, extinction, migration and adaptation), it is first important to have a good understanding of those past climate changes per se. This, however, is currently limited due to the lack of global‐scale climatic datasets with high temporal resolution spanning the Plio‐Pleistocene. To fill this gap, I here present Oscillayers, a global‐scale and region‐specific bioclim dataset, facilitating the study of climatic oscillations during the last 5.4 million years at high spatial (2.5 arc‐minutes) and temporal (10 kyr time periods) resolution. This dataset builds upon interpolated anomalies (Δ layers) between bioclim layers of the present and the Last Glacial Maximum (LGM) that are scaled relative to the Plio‐Pleistocene global mean temperature curve, derived from benthic stable oxygen isotope ratios, to generate bioclim variables for 539 time periods. Evaluation of the scaled, interpolated estimates of palaeo‐climates generated for the Holocene, Last Interglacial and Pliocene showed good agreement with independent general circulation models (GCMs) for respective time periods in terms of pattern correlation and absolute differences. Oscillayers thus provides a new tool for studying spatial‐temporal patterns of evolutionary and ecological processes at high temporal and spatial resolution. Main types of variable contained Nineteen bioclim variables for time periods throughout the Plio‐Pleistocene. Input data and R script to recreate all 19 bioclim variables. Spatial location and grain Global at 2.5 arc‐minutes (4.65 x 4.65 = 21.62 km2 at the equator). Time period and grain The last 5.4 million years. The grain is 10 kyr (= 539 time periods). Level of measurement Data are for terrestrial climates (excluding Antarctica) taking sea level changes into account. Software format All data are available as ASCII grid files.

To fill this gap, I here present Oscillayers, a ready to use global terrestrial palaeo-climatic dataset for all 19 conventional bioclim variables (Hijmans et al., 2005), spanning continuously from the beginning of the Pliocene (5.4 Myr) to the LGM (c. 20 kyr BP) in steps of 10 kyr plus input data (Δ layers) and an R script, to recreate those variables for the respective time periods (Figure 1).
All scaled and interpolated 19 palaeo-bioclim variables were also evaluated against independent GCMs for three time periods [i.e.
Oscillayers thus provides a novel tool for studying climatic fluctuations spanning the Plio-Pleistocene at high temporal and spatial resolution. Potential applications for eco-evolutionary studies are briefly discussed.
Step 1: Empirical Bayesian kriging (EBK; Krivoruchko, 2012) was used to calculate geographically interpolated surfaces of each variable for the present (interpolated surface present; ISP) and the LGM (interpolated surface LGM; ISL).
This step provides baseline climate estimates for areas that are now submerged but were subaerial during past glacial cycles when sea levels were lower under the assumption of spatial correlation between coastal and off coastal climates.
Step 3: Differences in surface temperature (Ts) between the LGM (c. 20 kyr) and those of the last 5.4 Myr in time steps of 10 kyr (= 539 steps), as derived from the isotope record (Hansen et al., 2013), were calculated (see Table S1 in Supporting Information Appendix S1).
Step 5: In a further step, the interpolated Δ layers were applied to the LGM layers to generate calibrated palaeo-bioclim layers of each time period (e.g. Bio1 LGM + Bio1 ΔT = Bio1 T ) as inspired by the Delta method (Ramirez-Villegas & Jarvis, 2010; see also http://www.world clim.org/downs caling). This implies that the underlying spatial pattern of each modelled time period is driven by the scaled differences between the LGM and the present.

| Data validation
The ability of Oscillayers to reproduce independent data (skill and va- F I G U R E 1 Flowchart for generating interpolated palaeo-bioclim layers (Oscillayers).
Step 1: Empirical Bayesian kriging interpolation of each variable for the present and the Last Glacial Maximum (LGM).
Step 2: Computation of the Δ layers.
Step 3: Calculation of surface temperature (Ts) differences between the LGM and proceeding time periods as derived from the isotope record for each time step (T).
Step 4: Scaling the Δ layers relative to those differences of Step 3.
Step 5: Application of the scaled Δ layers to the LGM variables for calibration.
Step 6: Clipping each layer with the corresponding palaeo-coastlines, as derived from reclassified digital elevation models (DEMs), to obtain the calibrated palaeo-bioclim layers of each time period (Oscillayers) [Colour figure can be viewed at wileyonlinelibrary.com] were compared with the Oscillayers generated for respective periods in terms of pattern correlation and absolute difference (see below).
Results were then compared (in similar terms) with inter-model comparisons between the GCMs for HOL and LGM, that is CCSM and MIROC, respectively (see also Lawing & Polly, 2011;Rödder et al., 2013). Pairwise correlations were calculated in SDMToolbox v. 2.2b (Brown, Bennett, & French, 2017), using Pearson's correlation coefficient (r), a commonly used metric for evaluating the skill of modelled climatic variables (Fordham et al., 2017). This coefficient can range between +1 and −1, indicating a positive or negative relationship, respectively, while a coefficient of 0 indicates that two layers are independent from each other (Brown et al., 2017).  Table 1). The same was true for the remaining 17 bioclim variables (mean r HOL = .978, range: .906-1.0; r LIG = .877, .597-.975; r PLIO = .904, .684-.963). The inter-model pattern correlations between CCSM and MIROC for the HOL and LGM, respectively, were generally smaller than those of Oscillayers-HOL CCSM but tendentially higher than Oscillayers-LIG and Oscillayers-PLIO (Table 1).

Absolute differences in the 19 bioclim variables between the
Oscillayers and the independent GCMs (for HOL, LIG and PLIO) compared favourably with those derived between the CCSM and MIROC variables for the HOL and LGM when judged by the 2.5-97.5% quantiles. In detail, Oscillayers-HOL CCSM absolute differences were smaller compared to those of HOL CCSM -HOL MIROC and LGM CCSM -LGM MIROC for nine and 19 variables, respectively (Table 1). Similarly, Oscillayers-LIG and Oscillayers-PLIO differences were smaller compared to those of LGM CCSM -LGM MIROC for 11 and 14 variables, respectively (Table 1).
Overall, the generated palaeo-bioclim layers showed good agreement with independent GCMs (HOL, LIG and PLIO), with differences being mostly smaller than those between the commonly used CCSM and MIROC models for the HOL and LGM, respectively. Hence, the current approach provides a sufficiently robust approximation of palaeoclimate conditions throughout the Plio-Pleistocene.

| Applications
Oscillayers provides climatic data for 19 bioclim variables (see Supporting Information Appendices S2 and S3 for representative animations through time for Bio1 and Bio12), plus input data (see Table S2 in Supporting Information Appendix S1) and an R script, to recreate those variables for time periods spanning the early Pliocene (5.4 Myr) to the LGM (c. 20 kyr) in steps of 10 kyr.
Oscillayers can be used for testing a variety of eco-evolutionary hypotheses over this time period, for example, about climate-induced range changes of taxa or climate-related patterns of diversification (speciation, extinction) and adaptation. This can be facilitated by projecting ecological niche models (ENMs) of ecosystems or species (extant or extinct) onto these palaeo-climatic layers (e.g. Espíndola et al., 2012;Roberts & Hamann, 2012), or by reconstructing ancestral climatic envelopes along molecular phylogenies (Lawing & Polly, 2011;Lawing et al., 2016;Meseguer et al., 2018;Rödder et al., 2013;Yesson & Culham, 2006). Such spatially explicit models through time, either derived from ENMs (assuming niche conservatism) or ancestral climatic envelope reconstructions (taking niche divergence into account) might also be used for phylogeographic inferences "as is", or for modelling and testing population demographic hypotheses within a coalescent framework (Collevatti et al., 2015(Collevatti et al., , 2013. Other potential applications include, for example, the generation of PalaeoENMs via georeferenced fossils (Myers et al., 2015); the testing of biodiversity-related hypotheses about palaeo-climatic stability in the tropics (e.g. Couvreur et al., 2015;Kissling et al., 2016;Rakotoarinivo et al., 2013); the testing of predictions of the glacial-sensitive model of island biogeography (Fernández-Palacios et al., 2016;Norder et al., 2019) or the facilitation of landscape connectivity (dispersal corridor) analyses over time in a conservation context (Eberle, Rödder, Beckett, & Ahrens, 2017;Yu et al., 2015).

| Limitations and caveats
The Oscillayers framework presented herein assumes that past climates can be described by relative differences between modern and Quaternary climates as guided by the isotope record (Collevatti et al., 2015;Lawing & Polly, 2011;Rödder et al., 2013). As a corollary, uncertainty of the interpolations may increase with time when this assumption becomes less likely to hold. Also, as the Oscillayers framework broadly assumes a modern continental configuration it cannot explicitly account for spatial effects of large-scale geological events (e.g. the Messinian Salinity Crisis, c. 5.96 to 5.33 Myr).
The current approach is therefore unlikely to be easily extended into pre-Pliocene time periods. Also, the underlying GCMs used for the generation, validation and evaluation of Oscillayers are fraught with uncertainty, too (e.g. downscaling artefacts, parameters and functions used; Hargreaves, 2010;Lima-Ribeiro et al., 2015;Wiens, Stralberg, Jongsomjit, Howell, & Snyder, 2009).
Here, this was accomplished by using simple land masks for the palaeo-coastlines of each of the 539 time periods based on reclassified F I G U R E 2 Oscillayers show good agreement with independent global circulation models (GCMs). Comparison of GCM-derived and interpolated layers (Oscillayers) for annual mean temperature (Bio1) and annual precipitation (Bio12) at global and regional (Madagascar; see also Figure S3 in Supporting Information Appendix S1) scales for the Holocene (HOL: a,   Indicates that the Oscillayers-GCM comparisons are more correlated (left in Table) or have lower absolute differences (right in Table) than the LGM CCSM -LGM MIROC inter-model comparisons, respectively.
Absolute differences are in the units of the respective bioclim variable. b Indicates that the Oscillayers-GCM comparisons are more correlated (left in Table) or have lower absolute differences (right in Table)