Microclimates can be accurately predicted across ecologically important remote ecosystems

Microclimate information is often crucial for understanding ecological patterns and processes, including under climate change, but is typically absent from ecological and biogeographic studies owing to difficulties in obtaining microclimate data. Recent advances in microclimate modelling, however, suggest that microclimate conditions can now be predicted anywhere at any time using hybrid physically- and empirically-based models. Here, for the first time, we test the utility of this approach across a remote, inaccessible, and climate change threatened polar island ecosystem at ecologically relevant scales. Microclimate predictions were generated at a 100 × 100 m grain (at a height of 4 cm) across the island, with models parameterised using either meteorological observations from the island’s weather station (AWS) or climate reanalysis data (CRA). AWS models had low error rates and were highly correlated with observed seasonal and daily temperatures (root mean squared error of predicted seasonal average Tmean ≤ 0.6 °C; Pearson’s correlation coefficient (r) for the daily Tmean ≥ 0.86). By comparison, CRA models had a slight warm bias in all seasons and a smaller diurnal range in the late summer period than in situ observations. Despite these differences, the modelled relationship between the percentage cover of the threatened endemic cushion plant Azorella macquariensis and microclimate varied little with the source of microclimate data (r = 0.97), suggesting that both model parameterisations capture similar patterns of spatial variation in microclimate conditions across the island ecosystem. Here, we have shown that the accurate prediction of microclimate conditions at ecologically relevant spatial and temporal scales is now possible using hybrid physically- and empirically-based models across even the most remote and climatically extreme environments. These advances will help add the microclimate dimension to ecological and biogeographic studies, which could be critical for delivering climate change-resilient conservation planning in climate-change exposed ecosystems.

for the first time, we test the utility of this approach across a remote, inaccessible, and climate 23 change threatened polar island ecosystem at ecologically relevant scales. Microclimate 24 predictions were generated at a 100 × 100 m grain (at a height of 4 cm) across the island, with 25 models parameterised using either meteorological observations from the island's weather 26 station (AWS) or climate reanalysis data (CRA). AWS models had low error rates and were 27 highly correlated with observed seasonal and daily temperatures (root mean squared error of 28 predicted seasonal average Tmean ≤ 0.6 °C; Pearson's correlation coefficient (r) for the daily 29 Tmean ≥ 0.86). By comparison, CRA models had a slight warm bias in all seasons and a 30 smaller diurnal range in the late summer period than in situ observations. Despite these 31 differences, the modelled relationship between the percentage cover of the threatened 32 endemic cushion plant Azorella macquariensis and microclimate varied little with the source 33 of microclimate data (r = 0.97), suggesting that both model parameterisations capture similar 34 patterns of spatial variation in microclimate conditions across the island ecosystem. Here, we 35 have shown that the accurate prediction of microclimate conditions at ecologically relevant 36 spatial and temporal scales is now possible using hybrid physically-and empirically-based 37 models across even the most remote and climatically extreme environments. These advances 38 will help add the microclimate dimension to ecological and biogeographic studies, which 39 could be critical for delivering climate change-resilient conservation planning in climate-40 change exposed ecosystems. 41

INTRODUCTION 44
Microclimates created by landscape and vegetation structures play a key role in facilitating 45 the persistence of species in locations that would otherwise be climatically inhospitable 46 8 pressure (Pa), wind speed at 2 m (m sec -1 ), wind direction (degrees from N), atmospheric 164 emissivity, multiple measures of direct and diffuse radiation, and cloud cover (%). In our 165 analysis, the ground and canopy albedo were fixed at 0.15 and 0.23 and habitat was specified 166 as 'Barren or sparsely vegetated'. The models were parameterised using either (Table 1a): 167 1) AWSthe hourly meteorological observations from the island's AWS for air 168 temperature, air pressure at sea level, wind direction, wind speed, and relative 169 humidity (with hourly estimates of the emissivity of the atmosphere, radiation, 170 and cloud cover downscaled from NCEP2 data). 171 2) CRAhourly meteorological data obtained solely from downscaled NCEP2 data. 172 The NCEP2 data were downloaded and temporally interpolated from six-hourly to hourly 173 values using the hourlyNCEP function from microclima. Models were run in hourly time 174 steps for two three-month periods that define the island's growing season: early summer (1 st 175 October to 31 st December) and late summer (1 st January to 31 st March). 176

177
Model predictions were evaluated against the in situ microclimate observation data ('Obs') 178 across the 62 sites (see Table 1b  and standard deviation of these within-site RMSE across the 62 sites (see below) are reported 199 (Table 2). We also evaluated the spatial distribution of prediction errors for the observed and 200 predicted temperature by mapping the observed minus predicted seasonal Tmin, Tmean, and 201 Tmax for each season and model (i.e. AWS and CRA). 202 Microclimate models were run from 1999 to 2019, which covers the majority of the 203 period for which hourly automated meteorological observation data exists for the island. 204 Models were run for each season separately. Across the 20-year time series, we calculated the 205 average seasonal mean temperature (Tmean) and the average seasonal growing degree days 206 > 5 °C (GDD5), which are both biologically relevant climate variables. Each of these 207 bioclimate variables was calculated for both the AWS and CRA model predictions. 208

Microclimate data in species distribution models 209
To assess the implications of the observed differences between the AWS and CRA driven 210 microclimate models in an applied setting, we model the percentage cover of a widespread 211 and recently threatened endemic plant species on the island, Azorella macquariensis, using 212 spatially contiguous microclimate estimates, rather than terrain proxies (c.f. Bricher et al., 213 2013). This is a critical step for understanding the effects of climate change on this threatened 214 ecosystem because by building cover models for species using microclimate data-rather 215 than terrain proxies-the sensitivity of the system to changes in the microclimate conditions 216 can be explored. Proportional cover data were collected across 90 sites using a stratified 217 random sampling strategy across the landscape of Macquarie Island. Sampling was conducted 218 across a 700 m 2 area at each site; for full details see Dickson et al. (2019). Here, we added an 219 additional ten sites drawn randomly from grid cells on the coastal fringes of the island (<50 220 m asl). A. macquariensis is extremely rare below this elevation (where it is outcompeted by 221 other species) and, therefore, the proportion of cover at these sites was set to zero. 222 Using these data, the proportion of A. macquariensis cover was modelled across the GDD5 was selected because plant cover is known to be affected by temperature extremes, 229 including the number of growing days in the short growing season of the region, and we also 230 included a quadratic term GDD5 2 because visualisation of the raw data suggested an apparent 231 nonlinear association between cover and GDD5. Bayesian beta regression models 232 (link = logit; link phi = identity) were fit using the rstanarm package in R (Goodrich, Gabry, 233 Ali, & Brilleman, 2018). The beta distribution is a continuous probability distribution defined 234 on the interval (0, 1) and, therefore, response variables must be transformed if they include 235 exact values of 0 or 1. We used the transformation ( × ( − 1) + 0.5)/ , where y is the 236 response variable and n is the sample size, to rescale the response variables onto the (0, 1) 237 interval (Cribari-Neto & Zeileis, 2009). Leave-one-out cross-validation (n = 1000) was used 238 to evaluate the model predictive performance. Models were built using both the AWS and 239 CRA driven microclimate datasets and the median posterior predictions for percentage cover 240 were compared by calculating the pair-wise Pearson's correlation coefficient. 241

243
Models driven by AWS data predicted the seasonal Tmean across the island with a RMSE ≤ 244 0.6 °C and MAE ≤ 0.5 °C in both the early and late summer (  Table 2). The daily RMSE for 250 Tmean indicates a mean daily error rate for predictions from the AWS models of c. 0.9-1.2 °C 251 (depending on the season) and higher error rates for the CRA model daily predictions (c. 1.2-252 1.7 °C). Prediction errors were comparable for daily Tmin and Tmax. The correlation between 253 observed and predicted values for daily Tmin and Tmean were consistently high for AWS 254 models (r = 0.85-0.93), but lower for predictions of daily Tmax (r = 0.61-0.76) from these 255 same models, and for Tmin, Tmean, and Tmax from CRA models (r = 0.48-0.77). Microclimate 256 models consistently added value over the macroclimate temperature time-series (NCEP2 257 derived), which had much higher RMSE rates (Tmean > 1.4 °C) and lower temporal 258 correlations for daily temperatures (r = 0.00-0.74) when benchmarked against the in situ 259 observation (Obs) data (Table 2). 260 Spatial patterns in prediction errors varied between seasons and between models 277 (Fig. 2). Seasonal Tmin was generally over predicted ( Fig. 2a-d), although prediction errors 278 from the AWS model in the late summer showed a slightly more random distribution around 279 zero (Fig. 2c). There was no strong pattern in the size of the Tmin error across the island for 280 either model or season. The prediction errors for seasonal Tmean are generally small and 281 distributed around zero for predictions from the AWS model in the early summer and CRA 282 model in the late summer (Fig. 2e,h), but show consistent directional errors for the other two 283 Tmean spatial predictions (Fig. 2f,g). The prediction errors for seasonal Tmax follow a similar 284 pattern to those of Tmean, but tend to be larger and to have greater spatial variation in the size 285 13 of the prediction errors (Fig. 2i,l). These characteristics are observable in a comparison of the 286 20-year averaged bioclimate variables, which shows that the CRA model consistently 287 predicts higher Tmean and GDD5 than the AWS model (Fig. 3). Even where predictions are 288 biased, the spatial correlation between the AWS and CRA derived bioclimate variables was 289 high (r ≥ 0.95), suggesting that both models predict similar spatial temperature patterns. 290

Species distribution modelling 291
The leave-one-out RMSE for percentage cover predictions using the bioclimate variables 292 percentage cover predictions using the two different microclimate datasets was r = 0.97, 296 indicating a very high correlation between predictions from the two models (Fig. 4). Both  Importantly, despite plant cover on the ground being highly variable, other areas of the island 301 with either high density cushion carpets or iconic terraces were also identified by the cover 302 model. While both microclimate models produced highly representative distributions for 303 most populations of A. macquariensis, the AWS derived bioclimate variables suggest wider 304 variation in cover in the south and more gradual spatial variation between neighbouring areas. 305

306
Here, we have evaluated the use of a hybrid physically-and empirically-based modelling 307 approach for predicting daily and seasonal fine-scale climate variation across a remote 308 climate-change threatened polar ecosystem at ecologically relevant spatial scales. When meteorological observations are not available, microclimate models driven only by climate 320 reanalysis data are able to capture the spatial variation in fine-scale climate conditions. These 321 results show that spatially and temporally explicit microclimates may be accurately predicted 322 across remote and exposed landmasses, which paves the way for novel and innovative 323 research on the vulnerabilities and conservation opportunities in these challenging and 324 threatened environments. 325 The models accurately predict the day-to-day variation in temperature, although there 326 is still error in these predictions and apparent bias in some of the CRA model predictions. 327 The latter is not unexpected as predictions that are unconstrained by in situ data are often 328 biased in some respect, and climate models typically go through a bias-correction process 329 prior to use (e.g. Navarro-Racines, Tarapues, Thornton, Jarvis, & Ramirez-Villegas, 2020). 330 Some of this error and bias will be caused by differences between the realised macro-, meso-331 and micro-scale conditions that influence climate variations across a landscape and those 332 captured by the data used to parameterise the models. For example, here we used cloud cover 333 estimates from the NCEP2 reanalysis data, which have a course spatial (c. 200 × 200 km) and 334 temporal (6 h) resolution and, thus, may underestimate variation in cloud cover, especially 335 over a small landmass situated in a vast ocean. Thus, the reanalysis data may not be entirely 336 representative of the island specific climatic conditions given the dominance of sea in the 337 region (Kearney et al., 2020). Similarly, the summary of orographic information below the 338 grain of the model (here, 100 × 100 m) will introduce variation into the model predictions.  In this study we have demonstrated the ability of physically-and empirically-based 382 microclimate models to predict ecologically meaningful microclimate conditions across a 383 remote, exposed, and climate change-vulnerable island ecosystem, particularly when in situ 384 meteorological data are available to parameterise the models. This paves the way for novel 385 ecological and biogeographic studies on the role of microclimate in determining biodiversity 386 patterns and trends. Projections of future changes in microclimate conditions are also possible 387 (i.e. Maclean, 2020) and these will be invaluable for understanding the plausible range of 388 changes in microclimate conditions across the region and for predicting the threats these 389