Spatial scale affects novel and disappeared climate change projections in Alaska

Abstract The formation of novel and disappeared climates between the last glacial maximum (LGM) and the present is important to consider to understand the expansion and contraction of species niches and distributions, as well as the formation and loss of communities and ecological interactions over time. Our choice in climate data resolution has the potential to complicate predictions of the ecological impacts of climate change, since climate varies from local to global scales and this spatial variation is reflected in climate data. To address this issue, we downscaled LGM and modern (1975–2005) 30‐year averaged climate data to 60‐m resolution for the entire state of Alaska for 10 different climate variables, and then upsampled each variable to coarser resolutions (60 m to 12 km). We modeled the distributions of novel and disappeared climates to evaluate the locations and fractional area of novel and disappeared climates for each of our climate variables and resolutions. Generally, novel and disappeared climates were located in southern Alaska, although there were cases where some disappeared climates existed within coastal and interior Alaska. Climate resolution affected the fractional area of novel and disappeared climates in three patterns: As the spatial resolution of climate became coarser, the fractional area of novel and disappeared climates (a) increased, (b) decreased, or (c) had no explainable relationship. Overall, we found the use of coarser climate data increased the fractional area of novel and disappeared climates due to decreased environmental variability and removal of climate extremes. Our results reinforce the importance of downscaling coarse climate data and suggest that studies analyzing the effects of climate change on ecosystems may overestimate or underestimate their conclusions when utilizing coarse climate data.

tion has the potential to complicate predictions of the ecological impacts of climate change, since climate varies from local to global scales and this spatial variation is reflected in climate data. To address this issue, we downscaled LGM and modern  30-year averaged climate data to 60-m resolution for the entire state of Alaska for 10 different climate variables, and then upsampled each variable to coarser resolutions (60 m to 12 km). We modeled the distributions of novel and disappeared climates to evaluate the locations and fractional area of novel and disappeared climates for each of our climate variables and resolutions. Generally, novel and disappeared climates were located in southern Alaska, although there were cases where some disappeared climates existed within coastal and interior Alaska. Climate resolution affected the fractional area of novel and disappeared climates in three patterns: As the spatial resolution of climate became coarser, the fractional area of novel and disappeared climates (a) increased, (b) decreased, or (c) had no explainable relationship. Overall, we found the use of coarser climate data increased the fractional area of novel and disappeared climates due to decreased environmental variability and removal of climate extremes. Our results reinforce the importance of downscaling coarse climate data and suggest that studies analyzing the effects of climate change on ecosystems may overestimate or underestimate their conclusions when utilizing coarse climate data.

K E Y W O R D S
Alaska, climate change, climate downscale, last glacial maximum, novel and disappeared climates, spatial resolution

| INTRODUC TI ON
Postglacial climate change provides a useful context and natural experimentation for assessing biotic responses to global climate change (Davis, 1990;Overpeck, Bartlein, & Webb, 1991;Webb III, 1992). The late quaternary, which includes the last glacial maximum (LGM) 21,000 years ago, matches the magnitude of predicted anthropogenic climate change and contains the largest manifestation of natural climate change preserved in the geologic record (Mix, Bard, & Schneider, 2001;Overpeck et al., 1991). Our awareness and understanding of past ecological responses to climate change is important as it enables ecologists to predict the potential responses of ecosystems to anthropogenic climate change now and in the future.
Niche theory predicts that n-dimensional changes in the environment (e.g., precipitation and temperature) will cause shifts in species distributions and the formation of novel species assemblages, since every species responds individualistically to its abiotic and biotic environment (Hutchinson, 1957;Jackson & Overpeck, 2000). This assumption has been supported by large changes in species ranges where past climates lacked modern analogs, leading to the formation of novel LGM species associations and biomes with no modern equivalent . Therefore, variations of climate in space and time (e.g., LGM vs. Modern n-dimensional environment) are thought to be an important factor in understanding the formation of novel and disappeared climates, as well as the expansion and contraction of a species' niche and distribution (Ackerly et al., 2010;Jackson & Overpeck, 2000;.
Novel climates are climatic environments with no-analog conditions in the past, whereas disappeared climates are climatic environments with no-analog conditions today (Ackerly et al., 2010;Chen, Hill, Ohlemuller, Roy, & Thomas, 2011;Fitzpatrick & Hargrove, 2009;Glassberg, 2014;Hobbs et al., 2006;Radeloff et al., 2015;. There are many ways climate data can be used to describe shifts in climate, with the most extreme description being the detection of no-analog climates. No-analog climate techniques identify a location or period the climate of which is dissimilar to that in another context (Ford et al., 2010), and then classify the climate into spatial and/or temporal categories (Mearns et al., 2001). For example, spatial and temporal classifications of no-analog climates typically describe local climate change using a dissimilarity metric (Grenier, Parent, Huard, Anctil, & Chaumont, 2013;e.g., Standardized Euclidean Distance) at the same grid cell between climate variables at two difference time periods, and then compare the climate realization for each land gridpoint from one time period the climate realization of all land gridpoints of the other time period, while retaining the minimum dissimilarity metric value (Ackerly et al., 2010;Fordham, Saltré, Brown, Mellin, & Wigley, 2018;Grenier et al., 2013;Ordonez, Williams, & Svenning, 2016;Williams, Jackson, & Kutzbacht, 2007). Other approaches have computed no-analog climates relying only on temporal classifications of novel and disappeared climates. For example, Fitzpatrick and Hargrove (2009) utilized techniques employed by species distribution models to identify no-analog climates in the future (areas where the current climate conditions of a study area do not exist in projected future climate), Burrows et al. (2014) calculated shifting climates using climate velocity (Burrows et al., 2014;Dobrowski et al., 2013;Loarie et al., 2009) to identify source (novel) and sink (disappeared) areas, and Wiens, Seavy, and Jongsomjit (2011) employed PCA analysis to collapse multidimensional climate into a single climate space to derive polygons encompassing current and future climate space to identify persisting climates, disappearing climates, and novel climates.
A key aspect of quantifying climate, including the identification of novel and disappeared climates, is the choice of spatial resolution. Spatial resolution has the potential to complicate the prediction of ecological impacts of climate change because climate varies dramatically at local scales and this variation is undetectable in coarse resolution climate data (Bellard, Bertelsmeier, Leadley, Thuiller, & Courchamp, 2012;Dobrowski, Abatzoglou, Greenberg, & Schladow, 2009;Franklin et al., 2013;Seo, Thorne, Hannah, & Thuiller, 2009).
Long-term climate patterns observed across the globe are a result of a combination of many different processes that occur at varying spatial scales. Most modeled climate data (regional and general circulation models) are coarse scale (>50 km), and these datasets are unlikely to incorporate many spatial features known to influence climate at finer scales (e.g., elevation gradients, coastal effects, temperature inversions, and rain shadows; Daly, 2006;Dobrowski et al., 2009;Levin, 1992).
Spatial variability in climate can be nested into macroclimates (global), mesoclimates (regional), topoclimates (landscape), and microclimates (local; Ackerly et al., 2010;Geiger, Aron, & Todhunter, 2009). Macroclimates are the broad patterns of atmospheric circulation across >50 km scales, such as the North-South rainfall gradient along the state of Alaska (Ackerly et al., 2010). Mesoclimates are variations at 1-50 km reflecting marine air and mountain range properties, such as rain shadow effects (Ackerly et al., 2010). Topoclimates include landscape scale effects such as aspect, slope, elevation, and terrain that affect surface radiation, wind, and cold-air drainage at the 10-m to 1-km scale (Ackerly et al., 2010). Lastly, microclimates have the finest scale variability and are determined by vegetation cover and fine-scale surface features (<10 m; Ackerly et al., 2010).
Finer spatial scales (topo and micro) create unique combinations of climate variables within a very limited area (Ackerly et al., 2010) and can provide significant processes that buffer against larger regional and global climate trends (e.g., temperature inversions; Randin et al., 2009;Willis & Bhagwat, 2009). For example, mesoscale PRISM mean global temperature variability is limited to 3°C, while a finer toposcale mean global temperature surface at 30 m found a global temperature variability as high as 8°C (Ackerly et al., 2010;Daly, 2006).
We selected climate surfaces commonly used in climate change and ecological analyses to identify and determine where novel, disappeared, and shared climates are distributed across the state of Alaska from the LGM to modern era. Previous studies have identified multivariate no-analog climates; however, we compute univariate no-analog climates as species have been found to be limited by a single environmental factor (Liebig's Law;De Baar, 1994;Berryman, 2003;Crimmins, Dobrowski, Greenberg, Abatzoglou, & Mynsberge, 2011), and we could identify which specific aspects of multivariate climate cause no-analog climates. Polar regions are currently experiencing the highest rates of warming (Larsen et al., 2014), making the state of Alaska, USA, an ideal location to study the impacts of spatial scale in climate change. To help clarify the effect of climate grid resolution on estimations of novel and disappeared climates, we analyzed how the amount of fractional area of novel, disappeared, and shared climates vary with climate grid spatial resolution by statistically downscaling coarse-scale GCM (General Circulation Model) climate (~100 km) at scales ranging from 60 m to 12 km. We compared modeled distributions and fractional area of novel, disappeared, and shared climates across nine Alaskan ecoregions to ask the following specific questions: 1. Where are novel, disappeared, and shared climates located in Alaska from the LGM to modern era? 2. How does the fractional area of novel, disappeared, and shared climates differ depending on the resolution of modeled climate grid data used for analysis?

| Overview
The first step of our analysis was to perform a topographically mediated downscaling of coarse-scale modern and LGM climate surfaces to a resolution of 60 m for the modern and LGM periods.
Next, we upsampled these surfaces to coarser resolutions. Finally, we performed our analysis investigating the impacts of scale on the distribution of shared, novel, and disappeared climates throughout Alaska.

| Study area
Alaska, USA, is an ideal location for understanding the potential impacts of scale on climate change as polar regions currently experience the highest rates of warming globally, and by the end of the 21st century will be at least 40% higher than the global mean (Larsen et al., 2014). There is agreement between several coupled atmosphere-ice-ocean climate models that global warming should be enhanced in the Arctic (Raisanen, 2001). Additionally, Alaska has a diverse and complex physiography and presence of long-term meteorological stations.
The total land area in Alaska is approximately 151,773.3 km 2 , with over 54,563 km of tidal shoreline, including islands, and stretches in latitude by nearly 20°. Seventeen of the 20 highest mountain peaks of North America are in Alaska, with the highest elevation at Mt.
McKinley (6,150 m.a.s.l), and lowest at the Pacific Ocean coastline (0 m.a.s.l). Presently, there are approximately 100,000 glaciers covering 75,109.9 km 2 (5% of the total land area of Alaska); however, during the LGM, glaciers covered an estimated 30% of the state. Present mean temperature is 16.8°C during the summer, and −11.0°C during the winter. standard deviations from the means of all observations across time for minimum temperature (T min ), maximum temperature (T max ), mean temperature (T ave ), and total monthly precipitation (P) (Aggarwal, 2016).

Digital elevation map and transforms
An elevation surface was assembled for the entire state of Alaska using U.S. Geologic Survey (USGS) 1 arc-second (~60 m) digital elevation maps (DEM; USGS, 2016). The digital elevation map (DEM) was resampled and reprojected to Alaska Albers Equal Area Conic, ensuring grid cells of exactly 60 m. From this dataset, we calculated slope, aspect, and the topographic convergence index (TCI) for Alaska (Wolock & McCabe, 1995). TCI was calculated with TauDEM using a D-infinity flow accumulation algorithm (Tesfa et al., 2011).
Regions where slope was 0° were given a placeholder value of 0.001 when calculating TCI to avoid divide-by-zero issues.

General circulation model climate data
Coarse-scale (1°) general circulation model (GCM) climate grids were collected from the National Center for Atmospheric Research's (NCAR) Community Climate System Model version 4 (CCSM4) for near surface (2 m) T min , T max , T ave , precipitation, short wave radiation (I cloud ), and wind (U, V; Gent et al., 2011;Kluzek, 2011). All climate surfaces were pooled into two 30-year averages, ~18,000-18,030 ya (years ago; t 1 ) for the LGM and1975-2005 (t 2 ) for modern climate. All GCM surfaces were reprojected to the Alaska Albers Equal Area Conic projection and resampled to 60-m resolution using bilinear interpolation. The wind speed vector surfaces were converted to wind speed and direction.

Shortwave irradiance
Solar radiation can regulate temperature in complex terrains as topography produces varying solar angles and can reduce terrain winds that diminish boundary layer mixing during the winter months (Daly, 2006;Dobrowski et al., 2009;Urban, Miller, Halpin, & Stephenson, 2000). Therefore, solar radiation is an important physiographic parameter to consider when downscaling temperature to topo and microscales. Mean monthly daily clearsky irradiance at a location x,y (I (x,y,t) , W/m 2 ) was modeled using the r.sun algorithm (Suri & Hofierka, 2004) running under GRASS GIS 7.1. This algorithm uses topographic elevation, slope, aspect, and geographic latitude as inputs referenced against solar angles. GRASS GIS's r.sun algorithm is a complex and flexible solar radiation model that has been found to outperform other similar products (SolarFlux, Solei, Solar Analyst, and SRAD) because it performs especially well for large areas at fine resolutions with complex terrain and can be used for long-term calculations at different scales (Hofierka & Suri, 2002;Ruiz-Arias, Tovar-Pescador, Pozo-Vázquez, & Alsamamra, 2009 Taylor, Stouffer, & Meehl, 2012). These groups calculate surface downwelling shortwave radiation using additional modeled climate surfaces produced from the same CCSM4 LGM and modern historical runs used in our temperature and precipitation downscale models. To transform our high-resolution "clear-sky" irradiance to "true-sky" irradiance, we calculated the ratio of true-sky to clear-sky irradiance for each month for both the LGM and today, and then applied each corresponding ratio to each monthly high-resolution clear-sky irradiance surface to create twelve monthly LGM and Modern "true-sky" surface radiation surfaces that are capable of considering coarse atmospheric properties (e.g., cloud cover) of each era with our clear-sky radiation surfaces. We then summarized our monthly radiation surfaces into an annual average radiation surface to compute novel and disappeared radiation climates for later analysis. Mean radiation was created by computing the yearly average of all monthly radiation surfaces for both the LGM and modern eras at 60-m resolution.

Temperature and precipitation
For minimum, maximum, and average temperature, as well as precipitation, we used an empirical downscaling approach as described in Dobrowski et al. (2009). This approach calibrates a downscaling model in which the high-resolution climate variable is a function of the coarse-scale climate and various topographic predictors (Equation 1). We used the following general downscaling approach: where observed Climate at location (x,y) at time t represents our high-resolution weather station data. CCSM Climate at the same (x,y) locations and time t represents our coarse GCM modeled climate.
Physiographic Inputs at the same (x,y) locations, and when applicable time t, represents our high resolution modeled physiographic surfaces (e.g., elevation, cold-air pooling, and radiation). We assume the topographic impacts on climate do not change in time, only in space, so the only varying predictors in our model are coarse-scale climate and radiation inputs. Our other topographic-based predictors remain constant (e.g., TCI).
We calibrated these models using weather station data to represent the high-resolution climate linked with the GCM climate grids that correspond to the date of the weather measurement. To determine variable selection, we used a random forest algorithm to derive variable importance statistics for all model parameters at a variety of spatial scales (60 m, 500 m, 1 km, and 5 km; Breiman, 2001). Using the most important variables, we produced three models for temperature (T min , T max , T ave ) calibrated using 80% of the available weather station data (43,768 observations; Madsen & Thyregod, 2010). The remaining 20% of weather station data (10,943 observations) was used for model validation, in which we calculated the root mean square error (RMSE), Pearson's correlation coefficient, and percent bias. We created a precipitation model employing similar methods previously used to create our temperature models; however, the precipitation model was trained using a stratified sampling procedure (vs. 80% dataset) to ensure an equal proportion of measured weather station precipitation values were represented in the training dataset since the majority of weather observations recorded 0 mm/month precipitation. The training dataset was stratified in ~310 mm width bins from randomly sampling 80% of the original precipitation dataset. We then randomly sampled 5,000 precipitation values from each bin to create the final training dataset. Our testing dataset was comprised from the remaining 20% of the original precipitation dataset.
Generalized linear models (GLM) were chosen to generate all downscale models. GLMs are advantageous for downscaling because they can extrapolate beyond the range of the model's training data, which is necessary when predicting LGM climates as the range of climates is likely different than the modern climate. Additionally, GLMs can handle more complicated situations; for example, they do not require a normal distribution of the response variable (Madsen & Thyregod, 2010). A variety of tests/exploratory models were created to check that all GLM assumptions were met. To ensure that all input predictors were independent of one another, all variables were plotted against one another to verify that there were no significant relationships between each variable. A correlation threshold of 0.95 was used as a cutoff value preventing too great a multicollinearity between any pair of predictors. Additionally, an exploratory generalized additive model (GAM) was created to produce partial plots for each independent variable to verify linear behavior of all model inputs, although this method was not chosen for the final models to avoid spline interpolations on our predictors (Venables & Ripley, 2002). Lastly, each model's residuals were examined to ensure that the residuals were normally distributed. The residuals were plotted against the predicted fitted values to ensure a homogenous structure of each model's variance.

Temperature model
The model we used to downscale T min , T max , and T ave was found to be a function of elevation at x,y (Z (x,y) ), mean monthly daily clear-sky irradiance at a location x,y (I (x,y,t) ), and TCI (C (x,y) ) which High Res x,y,t is a proxy for local convective forcing's such as cold-air pooling (Equation 2; Dobrowski et al., 2009;Katurji & Zhong, 2012 where T (x,y,t,t′) represents observed high-resolution temperature (min, max, average) at a given station (x,y) and month (t), and T′ (x,y,t,t′) is coarse modeled surface temperature (min, max, average). The coldest T min month (December for LGM, and January for modern) was used to represent the coldest annual temperatures. Likewise, the hottest month for T max (July for both LGM and modern) was used to represent the hottest annual temperatures.

Precipitation model
The model to downscale precipitation was found to require topographic predictors at a variety of scales, not just the high resolution 60 m predictors. Specifically, we used elevation at location x,y (Z (x,y) ) at 60 m; topographic slope at location x,y at 1-km resolution (m x,y ); wind speed (m/s, (x,y,t,t′) ) and wind direction at a given location (x,y) (degrees, (x,y,t,t′) ) and month (t) at 100 km; and the angular difference between geographic direction and topographic aspect to act as an orographic effect proxy (i.e., rain shadow) at 500-m resolution ( x,y,t,t′ ). Elevation and slope exhibited slight nonlinear trends which were corrected by applying a log-transformation to each of the two predictor variables. A cubic root transformation was applied to (P x,y,t,t′ ) and (P′ x,y,t,t′ ) so the model could not predict values below 0 mm/month (Equation 3). The final model form was

Potential evapotranspiration
Plants must use energy and water to grow and reproduce; therefore, the primary effects of climate on plants are regulated by the interactions of energy and water (Stephenson, 1990). Through water balance equations, energy is represented by potential evapotranspiration (PET) and available water (Stephenson, 1998). PET was calculated at monthly time steps at 60 m for the LGM and modern era using the Penman-Monteith method which utilizes our downscaled temperature, elevation, wind speed, and cloud-corrected irradiance surfaces as inputs (Allen, Pereira, Raes, & Smith, 1998;Monteith, 1965;Penman, 1948). This method determines a reference evapotranspiration from climate data and represents climatic water balance rather than physiological evapotranspiration. PET was calculated by using a standard hypothetical reference crop of height 0.12 m with a fixed surface stomatal resistance of 70 s/m and albedo of 0.23. Annual PET for the LGM and modern era was summarized by computing the summation of all monthly surfaces for each era.

Actual evapotranspiration and water deficit
The interactions of PET and available water can be described with additional water balance parameters: actual evapotranspiration (AET) and water deficit (DEF). AET and water deficit are biologically meaningful parameters that are well correlated with the distribution of vegetation types compared to other parameters such as temperature and precipitation (Stephenson, 1998). PET (i.e., evaporative demand) represents the total amount of energy available in the environment; essentially, the evaporative water loss from a site with unlimited water and is used to derive estimates of AET and water deficit (Stephenson, 1990). AET is the evaporative water loss given the actual water availability at a site and therefore represents the biologically usable energy and water in the environment (Stephenson, 1990). Water deficit refers to climatic water deficit, not soil water deficit, and represents the amount of evaporative demand that was not met by available water in other words: WD = PET − AET (Stephenson, 1990). While AET and water deficit are important for understanding the climatic controls of vegetation distributions, it is worth mentioning that all three parameters (PET, AET, and water deficit) are climatic products representing available water and energy at a site, not biological products dependent of specific vegetation types (Stephenson, 1998).
AET and DEF were calculated at monthly time scales at 60 m for the LGM and modern era, by using a snow hydrology model that models additional water and energy parameters that influence available water at a site (Dobrowski et al., 2013). Downscaled precipitation, mean temperature, and radiation were used to estimate the fraction of precipitation arising as either rainfall, snowfall, snowpack, or snowmelt during a given month at 60-m resolution for both the LGM and modern eras (Lutz, Van Wagtendonk & Franklin, 2010;Dobrowski et al., 2013). These products were then used to calculate AET and DEF in combination with PET and a coarse soil water-holding capacity surface (Dunne & Willmott, 1996) to determine the available plant extractable water from rainfall or snowmelt from the previous month (Dobrowski et al., 2013). A spatially constant available soil water-holding capacity (AWC) value of 5.0 cm (the mean AWC of Alaska) was used rather than a spatially varying AWC due to a lack of high-resolution AWC products in Alaska (Dunne & Willmott, 1996). The model first determines the amount of maximum potential available soil water at a site. If there is an excess of soil water, AET is the same as PET because evapotranspiration is not limited by water availability.
However, if available water is less than the sites maximum potential available water, AET will be less than PET (AET = PET − DEF), because maximum water evaporative demand cannot be met by the available amount of water present at the site. We created annual summaries of each surface (AET, DEF, snow, and rain) by computing the summation all monthly surfaces, for each climate variable, for each individual era.

| Climate aggregation and reclassification
To understand how climate grid resolution can potentially affect the distribution of novel and disappeared climates in Alaska, we aggregated each 60 m annual climate surface to coarser resolutions.
Aggregation surfaces were created using a standard average pixel Annual climate variables at each of the 11 spatial resolutions were binned to simplify the ranges of possible climate values for analysis.
We computed the 5th and 95th percentiles of each 60 m annual climate variable and removed these values as outliers in each surface. If a specific climate existed only during the LGM, the climate range was considered "disappeared" (blue pixels; Figure 1c). If a specific climate exists only in the modern era, the climate range was considered "novel" (red pixels; Figure 1b). If a specific climate existed during both time periods, the climate was considered "shared" (black pixels), as that specific range of climate could be found either during the LGM or modern era within the extent of Alaska (Figure 1b,c) TA B L E 1 Summary of annual downscaled (60 m) climates for the LGM and modern eras used in novel and disappeared climate analysis the LGM. If a specific climate was both novel and disappeared at the same location, the climate range was considered "both" (purple pixels; Figure 1d).
The locations of novel and disappeared climates were summarized by computing the fractional area of novel, disappeared, and shared climates for the entire state of Alaska, as well as the fractional area among nine "Level 2" Alaskan ecoregions defined by the U.S. Geologic Survey (Gallant, Binnian, Omernik, & Shasby, 1995).
We used modern AK ecoregions for both the modern and LGM eras, since, to the best of our knowledge, no LGM ecoregion boundary datasets currently exist. We quantified the amount of novel, disappeared, or shared climate at a given spatial resolution and used a Spearman's rank correlation coefficient to determine whether the relationship between novel, disappeared, or shared fractional climate and spatial resolution was positive, negative, or no apparent relationship. Wind speed was the least significant predictor for our precipitation model with a p-value of .0961; however, wind speed was still somewhat beneficial in predicting precipitation patterns across Alaska. Overall, elevation was the most significant predictor of precipitation, followed by slope, coarse modeled precipitation, orographic effects, wind speed, and lastly, wind direction as the weakest predictor as indicated by our random forest variable importance plots (Figure 3d).

| Downscale model evaluations
F I G U R E 1 Classification process for "novel," "disappeared," "shared," and "both" climates. 2 km climate surfaces were used for visualization purposes for this figure, as "both" climates are almost nonexistent at 60-m resolution.    While it is possible for downscaled novel and disappeared climates to exist at the same location from the LGM to modern era, only a very small fraction of T min climates displayed this phenomenon in Alaska at 60-m resolution and was determined to be insignificant as it only covered <0.001% (39.6 km 2 ) of Alaska. At coarser resolutions, there was a noticeable increase in climates classified as "both" novel and disappeared for some climate surfaces at a location; however, this portion of the results concentrates on our high-resolution downscaled climate surfaces ("both" climates vs. spatial resolution is discussed in the next section).

| No-analog climates versus spatial resolution
As spatial resolution becomes coarser, the fractional area of novel and disappeared climates generally increases, while the fractional area of shared climates generally decreases. All disappeared climate surfaces, except minimum temperature, had moderate (0.40-0.59) to very strong (0.80-1.00) Spearman's rank correlations between spatial resolution and fractional area, indicating that as the spatial resolution of climate becomes coarse, the fractional area of disappeared climates increases (Figure 7). All novel climate surfaces had a   The relationship between spatial resolution and fractional area of "both" novel and disappeared climates at the same locations was less clear than novel, disappeared, and shared climates ( Figure 10).
While "both" climates were virtually nonexistent at 60-m resolution for all climate surfaces except T min climates, the occurrence of "both" climates noticeably increased for rain and radiation climate surfaces at coarser resolutions. T min climate surfaces did not have any noticeable increase in both novel and disappeared T min climates occurring in the same location at any resolution. "Both" climates for radiation displayed a very strong Spearman's rank correlation between spatial scale and fractional area, indicating that as radiation climate becomes coarser, the amount of both novel and disappeared climates at the same location will increase. However, rain and T min had extremely low correlations, indicating that an increased occurrence of both novel and disappeared climates at the same will not always occur as climate resolution becomes coarser ( Figure 10).

| D ISCUSS I ON
As climate changes, either from natural processes and/or human activity, novel and disappeared climates will arise (Jackson & Overpeck, 2000;. Long-term climate patterns that affect observed novel and disappeared climates, whether in the past, present, or future, arise not only from geographic and temporal changes in the environment, but also as a result of atmospheric and topographic processes that occur across many spatial scales (Ackerly et al., 2010).
Determining how climate resolution affects climate change predictions is necessary to avoid under or overestimating the potential impacts of climate change for ecological analyses dependent on gridded climate data. We downscaled climate data at 11 spatial resolutions spanning three orders of magnitude to include climate data at resolutions used for previous climate change impact projections to identify the distribution and abundance of novel, disappeared, and shared climates across Alaska, as well as how the spatial resolution of climate data used affects these estimates of climate (e.g., PRISM, BioClim, NASA Nex, CMIP; Daly et al., 2004;Fick & Hijmans, 2017;Thrasher et al., 2013). We found that novel and disappeared climates primarily affected Southern Alaska. Additionally, climate data generally increased the fractional area of novel and disappeared climates as the resolution became coarser; however, it was also possible, although rare, to decrease the fractional area or have no apparent relationship between the fractional area of novel and disappeared climates and climate data resolution. T max climates may still fall within the optimal range of the species.

| Novel, disappeared, and shared climate distributions
However, while changes to LGM snow climates were low (2.3% disappeared snow), this change may be ecologically risky for the species if the changes fall outside its optimal snow range. In this case, disappeared snow climates are significant to a species, not disappeared T max climates, even though there was a far greater fractional area of disappeared T max climates. This suggests that disappeared or novel climates may only be important if they are a limiting factor or extend beyond the optimal niche for a species (De Baar, 1994;Vandermeer, 1972). If the change in climate is a limiting factor, the risk of distribution shifts or extirpation may increase for a species. However, if the change in climate is not limiting, the existence of novel and disappeared climates may have minimal impacts on a species.

| Climate grid resolution and fractional area patterns
In our study, the amount of novel and disappeared climate distributions was affected by climate grid resolution in two primary ways. First, as climate data becomes coarser, there is a positive relationship between the fractional area of novel and disappeared climates and spatial resolution. This pattern occurs for two reasons: (a) Coarser climate data reduced environmental variability, which in turn removes the extreme ranges of climate, and (b) there are fewer climates that fall within the shared climate bin definition, resulting in fewer climates classified as shared while simultaneously increasing the frequency of novel or disappeared climates simply due to the coarser scales ( Figure 11). As expected, all shared climates displayed a negative relationship between climate resolution and fractional area (Figure 9). Shared climates should display a negative relationship because as the fractional area of novel and disappeared climates increases with spatial resolution, the fractional area of shared climates must decrease because it accounts for areas not classified as either novel or disappeared. Previous studies by Franklin et al. (2013), Heikkinen, Luoto, Kuussaari, and Toivonen (2007), and Seo et al. (2009) Figure 12). At coarser resolutions, our T min climate aggregation scheme successively grouped similar temperatures together at each progressively coarser resolution (Wu, 1999 No significant relationship between fractional area of novel and disappeared climates and spatial resolution is possible, and this pattern did not visibly arise within our study. There were some cases where the Spearman's rank correlation was "moderate", although this does not necessarily indicate a relationship. However, while all surfaces displayed moderate to strong correlations, there was a fair amount of variation in fractional area as spatial resolution became coarser. This suggests that there may be significant thresholds occurring at various spatial scales. The exact causes of this variation are still unclear, but we hypothesize that this variation may be a result of distortion caused by aliasing artifacts when upsampling our 60 m downscaled climate surfaces to coarser scales (Haines-Young & Chopping, 1996;Kennie & McLaren, 1988), or may be caused by Moiré fringes (Gustafsson, 2000) producing F I G U R E 1 2 Coarser resolutions can decrease the amount of novel and disappeared climates in Alaska. Disappeared T min at 120 m displays regions with patchy, but random spatial distributions, while coarse disappeared T min climates display clustered distributions. Isolated patches of disappeared T min climates at high resolutions will be reclassified to shared T min climates when upsampling T min climates to coarser scales using standard pixel aggregation methods overlooked oscillations across a landscape. While Moiré fringes are a well-known issue affecting raster image processing and visualization, to the best of our knowledge, no studies have investigated whether naturally occurring processes and phenomenon known to influence climate or other Earth systems, generate spatial Moiré distributions.

| Limitations
Annual averages are a limitation because summer temperatures are not so different from today; however, the rest of the year was, so annual averages do not capture the seasonal climate variation over Alaska. This is only an issue in terms of when we are talking about the interpretation of novel and disappeared climates impact on plants, not on how scale affects the amount of novel and disappeared climates predicted since seasonality does not affect the definition of a novel or disappeared climate. Future studies should look into seasonal variation of novel and disappeared climates.

| CON CLUS IONS
It is important to understand how climate change through time affects the distribution and abundance of novel and disappeared climates, as well as how choice in gridded climate data resolution affects our estimations of climate change. If our 60 m downscaled climate surfaces accurately reflect physiographic processes affecting climate at multiple scales, our fine-scale climate products should improve estimations of novel and disappeared climate distributions, compared to coarser climate products, especially in areas with high topographic complexity.

CO N FLI C T O F I NTE R E S T
None declared.