Statistically downscaled climate data have been widely used to explore possible impacts of climate change in various fields of study. Although many studies have focused on characterizing differences in the downscaling methods, few studies have evaluated actual downscaled datasets being distributed publicly. Spatially focusing on the Pacific Northwest, we compare five statistically downscaled climate datasets distributed publicly in the US: ClimateNA, NASA NEX-DCP30, MACAv2-METDATA, MACAv2-LIVNEH and WorldClim. We compare the downscaled projections of climate change, and the associated observational data used as training data for downscaling. We map and quantify the variability among the datasets and characterize the spatio-temporal patterns of agreement and disagreement among the datasets. Pair-wise comparisons of datasets identify the coast and high-elevation areas as areas of disagreement for temperature. For precipitation, high-elevation areas, rainshadows and the dry, eastern portion of the study area have high dissimilarity among the datasets. By spatially aggregating the variability measures into watersheds, we develop guidance for selecting datasets within the Pacific Northwest climate change impact studies.
Projections of historical and future climate from general circulation models (GCM) have been widely used in studies that explore climate change impacts, such as ecology, biogeography, conservation planning, and natural resource management1–3. However, GCM outputs are at coarse spatial resolutions (~1 to 3 degrees), and are not sufficiently fine to capture spatial variations of climate in physiographically complex landscapes4. Many climate impacts appear only at local scales5, and local climate trends do not always follow global trends6,7. Consequently, the original GCM outputs may have significant biases in representing the statistical characteristics (e.g., mean, variance, and trend) of local climate conditions8. Therefore, the original GCM outputs cannot be directly applied to many fine-scale studies of climate change impacts9.
To overcome the inherent limitations of GCM outputs for fine-scale applications, climatologists have developed various techniques to generate fine-resolution climate projections from the coarse-resolution climate change projections output by GCMs10,11. The downscaling techniques aim to account for locally or regionally relevant factors that affect climate, such as orographic and coastal effects. Compared with the original GCM outputs, the downscaled climate projections are expected to better represent fine-scale weather patterns, especially for areas with heterogeneous landscape features12. In one class of downscaling methods called statistical downscaling, observational climate data from the recent past are used as “training data” to establish statistical relationships between coarse-resolution GCM output and the fine-resolution training data, and those relationships are used to generate fine-resolution future projections from coarse-resolution GCM projections of future climate. Statistical downscaling is relatively computationally efficient, and many statistically downscaled climate change projections have been published online. The published downscaled climate projections provide scientists and land managers an opportunity to explore climate change impacts at local and regional scales, and, in recent years, many statistically downscaled climate datasets have been used to assess the impact of climate change on hydrology13,14, species distributions15, biodiversity16, biological application17, and fire modeling18.
To date, several studies have compared statistical downscaling methodologies5,19–21, such as Constructed Analogues10, Bias Corrected Spatial Disaggregation22, and Bias Corrected Constructed Analog23. Because the primary focus of these comparison studies is the downscaling methodology itself, not the resulting downscaled datasets, the comparison studies use the same GCM output, the same training data, and spatial and temporal resolution and extents to eliminate them as possible sources of differences in the final downscaled data. While this is appropriate for studying the downscaling techniques, the studies do not provide direct assessments of the actual downscaled datasets that are being distributed to the research community and the public. The published datasets may exhibit differences arising from individual GCM bias, the characteristics of the training data, the specific implementation of downscaling algorithms, and data storage and representation issues. To our knowledge only one study, Wootten et al.24, has compared an ensemble of downscaled datasets that are being distributed to the public. Wootten et al.24 focuses on data downscaled for the Southeastern USA, and is limited to data downscaled from only one GCM. A similar study is needed for other regions, as well as for other GCMs and downscaling methods. A comprehensive evaluation of downscaled climate datasets can provide greater confidence in the projections of climate change25, and, in turn, provide greater confidence in the results of climate change impact studies that rely on them.
In this paper, we compare four downscaled climate datasets published online that cover the Pacific Northwest of USA: ClimateNA (https://adaptwest.databasin.org/pages/adaptwest-climatena/), NASA NEX-DCP30 (https://cds.nccs.nasa.gov/nex/), MACAv2-METDATA (https://climate.northwestknowledge.net/MACA/) and MACAv2-LIVNEH (https://climate.northwestknowledge.net/MACA/). We compare downscaled climate projections generated from six GCMs simulating the Representative Concentration Pathway (RCP) 8.5 climate change scenario. We identify areas of agreement and disagreement, quantify variation across the ensemble of projections, and assess the contribution of the respective training datasets to the differences. Additionally, we compare three of the datasets to WorldClim (http://www.worldclim.org/), another highly cited downscaled climate dataset for which only period averages are published, with no training data. We propose that characterizing the differences among the downscaled datasets provides data users some context for selecting a downscaled climate dataset for a particular application. Finally, we provide summarized data by watersheds at different scales, to serve as guidance for selecting one or more downscaled climate datasets.
Comparison of downscaled historical climate
Pair-wise comparisons of mean annual temperature and mean annual precipitation from 1961 to 1990 computed from the four downscaled climate datasets show differences among the downscaled datasets, as well as spatial patterns of those differences (Fig. 1). The differences are averaged across the six GCMs common to the four datasets. MACAv2-LIVNEH is the most dissimilar among the compared datasets, in terms of both temperature and precipitation. The mean annual temperature given by MACAv2-LIVNEH is 2-4° C cooler than that given by the other datasets across much of the study area, with the highest differences occurring over mountain ranges and peaks. In terms of mean annual precipitation, MACAv2-LIVNEH deviates from the other datasets across the study area, with values 50–80% higher on the windward slopes of the Cascades, and values 40–50% lower on the leeward slopes of the Cascades, as well as in the interior mountain ranges. These differences can be over 1,000 mm yr−1 in absolute precipitation amounts (Supplementary Fig. S1). MACAv2-METDATA exhibits less dissimilarity to NEX-DCP30 and ClimateNA than MACAv2-LIVNEH. Its mean annual temperature values are ~1° C higher than NEX-DCP30 and ClimateNA at all higher elevation portions of the study area; while its precipitation values are lower than those given by NEX-DCP30 and ClimateNA in the Cascades and to the west.
Compared with WorldClim for the period 1950-2000, NEX-DCP30 and MACAv2-METDATA are warmer and wetter, and MACAv2-LIVNEH is cooler and wetter (Table 1). The differences in temperature were due largely to differences in the monthly means of daily minimum temperature (Tmin), where the Tmin for NEX-DCP30 and MACAv2-METDATA were ~1° C warmer than WorldClim throughout the year, whereas the Tmin for MACAv2-LIVNEH was ~0.5° C lower than WorldClim. Monthly means of daily maximum temperature (Tmax) of the downscaled datasets deviated from WorldClim by as much as 0.5° C in winter and spring, but it was nearly identical to WorldClim for summer and autumn. All three downscaled datasets compared against WorldClim had higher mean annual precipitation, exceeding WorldClim by 21 to 60 mm per season in the spring, fall and winter.
Comparison of downscaled future climate projections
The magnitudes of change in temperature and precipitation projected by the datasets are generally similar to each other (Fig. 2). They vary by only +/− 0.3° C for projected increases in mean annual temperature from 1961-1990 to 2071–2100, which is less than a tenth of the overall climate warming projected for the region by the majority of GCMs. For example, MACAv2-METDATA’s projected warming is only about 0.1° C lower than those projected by NEX-DCP30 and ClimateNA for the eastern half of the study area. Projected magnitudes of change in precipitation generally vary less than 6% among the datasets for most of the study area, with isolated areas of higher differences occurring on the leeward slopes of the Cascades and in the dry, southeastern interior portions of the study area. Dissimilarities between all pairs of datasets are comparable in magnitude and spatial extent, although distinct spatial patterns of differences are displayed. For example, the dissimilarity between NEX-DCP30 and ClimateNA are clustered on the arid southeast portion of the study area, and the difference maps involving MACAv2-LIVNEH exhibit greater spatial heterogeneity. These differences in precipitation are minimized significantly when considered in terms of absolute amount of precipitation (Supplementary Fig. S2). Compared that way, areas of greatest differences in precipitation are generally confined to the windward side of the Cascades.
Pair-wise comparisons of downscaled future (2071–2100) climate projections (Supplementary Fig. S3) produce nearly identical results as the pair-wise comparisons of the downscaled historical climate data (Fig. 1) because the projected changes differ relatively little between datasets. As noted in the comparisons of downscaled data for the historical period, projected temperature can differ by as much as 4° C in mountains between two datasets, while projected mean annual precipitation can differ by as much as 2,000 mm yr−1 (Supplementary Fig. S3). Again, these are large and significant differences, relative to the 3-7° C warming projected by the ensemble of CMIP5 GCMs under RCP8.5 (Fig. 3) and relative to mean annual precipitation of 974 mm yr−1 for the study area.
Although the projected climates are dissimilar among the downscaled datasets, the magnitudes of climate change projected by the downscaled datasets are highly similar. In other words, even where two datasets project different mean annual temperatures for the 2071-2100 period, the amount of projected warming relative to the reference historical period (1961-1990) are similar. Standard deviations of projected magnitudes of change in annual means of Tmax and Tmin from a reference historical period (1961–1990) to the end of the century (2071–2100) is less than 0.25° C for the ensemble of four downscaled climate datasets (Fig. 4). Of the two temperature metrics, there was less agreement among the datasets for Tmin. For projected magnitudes of change in Tmax, the standard deviation of the ensemble was less than 0.05° C for the vast majority of the study area. For projected magnitudes of change in Tmin, the standard deviation was significantly higher in the Cascade Range to the east, with the highest disagreement in northern Washington, where the northern portions of the Cascades and the Northern Rockies are located.
For projected magnitudes of change in mean annual precipitation, the coefficient of variation across the downscaled datasets is generally low west of the Cascade Mountain Range where the precipitation is high. To the east Cascade crest where precipitation is generally lower, however, there are areas where the coefficient of variation ranged 30–60%, representing high levels of disagreement among the downscaled datasets. These areas are located on the leeward side of the Cascade Range and the Wallowa Mountains in Eastern Oregon, where rainshadows occur.
The differences among the dataset in their projections of change in climate, with respect to the reference historical period (1961–1990), exhibit distinct seasonal patterns that vary throughout the century (Fig. 5). With all four of the downscaled datasets the seasonal cycles become increasingly accentuated from the early- (2011–2040) to mid- (2041–2070) to the end-century (2071–2100): Tmax becomes increasingly warmer in late summer (July-August); Tmin increases similarly in both winter and summer; and mean monthly precipitation increases in the winter while decreasing in the summer. In all three time periods considered, ClimateNA’s projections show greater amplitude in seasonal variations compared to the other datasets. Its projections for change in Tmax are generally greater March to July, while its projections for change in Tmin are smaller for October and November, compared to the other datasets. ClimateNA also projects a significantly higher increase in November precipitation compared to the other datasets, while projecting smaller increases in winter and spring. These results represent projections of change averaged across six GCMs common to the four datasets. Projected changes by season by individual GCM are detailed in Supplementary Table S1.
Projected magnitudes of change in climate from the reference historical period (1961–1990) to the end of the century (2071–2100) vary greatly according to proximity to the coast and by elevation (Fig. 6). Projected magnitudes of change in Tmax averaged across all latitudes of the study area show a high level of agreement among the four downscaled climate datasets, except at the coast (Fig. 6a). Projected magnitudes of change in Tmax also disagree at the coast but also diverge significantly east of the Cascades. In terms of projected magnitudes of change in mean annual precipitation, NEX-DCP30 and MACAv2-METDATA project 2–5% greater increase than ClimateNA and MACAv2-LIVNEH in the eastern half of the study area.
Latitudinally averaged transects illustrate the differences among the datasets in representing orographic effects in greater detail. Along the northern transect (48.4–49 °N) (Fig. 6b), which spans the northern Cascades as well as the Rockies, MACAv2-LIVNEH exhibits greater variability in temperature increases over the complex, mountainous topography. NEX-DCP30 and ClimateNA project ~2% greater increases in mean annual precipitation at the crest of the Cascades, with intensifying rainshadow effects immediately east of the crest. Along the central transect (44.5–45.5 °N) (Fig. 6c), MACAv2-METDATA exhibits intensification of the marine effect on temperature, with its projected increase in Tmax and Tmin 0.3° C lower than the other three datasets. In terms of precipitation, NEX-DCP30 and MACAv2-METDATA project clearly larger increase in mean annual precipitation over the range of −121.0o–−119.0o. Along the southern transect (42oN–43oN) (Fig. 6d), variability in projected increases in Tmax is notably muted. NEX-DCP30 and ClimateNA show slightly higher increase in Tmin than the other two datasets over the eastern half of the study area. On the east side of Cascades and further inland, NEX-DCP30 and MACAv2-METDATA project greater increases in precipitation than the other two datasets.
Differences in training data used for downscaling
The observational data shows clear difference among datasets (Table 2). NEX-DCP30 and ClimateNA both use PRISM as the training data. The major difference among training datasets is that LIVNEH has significantly cooler winters, thereby contributing to cooler annual temperatures. In particular, for the period of 1961–1990, LIVNEH has 0.21 °C lower annual Tmax, and 1.39 °C lower Tmin than PRISM. For period 1981–2010, LIVNEH has 0.17 °C lower annual Tmax, 1.49 °C lower Tmin than PRISM. In terms of mean annual precipitation and seasonal precipitation, PRISM, METDATA and LIVNEH are similar, each varying from the ensemble average by less than 3%. The three training datasets have similar spatial patterns across the Pacific Northwest region (Fig. 7). The visible difference in Tmax occurs in Northern Cascades where LIVNEH has lower Tmax than the other datasets. PRISM and METDATA have similar annual Tmin in coastal areas, larger than that in LIVNEH. LIVNEH has substantially lower Tmin than other datasets in the northern Cascades and in the John Day watershed. LIVNEH also shows substantially lower precipitation in the Cascades than the other datasets.
Guidance for dataset selection
Variability in projected future (2071–2100) climate among the datasets is spatially aggregated to watersheds to summarize uncertainty of climate change projections across datasets and to aid their selection, application and interpretation (Fig. 8a). Four different scales of watersheds defined by the US Geological Survey26 are used. For all three climate variables the differences among datasets are minimized when aggregated to large, regional watersheds. Conversely the differences become more acute for small watersheds. For example, at the US Geological Survey hydrologic unit code level 4 (HUC4) and HUC6 levels26, the projected Tmax of all the watersheds differ by less than 0.4 °C. At the HUC10 level, some watersheds have projections that vary by as much as 2.24 °C. More watersheds are divergent on projections of Tmin than the other two climate metrics. At the HUC10 level, projections of Tmax are divergent along the coast and on the western slope of the Cascades, with some watershed having datasets that vary as much as 2.2 °C. Projections of Tmin are the most divergent in high-elevation watersheds across the study area. With projections of precipitation, several watersheds – but not all – in rainshadows of mountains are indicated as having highly divergent data. Detailed rankings of watersheds at the HUC8 level are given in Supplementary Table S2.
Variability in projected magnitudes of change to climate aggregated to watersheds may aid studies where the magnitudes of change with respect to a reference period, rather than absolute projections of climate, are more relevant (Fig. 8b). As with projected climate aggregated to watersheds, differences among the datasets are minimized when aggregated to large, regional watersheds. For example, differences in changes to Tmax from 1961–1990 to 2071–2100 among the four downscaled datasets are less than 0.06 °C at HUC4 and HUC6 levels26. As with projected climate, projected change in Tmin is the most divergent, with the vast majority of HUC10 watersheds on the east side of the cascades having relatively more divergent projections of change. For projected magnitudes of change in Tmax, the watersheds with the most divergent datasets are on the west side of the Cascades, especially in the Olympic Peninsula in the northwest corner of the study area. However, as noted previously, the differences in projected magnitudes of change to temperature are small (<0.16 °C). For projected magnitudes of change in precipitation, the drier watersheds in the rainshadows of the Cascades have the most divergent datasets, while the wetter watersheds on the west side of the Cascades exhibit better agreement.
While the differences among the downscaled datasets in their projections of change in climate are generally small for much of the study (Fig. 8b), the projections of future climates vary widely for many HUC8 and HUC10 level watersheds the study area (Fig. 8a). These differences may be critical where biological or physical responses to climate change are non-linear or depend on thresholds. For example, in Western Cascades, as mean annual temperature increases, there is an increased rate of reduction in vegetation carbon due to fire activity27. In that case, selecting a downscaled dataset that is 2 °C warmer than another has larger consequences than may be inferred from the 2 °C difference. In another example, MC1 dynamic global vegetation model triggers fire suppression efforts based on fireline intensity and rate of spread thresholds28. The use of a “warmer” downscaled climate dataset with MC1 instead of a “less warm” dataset would alter whether and when those thresholds are exceeded.
The variability of datasets aggregated by watersheds (Fig. 8) can serve as a practical guide for selecting downscaled datasets for sub-regional-scale studies. Where the variable(s) of interest (i.e., Tmax, Tmin or precipitation) have a high degree of variability among datasets, there is a greater need to use multiple downscaled datasets to capture the range of future climate projections and the ecological response to those projections. For example, in the Olympic Peninsula in the northwest corner of the study area, there is a relatively high degree of variability among the downscaled datasets, whether their direct climate projections are considered or the magnitude of changes is considered. The average Tmax and Tmin may vary by over 2 °C depending on the choice of downscaled climate dataset, while precipitation may vary by over 20%. Using only one downscaled dataset there may severely bias projections of climate change impacts. Conversely, the watershed maps identify areas where there are good agreements among datasets – this is an important benefit, as dataset selection is typically constrained by many criteria: the need for relevant climate variables, the need to seek the finest resolution data (e.g.,<4 km for plant species distributions29), maximize accuracy24, select high-performing GCMs among the 34 GCMs published by CMIP5, and select appropriate RCP climate change scenarios30. Because not every dataset (Table 3) has all the desired qualities, many study areas may be limited to one downscaled dataset. When only one dataset is suitable, yet there is high uncertainty in a study area due to variability among the downscaled datasets, using a large number of GCMs within the single dataset may help address the range of variability across datasets.
Dataset variability maps (Fig. 8) must be interpreted with some caution because large watersheds, e.g., HUC4 watersheds, include the smaller scale watersheds that have high variability of datasets. Therefore, even regional studies must remain cautious about uncertainty of projections in the smaller watersheds, as well as geographic and topographic features where divergence among datasets is observed (Fig. 6). The coast and high-elevation areas are subject to dissimilar temperature projections, and mountain peaks and their rainshadows may have divergent projections of precipitation.
Pair-wise comparisons of datasets for both the historical and future climate yield nearly identical results (Figs. 1 and 2), and the variability of projected magnitudes of changes in climate is small among the datasets (Fig. 4). These results suggest that variability among downscaled climate datasets is mainly controlled by differences in the observational climate data used as training data for the downscaling methods. Indeed, MACAv2-LIVNEH, which had the most dissimilar climate projections to the other datasets, uses LIVNEH training data, which is the most dissimilar to the other training data (Table 2). Comparisons of MACAv2-METDATA with MACAv2-LIVNEH, which uses the same method but different training data, further illustrate the important role of training data. Projected changes in temperature, averaged latitudinally, vary significantly between MACAv2-METDATA and MACAv2-LIVNEH (Fig. 6b), as do projected changes in precipitation. ClimateNA assimilates only one PRISM normal period (for 1971–2000) into its training data, and calculates normals for other periods using CRU-TS 3.22 dataset31. This is a potential source of differences between the ClimateNA and NEX-DCP30, and MACAv2-METDATA.
Downscaling methods do partly drive dissimilarity of datasets in some places. We identify two general areas where projections by NEX-DCP30 and MACAv2-METDATA diverge: one, east of the Cascades, projected changes Tmin diverge by more than 0.2 °C (Fig. 6a); and two, at the Cascade crest, projected changes in mean annual precipitation diverge by more than 2% (Fig. 6b). NEX-DCP30 and MACAv2-METDATA use PRISM and METDATA as training data, respectively, and METDATA is based on PRISM. The uncertainty in high-elevation area likely arises from sparse coverage of weather observations, and poorly understood mechanisms of climate change at fine scales.
The downscaled climate datasets we compare use observational climate data as training data, and those observational data have common weather stations. The datasets and the training data are not independent, and pair-wise comparisons are predisposed toward agreement. For the same reason, a truly independent observational dataset does not exist for evaluating downscaled climate datasets. WorldClim (http://www.worldclim.org/) does not serve as an independent observational dataset to compare the downscaled historical climate data because the weather stations from which it draws its data are partly in common with the training data used by the downscaled datasets. A bootstrapping statistical method may effectively address the lack of independent observational dataset for validation.
We compared a limited number of climate variables because they were the only ones common to the four selected downscaled datasets. All except NEX-DCP30 also include variables representing water vapour (e.g., relative or specific humidity) and wind. ClimateNA is available as a software package, which the user can use to downscale future climate projections for a given area, including 27 bioclimate variables. Comparison of additional variables may reveal other important areas of similarity and dissimilarity. We compared climate variables representing monthly averages and totals (temperature and precipitation, respectively). An important disadvantage of monthly data is that they may not sufficiently represent extreme climate events (e.g., drought, heat wave), which can have critical impact on vegetation32, such as forest die-off33. MACAv2-METDATA has been demonstrated to be modestly better than data downscaled using BCSD for simulating climate change impacts on wildfire dynamics11. The coarse temporal resolution of data hinders exploration of ecological processes with fast response to thermal and water stresses. Data downscaled using Multivariate Adaptive Constructed Analogs (MACA), where climate projections are constructed from a library of observed fine-scale weather patterns, are likely to better represent the occurrence of extremes than those that downscale using average climate surfaces.
We focused on the Pacific Northwest region of the United States, defined as the geographic area from −124.59o to −116.5o longitude, and from 42o to 49o latitude. The study area is 4.4×105 km2 in size and spans two states, Washington and Oregon, in the northwest corner of the conterminous U.S. The climate in the region is Mediterranean, with wet but mild winters, and warm and dry summers. The Cascade Range runs north to south, bisecting the study area. West of the Cascades, moist coniferous forests dominate the landscape. East of the Cascades, the significantly drier climate supports only shrublands and grasslands at lower elevations, while forests occupy higher elevations.
Downscaled climate datasets
We compared four downscaled climate datasets that are distributed publicly and cover the Pacific Northwest at a relatively high resolution, from 0.5′ to 3.75′: ClimateNA (https://adaptwest.databasin.org/pages/adaptwest-climatena/), NASA NEX-DCP30 (https://cds.nccs.nasa.gov/nex/), MACAv2-METDATA (https://climate.northwestknowledge.net/MACA/), and MACAv2-LIVNEH (https://climate.northwestknowledge.net/MACA/). Because the four selected datasets are described in detail elsewhere34–36, we provide only a brief overview of each dataset below, and summarize their key characteristic in Table 3. As a set, the four selected downscaled climate datasets represent multiple statistical downscaling methods applied by multiple institutions, using multiple observational data. Bureau of Reclamation distributes a suite of statistically downscaled climate datasets covering the Pacific Northwest37, as does NARCCAP38, an ensemble of dynamically downscaled climate datasets, but we excluded both datasets from our analysis because they have significantly coarser spatial resolution (7.5′).
Each statistical downscaling method used to create the four downscaled climate datasets requires the use of an observational climate data as training data. ClimateNA and NEX-DCP30 were created using Parameter-elevation Regressions on Independent Slopes Model (PRISM)39 as training data; MACAv2-METDATA used METDATA36; and MACAv2-LIVNEH used LIVNEH40. To understand the importance training data, we obtained each training dataset, and compared their characteristics.
For comparing future climate projections, we selected two representative concentration pathways30, RCP4.5 and RCP8.5, to represent the low and high levels of greenhouse gas forcings41. We selected downscaled data from the six Coupled Model Intercomparison Project Phase 5 (CMIP5)42 GCMs that are common to all of the four downscaled datasets: CanESM2, CCSM4, CNRM-CM5, HadGEM2-ES, INMCM4, and IPSL-CM5A-MR. The twelve climate projections we evaluated (two RCPs x six GCMs) project a wide range of climate change in the Pacific Northwest by the end of the century in terms of mean annual temperature and precipitation (Fig. 3). Under RCP8.5, the projected rise in temperature ranges over 3 °C among the six selected GCMs, while the projections of change in precipitation range from 5 to over 20%. Notably, all projections we evaluated project an increase in precipitation, with the exception of INMCM4 under RCP4.5, which projects a small decrease in mean annual precipitation. The six selected GCMs have been evaluated for their ability to simulate the recent historical climate of the Pacific Northwest43, and all except for INMCM4 are among the top twelve best performers out of the 41 GCMs evaluated. INMCM4 was ranked 30 out of 41. For each of the twelve climate projections we selected three climate variables that are common to all four climate datasets: monthly mean of daily maximum temperature (Tmax), monthly mean of daily minimum temperature (Tmin) and monthly precipitation.
ClimateNA uses the Delta method31, the simplest of the statistical downscaling methods, to downscale GCM output for North America35,44–46. Multiple training datasets are used to cover North America, with PRISM39 dataset being used for the conterminous U.S. portion of the dataset46. PRISM uses physiographic information to interpolate observations from over 10 000 weather stations into 30″ gridded data39. We evaluated the 1km ClimateNA dataset published by A Climate Adaptation Conservation Planning Database for Western North America (AdaptWest) (https://adaptwest.databasin.org/pages/adaptwest-climatena/). Although the downscaling method used by ClimateNA is relatively simple, ClimateNA prioritizes accessibility and utility by spanning the whole continent and including 27 bioclimatic variables. It is also distributed as a software package, so that end-users may generate their own data. ClimateNA dataset published by AdaptWest comprise eight high performing CMIP5 GCMs selected to represent clusters of similar GCMs47. The published dataset comprises climate normals for two historical periods, 1961-1990 and 1981-2010, and three future time periods under two climate change scenarios30, RCP4.5 and RCP8.5: 2020 s (2011–2040), 2050 s (2041–2070), and 2080 s (2071–2100).
NEX-DCP30 (https://cds.nccs.nasa.gov/nex/) comprises CMIP5 GCM outputs downscaled to 30″ resolution (approximately 800 m) for the conterminous US using the Bias-Correction Spatial Disaggregation (BCSD) method34. BCSD is more sophisticated than the Delta method43 in that it uses quantile-mapping to bias-correct the lower resolution GCM data to match the distribution of values in the training data before spatially disaggregates the GCM data to match the finer scale spatial patterns of the training data19,22,48. NEX-DCP30 used PRISM39 as the training dataset. Among the four datasets we compared, NEX-DCP30 comprised the greatest number of CMIP5 climate projections, including 34 GCMs under four different RCPs. The variables downscaled are limited to just three fundamental metrics – Tmax, Tmin and precipitation – but the values are given for every month from 1950 to 2099 for each GCM-RCP combination.
MACAv2-METDATA (https://climate.northwestknowledge.net/MACA/) was generated using the Multivariate Adaptive Constructed Analogs (MACA) method11, which builds the downscaled climate data time series by matching coarse-scale GCM patterns to a library of fine-scale observed weather patterns, along with a two-stage of bias correction procedure. The MACA method performs better than BCSD in regions of complex terrain11. MACAv2-METDATA uses METDATA36 as training data, which combines NASA North American Land Data Assimilation System (NLDAS)26 with PRISM39. METDATA covers the conterminous US at 2.5’ resolution (approximately 4km) from 1979 to present. The downscaled climate dataset comprises twenty CMIP5 GCMs under two emission scenarios, RCP4.5 and RCP8.5. The data covers the conterminous US at 2.5’ resolution. A suite of 9 climate variables, including maximum and minimum temperature and precipitation, are available at a daily and monthly time step from 1950 to 2099.
MACAv2-LIVNEH (https://climate.northwestknowledge.net/MACA/) was also generated using the MACA method11, but it was generated using LIVNEH observation dataset as training data. LIVNEH is a long-term hydrologically based dataset of land surface fluxes and states for the conterminous US40. It is gridded at a spatial resolution of 1/16 degrees (3.75′ or approximately 6 km), based on interpolated daily temperature and precipitation observations from ~20 000 NOAA Cooperative Observer stations. MACAv2-LIVNEH dataset includes the same ensemble of GCMs and RCPs as MACAv2-METDATA, and the data have the same spatial extent and temporal resolution and extent as MACAv2-METDATA; however, MACAv2-LIVNEH offers a more limited number of climate variables (Table 3).
WorldClim (http://www.worldclim.org/) covers global lands surfaces except for Antarctica at 30 arc-second resolution. WorldClim’s 1km resolution climate surfaces were created by interpolating weather station data from Global Historical Climatology Network (GHCN)49, World Meteorological Organization (WMO) Climatological normals (CLINO)50, Food and Agriculture Organization of the United Nations Agroclimatic Database (FAOCLIM)51, plus several regional databases. WorldClim uses 14,835 sites for maximum and minimum temperatures and 47,554 sites for precipitation, some of which may be used also in PRISM, METDATA and LIVNEH. The technique used to create WorldClim’s climate surfaces is unique: it uses a thin-plate smoothing spline algorithm to fit a second-order spline where latitude, longitude, and elevation are independent variables52. WorldClim publishes data only as period averages: 1960–1999 for the recent historical period, and two future periods, 2041–2060 and 2061–2080. Future data are available for RCP2.6, RCP4.5, RCP6.0 and RCP8.5, for 19 GCMs.
Comparing the downscaled datasets
NEX-DCP30, MACAv2-METDATA, and MACAv2-LIVNEH publish monthly data online, but ClimateNA and WorldClim publish only period averages. The periods published by ClimateNA and WorldClim do not align. Therefore, we chose to focus our comparison on just four datasets – NEX-DCP30, MACAv2-METDATA, MACAv2-LIVNEH, and ClimateNA – by aligning the periods of analysis with those published by ClimateNA. We chose ClimateNA over WorldClim because ClimateNA has a smaller spatial focus (North America vs. global).
We focus on climate variables that are common to the four downscaled climate datasets we focus on: monthly mean of daily maximum temperature (Tmax), monthly mean of daily minimum temperature (Tmin), and total monthly precipitation downscaled from six GCMs under RCP8.5. MACAv2-LIVNEH downscaled dataset and the LIVNEH training dataset has a spatial resolution of 3.75’ (1/16°), the coarsest among the four downscaled datasets. We interpolated all other datasets to 3.75’ (1/16°) to facilitate comparison.
We examine the differences among the downscaled climate datasets by calculating the differences between every possible pair of datasets for mean annual temperature and mean annual precipitation for two periods, 1961–1990 and 2071–2100 (Figs. 1,2). Mean annual temperature is calculated as the average of annual means of daily maximum and minimum temperatures. Differences are calculated per GCM, and then averaged across the six GCMs common to the four datasets. We examine projected magnitudes of change in climate variables by subtracting the average value from the reference period (1961–1990) from future (2011–2040, 2041–2070, and 2071–2100) period averages.
To aid in the selection and application of downscaled climate datasets in our study area, we created maps of dataset variability for each of the three climate variables we compared, to identify areas with highly divergent climate datasets. To calculate dataset variability, we first calculated the average of six climate projections for 2071–2100 within a dataset, one climate projection per GCM. Then, for Tmax and Tmin, we calculated standard deviation across the four datasets. The resulting standard deviation grids were aggregated into four levels of watersheds – HUC4, HUC6, HUC8 and HUC10 – defined by the US Geological Survey Watershed Boundary Dataset26. For precipitation, we calculated coefficient of variation across the four datasets, so that the measure of variance is scaled by the magnitude precipitation at a given location. We repeated the above steps for projected magnitudes of change in temperature and precipitation for the four levels of watersheds, to create a set of maps that represent variability of magnitudes of change.
To characterize the possible role played by the training datasets in engendering differences among the downscaled datasets, we compared the three training datasets used to create the four downscaled datasets. We calculated seasonal and annual averages for each training dataset and then compared it to the average of the ensemble of the training datasets (Table 2).
As an additional analysis, we compare NEX-DCP30, MACAv2-METDATA and MACAv2-LIVNEH against WorldClim52 (http://www.worldclim.org/) by comparing their characteristics for the 1950–2000 period (Table 1). As previously noted, ClimateNA is published online as period averages with different time periods, and is excluded from this comparison.
Matlab code used for data analysis may be obtained freely by contacting the authors, with no restrictions to access.
How to cite this article: Jiang, Y. et al. Inter-comparison of multiple statistically downscaled climate datasets for the Pacific Northwest, USA. Sci. Data 5:180016 doi: 10.1038/sdata.2018.16 (2018).
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Bachelet, D, Neilson, R. P., Lenihan, J. M. & Drapek, R. J. Climate change effects on vegetation distribution and carbon budget in the United States. Ecosystems 4, 164–185 (2001).
Hannah, L., Midgley, G. F. & Millar, D. Climate change-integrated conservation strategies. Global Ecol. Biogeogr. 11, 485–495 (2002).
Alo, C. A. & Wang, G. Hydrological impact of the potential future vegetation response to climate changes projected by 8 GCMs. J. Geophys. Res. 113, G03011 (2008).
Xie, S. et al. Towards predictive understanding of regional climate change. Nature Climate Change 5, 921–930 (2015).
Gutmann, E. et al. An intercomparison of statistical downscaling methods used for water resource assessments in the United States. Water Resour. Res. 50, 7167–7186 (2014).
Helmuth, B. et al. Climate change and latitudinal patterns of intertidal thermal stress. Science 298, 1015–1017 (2002).
Cordero, E. C., Kessomkiat, W., Abatzoglou, J. T. & Mauget, S. A. The identification of distinct patterns in California temperature trends. Climatic change 108, 357–382 (2011).
Knutti, R. & Sedlácek, J. Robustness and uncertainties in the new CMIP5 climate model projections. Nature Climate Change 3, 369–373 (2013).
Osmond, B. et al. Changing the way we think about global change research: scaling up in experimental ecosystem science. Global Change Biology 10, 393–407 (2004).
Hidalgo, H. G., Dettinger, M. D. & Cayan, D. R. Downscaling with constructed analogues: Daily precipitation and temperature fields over the United States. California Energy Commission PIER Final Project Report CEC-500-2007-123 (2008).
Abatzoglou, J. T. & Brown, T. J. A comparison of statistical downscaling methods suited for wildfire applications. Int. J. Climatol. 32, 772–780 (2012).
Wilby, R. L. et al. The guidelines for use of climate scenarios developed from statistical downscaling methods. Supporting material of the Intergovernmental Panel on Climate Change (IPCC), prepared on behalf of Task Group on Data and Scenario Support for Impacts and Climate Analysis (2004).
Meenu, R., Rehan, S. & Mujumdar, P. P. Assessment of hydrologic impacts of climate change in Tunga-Bhadra river basin, India with HEC-HMS and SDSM. Hydrol. Process 1085–1099 (2012).
Flint, A. L. & Flint, L. E. Downscaling future climate scenarios to fine scales for hydrologic and ecologic modeling and analysis. Ecological Processes 1, 2 (2012).
Kremen, C. et al. Aligning conservation priorities across taxa in Madagascar with high-resolution planning tools. Science 320, 222–226 (2008).
Loarie, S. R. et al. Climate Change and the Future of California's Endemic Flora. PLoS ONE 3, e2502 (2008).
Stoms, D. M., Dashiell, S. L. & Davis, F. W Siting solar energy development to minimize biological impacts. Renewable Energy 57, 289–298 (2013).
Abatzoglou, J. T. & Kolden, C. A. Relationships between climate and macroscale area burned in the western United States. International Journal of Wildland Fire 22, 1003–1020 (2013).
Maurer, E. P. & Hidalgo, H. G. Utility of daily vs. monthly large-scale climate data: An intercomparison of two statistical downscaling methods. Hydrol. Earth Syst. Sci. 12, 551–563 (2008).
Shrestha, R. R., Schnorbus, M. A., Werner, A. T. & Zwiers, F. W. Evaluating Hydroclimatic Change Signals from Statistically and Dynamically Downscaled GCMs and Hydrologic Models. J. Hydrometeor 15, 844–860 (2014).
Wang, L., Ranasinghe, R., Maskey, S., van Gelder, P. H. A. J. & Vrijling, J. K. Comparison of empirical statistical methods for downscaling daily climate projections from CMIP5 GCMs: a case study of the Huai River Basin, China. International journal of climatology 36, 145–164 (2016).
Wood, A. W., Leung, L. R., Sridhar, V. & Lettenmaier, D. P. Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs. Climatic Change 15, 189–216 (2004).
Maurer, E. P., Hidalgo, H. G. & Das, T. The utility of daily large-scale climate data in the assessment of climate change impacts on daily streamflow in California. Hydrol. Earth Syst. Sci. 14, 1125–1138 (2010).
Wootten, A. et al. Downscaled climate projections for the Southeast United States: evaluation and use for ecological applications (No. 2014-1190). US Geological Survey (2014).
Hall, A. Projecting regional change. Science 346, 1461–1462 (2014).
Mitchell, K. E. et al. The multi‐institution North American Land Data Assimilation System (NLDAS): Utilizing multiple GCIP products and partners in a continental distributed hydrological modeling system. Journal of Geophysical Research: Atmospheres 109 (D7) (2004).
Vano, J. A., Kim, J. B., Rupp, D. E. & Mote, P. W. Selecting climate change scenarios using impact‐relevant sensitivities. Geophysical Research Letters 42, 5516–5525 (2015).
Rogers, B. M. et al. Impacts of climate change on fire regimes and carbon stocks of the US Pacific Northwest. Journal of Geophysical Research: Biogeosciences 116 (G3) (2011).
Franklin, J. et al. Modeling plant species distributions under future climates: how fine scale do climate projections need to be? Global Change Biology 19, 473–483 (2013).
Van Vuuren, D. P. et al. The representative concentration pathways: an overview. Climatic change 109 (1-2) (2011).
Mitchell, T. D. & Jones, P. D. An improved method of constructing a database of monthly climate observations and associated high-resolution grids. Int. J. Climatol. 25, 693–712 (2005).
Jentsch, A., Kreyling, J. & Beierkuhnlein, C. A new generation of climate-change experiments: events, not trends. Front Ecol. Environ. 5, 365–374 (2007).
Breshears, D. D. et al. Regional vegetation die-off in response to global-change-type drought. Proceedings of the National Academy of Sciences of the United States of America 102, 15144–15148 (2005).
Thrasher, B. et al. Downscaled climate projections suitable for resource management. Eos Trans. AGU 94, 321–323 (2013).
Hamann, A., Wang, T., Spittlehouse, D. L. & Murdock, T. Q. A comprehensive, high-resolution database of historical and projected climate surfaces for western North America. Bulletin of the American Meteorological Society 94, 1307–1309 (2013).
Abatzoglou, J. T. Development of gridded surface meteorological data for ecological applications and modelling. Int. J. Climatol. 33, 121–131 (2013).
Brekke, L., Thrasher, B. L., Maurer, E. P. & Pruitt, T. Downscaled CMIP3 and CMIP5 climate projections: release of downscaled CMIP5 climate projections, comparison with preceding information, and summary of user needs. US Department of the Interior, Bureau of Reclamation, Technical Service Center (2013).
Mearns, L. O. et al. A regional climate change assessment program for North America. EOS 90, 311–312 (2009).
Daly, C., Gibson, W. P., Taylor, G. H., Johnson, G. L. & Pasteris, P. A knowledge-based approach to the statistical mapping of climate. Climate Research 22, 99–113 (2002).
Livneh, B. et al. A long-term hydrologically based data set of land surface fluxes and states for the conterminous U.S.: Update and extensions. J. Clim 26, 9384–9392 (2013).
Rogelj, J., Meinshausen, M. & Knutti, R. Global warming under old and new scenarios using IPCC climate sensitivity range estimates. Nature Climate Change 2, 248–253 (2012).
Taylor, K. E., Stouffer, R. J. & Meehl, G. A. An overview of CMIP5 and the experiment design. Bulletin of the American Meteorological Society 93, 485–498 (2012).
Rupp, D. E., Abatzoglou, J. T., Hegewisch, K. C. & Mote, P. W. Evaluation of CMIP5 20th century climate simulations for the Pacific Northwest USA. Journal of Geophysical Research: Atmospheres 118, 1–23 (2013).
Wang, T., Hamann, A., Spittlehouse, D. L. & Aitken, S. N. Development of scale-free climate data for western Canada for use in resource management. Int. J. Climatol. 26, 383–397 (2006).
Wang, T., Hamann, A., Spittlehouse, D. L. & Murdock, T. Q. ClimateWNA-High-resolution spatial climate data for western North America. J. Appl. Meteor. Climatol 51, 16–29 (2012).
Wang, T., Hamann, A., Spittlehouse, D. & Carroll, C. Locally downscaled and spatially customizable climate data for historical and future periods for North America. PLoS ONE 11, e0156720 (2016).
Knutti, R., Masson, D. & Gettelman, A. Climate model genealogy: Generation CMIP5 and how we got there. Geophys. Res. Lett. 40, 1194–1199 (2013).
Wood, A. W., Maurer, E. P., Kumar, A. & Lettenmaier, D. P. Long-range experimental hydrologic forecasting for the eastern United States. J. Geophys. Res. 107, 4429 (2002).
Peterson, T. C. & Vose, R. S. An overview of the Global Historical Climatology Network temperature data base. Bulletin of the American Meteorological Society 78, 2837–2849 (1997).
World Meteorological Organization (WMO). Climatological normals (CLINO) for the period 1961-1990. Pub. 847, 768 pp., Geneva (1996).
Food and Agriculture Organization of the United Nations (FAO). FAOCLIM 2.0 A World-Wide Agroclimatic Database. Food and Agriculture Organization of the United Nations (2001).
Hijmans, R. J., Cameron, S. E., Parra, J. L., Jones, P. G. & Jarvis, A. Very high resolution interpolated climate surfaces for global land areas. International Journal of Climatology 25, 1965–1978 (2005).
We thank Dr. Daniel Griffith for text edits on an early draft of this manuscript. Yueyang Jiang is supported by a joint-venture agreement between Oregon State University and the USDA Forest Service Pacific Northwest Research Station, 14-JV-11261952-100. John B. Kim is supported in part by USDA Forest Service Western Wildland Environmental Threat Assessment Center (WWETAC).
The authors declare no competing financial interests.
Supplementary information accompanies this paper at
About this article
Cite this article
Jiang, Y., Kim, J., Still, C. et al. Inter-comparison of multiple statistically downscaled climate datasets for the Pacific Northwest, USA. Sci Data 5, 180016 (2018). https://doi.org/10.1038/sdata.2018.16
A new statistical downscaling approach for global evaluation of the CMIP5 precipitation outputs: Model development and application
Science of The Total Environment (2019)
Assessments of downscaled climate data with a high‐resolution weather station network reveal consistent but predictable bias
International Journal of Climatology (2019)
How Do Modeling Decisions Affect the Spread Among Hydrologic Climate Change Projections? Exploring a Large Ensemble of Simulations Across a Diversity of Hydroclimates
Earth's Future (2019)
Future Predictions of Rainfall and Temperature Using GCM and ANN for Arid Regions: A Case Study for the Qassim Region, Saudi Arabia