TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015 — TEI-tagged view

John T. AbatzogloupersName, Solomon Z. DobrowskipersName, Sean A. ParkspersName, Katherine C. HegewischpersName (2018date). “TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015.” Scientific Datatitle·j.

Object of study — TerraClimateterm construction, monthly climate and water-balance variables, validation, uses, and uncertaintyterm

Research question — how to provide globally consistent, time-varying climate and water-balance data at ecologically meaningful spatial resolution

Contribution — a public four-kilometre monthly global terrestrial datasetterm for 1958date–2015date built from WorldClim normals, CRU and JRA-55term anomalies, and a water-balance model

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

TerraClimateterm, a high-resolution global datasetterm of monthly climate and climatic water balance from 1958date–2015date Abstract We present TerraClimateterm, a datasetterm of high-spatial resolution (1/24°, ~4-km) monthly climate and climatic water balance for global terrestrial surfaces from 1958date–2015date. TerraClimateterm uses climatically aided interpolationterm, combining high-spatial resolution climatological normals from the WorldClim datasetterm, with coarser resolution time varying (i.e., monthly) data from other sources to produce a monthly datasetterm of precipitationterm·var, maximum and minimum temperatureterm·var, wind speedterm·var, vapor pressureterm·var, and solar radiationterm·var. TerraClimateterm additionally produces monthly surface water balance datasetsterm using a water balance modelterm that incorporates reference evapotranspirationterm, precipitationterm·var, temperature, and interpolated plant extractable soil water capacity. These data provide important inputs for ecological and hydrological studies at global scales that require high spatial resolutionterm and time varying climate and climatic water balance data. We validated spatiotemporal aspects of TerraClimateterm using annual temperature, precipitationterm·var, and calculated reference evapotranspirationterm from station data, as well as annual runoffterm from streamflow gauges. TerraClimate datasetsterm showed noted improvement in overall mean absolute errorterm and increased spatial realism relative to coarser resolution gridded datasetsterm.

Background & Summary Global environmental questions that invoke climate as a driver of a specific phenomenon require spatially and temporally consistent datasetsterm. An extensive and growing collection of historical observed climate datasetsterm exist including those that interpolate data from surface observations 1 , reanalysis 2 , and some combination thereof 3–5 with each datasetterm offering advantages in the variables provided, temporal and spatial resolution, as well as the geographic extent and time period covered. Increased interest in applying climate data to multidisciplinary problems has prompted the development of climate datasetsterm at more meaningful spatial scales. While time varying (monthly data spanning multiple decades) and high spatial resolutionterm (<5-km) climate datasetsterm have been developed at national and continental scales 6–8 , and high-spatial resolution climate normals have been developed for the globe 9–11 , we are unaware of time varying (i.e., by month and year), high-resolution climate data that covers all global land surfaces and encompasses many of the essential surface climate variables.

There is increasing recognition of the mismatch between the scale of existing climate data and the scale at which organisms experience their environment 12 . This has motivated a large body of work aimed at downscaling gridded climate fields for use in ecological and global change studies, the results of which may be sensitive to the spatial scale of data used 13 . High spatial resolutionterm climate data is often necessary for ecological and hydrological analyses and modeling 14 . The newest version of WorldClim 15 used station data, remotely sensed land surface temperature, and topographic covariates to produce high spatial resolutionterm climate surfaces of temperature, precipitationterm·var, wind speedterm·var, vapor pressureterm·var, and downward shortwave flux. While WorldClim provides a complete set of variables to assess long-term monthly normalsterm for several climate variables, it does not lend itself to temporal analysis that may be important for linking climate variability and climate impactsterm in ecological, agricultural, and hydrological systems.

In addition to temperature and precipitationterm·var, there has been interest in measures of climate that are more directly linked to ecosystem productivityterm and water resources. Water balance modelsterm integrate climate and biophysical factors to produce a set of variables that account for the concurrence of water and energy. Water balance modelsterm have been used to estimate runoffterm 16 , 17 , ecosystem productivityterm 18 , ecosystem distribution 19 , and ecological disturbance 20 , 21 . Whereas some global water balance datasetsterm exist from simple water balance modelsterm 22–24 or more sophisticated land-surface hydrologic models 25 , these models typically provide output at relatively coarse spatial scales (>50-km) or lack extensive time series components. Higher resolution (<5-km) country and continental scale water balance datasetsterm have been developed 21 , yet global, time varying, high-resolution datasetsterm do not currently exist.

This paper outlines the development of a global monthly high-resolution climate datasetterm from 1958date–2015date, herein referred to as TerraClimateterm, that includes the requisite variables for calculating energy-based reference potential evapotranspiration and a water balance modelterm. We discuss the procedures for downscaling coarser resolution temporal anomalies from Climate Research Unit time series data version 4.0term (CRU Ts4.0term) 1 and the Japanese 55-year Reanalysisterm (JRA-55term) 2 from 1958date–2015date with high-resolution climatological fields from WorldClim 11 , 15 using climatic aided interpolation. Furthermore, we discuss the procedures used for developing the monthly water balance datasetterm for the period 1958date–2015date.

Methods Three global gridded climate datasetsterm were used to develop TerraClimateterm, each offering distinct spatial or temporal qualities ( Table 1 ).

The high-spatial resolution climatology from WorldClim version 2 ( http://worldclim.org/version2 ) used thin-plate splines to interpolate station observations using covariates including MODIS derived land surface temperature and cloud cover, along with topographic features to develop monthly climate normal surfaces for global land surface 15 . Monthly climate normals for the 1970–2000 period were developed for maximum and minimum temperatureterm·var, precipitationterm·var, solar radiationterm·var, vapor pressureterm·var and wind speedterm·var at four different spatial grains. We used the 2.5 arc-minute (1/24°) resolution to accommodate a high-spatial resolution (<5-km) product that is computationally tractable for developing longer time series.

We used the monthly diurnal temperature range (DTR) from WorldClim version 1.4 ( http://worldclim.org/version1 ) rather than the WorldClim version 2 datasetterm, as the latter showed widespread biasesterm in mid latitudes despite monthly mean temperature showing nominal biasterm ( Supplementary Figs 1 and 2 ). The monthly DTR from WorldClim version 1.4 for the 1960–1990 period was further modified to account for the change in DTR between the 1960–1990 base period and 1970–2000 base period of WorldClim v2.0. This was facilitated by interpolating differences in monthly DTR for these two base periods in the CRU Ts4.0term data. Finally, monthly average maximum and minimum temperatureterm·var climatologies for the 1970–2000 period were reconstructed using WorldClim v2.0 monthly mean temperatures and the DTR from WorldClim v1.4 as modified by CRU Ts4.0term.

Time varying climate data were derived from two sources: (i) Climate Research Unit (CRU) time series data version 4.0 (ref. 1 ) (Data Citation 1), and (ii) the Japanese 55-year Reanalysisterm (JRA-55term (ref. 2 ); http://jra.kishou.go.jp/JRA-55term/index_en.html ). The CRU Ts4.0term data provide monthly average maximum and minimum temperatureterm·var, vapor pressureterm·var, and cumulative precipitationterm·var at a 0.5° grid from 1901–2015date for global land surface and were developed based on station data. The CRU Ts4.0term data however does not cover all islands and does not include Antarctica. Furthermore, CRU Ts4.0term data assumes climatological values for pixels where there are no long-term stations within a decorrelation distance, resulting in voids of climate signals in poorly sampled land areas of the globe. The JRA-55term is the longest running full-observing-system modern reanalysis product, providing spatially and temporally complete data covering the period 1958-present. We calculated monthly average 10-m wind speedterm·var, downward surface shortwave flux, temperature, and precipitationterm·var data from 3-hourly products at a 1.25° grid. For both CRU Ts4.0term and JRA-55term datasetsterm, we calculated monthly anomaliesterm for all variables relative to the 31-year baseline period (1970–2000) defined by WorldClim v2.0.

Climatologically aided interpolation was used to superimpose monthly climate anomalies from CRU Ts4.0term and JRA-55term with monthly climate normals from WorldClim to estimate monthly time series from 1958date–2015date. Climatologically aided interpolation is a simple spatial downscaling approach that employs bilinear interpolationterm of temporal anomalies from a higher-temporal, lower-spatial resolution datasetterm to a lower-temporal, higher-spatial resolution datasetterm 26 , 27 . All anomalies were calculated relative to the 1970–2000 base period to adhere with the WorldClim climatology. Monthly average maximum temperature, minimum temperature, vapor pressureterm·var, and precipitationterm·var anomalies were prioritized from CRU Ts4.0term where these data cover land surface pixels. The CRU Ts4.0term data assigns pixels to their climatological normal when there are no stations within decorrelation distances 28 in a given month. These represent a very small fraction of the temperature and precipitationterm·var record from 1958date–2015date. Rather than use a potentially artificial null climate signal such instances, we bilinearly interpolate JRA-55term anomalies to infill data voidsterm. CRU Ts4.0term anomalies were further extrapolated to a 0.5° buffer from land masses by interpolating anomalies from the 8 nearest neighbors. Monthly anomaliesterm from JRA-55term were used for pixels on islands located outside this buffer, as well as in Antarctica which is not covered by CRU Ts4.0term. Anomalies of 10-m wind speedterm·var and downward surface shortwave flux were taken exclusively from JRA-55term. Anomalies were computed as additive (i.e., departure from normal) for temperature, wind speedterm·var, and solar radiationterm·var, and multiplicative (i.e., percent of normal) for precipitationterm·var and vapor pressureterm·var. We capped monthly precipitationterm·var anomalies at 2,000% of normal to avoid unrealistic precipitationterm·var in exceptionally dry regions and months where climatological precipitationterm·var from CRU Ts4.0term or JRA-55term approaches zero. There were rare occurrences (~0.1% of pixels) where monthly mean maximum temperature was less than 0.5 °C higher than minimum temperature due to our method independently accounting for anomalies in maximum and minimum temperatureterm·var. In such cases, we set the monthly minimum temperature equal to 0.5 °C below the monthly maximum temperature for internal consistency.

Monthly reference evapotranspirationterm (ET 0term ) was calculated using the Penman Montieth approach 29 . ET 0term has been argued to be a more appropriate measure of potential water lossterm over purely temperature based metrics for estimating potential evapotranspiration as it uses an energy balance approach. ET 0term assumes a reference grass surface across space, but can yield biasesterm where potential vegetation water use departs substantially from this assumption. Following prior studies, we modified ET 0term to account for reduced surface water flux when snow cover exists or prior to the onset of the growing season and active transpiration using an empirical relationship with temperature that accounts for precipitationterm·var phase changes 16 , 30 .

A one-dimensional modified Thornthwaite-Mather climatic water-balance modelterm (WBMterm) 22 , 31 was used to calculate monthly water balance from 1958date–2015date. The WBMterm is a single bucket modelterm applied consistently across global land surfaces that operates on a monthly time step and considers the interplay between precipitationterm·var, ET 0term , as well as soil and snowpack water storage. The WBMterm accounting scheme considers runoffterm as the excess of liquid water supply (precipitationterm·var and snowmelt) used by monthly ET 0term and soil moistureterm recharge. Soil water is extracted during months where ET 0term exceeds liquid water supply, with the extraction efficiency of soil water declining exponentially with the ratio of soil water to extractable soil water capacity. Under such conditions, actual evapotranspirationterm is counted as the liquid water supply plus the soil water utilized and climatic water deficitterm is the difference between ET 0term and actual evapotranspirationterm. The WBMterm requires data on the plant extractable soil water capacity. We used extractable soil water storage capacity data at a 0.5° grid from Wang-Erlandsson et al. 32 that were derived from estimates of satellite based evaporation, ET 0term , and precipitationterm·var. Wang-Erlandsson et al. 32 provided estimates for varying return intervals. We used the 2-year period given our use of ET 0term , as it best matched with evapotranspiration variability in grassland biomes, corresponding with our use of a reference grass surface in the WBMterm. To adhere to the spatial resolution of TerraClimateterm, we first extrapolated data using a 0.5° buffer from existing data using the 8 nearest neighbors, and then bilinearly interpolated plant extractable soil water from its native 0.5° grid. We set a lower bound on plant extractable soil water of 10 mm and set a default value of 50 mm for places void of data.

We use the WBMterm outlined by DobrowskipersName et al. 31 , with two modifications to better account for runoffterm and snowpack dynamics on a global scale. First, we use an empirical temperature-based transformation for precipitationterm·var phase to determine the proportion of precipitationterm·var falling as liquid, as well as the fraction of snowpack melted each month 30 . Although snowmelt is a more complicated process that could be improved by incorporating shortwave fluxes and snowpack ripeness, temperature based indices have been shown to be sufficient for macroscale monthly water balance calculations 33 . Secondly, we allow for 5% of the cumulative rainfall and snowmelt each month to occur as direct runoffterm irrespective of soil conditions 17 . The resultant datasetsterm from the WBMterm are monthly total runoffterm (Q), climatic water deficitterm, actual evapotranspirationterm, as well as snow water equivalentterm and soil moistureterm at the end of each month.

Data Records TerraClimateterm is available to the public through an unrestricted data repository hosted by the University of Idaho’s Northwest Knowledge Network (Data Citation 2). The full list of TerraClimate datasetsterm are provided in Table 2 . Files were created separately for each variable and year in NetCDFterm format following the Climate and Forecast metadata standardsterm. A compressed archive of NetCDFterm files for the period of record covering all variables can be accessed at http://doi.org/10.7923/G43J3B0R . However, due to the sheer size of the data once uncompressed (~800 GB), we provide access to individual annual files for each variable at https://climate.northwestknowledge.net/TERRACLIMATEterm/ . Temporal extensions to TerraClimateterm will be updated periodically as additional years of data become available.

Technical Validation We conducted a validation of the TerraClimateterm product for monthly temperature and precipitationterm·var using station-based data obtained from the Global Historical Climatology Networkterm (GHCNterm) database 34 . We examined time series of temperature data that have been corrected for climate inhomogeneities that arise due to non-climatic artifacts, such as station relocation or observational practice, from the GHCNv3 datasetterm; whereas we used monthly precipitationterm·var data from the GHCNv2 datasetterm. We constrained our validation to GHCNterm stations that had complete monthly data for at least 30 calendar years from 1958date–2015date resulting in a total of 3,230 stations for temperature and 6,102 stations for precipitationterm·var. As a caveat, we note that many of these stations were used in both the WorldClim and CRU datasetsterm making the comparison not completely independent.

Data from TerraClimateterm were extracted at grid cells co-located with individual stations. We calculated linear validation statistics of Pearsons’ correlation coefficients, mean absolute errorterm (MAEterm), and biasterm for each station. We constrained our validation to calendar year summaries rather than individual months to avoid artificially inflating the correlation by incorporating the seasonal cycle. To demonstrate potential added value from TerraClimateterm, we compare our validation statistics to CRU Ts4.0term, acknowledging that by using CRU Ts4.0term in climatically aided interpolationterm, we should expect similar results for temporal variability.

The TerraClimate datasetterm largely captured interannual variabilityterm in mean annual temperature across GHCNterm stations with a median Pearson’s correlation coefficient (ρ) of 0.95, with ρ>0.9 for approximately 80% of the stations ( Fig. 1 ). Most of the subpar correlations (ρ<0.8) were found equatorward of 30° latitude, where interannual variabilityterm in mean temperature is lesser in magnitude. Similarly, TerraClimateterm precipitationterm·var fields captured interannual variabilityterm observed at GHCNterm stations globally with a median correlation of ρ=0.90. Subpar correlations for annual precipitationterm·var were often found adjacent to stations with high correlation suggesting that interannual precipitationterm·var variability can vary across relatively short spatial scales, particularly where convective or orographic features are involved that influence the spatial variability of precipitationterm·var anomalies at scales finer than those resolved by CRU Ts4.0term. Alternatively, subpar correlations may also derive from potential inhomogeneities or errors in station-based precipitationterm·var records.

Correlation statistics were nearly identical to those found using the CRU Ts4.0 datasetterm, as expected, given its use in the development of the TerraClimate datasetterm ( Table 3 ). However, the MAEterm was substantially improved in the TerraClimate datasetterm, a consequence of biasterm correcting the data using the WorldClimv2 data. For mean temperature, the median MAEterm fell from 0.53 °C for CRU to 0.32 °C for TerraClimateterm, whereas for precipitationterm·var the median MAEterm fell from 11% to 9%. We also note that unaccounted-for error and associated biasterm inherent in comparing gridded climate data to station data is a function of mismatches in the scale of the data being compared. This is particularly apparent in regions of complex terrainterm where the elevation of a station may differ substantially from the average elevation of a co-located grid cell. For example, approximately 40% of the variance in the mean annual temperature biasterm was linearly explained by biasesterm in elevation (i.e., grid cell elevation minus station elevation).

Given the potential fidelity problems when validating TerraClimateterm with GHCNterm stations, we perform a complementary validation of temperature and precipitationterm·var data using a network of 2,100 automated climate stations located across mountains of the western United States from the Snowpack Telemetry (SNOTELterm) and Remote Automated Weather Stations (RAWSterm) networks. Due to suspected inhomogeneities with minimum temperature records from SNOTELterm stations arising from instrumentation changes 35 , we used a set of homogeneized daily temperature data for both RAWSterm and SNOTELterm stations 36 . Additionally, we used daily records of accumulated precipitationterm·var from a more limited set of 513 SNOTELterm stations that were quality controlled 37 . Station-years where more than 5% of daily observations were missing were omitted from the validation. Due to the short period of record for most stations, we used a more liberal criterion of at least 10 years of complete data to compute interannual correlations, resulting in 1,292 stations for temperature and 503 stations for precipitationterm·var.

TerraClimateterm largely captured interannual variabilityterm in mean annual temperature across the Snotelterm and Rawsterm network with a median Pearson’s correlation coefficient (ρ) of 0.90 ( Fig. 2a ). Interannual correlations between calendar year precipitationterm·var from TerraClimateterm and recorded precipitationterm·var at SNOTELterm sites was also lower than seen for neighboring GHCNterm stations across the western US with a median ρ of 0.79 ( Fig. 2b ), with a dry biasterm (−30%) across interior mountains where correlations were generally lower. The lower correlations compared with GHCNterm could be a function of station quality control measures, but also factors that influence temperature and precipitationterm·var variability in mountainous areas such as inversions, snow-albedo feedbacks, and orographic precipitationterm·var ratios. In addition, the incongruence of the scales of gridded data and in-situ measurements are heightened in regions of complex terrainterm thus leading to an additional challenge in validation efforts. Validation statistics using CRU data showed generally similar values for correlation as those for TerraClimateterm, but with CRU showing slightly larger errors as noted with MAEterm ( Table 3 ).

Comparisons of interannual variabilityterm in annual ET 0term from TerraClimateterm were made using a global network of FLUXNET stationsterm 38 . Hourly and bi-hourly observations of temperature, humidity, wind speedterm·var, and solar radiationterm·var from these stations were used to calculate ET 0term . Observational data were aggregated to monthly temporal scales to ensure that we used the same procedures for calculating ET 0term as TerraClimateterm. Interannual correlations were computed for stations that had at least 10 years of complete data, missing no more than 10% of data in any calendar year.

Pearson’s correlations of annual ET 0term from TerraClimateterm showed a median ρ of 0.77 over the 50 FLUXNET stationsterm ( Fig. 2c ), with a MAEterm of approximately 57 mm (8.5%). The spatial correlation between observed annual mean ET 0term and ET 0term as modeled by TerraClimateterm was ρ=0.98 highlighting that the spatial pattern was well captured. A slight overall biasterm (median +8%) in TerraClimateterm ET 0term was seen relative to ET 0term calculated at FLUXNET sites, which may arise due to the non-standard heights relative to canopies measured at FLUXNET sites. Overall validation statistics were slightly better with TerraClimateterm than those seen for CRU ( Table 3 ). In addition, less of the spatial variability in climatological ET 0term was captured with CRU (ρ=0.94).

To demonstrate potential utility of the water balance component of the datasetterm, we compared interannual variabilityterm in water-year runoffterm from our water balance modelterm (WBMterm) to streamflow from 587 pristine river basins that are part of the Climate Sensitive Stations Datasetterm from the Global Runoffterm Data Centre (The Global Runoffterm Data Centre, 56068 Koblenz, Germany, 2013). We considered water-year (Oct-Sep) rather than calendar year comparisons to account for the contributions of snowmelt runoffterm from watersheds in the northern hemisphere. These stations were selected based on having nominal water extraction or diversions, consistency of data records, at least 30 complete water-years of data from 1958date–2015date, and existing watershed boundary data. We extracted raster based runoffterm from the WBMterm for each gauge by delineating pixels that occurred within the upstream area of the respective watershed. The identical WBMterm was also run using CRU Ts4.0 datasetsterm for comparative purposes.

Pearsons’ correlation coefficients were calculated between water-year cumulative runoffterm (Q) simulated by the WBMterm and observations. Streamflow data were scaled to units of mm per year by accounting for the size of the upstream contributing watershed. We considered temporal correlations to assess interannual variabilityterm, as well as spatial correlations to assess the performance of the datasetterm to capture the spatial variability in annual Q. Interannual variabilityterm in estimated runoffterm was reasonably captured by the WBMterm, with a median ρ=0.80 ( Fig. 3a ). The WBMterm captured over 70% of the interannual variabilityterm in Q of the mid-latitude basins in the United States, Europe and southern Australia. Lower explained variance was seen for parts of the tropics, boreal systems, and across the Great Plains of North America. Overall, the WBMterm exhibited a tendency for less runoffterm than was observed with a mean biasterm of −24%. The median relative MAEterm of water-year runoffterm was 36% (101 mm; Fig. 3b ). However, MAEterm exceeded 100% for 6% of the watersheds, predominantly arid watersheds across the Great Plains and Australia where annual mean Q<25 mm. The spatial correlation between observed and simulated annual mean Q was ρ=0.87, suggesting that the patterns were well represented by the WBMterm ( Fig. 3c ). By comparison, these validation results were improvements over the CRU Ts4.0 datasetterm using the same WBMterm framework in terms of the median interannual correlation (ρ=0.765), relative MAEterm (42%, 118 mm), and the spatial correlation (ρ=0.84).

Water balance modelsterm can be biased due to several factors. Errors can arise through inaccurate estimates of climate forcings, by not accounting for vegetation dynamics within a watershed, and by the overall simplicity of the single bucket modelterm approach which does not account for subsurface routing of water. Notably, by using ET 0term , we assume a constant reference vegetation type that may substantially deviate from observations. Additionally, plant water soil storage capacity likely deviates at scales finer than those mapped here, which could contribute to uncertaintyterm. We conducted an additional validation analysis by comparing cumulative water-year precipitationterm·var data to observed Q for each watershed. The WBM runoff dataterm explained substantially more variance of the interannual variabilityterm in Q than using precipitation data aloneterm (median ρ=0.68). Observed mean runoffterm exceeded estimated mean annual precipitationterm·var for 30 of the 587 watersheds, further illustrating the challenges of interpolating estimates of precipitationterm·var data, particularly in mountainous watersheds as highlighted by previous studies 15 , 39 . Underestimates of runoffterm and precipitationterm·var in some mountain watersheds highlights the need for continued improvement of mapping climate at fine spatial scales globally 10 , 15 and investment in observational systems in historically data sparse regions such as mountains.

We illustrate some of the value-added information from the TerraClimate datasetterm compared to CRU Ts4.0term for a portion of northwestern North America in 2015date ( Fig. 4 ). Specifically, we highlight the spatial detail added by the TerraClimate datasetterm that is absent in CRU across a region with substantial geographic gradients in energy and moisture that are apparent in July 2015date average maximum temperature ( Fig. 4b,c ), and 2015date annual precipitationterm·var totals ( Fig. 4d,e ). Integrating climate metrics through the WBMterm improves the biophysical relevanceterm of the gridded products for use in global-change and ecological studiesterm, as can be seen by the additional detail in TerraClimateterm for annual climatic water deficitterm ( Fig. 4f,g ), and annual Q in 2015date ( Fig. 4h,i ).

The use of climate data has increased in the past decade due to increasing availability of data, broadening of the disciplines that use such data, and increased interest in assessing local impacts from climate change. The potential applications of TerraClimateterm span a variety of spatial scales from users wanting local climate and water balance time series in regions void of easily accessible or sufficiently long station-based data, to users conducting broader spatial analyses to better understand and refine regional and global climate impactsterm to water resources and agriculture 40 . In addition, TerraClimateterm has many potential uses in ecology, macroecology, and global change science. For example, water balance data are increasingly used in explaining and predicting patterns of wildfire activity 20 , 41 , 42 , vegetation distribution 19 , 43 and structure 44 , climate velocity 31 , and drought induced tree mortality 45 .

Usage Notes Climatological data for several voxels from the WorldClim datasetterm near perennial sea ice or glaciers primarily in Antarctica and Greenland had exceptionally large, and likely unrealistic, vapor pressureterm·var deficits (>0.5 kPa). These voxels were primarily located near the coastline and had nominal differences in diurnal temperature range (<2 °C). Moreover, we caution users that output from the water balance modelterm, as well as estimated ET 0term , are likely suspect in perennially glaciated landscapes including Greenland and Antarctica. Secondly, there are a few voxels near the South Pole and in the Sahara Desert where climatological 1970–2000 precipitationterm·var from WorldClim v2.0 is zero. The use of multiplicative anomalies in climatologically aided interpolation sets all values to zero, which may be suspect.

Uncertaintyterm in observational gridded climate datasetsterm that use station based data can arise from a variety of sources ranging from procedural choices for interpolating and extrapolating data spatially, to changes in data availability through time, to climate inhomogeneities that lead to structural uncertaintyterm in climate datasetsterm 46 , 47 . TerraClimateterm uses three different datasetsterm, each of which contain their own sources of uncertaintyterm. We provide a measure of uncertaintyterm in TerraClimateterm data using estimates of the number of stations that contribute to anomaly fields from CRU Ts4.0term based on decorrelation distances 28 . CRU Ts4.0term provides estimates of the number of stations that contribute to estimates of monthly anomaliesterm at the native scale of the data and provides these data for temperature, precipitationterm·var, and vapor pressureterm·var fields. The number of stations (between 0 and 8) contributing to monthly CRU Ts4.0term precipitationterm·var, mean temperature, and vapor pressureterm·var fields were interpolated to the TerraClimateterm grid and provided as ancillary dataterm for precipitationterm·var, monthly maximum and minimum temperatureterm·var, and vapor pressureterm·var. Anomalies from JRA-55term were used for voxels where zero stations contributed to the anomaly fields in CRU Ts4.0term data. The fraction of months from 1958date–2015date with 8 contributing stations (maximum number of stations influencing) and 0 contributing stations (anomalies strictly derived from JRA-55term) from CRU Ts4.0term for temperature, precipitationterm·var, and vapor pressureterm·var is shown in Fig. 5 . These ancillary dataterm are provided to help users of TerraClimateterm identify the robustness of interannual variabilityterm in data for specific geographic locations, as well as the data source of the interannual variabilityterm used in TerraClimateterm for these four variables. For example, the low fraction of months with high station density from CRU may limit the robustness of TerraClimateterm across regions such as Africa and South America, where temporal anomalies instead rely more heavily on JRA-55term. Ancillary dataterm are not provided for downward shortwave radiation or 10-m winds which exclusively use JRA-55term anomalies, or for ET 0term and the water balance variables which use combinations of the essential climate variables.

Long-term trends in primary climatological fields of TerraClimateterm are preserved from their parent datasetsterm in JRA-55term and CRU Ts4.0term and hence TerraClimateterm should not be used to provide an independent estimate of trends. Interannual variabilityterm in primary climate variables from TerraClimateterm will not incorporate spatial variability in anomalies at scales finer than their parent datasetsterm, including those associated with orographic precipitationterm·var ratios, inversions, and near coastal maritime influences. Likewise, while WorldClim data incorporated into TerraClimateterm allows for a high-resolution datasetterm, such approaches may not adequately resolve microclimate featuresterm, particularly in complex terrainterm or heterogeneous land-cover. Some approaches have been used to better resolve microclimate featuresterm to finer spatial scales 48 , yet such relationships have been typically applied to regional domains.

Additional information How to cite this article: AbatzogloupersName, J. T. et al. TerraClimateterm, a high-resolution global datasetterm of monthly climate and climatic water balance from 1958date–2015date. Sci. Data 5:170191 doi: 10.1038/sdata.2017.191(2018).

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Material