Hydroclimate extremes critically affect human and natural systems, but there remain many unanswered questions about their causes and how to interpret their dynamics in the past and in climate change projections. These uncertainties are due, in part, to the lack of long-term, spatially resolved hydroclimate reconstructions and information on the underlying physical drivers for many regions. Here we present the first global reconstructions of hydroclimate and associated climate dynamical variables over the past two thousand years. We use a data assimilation approach tailored to reconstruct hydroclimate that optimally combines 2,978 paleoclimate proxy-data time series with the physical constraints of an atmosphere—ocean climate model. The global reconstructions are annually or seasonally resolved and include two spatiotemporal drought indices, near-surface air temperature, an index of North Atlantic variability, the location of the intertropical convergence zone, and monthly Niño indices. This database, called the Paleo Hydrodynamics Data Assimilation product (PHYDA), will provide a critical new platform for investigating the causes of past climate variability and extremes, while informing interpretations of future hydroclimate projections.
Machine-accessible metadata file describing the reported data (ISA-tab format)
Background & Summary
Hydroclimate extremes, including persistent droughts and pluvials, can have extensive effects on societies and ecosystems. For example, multi-year droughts in California have caused significant agricultural losses, tree mortality, forest fires, and other impacts (e.g., refs 1,
This work uses a data assimilation (DA) method to derive the first global reconstructions of hydroclimate and associated dynamical variables that span the last two millennia. In contrast to traditional reconstruction approaches and existing paleo-hydroclimate products, DA-based reconstruction methods can simultaneously estimate both hydroclimate fields and corresponding atmosphere-ocean states16; having these two components is critical for analyses of the causes of hydroclimate extremes. DA for paleoclimate works by optimally fusing proxy information with the dynamical constraints of climate models17,
This Paleo Hydrodynamics Data Assimilation product (PHYDA), represents the first collection of global hydroclimate reconstructions along with their associated dynamical variables. It is also the first DA-based paleoclimate reconstruction to explicitly include the location of the ITCZ and monthly SST indices. We have also made several innovations in the reconstruction methodology compared to previous approaches19,20,36,
We employ a DA technique that optimally combines proxy data or observations with climate model states. The model provides an initial, or prior, state estimate that is updated based on the proxy observations and an estimate of the errors in both the observations and the prior. The general state update equations of DA41 can be written as
In these equations, xb is the prior (or ‘background’) estimate of the state vector and xa is the posterior (or ‘analysis’) state vector; the state vector contains all of the variables that are to be reconstructed. Observations (or proxies) are contained in y. The observations are estimated by the prior through , which is, in general, a nonlinear vector-valued observation operator that maps xb from the state space to the observation space. B is the prior covariance matrix, R is the error covariance matrix for the proxy data, and H represents a linearization of . In a general sense, the reconstruction process works by computing an optimal linear fit between the initial guess of the climate state, the prior xb, and the proxies y. Because proxies are available at annual time steps, a reconstruction is made by iteratively computing equations (1) and (2) for each year (or season within each year) of the existing proxy data.
We implement the general DA equations described above by using an ensemble square root filter from ref. 42. Though the authors of ref. 42 recommend the sequential assimilation of observations for computational reasons, the simultaneous assimilation of observations is actually an order of magnitude faster in contemporary matrix-optimized computing software, such as MATLAB. We therefore modified previous approaches used for paleoclimate DA19,20,36,
After computing equations (3), (4), (5), (6), (7), (8) the full posterior ensemble is then recovered through where is added to each column of . Collectively, equations (3), (4), (5), (6), (7), (8), (9) are computed for each year (or a particular season of each year) to arrive at a series of posterior ensemble state estimates that together constitute the probabilistic spatiotemporal reconstruction. In the reconstruction files for all variables (Data Citation 1: Zenodo https://doi.org/10.5281/zenodo.1154913) we have included the posterior ensemble mean, 1 standard deviation of the posterior ensemble as well as its 5th, 50th, and 95th percentiles; this error estimate explicitly includes uncertainty information from the spread in the climate model prior (HBHT) as well as the error in the proxy models (R).
In our implementation of equations (3), (4), (5), (6), (7), (8), (9), R is assumed to be a diagonal covariance matrix (uncorrelated errors) where the entries are the error variance of each proxy (defined in the ‘Proxy system models’ section). If the proxy errors are correlated then equations (3), (4), (5), (6), (7), (8), (9) can be computed in the same way using a non-diagonal R.
Climate model data and reconstruction variables
As in previous studies19,20,36,
We construct the prior xb using the CESM LME22, which used atmosphere and land components with ~2-degree resolution and ocean and sea ice components with ~ 1-degree resolution. The simulations were run from the years 850 to 1850 CE using estimates of the transient evolution of solar intensity, volcanic emissions, greenhouse gases, aerosols, land-use conditions, and orbital parameters44. The simulations were given identical forcings but differed by round-off error in the initial atmospheric state; this difference was sufficient to generate simulations with different internal ocean-atmosphere variability and therefore different time histories (e.g., annual Niño 3.4 indices from the simulations are uncorrelated). For the reconstruction prior, we used a single simulation, number 10 from the full-forcing ensemble, to generate our prior ensemble; specifically, we used the middle 998 years of the CESM simulation excluding the two simulation endpoints to create a static 998 member prior ensemble that was used to estimate the climate state in each year of the reconstruction (the last year cannot be used because of the particular annual averaging we used here and the first year cannot be used because the variable SPEI integrates the previous 12 months of climate information, therefore only the second year of an SPEI time series is meaningful). This prior is consistent with previous work that has established that the prior is not required to contain year-specific forcing or boundary condition information, rather it must merely be statistically representative of the reconstruction period19,20,36,
Reconstructions were performed from the years 1–2000 CE targeting three different temporal windows: annual means (defined as April to the next calendar year March), the boreal growing season of JJA, and the austral growing season of DJF. The particular annual average used herein was chosen to account for the seasonal cycle of a global network of proxies as well as climate phenomena like the El Niño—Southern Oscillation, the continuity of which would be ignored with a calendar year average. Except for the monthly Niño SST indices described below, all other variables were reconstructed over the annual, JJA, or DJF windows.
Each reconstruction contains the following gridded fields over the global domain: 2 m air temperature, PDSI, and SPEI using a 12-month decaying exponential weighting kernel47 chosen to closely resemble the time scale of PDSI; the potential for skillful reconstruction of these fields was previously demonstrated using pseudoproxy experiments16. Both PDSI and SPEI were computed using the Penman-Monteith equation for potential evapotranspiration and monthly climate model output of precipitation, 2 m temperature, vapor pressure, net surface radiation, surface pressure, and surface wind (estimated from 10 m down to 2 m using the wind profile power law); the climatologically bias-corrected temperature and precipitation fields were used in the calculations. PDSI was computed using the MATLAB code from ref. 48, which produces the standard formulation of PDSI as opposed to self-calibrating versions (e.g., ref. 49). Both PDSI and SPEI are broadly used in drought monitoring50 and historical drought reconstructions14,15,51,
We also reconstruct the following index variables: the area-weighted global mean temperature, the North Atlantic SST index which is the non-detrended and non-smoothed version of the Atlantic multidecadal oscillation (AMO), the monthly Niño SST indices (Niño 1+2, 3, 3.4, 4), the monthly equatorial Pacific zonal SST gradient24, and the location of the intertropical convergence zone (ITCZ) in 11 longitudinal zones. Because there are different smoothing and standardization conventions in computing Niño SST indices, we have simply computed the area-averaged monthly SST values in each Niño region. We use the definition of the location of the ITCZ from23, which is the expected value of precipitation (P) using a 10th-power area weighting, integrated over the tropical latitudes ϕ1 and ϕ2, In equation (10), we use the annual or seasonally averaged precipitation and also ϕ1=30°S and ϕ2=30°N to account for Monsoon regions where the seasonal precipitation maximum can extend far beyond the equator. Following refs. 23,55 we interpolate the tropical precipitation to a 0.1 degree latitudinal grid before computing equation (10). The 11 longitudinal zones are listed in the tables and include all major ocean and land regions in the tropics (e.g., continental Africa and the Atlantic) and for different definitions of these regions (e.g., different definitions of the Pacific ITCZ sector). Note that for computation simplicity, all of the reconstructed index variables are included in the prior state vector rather than being post-processed from reconstructed spatial climate fields. The monthly indices are reconstructed by the appended state method where here for example, each monthly index occupies 12 elements in the state vector of a given ensemble member.
Two proxy databases form the foundation of the database: the updated PAGES2k database56 and the tree-ring width collection of ref. 57. Additionally, 59 publicly available proxy records including ice cores, speleothems, and lake sediments were also included. In total this database includes 2,978 annually resolved proxies after removing duplicates; only annually or seasonally resolved proxy data values are used such that we include only the annually resolved portions of mixed-resolution proxies. This constitutes the largest multiproxy database employed thus far in a global reconstruction. Figure 1 shows the spatial and temporal distribution of the combined proxy network, with the numbers of each proxy type indicated in the caption. Note that the ice core, speleothem, and sediment records have been grouped together because they are modeled similarly in the proxy system modeling framework (see the following section). Age model uncertainties for the relevant proxy types are only accounted for through our use of the best estimate of the annual ages as determined by the authors of each proxy dataset. The full proxy database and additional proxy metadata is publicly available (Data Citation 2: Zenodo https://doi.org/10.5281/zenodo.1189006).
Proxy system models
DA-based reconstructions must use climate model variables to estimate proxy observations ( in Equation 1). For example, a given climate model's temperature and precipitation can be used to estimate tree-ring width through a sub-model. Such ‘forward models’ are referred to in paleoclimatology as proxy system models (PSMs)58. Here we employ statistical, regression-based PSMs that are specific to each proxy; this improves on previous DA-based reconstructions that used only a univariate linear regression with temperature for all proxies20. We first illustrate this procedure for the ‘other records’ in Fig. 1. The PSMs for these proxies are derived from linear regressions between the ith annual proxy time series, pi, and the local instrumental temperature series from ref. 45 indicated by Xi. Each pi is standardized to unit variance and for the three different reconstructions discussed previously, Xi is either an annual, JJA, or DJF average. The regression equation is
which is calculated over the available temporal overlap between the proxy and the instrumental time series within the calibration period 1920–2000 (leaving approximately 5 decades for a verification interval of 1871–1919, consistent with many previous studies, e.g., refs 59,
The PSM for tree rings is modeled similar to the approach above but with either local instrumental temperature or local instrumentally-derived PDSI62, depending on which instrumental data type has the highest absolute correlation with the proxy over the calibration period. This correlation is computed using the averaging time scale of the reconstruction such that it is possible for a given site to be modeled with temperature for one time average and PDSI for another time average. For the annual reconstruction, 1719 tree-ring chronologies were modeled with PDSI while 872 were modeled with temperature; for the JJA reconstruction 1579 were modeled with PDSI and 1012 with temperature; and for the DJF reconstruction 1572 were modeled with PDSI and 1019 with temperature. Using both temperature and PDSI in the PSMs is necessary because of the heterogenous sensitivities of different tree-ring sites and the inclusion of both tree ring-width and density; additionally, using both temperature and moisture sensitive trees is essential for producing a skillful DA-based reconstruction of both temperature and moisture fields16.
We employ a bivariate regression-based PSM based on ref. 63 for the coral and sclerosponge δ18O proxies. This PSM uses both SST and sea surface salinity to estimate proxy δ18O. Here we calculate regression parameters for each site individually using the long-term ocean reanalysis from ref. 64 instead of basin or region-wide parameter values as used in ref. 63. For non-δ18O coral proxies, we employ a linear univariate regression with SST.
The MATLAB code (https://www.mathworks.com/products/matlab.html) necessary to perform the reconstructions discussed in this data descriptor are available at https://github.com/njsteiger/PHYDA under a free BSD license. The reconstructions were performed using MATLAB version R2015a.
Each of the three reconstructions constituting the first version of PHYDA are publicly available at the Zenodo data repository as NetCDF4 files (Data Citation 1: Zenodo https://doi.org/10.5281/zenodo.1154913), which include all of the reconstructed variables and their uncertainties; specifically this includes the posterior ensemble mean, 1 standard deviation of the posterior ensemble as well as its 5th, 50th, and 95th percentiles. The NetCDF4 format also incorporates all of the associated variable metadata. The paleoclimate proxy database used herein is also publicly available at the Zenodo data repository (Data Citation 2: Zenodo https://doi.org/10.5281/zenodo.1189006).
We validate the reconstructions against observations primarily using two skill metrics: Pearsons' correlation (r) and the mean continuous ranked probability skill score (CRPSS). Correlation is computed using only the reconstruction mean time series at each grid point while the CRPSS metric accounts for both the mean grid point time series as well as the grid point uncertainty estimates. CRPSS is based on the continuous ranked probability score (CRPS), which is a ‘strictly proper’ scoring rule that accounts for the skill of the entire posterior reconstruction distribution65. CRPS penalizes bias, incorrect variance, incorrect phasing, and an ensemble spread that is either too wide or overconfident. Because the posterior ensemble estimates are approximately normally distributed we use equation (5) from ref. 66, where yn=(y−μ)/σ, with y being the observed value, μ the mean of the posterior ensemble estimate, and σ the standard deviation of the posterior ensemble, and where ϕ(yn) and Φ(yn) are respectively the normal probability density function and the normal cumulative distribution function of yn. Note that this implementation assumes that there is no error in the observations. All of our uses of equation (13) are for time series, either individual time series or grid point time series. We therefore compute the mean of all the time-step values of equation (13) and denote it as CRPS. The skill score version, CRPSS, is the reconstructed CRPS computed with respect to the CRPS of a reference distribution, CRPSS≡1−CRPSrec/CRPSref, here the initial uninformed prior. We use CRPSS instead of CRPS because CRPS has the referenceless range of [0, ∞) while CRPSS has the range (−∞, 1] with positive CRPSS indicating that the reconstructed distribution is more skillful for this metric than the uninformed prior. CRPSS is generally a more stringent skill metric than correlation, so we focus here primarily on CRPSS. Additionally, for validating the time series reconstructions we use the metrics of the coefficient of efficiency67 and the cross-spectral coherence computed using a multi-taper method68.
The top two rows of Fig. 2 show the skill of the reconstructed 2 m temperature and SPEI at each grid point using the CRPSS skill metric. The bottom row of Fig. 2 summarizes the spatial skill in box plots for all the spatial variables using r in addition to CRPSS. Seasonal (JJA and DJF) and annual reconstructions are organized by column. The skill metrics are computed for the years 1901–2000 against Berkeley Earth45 for temperature and an observational SPEI computed with a 12-month decaying exponential kernel and using the CRU TS3.23 land surface datasets69; the interval of 1901–2000 is chosen because CRU TS3.23 only extends back to the year 1901. The skill assessments do not include Antarctica because of the sparsity of observational data in this region and because hydroclimate indices are not suited for use over ice-covered landscapes. Assessments of PDSI are included in the bottom row of Fig. 2. We note that the reconstruction uses standard PDSI while the observational verification data62 uses the slightly different self-calibrating version of PDSI.
It is important to note that in Fig. 2 we compute the skill metrics over the interval 1901–2000 while the parameters of the PSMs are fit to observations over the interval of 1920–2000. In traditional reconstruction techniques (e.g., ref. 70) it would not be suitable to show validation statistics over the calibration interval because the instrumental data are used to both fit the proxy data and also for the reconstruction target. In contrast, the PSM parameter fitting here is not an equivalent process because the target field is a pre-industrial climate model simulation and the temporal information is only derived from the proxies. However, in validating these reconstructions we do not rely solely on skill metrics computed over a significant fraction of the PSM calibration interval. In Tables 1, 2, 3 we additionally compute skill metrics for the temperature-based climate indices over the period 1871–1919.
Skill tends to be highest in the tropics and nearby the proxy locations (cf. Fig. 1) during the summer growing season, as is evident, for instance when comparing JJA and DJF reconstruction skill over North America. The annual results also compare favorably with the seasonal reconstructions, particularly over the tropics, showing that it is possible to provide skillful results across a range of time intervals using this DA approach, thus verifying the theoretical results of previous pseudoproxy experiments16. The box plots in the bottom row of Fig. 2 show generally consistent results across the seasons and variables, though the temperature reconstructions are generally more skillful than the PDSI or SPEI reconstructions, while PDSI and SPEI are fairly comparable. Though not shown, the spatial patterns of r are similar to those of CRPSS: for example, the JJA SPEI spatial correlation between the r and CRPSS maps is 0.70 while the corresponding spatial correlation between r and CRPSS for JJA 2 m temperature is 0.68. However, unlike CRPSS, r is consistently high across regions that possess many proxies; this can be seen, for example, when contrasting the CRPSS metric of JJA SPEI (Fig. 2) with r of JJA PDSI in northern Mexico (Fig. 3).
As further validation, we also compare the PDSI reconstructions to the available Drought Atlases14,15,51,
The reconstructed dynamical climate indices span the globe and include many drivers of hydroclimate variability. Figure 4 shows representative verifications of three of these indices: (a) the AMO, (b) the location of the ITCZ over the South Asian monsoon region, and (c) the monthly Niño 3.4 index, with Fig. 4(d) showing the cross-spectral coherence of the reconstructed and observation-based Niño 3.4 index. The panels in Fig. 4 illustrate that a range of important climate indices from different regions are skillfully reconstructed, with high positive correlations, CRPSS, and coherence values. We highlight in particular that these are the first DA-based paleoclimate reconstructions of the location of the ITCZ and monthly Niño indices.
We have additionally performed an exhaustive verification of all the reconstructed index variables. Tables 1, 2, 3 present several skill metrics for each of the temperature-based variables: r and CRPSS over the interval 1871–2000, r and CRPSS over the interval 1871–1919, cross-spectral coherence at the specific periods of 2.5, 5, and 10 years (as in Fig. 4 where the full range of coherence is shown), and CE using the verification mean of 1871–1919 (mimicking a traditional calibration—validation skill test, where here the calibration period is the period over which the PSMs were trained). We note, however, that unlike CRPSS, CE is not a strictly proper scoring metric for ensemble reconstructions65; we also note that both r and CRPSS do not incorporate two time periods as in CE, so r (or r2) is not directly comparable with CE in the manner traditionally used in statistical dendroclimatology71. These skill values are all shown for the annual (Table 1), JJA (Table 2), and DJF (Table 3) reconstructions. Many variables show skill (positive values) across many or all metrics with some variables having particularly high values, such as global mean temperature (GMT) with r=0.88, CRPSS=0.56, and CE=0.77 (Table 1). A few variables, such as the monthly Niño 1+2 index, appear to have skill only at multiyear time scales, low r and negative CRPSS and CE yet high coherence at 2.5, 5, and 10 year periods (Table 1). We note that the negative CE values for the monthly Niño indices (Tables 1, 2, 3) are the result of an annual cycle that shows up too strongly in the reconstructions; r does not account for variance, coherence is looking at multiyear time scales where there is not a variance issue, and CRPSS considers several factors in the reconstruction that outweigh too much variance in this instance. At an annual average of the indices, when the annual cycle is averaged out, all CE values improve; for the reconstruction using annual PSMs and using the tropical annual average defined previously, Niño 1+2 CE=0.37, Niño 3 CE=0.34, Niño 3.4 CE=0.35, Niño 4 CE=0.07, and ΔSST Pacific CE=0.12 (cf. the corresponding column in Table 1). Tables 4, 5, 6 show the r and CRPSS metrics for all of the ITCZ reconstructions, which are limited to the period of 1979–2000 because the Global Precipitation Climatology Project version 2.3 (ref. 46) is only available back to the year 1979. All of the ITCZ reconstructions are skillful in at least one season, though the skill in some regions is strongly dependent on the season, e.g., the Tropical East Africa annual mean versus JJA and DJF (top row in Tables 4, 5, 6).
The series of validation tests presented in this section include 2 measures of spatial skill, r and CRPSS, 18 box plots summarizing the spatial skill of all spatially-resolved variables, and 6 tables with a total of 168 entries verifying the skill of the reconstructed climate indices. These assessments have been done with different skill metrics over three different time intervals (both observational and paleo time intervals) to ensure that a robust picture of each variable's skill can be seen. These validation tests show that many variables are skillfully reconstructed but the level of skill is dependent on the region, the variable, the season (e.g., JJA versus DJF), and the timescale (e.g., annual versus decadal). Future versions of PHYDA will include high-resolution climate model simulations for the prior from the upcoming Paleoclimate Model Intercomparison Project phase 4 (ref. 72), including a more sophisticated bias-correction scheme (e.g., refs 73,74), updates to the proxy network (such as the inclusion of all the tree-ring records used in the drought atlases), and updates to the PSMs as they become further developed.
Paleoclimate reconstructions rely on a relatively sparse network of noisy proxy data time series and the reconstruction may have significant uncertainty depending on the variable, the location, and the time period of interest75. Before using PHYDA for analyses, users should consult the relevant spatial verification Figs. 2,3 or Tables (1, 2, 3, 4, 5, 6) to determine whether the variables of interest can provide useful information. It is also important to consider the range of uncertainty on the variable of interest (included in the NetCDF4 files) and not just the ensemble mean. Because of the decreasing proxy availability further back in time (Figure 1b) the uncertainty in the reconstruction correspondingly increases. Because of how the DA reconstruction methodology is formulated, decreasing amounts of information from proxies will yield a corresponding decrease in the variance of the ensemble mean reconstruction as the prior becomes more heavily relied upon; this gradual reduction in variance of the ensemble mean should not be interpreted as a reduction in the variance of the historical climate.
How to cite this article: Steiger, N. J. et al. A reconstruction of global hydroclimate and dynamical variables over the Common Era. Sci. Data 5:180086 doi: 10.1086/sdata.2018.86 (2018).
Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Steiger, N.J. et al. Zenodo https://doi.org/10.5281/zenodo.1189006 (2018)
We acknowledge the CESM1 (CAM5) Last Millennium Ensemble Community Project and the supercomputing resources provided by NSF/CISL/Yellowstone. This work was supported by the NOAA Climate and Global Change Postdoctoral Fellowship Program administered by UCAR's Visiting Scientist Programs. This work was also supported in part by the National Science Foundation under grants, AGS-1243204, AGS-1401400, AGS-1602581, AGS-1602920, and OISE-1743738. LDEO contribution number 8214. We also thank Mark Cane, Alexey Kaplan, and A. Park Williams for very helpful discussions in the development of this product.
About this article
Local Regions Associated With Interdecadal Global Temperature Variability in the Last Millennium Reanalysis and CMIP5 Models
Journal of Geophysical Research: Atmospheres (2019)