Global surface temperatures: BEST: Berkeley Earth Surface Temperatures

The following was contributed by Dr. Kevin Cowtan, December 2018:

The Berkeley Earth project provides a set of historical temperature products consisting of estimates of the monthly mean of average, maximum and minimum surface air temperature anomaly over land (Rohde et al., 2013a,b), as well as a pair of blended land-ocean products (Cowtan et al., 2015, supporting information) which provide near global coverage by combining land air temperatures with the HadSST3 sea surface temperature reconstruction (Kennedy et al., 2011a,b). Land temperature data cover the period from 1753 (or 1833 for minimum/maximum temperatures) to the present. The land-ocean reconstructions cover the period from 1850 to the present. Gridded data are provided on a 1 degree latitude/longitude grid. Experimental products are also available at daily resolution and at higher spatial resolution for the USA and Europe.

The distinctive features of the Berkeley land temperature product include:

• A larger inventory of weather station observations than existing products, comprising of over 30,000 weather stations, although some of the stations not included in existing datasets only provide short spans of data (Rohde et al., 2013a).

• Automated identification of ‘breakpoints’ in weather station records corresponding to likely changes in station location or equipment, allowing each station record to be broken down into a series of internally homogeneous fragments (Rohde et al., 2013b). Breakpoint detection is similar to a method used in the Global Historical Climatology Network dataset (Menne and Williams, 2009).

• Reconstruction of a gridded temperature field from the weather station data using a statistical method which incorporates both information about the spatial variability of temperature data and estimates of the reliability of individual weather stations (Rohde et al., 2013b).

• Estimation of local climatology for each station record, allowing the estimation of absolute temperatures as well as temperature anomalies. This enables the use of information from short station records or fragments without the requirement for a significant overlap with either a nearby longer weather station (as in the case of NASA GISTEMP) or a climatology period (as in the case of the Climatic Research Unit CRUTEM) (Hansen et al., 2010; Osborn and Jones, 2014). 

The resulting gridded data are spatially complete over land for recent decades, while in the early record temperature estimates are provided for regions which are close enough to a weather station to provide a meaningful local temperature estimate. The homogenization and reconstruction algorithms lead to a temperature field which is spatially smooth, especially in regions of poor station coverage.

Comparisons of the Berkeley land temperature average to similar products from NASA GISTEMP (Hansen et al., 2010) and the Climatic Research Unit of the University of East Anglia CRUTEM4 (Osborn and Jones, 2014) are shown in figure 1. Figure 1(a) shows a comparison of temperature series from the three providers, and shows large systematic differences over time between the three products, however these differences arise in large part from the different spatial coverage of the products due to different weather station inventories and reconstruction algorithms. Comparison of the temperature average over only the region of common coverage shows much greater consistency between the three datasets. The largest remaining difference arises in the 19th century when the Berkeley reconstruction is cooler than CRUTEM4.

Tests of the Berkeley temperature reconstruction algorithm using synthetic data suggest that it is likely to produce a better estimate of global mean temperature than methods used by other providers, as well as a more detailed reconstruction in regions of higher station density (Rohde, 2013). The greater number of weather stations records than in other products should also contribute to a more reliable result, subject to the downweighting of poor-quality stations performing effectively. The identification of instrumental changes in station records appears to be comparable to the more widely used NOAA alg rithm (Menne and Williams, 2009), at least for the period from 1850, however exposure bias prior to the widespread use of Stephenson screens is likely to impact earlier results (B ̈ohm et al., 2010).

The two Berkeley land-ocean data sets incorporate sea surface temperature (SST) data from the HadSST3 dataset (Kennedy et al., 2011a,b). The sea surface temperature data are infilled using a statistical method which incorporates information about local errors as well as spatial correlation to produce a smoothly varying temperature field. The two versions differ in whether temperature anomalies over sea ice are determined from weather station data (the ‘recommended’ reconstruction) or sea surface temperatures. While the HadSST3 dataset currently incorporates the most extensive analysis of observational biases of current SST products, SSTs are still subject to large systematic biases which vary over time (Kent et al., 2017), so the resulting temperature series will be impacted by uncorrected biases in HadSST3 data (Cowtan et al., 2017; Carella et al., 2018; Davis et al., 2018).

Additional commentary Homogenization of weather station data

Accurate assessment of land air temperature records from weather stations are contingent on the identification and correction of observational biases arising from changes in weather station location, instrumentation or operation practices. Changes in equipment include the introduction of the Stephenson screen in the 19th century and increasing use digital thermometers in recent decades (Menne and Williams, 2009).

The CRUTEM temperature record is largely based on weather station records which have already been corrected for station changes by national weather services, while weather station data from the Global Historical Climatology Network (GHCN) and Berkeley Earth are corrected by automatic algorithms which use both data and metadata to reconcile inconsistencies between nearby stations (Menne and Williams, 2009; Rohde et al., 2013b). Station changes may be identified from metadata, from breaks in the record, or from inconsistencies between a station and its surrounding neighbours.

The GHCN station data are corrected using pairwise homogenization alignment (PHA). The largest difference between the PHA and Berkeley approaches is that in PHA the local station network is used to both identify and correct for weather station changes, and the corrected records are then used in the determination of a gridded temperature reconstruction. In the Berkeley approach the local station network is used to identify station changes, however no correction is calculated at this stage: rather the station record is split into two sub-records. These are then used in the determination of the gridded temperature reconstruction in the same way as records from distinct non-overlapping weather stations.

A comparison of the PHA and Berkeley approaches using synthetic data (Hausfather et al., 2012) suggests that while the PHA method produces better results when no metadata concerning station changes are available, the Berkeley approach may perform better when metadata are available. This may mean that while the PHA algorithm is better at detecting station changes, the Berkeley approach is better at determining the appropriate correction to the temperature observations.

Homogenization of weather station data remains challenging in data sparse periods and regions, and in the presence of rapid local temperature change. Way et al. (2017) found that the Berkeley data show less warming in Canada than a larger archive of station data maintained by the national weather service. Neither the PHA or Berkeley algorithms make use of information about the proximity of a weather station to a coast: coastal regions respond more slowly to climatic changes, and as a result automated homogenization may inflate warming at coastal stations while suppressing it at inland stations in order to reconcile the two sites. The introduction of the Stephenson screen in the mid 19th century reduced the exposure of thermometers to sun and wind, as a result of which earlier records are likely to be biased warm; however since the sign of the bias is seasonally varying it may not be easily detected by the Berkeley algorithm (B ̈ohm et al., 2010; Domonkos and Coll, 2017).

Differences between the two Berkeley land-ocean reconstructions

Berkeley provides two blended land-ocean datasets which provide global coverage for recent decades. The two analyses differ in the way temperatures are determined for sea-ice covered regions: the ‘recommended’ reconstruction infers air temperatures over sea ice from the closest land-based weather stations, while the alternative reconstruction uses the nearest available sea surface temperature observations. The recommended reconstruction shows more warming in the 21st century, with the principal differences arising from the Arctic. A temperature data user must make a choice as to which of these reconstructions is more suitable for their experiment.

For the purposes of estimating air temperatures, the two Berkeley reconstructions can be evaluated on the basis of comparisons to independent measures of surface air temperature. The MERRA2 reanalysis (Gelaro et al., 2017) is a recently developed atmospheric reanalysis spanning 1980 to the present, and is of particular interest since it does not assimilate weather station temperature observations, but does assimilate satellite observations for the whole period. Figure 2 shows per-grid-cell time series correlation between the two Berkeley Earth reconstructions (with land temperatures or SSTs over ice). The land temperature-over-ice reconstruction shows significant correlation into the high Arctic, while SST-over-ice reconstruction shows almost no correlation with MERRA2 for the central Arctic.

Daily variations in Arctic temperature also provide evidence for the use of infilling from land air temperatures for the central Arctic. A key controlling factor in Arctic temperature evolution is melting/freezing with the corresponding transfer of the latent heat of fusion. This plays a key role in the reanalysis, but is absent from historical temperature records because they work with anomalies. This suggests that reanalyses should provide useful independent information concerning Arctic temperature change, and the greater agreement between the reanalysis and the recommended reconstruction supports the use of this reconstruction in the estimation of global surface air temperature change.

The International Arctic Buoy Programme (IABP) provides surface observations of air temperature over sea ice from ice buoys, primarily during the Arctic winter/spring, which are also when the impact of infilling the Arctic is greatest (Rigor et al., 2000). The IABP gridded data are not independent from the Berkeley data due to the inclusion of land stations, however the correlations between the Berkeley land- and SST-based Arctic reconstructions and IABP (Figure 3) provide further support for the use of weather station data to reconstruct temperatures over the Arctic ocean.##