Page 244 - Contributed Paper Session (CPS) - Volume 2

CPS1846 Maryam I. et al.
from ships and drifting buoys (e.g. Kennedy et al., 2011b). The raw data initially
undergo quality control process followed by a homogenization assessment
(Dunn et al., 2014). In quality control, the data are subject to removing the
random noise resulting from instrumental or observer error (Dunn et al., 2016).
After this, systematic biases are removed (Dunn et al., 2016) that arise from
station movements or incorrect station merges, changes in instruments and
observing practices and land use changes around stations (more commonly
known as urbanization impacts). The process of getting rid of these systematic
biases is referred as homogenization (Domonkos and Coll, 2017). The purpose
of homogenization techniques is to remove or at least reduce the non-climatic
signals that will likely affect the true data characteristics (e.g. Hausfather et al.,
2016; Cao et al., 2017). Blended land and sea surface temperature data are
generated by a variety of organizations. These include Merged Land-Ocean
Surface Temperature (MLOST) by the National Oceanic and Atmospheric
Administration (NOAA) (Smith et al., 2008), Goddard Institute for Space
Studies (GISS) surface temperature anomalies by the National Aeronautics and
Space Administration (NASA) (Hansen et al., 2010), temperature anomalies
provided by Japanese Meteorological Agency (JMA) (Ishii et al., 2005),
HadCRUT temperature anomalies by the Met Office Hadley Centre and the
University of East Anglia Climatic Research Unit (Morice et al., 2012), and
Berkeley Earth Surface Temperature (BEST) by Rhode et al. (2013). Each group
compiles these monthly temperature products using somewhat different input
data, and extensively different quality control and homogenization procedures
(e.g. Rohde, 2013; Jones, 2016). For instance, GISS makes substantial use of
satellite data (Hansen et al., 2010); MLOST only uses satellite data in a limited
capacity (Smith et al., 2008); and HadCRUT and BEST use no satellite data at
all (Morice et al., 2012; Rhode et al., 2013). These data sets are different in
terms of their starting years: 1850-present for HadCRUT and BEST; 1880-
present for GISS and NOAA; and 1891-present for JMA. The spatial resolution
is different as well. Each group also employ different methods of averaging to
derive gridded temperature products from in situ measurements (Jones, 2016;
McKinnon et al., 2017). In addition to these methodological differences, spatial
coverage is also being treated differently by these groups.
The HadCRUT4 and JMA datasets do not interpolate over grid boxes
having missing observations. It is important to note that JMA records are
produced using optimally interpolated (i.e. kriging) sea surface temperature
anomalies (Ishii et al., 2005; Kennedy, 2014). On the other hand, no spatial
interpolation is performed on HadSST3 (Rayner et al., 2006) and CRUTEM4
(Jones et al., 2012) that are the land and sea components of HadCRUT4 data
set. The MLOST performs linear spatial interpolation using nearby stations in
areas lacking stations (Smith et al., 2008). For broader spatial coverage, the
GISS uses a linear inverse distance weighting with data from all the stations up

233 | I S I W S C 2 0 1 9