Editing IPCC:AR6/WGI/Chapter-10 (section)

==== 10.2.2.2 Homogenization ====

<div id="h3-9-siblings" class="h3-siblings"></div>

Homogenization aims to make data spatially and temporally ‘homogeneous’. Changes in a homogeneous time series are solely due to large-scale climatic changes (whether forced or due to internal variability). Station data are influenced by factors that act at regional scales, from the mesoscale and local scale down to the microscale ( [[#WMO--2019|WMO, 2019]] ). Station time series contain inhomogeneities such as artificial jumps or trends, which hamper assessments of regional long-term trends. Typical reasons for this are the urbanization of a station’s surroundings, which can lead to warming ( [[#Hamdi--2010|Hamdi, 2010]] ; [[#Hansen--2010|Hansen et al., 2010]] ; [[#Adachi--2012|Adachi et al., 2012]] ; [[#Jones--2016|Jones, 2016]] ; Y. [[#Sun--2016|]] [[#Sun--2016|Sun et al., 2016]] ), or relocations outside of the urban area, which could lead to cooling ( [[#Tuomenvirta--2001|Tuomenvirta, 2001]] ; [[#Yan--2010|Yan et al., 2010]] ; [[#Xu--2013|Xu et al., 2013]] ; [[#Dienst--2017|Dienst et al., 2017]] , 2019). Another potential source of inhomogeneity is a change in measurement methods that affect most instruments of an observational network over a limited time span, such as the transition to Stevenson screens ( [[#Parker--1994|Parker, 1994]] ; [[#Böhm--2010|Böhm et al., 2010]] ; [[#Brunet--2011|Brunet et al., 2011]] ; [[#Auchmann--2012|Auchmann and Brönnimann, 2012]] ) or to automatic weather stations ( [[#WMO--2017b|WMO, 2017b]] ).

The above examples have been selected as they are present in many stations and without going through homogenization they could potentially have influenced global land warming estimates ( [[IPCC:Wg1:Chapter:Chapter-1#1.5.1|Section 1.5.1]] ). Single-break inhomogeneities tend to have a magnitude comparable to global climate change ( [[#Tuomenvirta--2001|Tuomenvirta, 2001]] ; [[#Venema--2012|Venema et al., 2012]] ) and are thus important for analyses of small regions. Also station records in national networks often have similar changes, making them important for national climate change estimates, but many of these influences are averaged out at the global scale ( [[#Jones--2016|Jones, 2016]] ).

The main approach to reduce the influence of inhomogeneities in station observations is statistical homogenization by comparing the data from a candidate station with those of neighbouring reference stations in conjunction with the use of metadata ( [[#Trewin--2010|Trewin, 2010]] ). This is a challenging task because both reference and candidate records normally have multiple inhomogeneities. Three challenges should be considered. First, most of our understanding of statistical homogenization stems from the homogenization of temperature observations from dense networks. Recent studies suggest that our ability to remove biases quickly diminishes for sparse networks ( [[#Gubler--2017|Gubler et al., 2017]] ; [[#Lindau--2018a|Lindau and Venema, 2018a]] ). This affects early instrumental data and observations that are not strongly correlated between stations, such as wind and humidity ( [[#Chimani--2018|Chimani et al., 2018]] ).

Second, in addition to systematic errors, homogenized data also suffer from random errors, introduced by the homogenization process. These errors are largest at the station level but are also present in network-averaged signals ( [[#Lindau--2018b|Lindau and Venema, 2018b]] ). These errors are determined by the break time series, as well as the noise series and the performance of the homogenization method, are spatially correlated, and have an impact on activities such as interpolation and statistical post-processing of climate simulations ( [[#10.2.3.1|Section 10.2.3.1]] ). Third, the above discussion pertains to the homogenization of monthly and annual means. Homogenization of daily variability around the mean is more difficult. For daily data, specific correction methods are used ( [[#Della-Marta--2006|Della-Marta and Wanner, 2006]] ; [[#Mestre--2011|Mestre et al., 2011]] ; [[#Trewin--2013|Trewin, 2013]] ; [[#Zhou--2021|]] [[#Zhou--2021|C. Zhou et al., 2021]] ) that are able to improve the homogeneity of test cases, although recent independent validation efforts were not able to show much improvement ( [[#Chimani--2018|Chimani et al., 2018]] ). The difference with homogenization methods of monthly and annual means may stem from assumptions on the nature of inhomogeneities for daily data, which are not yet well understood ( [[#Chimani--2018|Chimani et al., 2018]] ).

It is ''virtually certain'' that statistical homogenization methods reduce the uncertainties of long-term estimates. Considering a decomposition of the long-term warming error into a bias and a noise uncertainty around the bias, the (trend) bias especially will be reduced, but also most of the noise uncertainty. This conclusion is based on our understanding of the causes of inhomogeneities and their statistical nature combined with the design principles of statistical homogenization methods, as well as on analytical ( [[#Lindau--2018b|Lindau and Venema, 2018b]] ), numerical ( [[#Venema--2012|Venema et al., 2012]] ; [[#Williams--2012|Williams et al., 2012]] ) and empirical validation studies ( [[#Hausfather--2016|Hausfather et al., 2016]] ; [[#Gubler--2017|Gubler et al., 2017]] ; [[#Killick--2020|Killick et al., 2020]] ).

The above section is about the homogenization of land stations. Satellite data has its own issues and methods for homogenization ( [[#Brinckmann--2013|Brinckmann et al., 2013]] ; [[#Huang--2015|Huang et al., 2015]] ; [[#Brogniez--2016|Brogniez et al., 2016]] ). The homogenization of radiosonde data and land station data use similar methods ( [[#Haimberger--2012|Haimberger et al., 2012]] ; [[#Jovanovic--2017|Jovanovic et al., 2017]] ).

<div id="10.2.2.3" class="h3-container"></div>

<span id="data-scarcity"></span>