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

==== 10.2.3.1 Observations for Calibrating Statistical Methods ====

<div id="h3-14-siblings" class="h3-siblings"></div>

Statistical downscaling, bias adjustment and weather generators are post-processing methods used to derive climate information from climate simulations. They all require observational data for calibration as well as evaluation ( [[#10.3.3.1|Section 10.3.3.1]] ). Typically, the so-called perfect prognosis methods use quasi-observations for the predictors (i.e., reanalyses) and actual observations for the predictands (the surface variables of interest). By contrast, bias adjustment methods use observations only for the predictands. Weather generators typically require only observed predictands, although some are conditioned on observed predictors as well. Very often these methods are based on daily data, because of user needs, but also because of the limited availability of sub-daily observations and the limited ability of climate models to realistically simulate sub-daily weather ( [[#Iizumi--2012|Iizumi et al., 2012]] ). Some methods are calibrated on the monthly scale, but some of the generated time series are then further disaggregated to the daily scale (e.g., [[#Thober--2014|Thober et al., 2014]] ). A few methods, mainly weather generators, represent sub-daily weather ( [[#Mezghani--2009|Mezghani and Hingray, 2009]] ; [[#Kaczmarska--2014|Kaczmarska et al., 2014]] ). Many methods simulate temperature and precipitation only, although some also represent wind, radiation and other variables. The limited availability of high quality and long observational records typically restricts these applications to a few cases ( [[#Verfaillie--2017|Verfaillie et al., 2017]] ; [[#Pryor--2019|Pryor and Hahmann, 2019]] ). Overall, there is ''high confidence'' that limited availability of station observations, including variables beyond temperature and precipitation as well as sub-daily data, limit the use of statistical modelling of regional climate.

All the limitations and challenges of observational data discussed in [[#10.2.2|Section 10.2.2]] also apply to its use for post-processing of climate model data. High quality and long observational data series are particularly relevant to quantify uncertainties. Different reanalyses present significant discrepancies when used as key predictor variables at the daily scale and may even affect the downscaled climate change signal ( [[#Brands--2012|Brands et al., 2012]] ; [[#Dayon--2015|Dayon et al., 2015]] ; [[#Manzanas--2015|Manzanas et al., 2015]] ; [[#Horton--2019|Horton and Brönnimann, 2019]] ). There is ''high confidence'' that reanalysis uncertainties limit the quality of statistical downscaling in some regions, although no assessment has been made for the most recent reanalysis products.

An important issue for bias adjustment is the correct representation of the required spatial scale. Ideally, bias adjustment is calibrated against area-averaged data of the same spatial scale as the climate model output. Hence, high-quality observed gridded datasets with an effective resolution close to the nominal model resolution are required. Driven by the need to also generate regional-scale information in station-sparse regions, researchers have considered derived datasets that blend in situ and remote-sensing data to produce high-resolution observations to be used as predictands (Sections 10.2.1.2 and 10.2.2.4; [[#Haiden--2011|Haiden et al., 2011]] ; [[#Wilby--2013|Wilby and Yu, 2013]] ).

<div id="10.2.3.2" class="h3-container"></div>

<span id="observation-for-paleoclimate-data-assimilation"></span>