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

==== 10.3.3.7 Statistical Downscaling, Bias Adjustment and Weather Generators ====

<div id="h3-30-siblings" class="h3-siblings"></div>

The performance of statistical downscaling models, bias adjustment and weather generators is determined by the chosen model structure (e.g., to represent variability and extremes or spatial dependence) and, if applicable, the predictors selected ( [[#Maraun--2019a|Maraun et al., 2019a]] ). The VALUE initiative has assessed a range of such methods in a perfect-predictor experiment where the predictors are taken from reanalysis data ( [[#Maraun--2015|Maraun et al., 2015]] , 2019a; [[#Gutiérrez--2019|Gutiérrez et al., 2019]] ). Table 10.2 shows an overview comprising performance results from VALUE and other studies. These results isolate the performance of the statistical method in the present climate. The overall performance in a climate change application also depends on the performance of the driving climate model (Sections 10.3.3.3–10.3.3.6) and the fitness of both the driving model and the statistical method for projecting the climatic aspects of interest ( [[#10.3.3.9|Section 10.3.3.9]] ).

<div id="_idContainer024" class="Basic-Text-Frame"></div>

'''Table''' '''10.2 |''' '''Performance of different statistical method types in representing local weather at daily resolution.''' Individual state-of-the-art implementations may perform better. ‘+’: should work reasonably well based on empirical evidence and/or expert judgement; ‘o’: problems may arise depending on the specific context; ‘–’: weak performance either by construction or inferred from empirical evidence; ‘?’: not studied. The categorisation assumes that predictors are provided by a well-performing dynamical model. Statements about extremes refer to moderate events occurring at least once every 20 years. Adopted and extended from Maraun and Widmann (2018b).

[[File:8cd8d3e180d27cffffe2f61936399f18 IPCC_AR6_WGI_Chapter_10_Table_10_2.png]]
<div id="10.3.3.7.1" class="h4-container"></div>

<span id="performance-of-perfect-prognosis-methods"></span>
===== 10.3.3.7.1 Performance of perfect prognosis methods =====

<div id="h4-12-siblings" class="h4-siblings"></div>

Perfect prognosis methods can perform well when the synoptic forcing (i.e., the explanatory power of large-scale predictors) is strong ( [[#Schoof--2013|Schoof, 2013]] ). Using this approach, downscaling of precipitation is particularly skilful in the presence of strong orographic forcing. The representation of daily variability and extremes requires analogue methods or stochastic regression models, although the former typically do not extrapolate to unobserved values ( [[#Gutiérrez--2019|Gutiérrez et al., 2019]] ; [[#Hertig--2019|Hertig et al., 2019]] ). Temporal precipitation variability is well-represented by analogue methods and stochastic regression, but analogue methods typically underestimate temporal dependence of temperature ( [[#Maraun--2019b|Maraun et al., 2019b]] ). Spatial dependence of both temperature and precipitation is only well-represented by analogue methods, for which analogues are defined jointly across locations, and by stochastic regression methods explicitly representing spatial dependence ( [[#Widmann--2019|Widmann et al., 2019]] ). Overall, there is ''high confidence'' that analogue methods and stochastic regression are able to represent many aspects of daily temperature and variability, but the analogue method is inherently limited in representing climate change ( [[#Gutiérrez--2013|Gutiérrez et al., 2013]] ).

<div id="10.3.3.7.2" class="h4-container"></div>

<span id="performance-of-bias-adjustment-methods"></span>
===== 10.3.3.7.2 Performance of bias adjustment methods =====

<div id="h4-13-siblings" class="h4-siblings"></div>

This subsection assesses the performance of bias adjustment in a perfect predictor context. In practice, climate model imperfections may cause substantial additional issues in the application of bias adjustment. These are assessed separately in Cross-Chapter Box 10.2.

Bias adjustment methods, if driven by reanalysis predictors, in principle adjust well all the aspects that they intend to address ( [[#Maraun--2018b|Maraun and Widmann, 2018b]] ). For temperature, all univariate methods are good for adjusting means, variance, and high quantiles ( [[#Gutiérrez--2019|Gutiérrez et al., 2019]] ; [[#Hertig--2019|Hertig et al., 2019]] ). For precipitation, means, intensities, wet-day frequencies, and wet–dry and dry–wet transitions are well-adjusted ( [[#Gutiérrez--2019|Gutiérrez et al., 2019]] ; [[#Maraun--2019b|Maraun et al., 2019b]] ). The representation of high quantiles depends on the chosen method, although flexible quantile mapping performs best ( [[#Hertig--2019|Hertig et al., 2019]] ). Empirical (non-parametric) methods perform better than parametric methods over the observed range, but it is unclear how this translates into extrapolation to unobserved values (IPCC, 2015; [[#Hertig--2019|Hertig et al., 2019]] ). Many quantile mapping methods overestimate interannual variability ( [[#Maraun--2019b|Maraun et al., 2019b]] ). Temporal and spatial dependence are usually not adjusted and thus inherited from the driving model ( [[#Maraun--2019b|Maraun et al., 2019b]] ; [[#Widmann--2019|Widmann et al., 2019]] ). Spatial fields are thus typically too smooth in space, even after bias adjustment ( [[#Widmann--2019|Widmann et al., 2019]] ).

Several studies show improved simulations of present-day impacts, when the impact model is fed with bias-adjusted climate model output, including the assessment of river discharge ( [[#Rojas--2011|Rojas et al., 2011]] ; [[#Muerth--2013|Muerth et al., 2013]] ; [[#Montroull--2018|Montroull et al., 2018]] ), forest fires ( [[#Migliavacca--2013|Migliavacca et al., 2013]] ), crop production ( [[#Ruiz-Ramos--2016|Ruiz-Ramos et al., 2016]] ), and regional ocean modelling ( [[#Macias--2018|Macias et al., 2018]] ).

There is ''high confidence'' that bias adjustment can improve the marginal distribution of simulated climate variables, if applied to a climate model that adequately represents the processes relevant for a given application (Cross-Chapter Box 10.2).

<div id="10.3.3.7.3" class="h4-container"></div>

<span id="performance-of-weather-generators"></span>
===== 10.3.3.7.3 Performance of weather generators =====

<div id="h4-14-siblings" class="h4-siblings"></div>

Weather generators represent well most aspects that are explicitly calibrated. This typically includes mean, variance, high quantiles (for precipitation, if explicitly modelled), and short-term temporal variability for both temperature and precipitation, whereas interannual variability is strongly underestimated ( [[#Frost--2011|Frost et al., 2011]] ; [[#Hu--2013a|Hu et al., 2013a]] ; [[#Keller--2015|Keller et al., 2015]] ; [[#Dubrovsky--2019|Dubrovsky et al., 2019]] ; [[#Gutiérrez--2019|Gutiérrez et al., 2019]] ; [[#Hertig--2019|Hertig et al., 2019]] ; [[#Maraun--2019b|Maraun et al., 2019b]] ; [[#Widmann--2019|Widmann et al., 2019]] ). There is growing evidence that some spatial weather generators fairly realistically capture the spatial dependence of temperature and precipitation ( [[#Frost--2011|Frost et al., 2011]] ; [[#Hu--2013a|Hu et al., 2013a]] ; [[#Keller--2015|Keller et al., 2015]] ; [[#Evin--2018|Evin et al., 2018]] ; [[#Dubrovsky--2019|Dubrovsky et al., 2019]] ). There is ''high confidence'' that weather generators can realistically simulate a wide range of local weather characteristics at single locations, but there is ''limited evidence'' and ''low agreement'' of the ability of weather generators to realistically simulate the spatial dependence of atmospheric variables across multiple sites.

<div id="10.3.3.8" class="h3-container"></div>

<span id="performance-at-simulating-historical-regional-climate-changes"></span>