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

==== Atlas.1.4.5 Bias Adjustment ====

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

Bias adjustment is often applied to data from climate model simulations to improve their applicability for assessing climate impacts and risk (e.g., in the Inter-Sectoral Impact Model Intercomparison Project, ISIMIP; [[#Rosenzweig--2017|Rosenzweig et al., 2017]] ). Bias-adjustment approaches ( [[IPCC:Wg1:Chapter:Chapter-10#10.3.1.3|Section 10.3.1.3]] ) are particularly beneficial when threshold-based indices are used, but they can introduce other biases, in particular when applied directly to coarse-resolution GCMs (Cross-Chapter Box 10.2). Bias-adjustment techniques should be chosen carefully for a specific application. In the Atlas, bias adjustment is not applied systematically (in particular, it is not applied for the variables assessed in the Atlas chapter), and only some threshold-dependent extreme indices and climatic impact-drivers (CIDs) included in the Interactive Atlas are bias adjusted (in particular TX35 and TX40 in coordination with Chapter 12). To facilitate integration with WGII, the Atlas uses the same bias-adjustment method as in ISIMIP3 ( [[#Lange--2019a|Lange, 2019a]] ) and the same observational reference (W5E5, see [[#Atlas.1.4.2|Atlas.1.4.2]] ), upscaled to the same resolution as the model to avoid downscaling artefacts (Cross-Chapter Box 10.2). The ISIMIP3 bias-adjustment method is a trend-preserving approach that is recommended for general applications, as it reduces biases while preserving the original climate change signal ( [[#Casanueva--2020|Casanueva et al., 2020]] ). Following the recommendations given in Chapter 10, results in the Interactive Atlas are displayed for both the adjusted and the raw model output.

<div id="box-atlas.1" class="h2-container box-container"></div>

'''Box Atlas.1 | CORDEX-CORE'''

<div id="h2-7-siblings" class="h2-siblings"></div>

[[File:34dc524f1f3545ee6f9042b6bfa01860 IPCC_AR6_WGI_Atlas_Box_Figure_1.png]]

'''Box Atlas.1, Figure'''

'''1 |'''

'''Temperature and precipitation climate change signals at the end of the century (2070–2099).''' The top panels show climate change signals for '''(a)''' temperature and '''(b)''' precipitation for the entire CMIP5 ensemble (box-whisker plots) and the CORDEX-CORE driving GCMs (grey symbols) of the respective CORDEX-CORE results (non-grey symbols) in the South Asia (SAS) reference region. The shape of the grey symbols represents the climate sensitivity of the driving GCMs: triangles pointing upwards (low equilibrium), circles (medium equilibrium), triangles pointing downwards (high equilibrium). The corresponding RCM results are drawn using the same symbols, but in orange for REMO and in blue for RegCM. The bottom panels show the warming signal by 2070–2099 over the CORDEX regions for RCP2.6 '''(c)''' and RCP8.5 '''(d)''' (Figure from [[#Teichmann--2021|Teichmann et al., 2021]] ).
Box Atlas.1

The main objective of CORDEX-CORE is to provide a global homogeneous foundation of high-resolution regional climate model (RCM) projections to improve understanding of local phenomena and facilitate impact and adaptation research worldwide ( [[#Gutowski%20Jr.--2016|Gutowski Jr. et al., 2016]] ). The experimental framework is designed to produce homogeneous regional projections for most inhabited land regions using nine CORDEX domains at 0.22° resolution (Figure Atlas.6): North, Central and South America (NAM, CAM, SAM); Europe (EUR); Africa (AFR); East, South and Southeast Asia (EAS, WAS, SEA); and Australasia (AUS). Due to computational requirements, three GCMs were selected to drive the simulations, HADGEM2-ES, MPI-ESM and NorESM, covering, respectively, the spread of high-, medium- and low-equilibrium climate sensitivities from the CMIP5 ensemble at a global scale (with MIROC5, EC-Earth and GFDL-ES2M as secondary GCMs), focusing on two scenarios RCP2.6 and RCP8.5 (see Box Atlas.1, Figure 1). Two RCMs have contributed so far to this initiative (REMO and RegCM4) constituting an initial homogeneous downscaled ensemble to analyse mean climate change signals and hazards ( [[#Coppola--2021b|Coppola et al., 2021b]] ; [[#Teichmann--2021|Teichmann et al., 2021]] ), and there are ongoing efforts to extend the CORDEX-CORE ensemble with additional regional simulations (e.g., the COSMO-CLM community) to increase the ensemble size. CORDEX-CORE simulations are distributed as part of the information available for the different CORDEX domains at the Earth System Grid Federation (ESGF).

CORDEX-CORE spans the spread of the CMIP5 climate change signals for interquartile ranges of annual mean temperature and precipitation for most of the reference regions covered (Box Atlas.1, Figure 1; [[#Teichmann--2021|Teichmann et al., 2021]] ). However, it is still a small ensemble and for other variables like extremes or climatic impact-drivers it has only been partially investigated in [[#Coppola--2021b|Coppola et al. (2021b)]] and needs further analysis.

<div id="cross-chapter-box-atlas.1" class="h2-container box-container"></div>

'''Cross-Chapter Box Atlas.1 | Displaying Robustness and Uncertainty in Maps'''

<div id="h2-8-siblings" class="h2-siblings"></div>

'''Coordinators:''' José Manuel Gutiérrez (Spain), Erich Fischer (Switzerland)

'''Contributors:''' Alessandro Dosio (Italy), Melissa I. Gomis (France/Switzerland), Richard G. Jones (UK), Maialen Iturbide (Spain), Megan Kirchmeier-Young (Canada/USA), June-Yi Lee (Republic of Korea), Stéphane Sénési (France), Sonia I. Seneviratne (Switzerland), Peter W. Thorne (Ireland/UK), Xuebin Zhang (Canada)

Spatial information on observed and projected future climate changes has always been a key output of IPCC reports. This information is typically represented in the form of maps of historical trends (from observational datasets) and of projected changes for future reference periods and scenarios relative to baseline periods (from multi-model ensembles). These maps usually include information on the robustness or uncertainty of the results such as the significance of trends or the consistency of the change across models. Visualization of this information combines two aspects that are intertwined: the core methodology (measures and thresholds) and its visual implementation. For observed trends, robustness can be simply ascertained by using an appropriate statistical significance test. However, for multi-model mean changes, the consistency across models for the sign of change (model agreement) and the magnitude of change relative to unforced climate variability (signal-to-noise ratio) provide two complementary measures allowing for simple or more comprehensive approaches to represent robustness and uncertainty. While they can be visually represented in various ways with more or less complexity ( [[#Retchless--2016|Retchless and Brewer, 2016]] ), the most common implementation for maps in the climate science community remains the overlay of symbols and/or masking of the primary variable. This Cross-Chapter Box reviews the approaches followed in previous IPCC reports and describes the methods used across this WGI report, presenting the rationale and discussing its relative merits and limitations.

The objectives in AR6 for representing robustness and uncertainty in maps are: 1) adopting a method that can be as coherent as possible across the different global/regional chapters while accommodating different needs, 2) being visually consistent across WGs, and 3) making the different layers of information on the maps as accessible as possible for the reader. As a result, a single approach is selected for observations and two alternative approaches (simple and advanced) are adopted for projected future changes. It is important to highlight that, as in previous reports, these approaches are implemented in maps at a grid-box level and, therefore, are not informative for larger spatial scales (e.g., over AR6 reference regions) where the aggregated signals are less affected by small-scale variability leading to an increase in robustness. This is particularly relevant for the AR6 regional assessments and approaches (e.g., for trend detection and attribution; Cross-Chapter Box 1.4, [[IPCC:Wg1:Chapter:Chapter-11#11.2.4|Section 11.2.4]] ) which are performed for climatological regions and not at grid-box scale (Chapters 11 and 12, and Atlas). Both small and large scales are relevant (e.g., adaptation occurs at smaller scales but also at the level of countries, which are typically larger than a few grid boxes). They are both addressed in the Interactive Atlas, which implements the above approaches for representing robustness in maps at the grid-box level, but also enables the analysis of region-wide signals (e.g., AR6 WGI reference regions, monsoon regions, etc.), helping to isolate background changes happening at larger scales ( [[#Atlas.2.2|Atlas.2.2]] ).

'''Approaches used in previous reports'''
Recent IPCC reports adopted different approaches for mapping uncertainty/robustness, including their calculation method and/or their visual implementation. In AR5 WGI ‘+’ symbols were used to represent significant trends in observations at grid-box level. For future projections, different methods for mapping robustness were assessed (AR5 Box 12.1, [[#Collins--2013|Collins et al., 2013]] ), while proposing as a reference an approach based on relating the multi-model mean climate change signal to internal variability, calculated as the standard deviation of non-overlapping 20-year means in the pre-industrial control runs. Regions where the multi-model mean change exceeded two standard deviations of the internal variability and where at least 90% of the models agreed on the sign of change were stippled (as an indication of a robust signal). Regions where the multi-model change was less than one standard deviation were hatched (small multi-model mean signal). However, this category did not distinguish areas with consistent small changes from areas of significant but opposing/divergent signals. In addition, the unstippled/unhatched areas were left undefined, since the categories were not mutually exclusive.

The AR5 WGII ( [[#Hewitson--2014|Hewitson et al., 2014]] ) used hatching to represent non-significant trends in observations. For future projections, an elaborated approach with four mutually exclusive and exhaustive categories was proposed (to avoid some of the limitations of the AR5 WGI approach): very strong agreement (same as in WGI); strong agreement; divergent change; and little or no change. These depended on the percentage of models showing change greater than the baseline variability and/or agreeing on sign of change (using a 66% agreement threshold). Leaving the robust regions uncovered minimized any interference with the perception of underlying colours that encoded the primary information of the figure.

The two special reports IPCC SR1.5 ( [[#Hoegh-Guldberg--2018|Hoegh-Guldberg et al., 2018]] ) and SROCC ( [[#IPCC--2019a|IPCC, 2019a]] , c) adopted a simplified approach, using only model agreement (≥66% of models agree on sign of change) to characterize robustness. However, cross-hatching was used in SR1.5 to highlight robust areas where models agree, whereas the SROCC used hatching/shading to represent regions where models disagree. Similarly, stippling was used in SR1.5 to indicate regions with significant trends, whereas it was used in SROCC to represent regions where the trends were not significant.

'''Recent methodologies'''
Since AR5 there has been a growing interest for disentangling small consistent climate change signals from significant divergent opposite changes resulting in conflicting information ( [[#Tebaldi--2011|Tebaldi et al., 2011]] ), and different statistical tests have been applied to assess the significance of signals working with the individual models forming the ensemble ( [[#Dosio--2018|Dosio and Fischer, 2018]] ; [[#Yang--2018|Yang et al., 2018]] ; [[#Morim--2019|Morim et al., 2019]] ). Moreover, new approaches have been proposed to identify large changes of opposite sign that compensate in the mean ( [[#Zappa--2021|Zappa et al., 2021]] ). Recent literature has also highlighted the respective risksof Type I vs Type II errors, which can be associated with the determination of robustness in analysed signals ( [[#Lloyd--2018|Lloyd and Oreskes, 2018]] ; [[#Knutson--2019|Knutson et al., 2019]] ). Type I errors are identifying signals when there are none, while Type II errors are concluding there is no signal when there is one. In the case of grid-box level analysis, the focus on small-scale features with inherently large signal-to-noise ratio may emphasize noise even though signals are

Cross-Chapter Box Atlas.1

present when aggregated at larger scale (Sections 11.2.4 and 11.2.5). Consequently, changes averaged over regions or a number of grid boxes emerge from internal variability at a lower level of warming than at the grid-box level (e.g., Cross-Chapter Box Atlas.1, Figure 2). Hence, focus on grid-box significance enhances the risk of Type II errors for overlooking signals significant at the level of AR6 regions. The significance of signals is also affected by the interdependence of single simulations considered in a given ensemble, for example when several come from the same modelling group and share parametrizations or model components ( [[#Knutti--2013|Knutti et al., 2013]] ; [[#Maher--2021|Maher et al., 2021]] ). The risk of Type II errors increases when a model ensemble includes several related simulations showing no signal.

'''The AR6 WGI approach'''
The AR6 WGI adapts the approaches applied in previous IPCC reports into a comprehensive framework based on the two general principles followed by AR5 WGII: 1) not obscuring (with stippling or hatching) the areas where relevant/robust information needs to be highlighted (since stippling and hatching obstruct the visualization of the colours, which can affect the perception/interpretation of the underlying data); 2) using mutually exclusive and exhaustive categories to avoid leaving areas undefined. The three adopted approaches (one for observations and two for model projections) are described in Cross-Chapter Box Atlas.1, Table 1. This framework integrates as much as possible the specificities of each WGI Chapter, proposing in some cases alternative thresholds.

'''Approach A''' is intended for observations and consists of two categories, one for areas with significant trends (colour, no overlay) and one for non-significant ones (coloured areas overlaid with ‘x’), typically using a two-sided test for a significance level of 0.1; [[IPCC:Wg1:Chapter:Chapter-2|Chapter 2]] and Atlas trends have been calculated using ordinary least squares regression accounting for serial correlation ( [[#Santer--2008|Santer et al., 2008]] ).

'''Approach B''' is the simple alternative for model projections. It consists of two categories, one for model agreement (at least 80% of the models agree on the sign of change; colour, no overlay) and the other one for non-agreement (hatching). It is noted that model agreement is computed using ‘model democracy’ (i.e., without discarding/weighting models), since quantifying and accounting for model interdependence (shared building blocks) still remains challenging ( [[IPCC:Wg1:Chapter:Chapter-4#4.2.6|Section 4.2.6]] ). Different thresholds have been used in previous reports and in the literature. In CORDEX studies, 80% has been widely used ( [[#Dosio--2018|Dosio and Fischer, 2018]] ; [[#Kjellström--2018|Kjellström et al., 2018]] ; [[#Nikulin--2018|Nikulin et al., 2018]] ; [[#Yang--2018|Yang et al.,2018]] ; [[#Akperov--2019|Akperov et al., 2019]] ; [[#Rana--2020|Rana et al., 2020]] ), partially due to the small ensemble sizes available in some cases; this also helps to reduce the impact of model interdependence in the final results. Although 90% (used in AR5 WGI) provides high confidence on the forced change, it is deemed too stringent for precipitation-like variables and regional assessments and was therefore not included (see Cross-Chapter Box Atlas.1, Figure 1). The 66% threshold, which has been used in previous reports (e.g., SR1.5 and SROCC) and in the literature, is not used to avoid communicating weak confidence. Cross-Chapter Box Atlas.1, Figure 1 illustrates the application of this approach.

'''Approach C''' is a more advanced alternative for model projections, extending the AR5 WGI and simplifying the AR5 WGII approaches (fewer categories). It consists of three categories: ‘robust change’, ‘conflicting change’, and ‘no change or no robust change’ (see the details in Cross-Chapter Box Atlas.1, Table 1). The first two categories can be interpreted as areas where the climate change signal likely emerges from internal variability (i.e., it exceeds the variability threshold in ≥66% of the models). The variability threshold is defined as

γ = <code> </code> √ <code> 2 </code> * 1.645 * δ <sub>20yr</sub>

, where

δ <sub>20yr</sub>

is the standard deviation of 20-year means, computed from non-overlapping periods in the pre-industrial control (after detrending with a quadratic fit as in AR5 WG1); in cases where this information is not available (e.g., for CORDEX or HighResMIP), the following approximation is used instead:

γ = <code> </code> √ <code> 2/20 </code> * 1.645 * δ <sub>1yr</sub>

, where

δ <sub>1yr</sub>

is the interannual standard deviation measured in a linearly detrended modern period (note that for white noise

δ <sub>20yr</sub> = δ <sub>1yr</sub> / √ <code> 20 </code>

). The factor

√ <code> 2 </code>

is used as in the AR5 WGI approach to account for the fact that the variability of a difference in means (the climate change signal) is of interest. This approach is an evolution of the AR5 WGI method with three notable differences: (a) AR6 uses a lower threshold for internal variability (1.645 corresponding to a 90% confidence level, instead of 2 as used in AR5 WG1); (b) the threshold on agreement in sign is lowered from ≥90% to ≥80%, leading to more grid boxes classified as robust as opposed to conflicting signal; (c) the AR6 method compares signal to variability in each individual model and consequently introduces a 66% cut-off on significant changes, implying that the climate change signal ''likely'' emerges from internal variability in the baseline period.
<br/><br/>
Cross-Chapter Box Atlas.1, Figure 1 illustrates the application of this method considering the effect of the baseline period (1850–1900 versus 1995–2014) and shows that it provides similar results to related approaches proposed in the literature ( [[#Zappa--2021|Zappa et al., 2021]] ).

The two alternative approaches discussed above allow visualization of differentlevels of detail of information on the projected change and are intended for different communication purposes. Approach B just informs on the consistency of the sign of change independent

of its significance relative to internal variability, whereas approach C puts the projected changes into the context of internal variability and allows the highlighting of areas of conflicting signals. It is important to note that different approaches can be applied to the same variable between different chapters for different communication purposes. For example, in maps showing multi-model mean changes of precipitation, [[IPCC:Wg1:Chapter:Chapter-4|Chapter 4]] adopts approach C but [[IPCC:Wg1:Chapter:Chapter-8|Chapter 8]] applies approach B.

In terms of visual implementation, the approach follows recommendations resulting from conversations with IPCC national delegations: 1) having a consistent approach across WGs would aid consistency and reduce the risk of confusion; 2) defining ‘hatching’ as ‘diagonal lines’ in the caption would aid accessibility for non-expert audiences; 3) a clear and concise legend that explains what these patterns represent should be included directly in the figure; 4) information about model uncertainty should be overlaid such that it does not detract from the data underneath.

Since stippling is commonly used to represent statistical significance, diagonal lines were chosen to ‘obscure’ the problematic categories in the above approaches; it also facilitates the visualization of uncertainty in the Interactive Atlas when zooming in. To avoid confusion, methods or thresholds that were unrelated to the three approaches hereby presented were visualized with a different pattern (i.e., model improvement between low- and high-resolution simulations in Chapter 3; agreement between observation-based products in Chapter 5; correlation between two variables in Chapter 6).

'''Cross-Chapter Box Atlas.1, Table'''

'''1 |'''

'''Approaches for representing robustness (uncertainty) in maps of observed (approach A) and projected (approaches B and C) climate changes.'''
[[File:387c6676c7b7a374eb2a57f27de08a87 IPCC_AR6_WGI_Atlas_Table_Box_Atlas_1_Table_1.jpg]] '''Uncertainty at the grid-box and regional scales: interpreting areas with diagonal lines'''
There is no one-size-fits-all method for representing robustness or uncertainty in future climate projections from a multi-model ensemble. One of the main challenges is the dependence of the significance on the spatial scale of interest: while a significant trend may not be detected at every location, a fraction of locations showing significant trends can be sufficient to indicate a significant change over a region, particularly for extremes (e.g., it is ''likely'' that annual maximum one-day precipitation has intensified over the land regions globally even though there are only about 10% of weather stations showing significant trends; Figure 11.13). The approach adopted in WGI works at a grid-box level and, therefore, is not informative for assessing climate change signals over larger spatial scales. For instance, an assessment of the amount of warming required for a robust climate change signal to emerge can strongly depend on the considered spatial scale. A robust change in the precipitation extremes averaged over a region or a number of grid boxes emerge at a lower level of warming than at the grid-box level because of larger variability at the smaller scale (Cross-Chapter Box Atlas.1, Figure 2).

[[File:19b5382b14e156b9e70a2eaa74cddf0f IPCC_AR6_WGI_Atlas_CCBox_Figure_1.png]]

'''Cross-Chapter Box Atlas.1, Figure'''

'''1 |'''

'''Illustration of the simple, (a) and (b), and advanced, (c–f), approaches (B and C in Cross-Chapter Box Atlas.1, Table 1) for uncertainty representation in maps of future projections.''' Annual multi-model mean projected precipitation change (%) from CMIP6 for the period 2040–2060 (left) and 2080–2100 (right) relative to the baseline periods 1995–2014 (a–d) and 1850–1900 (e and f) under a high-emissions (SSP3-7.0) future. Diagonal and crossed lines follow the indications in Cross-Chapter Box Atlas.1, Table 1. Further details on data sources and processing are available in the chapter data table (Table Atlas.SM.15).
[[File:a9539dd991d805756f63d4a08243b07f IPCC_AR6_WGI_Atlas_CCBox_Figure_2.png]]

'''Cross-Chapter Box Atlas.1, Figure 2'''

'''|'''

'''Climate change signals are more separable from noise at larger spatial scales.''' The figure shows the global warming level associated with the emergence of a significant increase in the probability, due to anthropogenic forcing, in the 1-in-20-year daily precipitation event. It uses a 500-year sample from the CanESM2 large ensemble simulations. The left panel uses data analysed over a single grid box, with no spatial aggregation, while the right box uses data averaged over 25 grid boxes to represent the regional scale, with moderate spatial aggregation. Aggregation over 25 grid boxes reduces natural variability, resulting in a smaller warming required for a clear separation between the signal and noise (after Kirchmeier‐Young et al., 2019).

<div id="Atlas.2" class="h1-container"></div>

<span id="atlas.2-the-online-interactive-atlas"></span>