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

=== 10.4.1 Methodologies for Regional Climate Change Attribution ===

<div id="h2-20-siblings" class="h2-siblings"></div>

Attribution at sub-continental and regional scales is usually more complicated than at the global scale due to various factors: a larger contribution from internal variability, an increased similarity among the responses to different external forcings leading to a more difficult discrimination of their effects, the importance at regional scale of some omitted forcings in global model simulations, and model biases related to the representation of small-scale phenomena ( [[#Zhai--2018|Zhai et al., 2018]] ). Since AR5 and in addition to standard optimal fingerprint regression-based approaches ( [[IPCC:Wg1:Chapter:Chapter-3#3.2.1|Section 3.2.1]] and Zhai et al. 2018), several emerging methodologies have been increasingly used for regional-scale climate change attribution. These include several statistical approaches that differ in their use or omission of spatiotemporal co-variance information. Dynamical adjustment and pattern recognition techniques fall into the category of spatiotemporal methods while univariate detection and attribution methods rely on single grid-point analysis. Finally, the development, evaluation and use of all these methodologies rely upon the availability of multiple and high-quality observational datasets ( [[#10.2|Section 10.2]] ) as well as multi-model simulations of the historical period constrained by different external forcing combinations, including single-forcing experiments and single-model initial-condition large ensembles (SMILEs).

<div id="10.4.1.1" class="h3-container"></div>

<span id="optimal-fingerprinting-methods"></span>
==== 10.4.1.1 Optimal Fingerprinting Methods ====

<div id="h3-38-siblings" class="h3-siblings"></div>

Optimal fingerprint regression-based methods have been applied to detection and attribution of mean temperature anthropogenic signal in several regions of the world such as Canada, India, central Asia, northern and western China, Australia, and North Africa ( [[#Xu--2015|Xu et al., 2015]] ; [[#Li--2017|]] [[#Li--2017|C. Li et al., 2017]] ; [[#Dileepkumar--2018|Dileepkumar et al., 2018]] ; Y. [[#Wang--2018|]] [[#Wang--2018|Wang et al., 2018]] ; [[#Peng--2019|Peng et al., 2019]] ; [[#Wan--2019|Wan et al., 2019]] ). The influence of anthropogenic forcing, and in particular that of greenhouse gases (GHGs), is robustly detected in annual and seasonal mean temperatures for all considered regions. Most of the observed regional temperature changes since the mid-twentieth century can only be explained by external forcings, with anthropogenic influence being the dominant factor. GHG increase is found to be the primary factor of the anthropogenic-induced warming while the aerosol forcing leads to a cooling offsetting a fraction of the GHG change ( [[#Li--2016|]] [[#Li--2016|]] [[#Li--2016|]] [[#Li--2016|C. Li et al., 2016]] , 2017). While the influence of external natural forcing can often be detected as well, its contribution to observed changes is usually much smaller ( [[#Li--2017|]] [[#Li--2017|C. Li et al., 2017]] ; [[#Wan--2019|Wan et al., 2019]] ). Temperature detection results are found to be robust to the use of different observational datasets and detection methodologies ( [[#Dileepkumar--2018|Dileepkumar et al., 2018]] ).

Detection of mean precipitation changes caused by human influence is much more difficult, due to a larger role of internal variability at regional to local scales, as well as substantial modelling and observational uncertainty ( [[#Wan--2015|Wan et al., 2015]] ; [[#Sarojini--2016|Sarojini et al., 2016]] ; [[#Li--2017|]] [[#Li--2017|C. Li et al., 2017]] ). However, multi-decadal precipitation changes due to anthropogenic forcing have been detected for several regions. [[#Ma--2017b|Ma et al. (2017b)]] show that anthropogenic forcing has strongly contributed to the observed shift of China daily precipitation towards heavy precipitation. The observed weakening of the East Asia summer monsoon, also known as the southern flooding and northern drought pattern has been partially linked to anthropogenic forcing ( [[IPCC:Wg1:Chapter:Chapter-8#8.3.2.4.2|Section 8.3.2.4.2]] ; [[#Song--2014|Song et al., 2014]] ; [[#Zhou--2017|Zhou et al., 2017]] ; [[#Tian--2018|Tian et al., 2018]] ). Changes in GHGs lead to increasing precipitation over southern China, while changes in anthropogenic aerosols over East Asia are the dominant factors determining drought conditions over northern China ( [[#Song--2014|Song et al., 2014]] ; [[#Tian--2018|Tian et al., 2018]] ). Based on all-forcing and single-forcing simulation ensembles with a high-resolution model, [[#Delworth--2014|Delworth and Zeng (2014)]] found that the observed long-term regional austral autumn and winter rainfall decline over southern and particularly south-west Australia is partially reproduced in response to anthropogenic changes in GHGs and ozone in the atmosphere, whereas anthropogenic aerosols do not contribute to the simulated precipitation decline. In contrast, the observed increase of north-west Australian summer rainfall since 1950 has been partially attributed to anthropogenic aerosol based on CMIP5 detection and attribution single-forcing simulations ( [[IPCC:Wg1:Chapter:Chapter-8#8.3.2.4.6|Section 8.3.2.4.6]] ; [[#Dey--2019a|Dey et al., 2019a]] , [[#Dey--2019b|b]] ).

It is noteworthy that these methods require a very significant reduction of spatial and temporal dimensions in order to reliably estimate the co-variance matrix of internal variability (an entire region is thus often considered as being only one or a few spatial points that represent the spatial average of the whole region or a few sub-regions; time samples are often 5- or 10-year averages). Finally, model bias is rarely considered in statistical models used in detection and attribution regional studies, while it has been shown to have a strong impact on the stability of detection results and their associated confidence intervals when increasing the spatial dimension ( [[#Ribes--2013|Ribes and Terray, 2013]] ). New statistical methods are emerging to provide some alternative to standard optimal fingerprinting but they have not yet been evaluated and applied at regional scales ( [[IPCC:Wg1:Chapter:Chapter-3#3.2.2|Section 3.2.2]] ).

<div id="10.4.1.2" class="h3-container"></div>

<span id="other-spatiotemporal-statistical-methods-for-isolating-regional-climate-responses-to-external-forcing"></span>
==== 10.4.1.2 Other Spatiotemporal Statistical Methods for Isolating Regional Climate Responses to External Forcing ====

<div id="h3-39-siblings" class="h3-siblings"></div>

The primary objective of any attribution method is to optimally separate the influences of external forcing and internal variability on a global or regional climate record. In a multi-model ensemble context, the estimation of the externally-forced climate response has been typically performed by ensemble averaging of linear trends or regional domain spatial average, thus not taking into account the available and complete space and time co-variance information. Since AR5, methods using spatiotemporal information have been further developed and used to improve the separation between external and internal drivers in multiple or single historical climate realizations performed by a given global model.

The typical ensemble size of CMIP historical climate simulations for a given model traditionally range between one and ten members, with three often being the default choice. At the regional scale, a simple ensemble average with such sample sizes does not provide robust estimates of the response patterns to external forcing ( [[#Maher--2019|Maher et al., 2019]] ; [[#Deser--2020|Deser et al., 2020]] ). Since AR5, pattern filtering methods such as signal-to-noise maximizing empirical orthogonal functions ( [[#Ting--2009|Ting et al., 2009]] ) have been shown to improve the identification of forced response patterns when few model members are available ( [[#Wills--2020|Wills et al., 2020]] ). Using SMILEs as a test bed, it has been shown that pattern filtering strongly reduces the number of ensemble members needed to estimate the forced response pattern compared to simple ensemble averaging. Pattern filtering allows the identification of low signal-to-noise signals such as the El Niño-like response to volcanic eruptions ( [[#Khodri--2017|Khodri et al., 2017]] ; [[#Wills--2020|Wills et al., 2020]] ).

Methods to extract the response to external forcing in an observed or simulated single realization include dynamical adjustment ( [[#Smoliak--2015|Smoliak et al., 2015]] ; [[#Deser--2016|Deser et al., 2016]] ; [[#Sippel--2019|Sippel et al., 2019]] ) and time scale separation methods ( [[#DelSole--2011|DelSole et al., 2011]] ; [[#Wills--2018|Wills et al., 2018]] , 2020). Dynamical adjustment seeks to isolate changes in surface air temperature or precipitation that are due purely to atmospheric circulation changes. The residual can then be analysed and attributed to internal changes in both land or ocean surface conditions and the thermodynamical response to external forcing. [[#Smoliak--2015|Smoliak et al. (2015)]] performed their dynamical adjustment using partial least squares regression of temperature to remove variations arising from sea level pressure changes. [[#Deser--2016|Deser et al. (2016)]] used constructed atmospheric circulation analogues and resampling to estimate the dynamical contribution to changes in temperature. [[#Sippel--2019|Sippel et al. (2019)]] used machine learning techniques known as regularized linear regression to provide estimates of circulation-induced components of precipitation and temperature variability from global to local scales. It is noteworthy that the dynamical adjustment method by itself cannot account for the component of the forced response associated with circulation changes that project onto atmospheric internal variability. However, this component can be estimated within a model framework by averaging the dynamical contribution across multiple members of a SMILE ( [[#Deser--2016|Deser et al., 2016]] ).

Dynamical adjustment methods have been used by, for instance, [[#Deser--2016|Deser et al. (2016)]] , [[#Saffioti--2016|Saffioti et al. (2016)]] , [[#O’Reilly--2017|O’Reilly et al. (2017)]] , [[#Gong--2019|Gong et al. (2019)]] , and R. [[#Guo--2019|]] [[#Guo--2019|Guo et al. (2019)]] . [[#Deser--2016|Deser et al. (2016)]] focused on the causes of observed and simulated multi-decadal trends in North American temperature. They demonstrated that the main advantage of this technique is to narrow the spread of temperature trends found by the model ensemble and to bring the dynamically-adjusted observational trend much closer to the forced response estimated by the model ensemble mean. Similar results were obtained by [[#Saffioti--2016|Saffioti et al. (2016)]] regarding recent observed winter temperature and precipitation trends over Europe. Similarly, [[#O’Reilly--2017|O’Reilly et al. (2017)]] applied dynamical adjustment techniques to more carefully determine the influence of the Atlantic Multi-decadal Variability (AMV; Annex IV.2.7) on continental climates. Over Europe, summer temperature anomalies induced thermodynamically by the warm phase of the AMV are further reinforced by circulation anomalies; meanwhile, precipitation signals are largely controlled by dynamical responses to the AMV. Based on a partial least-squares approach, [[#Gong--2019|Gong et al. (2019)]] showed that recent winter temperature 30-year trends over northern East Asia are strongly influenced by internal variability linked to decadal changes of the Arctic Oscillation. Using dynamical adjustment purely on precipitation observations, R. [[#Guo--2019|]] [[#Guo--2019|Guo et al. (2019)]] showed that human influence has led to increased winter precipitation across north-eastern North America, as well as a small region of north-western North America, and to an increase in precipitation across much of north-western and north central Eurasia. The latter results confirm previous findings obtained by standard optimal fingerprinting methods ( [[#Wan--2015|Wan et al., 2015]] ).

Time scale separation methods such as the low-frequency component analysis and ensemble empirical mode decomposition methods take advantage of the longer time scale associated with anthropogenic external forcing compared to that of most internal modes of variability. The low-frequency component analysis method tries to find low-frequency variability patterns by searching for linear combinations of a moderate number of empirical orthogonal functions that maximize the ratio of low-frequency to total variance. It has first been used to separate internal modes of interannual and decadal variability from slowly varying and externally-forced variability in the Pacific and Atlantic oceans ( [[#Wills--2018|Wills et al., 2018]] , 2019). The methodology has also been applied to patterns of observed surface air temperature to isolate the slow components of observed changes that are consistent with the expected response to anthropogenic greenhouse gas and aerosol forcing ( [[#Wills--2020|Wills et al., 2020]] ).

The ensemble empirical mode decomposition method ( [[#Wu--2009|Wu and Huang, 2009]] ; [[#Wilcox--2013|Wilcox et al., 2013]] ; [[#Ji--2014|Ji et al., 2014]] ; [[#Qian--2014|Qian and Zhou, 2014]] ) decomposes data, such as time series of historical temperature and precipitation, into independent oscillatory modes of decreasing frequency. The last step of the method leaves behind a smooth and low-frequency residual time series. Typically, the non-linear anthropogenic trend (e.g., of 20th-century temperature) can be reconstructed by summing the long-term mean, the residual, and eventually the lowest-frequency mode to account for a multi-decadal forced signal, for instance associated with anthropogenic aerosol forcing. The ensemble empirical mode decomposition method is an example of a data-driven, non-parametric approach that can be used to directly provide an estimate of the forced response without the need for model data ( [[#Qian--2016|Qian, 2016]] ).

<div id="10.4.1.3" class="h3-container"></div>

<span id="other-regional-scale-attribution-approaches"></span>
==== 10.4.1.3 Other Regional-scale Attribution Approaches ====

<div id="h3-40-siblings" class="h3-siblings"></div>

The univariate detection method does not use spatial pattern information, but compares observed trends in gridded datasets with distributions of trends from ensembles of simulations during the historical period ( [[#Knutson--2013|Knutson et al., 2013]] ; [[#Knutson--2018|Knutson and Zeng, 2018]] ). The trends arising from simulations constrained by natural forcing-only and all-forcing are compared with distributions of trends purely due to internal variability and derived from long simulations with constant pre-industrial external forcing. Consistency between observed and simulated historical trends is also assessed with statistical tests that can be applied independently over a large number of grid points. The fraction of area over a given region where the change is classified as detectable, attributable, or consistent/inconsistent, is then finally estimated. The method can be viewed as a simple consistency test for both amplitude and pattern of observed versus simulated trends. Its application to CMIP3 and CMIP5 models suggests that 80% of the Earth’s surface has a detectable anthropogenic warming signal ( [[#Knutson--2013|Knutson et al., 2013]] ). Regarding regional land precipitation changes over the 1901–2010 and 1951–2010 periods, application of the univariate detection method based on CMIP5 models suggests attributable anthropogenic changes at several locations such as increases over regions of the north-central USA, southern Canada, Europe, and southern South America and decreases over parts of the Mediterranean region, northern tropical Africa and south-western Australia ( [[#Delworth--2014|Delworth and Zeng, 2014]] ; [[#Knutson--2018|Knutson and Zeng, 2018]] ).

Another regional attribution technique is based on the similarity of past changes between observations and one or several simulations of a large ensemble that share the same time evolution for a suggested driver of these changes. [[#Huang--2020b|Huang et al. (2020b)]] used a perturbed physics ensemble to attribute the drying trend of the Indian monsoon over the latter half of the 20th century to decadal forcing from the Pacific Decadal Variability (PDV; Annex IV.2.6). The ensemble members predicted different trends in PDV behaviour across the 20th century and the negative precipitation trend was only replicated in those members with a strong negative-to-positive PDV transition across the 1970s, consistent with the observed PDV behaviour (see also the detailed case study in [[#10.6.3|Section 10.6.3]] ). In a similar manner, [[#Cvijanovic--2017|Cvijanovic et al. (2017)]] addressed the possible influence of Arctic sea ice loss on the North Pacific pressure ridge and, consequently, on south-western USA precipitation. They sampled the uncertainties in selected sea ice physics parameters to achieve a ‘low Arctic sea ice’ state in their perturbed simulations. They then compared the latter with control simulations representative of sea ice conditions at the end of the 20th century to assess changes purely due to sea ice loss.

New methods aiming to remove underlying model biases before performing detection and attribution, for instance related to precipitation changes, are emerging based on image transformation techniques such as warping ( [[#Levy--2014a|Levy et al., 2014a]] ). By correcting location and seasonal precipitation biases in CMIP5 models, [[#Levy--2014b|Levy et al. (2014b)]] showed that the agreement between observed and fingerprint patterns can be improved, further enhancing the ability to attribute observed precipitation changes to external forcings. The improvement mainly relies on the assumption that precipitation changes are tied to the underlying climatology, which has been shown to be a reasonable assumption in regions of the world where intensification of the hydrological cycle is expected ( [[#Held--2006|Held and Soden, 2006]] ).

Importantly, evidence that the models employed in regional-scale attribution are fit for purpose is essential in order to estimate the degree of confidence in the attribution results ( [[#10.3.3|Section 10.3.3]] ). For example, models need to be evaluated and assessed in their ability to simulate internal variability modes that are known to be important drivers of regional climate change (Sections 3.7 and 10.3.3.3 and Annexes IV.2 and IV.3). Models are likely to have different performance in different regions and therefore their evaluation needs to be performed in terms of key physical processes and mechanisms relevant to the climate of the region under consideration ( [[#10.3.3|Section 10.3.3]] ).

To conclude, there is ''very'' ''high confidence'' ( ''robust evidence'' and ''high agreement'' ) that the use of diverse and independent attribution methods, multiple model ensemble types and observed datasets strengthens the robustness of results of regional-scale attribution studies. Since AR5, multiple SMILEs have provided an adequate testbed for new attribution methodologies aimed at separating forced signals from internal variability in observational records as well as small-size single-model ensembles.

<div id="10.4.2" class="h2-container"></div>

<span id="regional-climate-change-attribution-examples"></span>