Detectable anthropogenic changes in daily-scale circulations driving summer rainfall shifts over eastern China

With growing concerns about the regional impacts of the expected increase in flood and drought frequency under global warming, understanding how precipitation responds to anthropogenic forcing in a particular region is becoming immensely important for informing local adaptation and mitigation planning (Field et al 2012, Sarojini et al 2016). Although the evidence for anthropogenically forced precipitation change on the global and continental scales continues to strengthen, attribution of regional changes in precipitation remains a huge challenge (Min et al 2008, Stott et al 2010, Hegerl and Zwiers 2011, Sarojini et al 2016). At the regional scale, changes in atmospheric circulation exert strong control on local climate, which could reinforce or counteract the response of regional precipitation to external forcing (Bindoff et al 2013, Shepherd 2014, Pfahl et al 2017). Strong spatial variation, large internal variability and model errors arise, thus precluding the detection of anthropogenic signals in precipitation change (Stott et al 2010, Hegerl and Zwiers 2011, Bindoff et al 2013, Ma et al 2017). However, it is believed that improved process understandings will provide greater opportunity for studies on the attribution of regional precipitation change (Sarojini et al 2016).

During the past few decades, eastern China's summers have become wetter in the south but drier in the north, and the region has witnessed the increasing exposure of agriculture, infrastructure and human life to the devastating damage caused by severe floods and droughts (Ding et al 2008, Huang et al 2013, Qian and Zhou 2014, Nigam et al 2016, Li and Lu 2017, Day et al 2018). This well-known 'southern flood–northern drought' (SFND) precipitation pattern, associated mainly with the weakening East Asia summer monsoon (EASM), has therefore motivated multiple detection and attribution (D&A) studies (Zhou et al 2009, 2017). Previous studies, that focused on changes in the decadal-scale mean state of circulation and precipitation pattern related to the EASM, have detected impacts of anthropogenic forcing and internal variability on the SFND using climate model simulations (Wang et al 2013, Song et al 2014, Jiang et al 2017, Yang et al 2017, Tian et al 2018, Dong et al 2019). However, when examining trends in summer precipitation and confronting model simulations with observations, D&A results remain controversial and progress is confined to identifying detectable human influence on the changing distribution of precipitation in eastern China (Liu et al 2015, Burke and Stott 2017, Ma et al 2017, Lu et al 2020). More specifically, different climate models with imperfect simulations of precipitation trends give nearly opposite signal of changes in total summer precipitation over eastern China, which consequently lead to equivocal D&A conclusions (Burke and Stott 2017, Lu et al 2020). Therefore, uncertainty in model simulations of trends in summer precipitation over eastern China obscures the detectability of human influence (Zhang 2015, Lu et al 2020). Additionally, precipitation trends of the opposite sign offsetting each other induced by averaging simplistically over eastern China where north–south contrasting precipitation changes are located may be another barrier to disentangling anthropogenic signals (Zhang et al 2007).

Recent studies have sought to improve understanding of the dynamic and thermodynamic mechanisms behind the EASM response to external forcing using moisture budget analysis, while changes in precipitation and evaporation cannot be isolated (Li et al 2015, Zhou et al 2020b). Through tracking the occurrence of daily circulation patterns (CPs) using self-organizing map (SOM) cluster analysis, Zhou et al (2020a) demonstrated the changing prevalence of CPs is robust and can well explain SFND over eastern China. The dynamic effect stemming from changes in CPs along with the dynamic–thermodynamic interactions dictate the latitudinal boundary of SFND. The framework of combining the dynamic (related to atmospheric circulation) and thermodynamic (linked to temperature change) decomposition with targeted model experiments has shown some success in the attribution of continent-scale precipitation change and probabilistic event attribution (Marvel and Bonfils 2013, Schaller et al 2016, Vautard et al 2016) and is promising to shed light on the attribution of the contrasting precipitation trends over eastern China.

As noted above, the critical but poorly understood role of dynamic processes highlights the importance of identifying dynamic responses to external forcing, which could subsequently facilitate the D&A of regional precipitation changes (Corti et al 1999, Shepherd 2014, Sarojini et al 2016). Using a process-based approach, we build on the SOM cluster analysis in Zhou et al (2020a) to address whether anthropogenic forcing impacts the frequency of large-scale daily CPs. Based on the link between CPs and summer precipitation, we infer if anthropogenic influence is detectable on the SFND pattern.

For the attribution analysis, we used a pair of model ensembles from the latest Met Office HadGEM3-GA6-based attribution system at N216 resolution (0.56° × ∼0.83° horizontal resolution, HadGEM3-A-N216 hereafter; Christidis et al 2013, Ciavarella et al 2018). One of these simulates the 'actual' world with both anthropogenic and natural (ALL) forcing and the other represents a counterfactual 'natural' world with only natural (NAT) forcing. Boundary conditions for ALL forcing are given by the observed sea surface temperatures (SSTs) and sea ice (SIC) fields from HadSST1 (Rayner et al 2003). For NAT forcing, the anthropogenically induced pattern of SSTs is removed from the observed SSTs. This is calculated as the multi-model mean difference between simulations with and without anthropogenic influence from the CMIP5 ensemble (Taylor et al 2012). Sea ice fields are also adjusted for the NAT forcing using an empirical relationship between historically observed polar SSTs and SIC (Christidis et al 2013, Ciavarella et al 2018). Each ensemble comprises 15 members during 1961–2013. As reflected in the more realistic precipitation globally and regionally benefited from high-resolution and forcing from prescribed SST and SIC, HadGEM3-A-N216 well captures many ongoing large-scale modes of variability (e.g. the El Niño–Southern Oscillation) and monsoon systems (Johnson et al 2016, Burke and Stott 2017, Ciavarella et al 2018, Li et al 2018). It also reproduces well the main large-scale CPs associated with local climate over Europe and their frequencies (Vautard et al 2019), as well as characteristics of circulation related to East Asian winter monsoon (Hao et al 2019). Despite the lack of ocean coupling, atmospheric general circulation models possess comparable skill to the coupled ones in simulating anthropogenic climate change and specifically the monsoon climatology over East Asia (He and Soden 2016, Li and Ting 2017). These aspects all support the utilization of HadGEM3-A-N216 to understand the large-scale dynamic responses to anthropogenic climate change.

The 20 CPs obtained from the SOM analysis (described later in this section) by Zhou et al (2020a), considered here as the reference circulation types for model simulations, were based on the reanalysis product (ERA reanalysis hereafter) from a combination of ERA-40 covering 1961–1978 and ERA-Interim spanning 1979–2015, at a spatial resolution of 2° (Uppala et al 2005, Dee et al 2011). The concatenated ERA reanalysis has been widely used in atmospheric circulation typing and validated against other reanalysis datasets (Bao and Wallace 2015, Huguenin et al 2020, Zhou et al 2020a). Though not as high resolution as the latest ERA5 reanalysis dataset, the ERA reanalysis is demonstrated to be sufficient in resolving and discriminating the characteristics of various CPs as atmospheric circulation is characterized by large spatial scales (Boé and Terray 2008, Belleflamme et al 2013, Bao and Wallace 2015, Hersbach et al 2018, Huguenin et al 2020). Additionally, the computational cost of SOM is greatly reduced with a smaller size of input elements using the coarse resolution dataset (Tymvios et al 2010, Lennard and Hegerl 2015). We also used the observed daily precipitation data for the period 1961–2015 from 519 stations over eastern China (112°–121° E, 26°–42° N, red rectangle in figure S1 (available online at stacks.iop.org/ERL/16/074044/mmedia)) from the National Meteorological Information Center, China Meteorological Administration.

SOM consists of a two-dimensional array of nodes, each of which is represented by a reference vector equal in dimension to the input data sample (Kohonen 2001, Hewitson and Crane 2002). After initialization with random values, the reference vectors of the 'winning' node with the minimum Euclidean distance to the input vector and its surrounding nodes are updated to better resemble the input data space in an iterative training process. A two-phase training process, involving a first broad capture of the large-scale structure in the input data and a further fine-tuning of the nodes to obtain final convergence, has been demonstrated to allow for better separation between clusters (Jiang et al 2015, Lennard and Hegerl 2015). A 5 × 4 SOM containing 20 CPs was produced using the two-phase approach with the daily circulations represented by the temporally standardized anomalies of 500 hPa geopotential height (covering 30°–160° E, 10°−70° N) and mean sea level pressure (covering 105°–135° E, 15°–45° N) in summers during 1961–2015 from ERA reanalysis (figure S1; see also Zhou et al 2020a). The effect of thermal expansion on geopotential height related to global warming is not accounted for here. As discussed in Horton et al (2015) and Zhou et al (2020a), substantial regional variations in the direction and magnitude of the trends in 500 hPa geopotential heights suggest the inappropriateness of detrending either with the removal of uniform domain average trend (commonly used in previous clustering studies which assume uniform thermal dilation, Cattiaux et al 2013) or the removal of local geopotential height trends (the signal of changes in CPs would also be removed). Moreover, the same SOM analysis using the detrended geopotential heights with domain average linear trend removed from each grid has demonstrated no significant impact on the sign and magnitude of trends in CP occurences (Zhou et al 2020a).

The daily circulation fields from reanalysis and model simulations with and without anthropogenic forcing are projected onto the SOM and ascribed to the single best matching node (CP) according to pattern similarity measured by the Euclidean distance (Huth 2000, Belleflamme et al 2013, Rousi et al 2021), which easily allows characterization of the occurrence frequency of each CP (Reusch et al 2005, 2007). Simulations from HadGEM3-A-N216 are interpolated to the 2° × 2° grid of the ERA reanalysis prior to CP classification. CP frequencies in the HadGEM3-A-N216 simulations are scaled up by 92/90 as model months are 30 d giving 90 summer days. To measure the possible influence of resolution on clustering and the robustness of the clusters in the model simulations, the variance ratio defined as the ratio of the average inter-cluster squared distance to the mean intra-cluster variance is used (Fabiano et al 2020, Rousi et al 2021), which is calculated as:

$$\begin{equation*}{R_{\text{v}}} = \frac{{2 \times \sum \nolimits _{{C_{mn}}}{{\left| {{C_m} - {C_n}} \right|}^2}}}{{\left( {k - 1} \right) \times \sum \nolimits_m{\sigma _m}^2}}\end{equation*}$$

where the summation in the numerator is done without repetition over all combinations of the clusters, k is the number of clusters and σ_m ² is the intra-cluster variance. C_m and C_n are the respective cluster centroids represented here by the composite mean of daily circulations assigned to that node. Larger variance ratio indicates better separation between clusters with stronger intra-type compactness. As projecting daily circulation simulations with the original high resolution to the SOM derived from coarse resolution data is not practical, we assume for the same day the high-resolution circulation could be classified to the same type as the low-resolution one has been assigned to. We compare the variance ratios for the clusters derived from model simulations before and after interpolation to validate the above assumption and examine the influence of resolution on CP classification. If the more detailed circulation information brought from high-resolution simulations lowers the variance ratio a lot, it means the high-resolution simulations are much less clustered and the CP typing is sensitive to resolution. Nevertheless, for the multimember simulations from HadGEM3-A-N216, variance ratios show very slight decrease in the high-resolution clusters, which demonstrates that the resolution has indiscernible effect on the clustering of CPs (figure S2). Moreover, the comparable variance ratios between clusters from ERA reanalysis and the ALL forcing ensemble show that the clustering of CPs in the model simulations is as robust as that in the reanalysis (figure S2).

The linkage between the frequency changes in CPs and the precipitation response over eastern China are investigated through the composite observational precipitation anomalies calculated by averaging over the same days as those assigned to each CP. Considering the north–south dipole pattern of precipitation trends over eastern China (divided by 34° N), CPs are identified as neutral, wet and dry types for the north and south respectively, when the domain-averaged composite precipitation anomalies under each CP is within ±0.5σ, exceeding +0.5σ and below −0.5σ (σ is the standard deviation of domain-averaged precipitation during 1961–2010; figure S3). Assuming the composite precipitation anomaly under each CP is temporally invariant, the dynamic contribution to trends in summer total precipitation anomalies can be calculated as the product of the composite precipitation anomaly and trends in the frequency of CPs (Horton et al 2015, Zhou et al 2020a). Our analyses focus on this dynamically induced precipitation change, which gives the precipitation trends purely due to changes in CP occurrence.

Risk ratio (RR)—a widely used metric in probabilistic event attribution calculated as the ratio between the two probability estimates in the ALL and NAT forcing (RR= P(ALL)/P(NAT))—is used here to measure the likelihood of human-induced changes (Stott et al 2016). By fitting probability density functions (PDFs) to the histogram of the variable considered, the probability of an event defined by the variable exceeding (or not exceeding) a prespecified threshold is calculated as the area under the PDF curve above (below) this threshold. All of the fitted PDFs are derived from a rectangular kernel density estimation, which applies weak smoothing and retains most of the information in the PDFs estimated from histograms (Fischer et al 2013). Unlike a histogram, which estimates PDF by assigning equal density value to every data point within discrete bins, the kernel density estimation fits individual probability density curve for each data value and then sums all these small PDFs together to obtain a smooth and continuous PDF curve. Prior to fitting PDFs, frequencies of CPs in both ERA reanalysis and each member from model simulations are normalized by their respective means and inter-annual standard deviations obtained from 1961 to 2010 for robust comparison between them. The NAT forcing simulations are normalized using the same statistics from ALL forcing experiment. The 95% CI is derived from the 1000 times resampling using bootstrapping (Efron and Tibshirani 1998). Each bootstrap creates a set of new data with the same length as the original by randomly resampling with replacement. Statistics of interest such as RR, return period and the corresponding threshold are computed 1000 times via the PDFs generated from the 1000 resamples. The 2.5th and 97.5th percentiles are then used to indicate the 95% CI.

On the topologically ordered SOM plane, similar CPs are mapped onto adjacent nodes, while very distinct CPs are placed far apart (figure S1). The 20 CPs are named with the combination of column label denoted by capital letters (A to D) and the row label indicated by numbers (1–5). Patterns A1 and A2 in the top-left feature intense and westward extended western Pacific subtropical high (WPSH), in contrast to patterns D4 and D5 in the bottom-right with widespread negative anomalies at low latitudes and very weak eastward-located WPSH. Patterns A5 and B5 in the lower-left characterize pronounced deep troughs at middle-high latitudes and southward displaced WPSH, while patterns C1, D1 and D2 in the top-right present very weak and northward shifted WPSH influenced by activities of tropical cyclones. Among the patterns in the lower-left portion, patterns A3, A4, B3 are characterized by East Asia-Pacific (EAP) teleconnection at positive phase and double blockings at high latitudes. In contrast, patterns C2, D3 and D4 on the right half of SOM feature the northward displaced WPSH or the negative phase of EAP teleconnection. Other intermediate CPs sharing similarities to the above-mentioned CPs such as the EAP-like patterns (B4, C3, C4 and C5) and northward-extended WPSH patterns (B1 and B2) are also identified (figure S1).

With different configurations of the intensity and positioning of synoptic systems, the 20 CPs produce diverse precipitation pattern over eastern China. Some CPs, which are assigned to distinct attributes, produce an opposite precipitation response for the south and north regions (figure S3). Among these CPs, patterns A3, A4 and B3 favors SFND, while patterns C2, D3 and D4 tend to produce precipitation pattern opposite to SFND. Changes in occurrence of these CPs will also bring a correspondingly north–south converse pattern of dynamic contribution (figures 1(a) and (c)). However, the dynamic effect induced by changes in the frequency of these CPs contributes positively both to the wetting in the south and drying in the north. This implies that the three CPs facilitating SFND are occurring more often and that concurrently the three CPs unfavorable to SFND are decreasing in frequency (also corroborated in figure 2). Noteworthily, amongst the CPs contributing positively to both the south wetting and north drying, the above-mentioned six CPs (except pattern B3 for the north region) have provided the largest positive dynamic contributions (hatched bars in figures 1(a) and (c)). Of the total positive dynamic contribution to the south wetting and north drying, the six CPs account for 84% and 68% respectively. In addition, the wetting-drying contrasts between the south and north region dynamically caused by the frequency changes of these six CPs are the strongest, compared to the other CPs (figure S4). In other word, the altering occurrences of these CPs create the most favorable dynamical conditions for SFND. Therefore, changes in these CPs play a key role in the contrasting precipitation trends over eastern China. Whether the frequency changes in these CPs are a manifestation of anthropogenic climate change is further explored.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Since the confidence of subsequent D&A analysis is contingent on the capabilities of the models employed, an evaluation of HadGEM3-A-N216's performance in simulating the occurrence of various CPs during 1961–2013 is performed. The PDFs for the frequency of the selected CPs constructed with ERA analysis and ALL forcing simulations over 1961–2013 show no visible difference (figures 1(b) and (d)). The two-sample Kolmogorov–Smirnov (K–S) test (Massey 1951, Wilks 2006) does not reject the null hypothesis that the frequency of the selected CPs favorable (P = 0.17) or unfavorable to SFND (P = 0.59) in the reanalysis and ALL forcing experiment come from populations with the same distribution. The interannual variability of the frequency of the selected CPs facilitating or counteracting SFND in ERA reanalysis is generally encompassed by the ensemble range of HadGEM3-A-N216 (figures 2(a) and (c)). For the ERA reanalysis, the standard deviations of the frequency of the selected CPs facilitating and counteracting SFND are 9.31 and 12.57 d, while for the ALL forcing multimember ensemble the median standard deviations are 8.38 and 11.8 d respectively. Overall, the interannual variabilities of the frequencies of the selected CPs from ALL forcing simulations resemble those in ERA reanalysis to a large extent. HadGEM3-A-N216 reproduces the trend in the occurrence of the CPs facilitating SFND reasonably well, but tends to slightly underestimate the negative trend of CPs counteracting SFND (figures 2(b) and (d)). For the PDFs and long-term change of each individual CP, HadGEM3-A-N216 also performs well for the most part (figures S5 and S6). These evidences demonstrate that HadGEM3-A-N216 is adequate for examining the trend in the frequency of CPs.

The occurrence of CPs facilitating SFND has increased by 1.55 and 1.87 d per 10 years (d 10 yr⁻¹), respectively, in ERA reanalysis and the ensemble mean of ALL forcing during 1961–2013, statistically significant at the 95% confidence level (figure 2(b)). In contrast, the ensemble mean of the NAT forcing simulations exhibits a weak negative trend in the occurrence of CPs facilitating SFND. Similarly, compared with the significant negative trends in ERA reanalysis (−4.65 d 10 yr⁻¹) and the ensemble mean of ALL forcing (−3.62 d 10 yr⁻¹), trends in the CPs counteracting SFND from NAT forcing simulations are weak and of opposite sign (figure 2(d)). Although the interannual variability driven by only natural forcing generally keeps pace with that in ERA reanalysis and the ALL forcing experiment, the significant trends in frequency of the selected CPs would not occur without anthropogenic influence. This is also validated by the distribution shifts between the recent 20 year period of 1994–2013 and the earlier 20 year period of 1961–1980 in the model experiments with and without anthropogenic forcing (figure S7). Inconsistent with the shift of entire distributions in ERA reanalysis and the ALL forcing experiment to higher (lower) values for the CPs favorable (unfavorable) to SFND, distributions in the naturally forced experiment stay statistically unchanged at the 95% confidence level (P > 0.05 based on the K–S test). In short, the above evidences confirm that anthropogenic influence is clearly detectable in the long-term frequency changes of CPs favorable and unfavorable to SFND.

To address the probability change in the occurrence of CPs, the distribution shift from the counterfactual natural to the factual world in the recent 20 year period (1994–2013) is explored. Relative to the natural forcing experiment, PDFs of ALL forcing shift to the positive and negative sides, respectively, for CPs favorable or unfavorable to SFND (figures 3(a) and (c)). This signifies that anthropogenic influence has increased the probability of favorable CPs (figure 3(a)) but decreased the likelihood of unfavorable CPs (figure 3(c)). Quantitatively speaking, for the 90th percentile threshold defined from the NAT experiment (one-in-ten year event), the occurrence of CPs facilitating SFND has increased about twofold (RR = 1.89, 95% confidence interval (CI): 1.30–2.81) while CPs counteracting SFND have become 7.75 times less frequent due to anthropogenic influence (RR = 0.129, 95% CI: 0.026–0.28). Human influence has shortened the return period of the CPs facilitating SFND at the natural 90th-percentile threshold to 4.41 years (CI: 3.57–5.61; figure 3(b)). In contrast, the return period of the CPs counteracting SFND which occur once a decade in the natural world has been lengthened to 75 years (CI: 37.5–300; figure 3(d)). Thus, from a probabilistic perspective, anthropogenic influence has played a critical role in increasing the likelihood of CPs facilitating SFND while decreasing that of CPs unfavorable to SFND.

Zoom In Zoom Out Reset image size

A least-squares multiple linear regression analysis is performed to explore the link between the frequency of CPs and summer precipitation for each station. The frequencies of all the 20 CPs except patterns A1 and A2 are used as predictors. Due to the invariant total frequencies of the 20 CPs (92 d in summer), the frequency of each CP is fully correlated with the total frequencies of the remaining CPs. Furthermore, the weak dynamic contributions caused by the frequency changes in patterns A1 and A2 for both the north and south region indicate a weak link between SFND and patterns A1 and A2 (figures 1(a) and (c)). To avoid multicollinearity caused by repetitive information from predictors whilst minimizing the loss of information, the frequency of patterns A1 and A2 are eliminated from the predictors.

The reconstructed trends using CP frequencies from ERA reanalysis and the ALL forcing experiment give a good representation of the converse north–south precipitation trends over eastern China during 1961–2013 (figures 4(a) and S8(a), S8(c), S8(d)). The underestimation of strong positive trends and overestimation of several negative trends from ALL forcing indicate that other processes such as thermodynamics could also be at play (figures 4(a) and (b)). Despite this, changes in CP occurrences could explain a large part of the trends in precipitation over eastern China during 1961–2013. In comparison, the precipitation regressed from naturally forced CP frequencies is characterized by very weak trends centered around zero (figures 4(a) and S8(b)). The single-peak PDF of precipitation trends reconstructed from the NAT forcing experiment is statistically distinguishable from other two-peak PDFs of observed and regressed precipitation trends via ERA reanalysis and the ALL forcing experiment (P < 0.05, based on the K–S test; figure 4(b)). The mismatch in precipitation trends reconstructed from CP frequencies with and without anthropogenic forcing has two main implications. First, the occurrence of CPs in the factual world driven by ALL forcing is different from that in the counterfactual natural world, which further confirms anthropogenic influence on the changing prevalence of the clustered CPs. Second, the north–south dipole pattern was evident when considering CP frequencies from the ALL forcing experiment, but not from CPs frequencies driven by purely natural forcing. Thus, anthropogenic forcing has played a critical role in the contrasting north–south precipitation trends over eastern China during 1961–2013.

Zoom In Zoom Out Reset image size

Changes in the occurrence of CPs associated with precipitation anomalies characterized by either SFND or the opposite pattern reveal a large part of the positive dynamic contribution to the wetting in the south and drying in the north over eastern China during 1961–2013. Both the trends and long-term distribution shift in the frequency of these CPs in ERA reanalysis are generally consistent with the model experiments that include anthropogenic forcing, but are inconsistent with the naturally forced experiments. This demonstrates that the increasing trend in CPs favoring SFND and the decreasing trend in CPs disfavoring SFND are attributable to human influence. From the perspective of probabilistic event attribution, anthropogenic influence has increased the likelihood of the occurrence of CPs facilitating SFND, but decreased that of CPs unfavorable to SFND.

The regression analysis shows that CP frequencies driven by both anthropogenic and natural forcing succeed in reproducing the dipole pattern of precipitation trends over eastern China during 1961–2013. In contrast, the precipitation trends reconstructed from naturally forced CP frequencies alone appear to be very weak and unable to capture the north–south contrast. The difference between the reconstructed precipitation trends not only confirms the differential responses of CPs to forcing with and without human influence, but also signifies the critical role of human influence in shaping the contrasting north–south precipitation trends over eastern China during 1961–2013.

This study was jointly supported by the National Key R&D Program of China (2018YFC1507700), the National Natural Science Foundation of China (41905082) and the Basic Research to Operation Funds of the Chinese Academy of Meteorological Sciences (2019Y009). SFBT & FCL were funded by the UK–China Research & Innovation Partnership Fund through the Met Office Climate Science for Service Partnership (CSSP) China as part of the Newton Fund.

The data that support the findings of this study are openly available. The links are listed below. Observed precipitation data is available at www.nmic.cn/site/index.html. ERA-Interim and ERA-40 reanalyses are available at https://apps.ecmwf.int/datasets/, simulations with and without anthropogenic influence from HadGEM3-A-N216 at https://data.ceda.acuk/badc/eucleia/data/EUCLEIA/output/MOHC/HadGEM3-A-N216.