Interdecadal aridity variations in Central Asia during 1950–2016 regulated by oceanic conditions under the background of global warming

Abstract

As a typical inland arid and semiarid region, Central Asia (CA) is vulnerable to the forced global warming (FGW) due to anthropogenic activity. Aiming at the interdecadal variation of the FGW-forced aridity pattern (FAP) in CA, we try to extract the associated oceanic and atmospheric modes by analyzing observations, reanalysis data and multi-model simulations during 1950–2016. The FAP in CA features a tripolar pattern with wetting–drying-wetting responses arranging from southeast to northwest and shows strong interdecadal-to-interannual amplitude variations. It is found that the sea surface temperature (SST) in the tropical South Atlantic (TSA) well correlates with the amplitude variation of FAP on interdecadal time scale, possibly through modulating the interannual SST modes characterized by the North Atlantic horseshoe-like dipole (NAHD) and the El Ninõ and South Oscillation (ENSO). Corresponding to the enhancing FAP from the middle 1970s to early 2000s, the TSA-modulated NAHD and ENSO, together with the Pacific Decadal Oscillation-modulated Indian Ocean Dipole-like mode, show connections with an Eurasian middle-latitude wave train coupled with the North Arctic Oscillation and equatorial low, which favors the moisture transport to strengthen the tripolar FAP by forming a local circulation dipole with positive/negative anomaly over the northwest/southeast CA. But after the early 2000s, the increasing FAP amplitude is decelerated due to the interdecadal decline of TSA accompanied by the weakened/reversed relationship between FAP and the NAHD/ENSO. Because of the corresponding breakdown of the wave train, the favorable local circulation is unavailable to support the sustained enhancement of FAP. Therefore, the multiscale coupling between the above oceanic and atmospheric modes is significantly related to the characteristic of stage of the forced aridity change in CA under the background of global warming.

Appendix

1.1 Sensitivities to data and analyzed time period

The analysis based on different data and over different study period are carried out to verify the treatment of this study. The Dai’s scPDSI obtained from https://rda.ucar.edu/datasets/ds299.0 is a widely used dataset for drought research. Figure S1 presents spatial mean of the CA scPDSI, the FGW-induced spatial-mean and the detrended (residual) time series based on CRU scPDSI and Dai’s scPDSI over 1920–2016 and 1950–2016. From Fig. S1 and Fig. 2a, the two spatial means (black lines with diamonds), respectively, derived from CRU and Dai’s scPDSI have the correlation of 0.87 (p = 0.01) during 1920–2016 (Figs. S1a, b,) and 0.90 (p = 0.01) during 1950–2016 (Fig. S1c and Fig. 2a). The detrended spatial means (blue lines with squares) have even high correlation up to 0.90 (p = 0.01) during 1920–2016 and 0.94 (p = 0.01) during 1950–2016. The relatively higher consistence between the two data over 1950–2016 may be partly due to the paucity of Dai’s data in CA before 1950 as shown by Fig. S3b. It should also be noted that Dai’s scPDSI had a weaker mean FGW response but a stronger FGW-induced trend in spatial mean since mid-1970s, as shown by the red lines with filled circles in Figs. S1b, c. This may be caused by the different calibration intervals used by these two datasets (van der Schrier et al. 2013). But this spatial-mean response will be ruled out from the detrended ${P}_{t}$ that is finally used to extract the forced pattern-related climate modes. According to Eqs. (9), (10) and (11), only the detrended spatial mean series (${\stackrel{-}{S}}_{var}$), i.e., the blue lines with squares in Fig. S1b, c, has contribution to ${P}_{t,var}$. Therefore, the high consistence (r = 0.94) of ${\stackrel{-}{S}}_{var}$ between the two scPDSI datasets suggests the difference between the two datasets has tiny impact on the spatial-mean response to FGW.

On the other hand, different lengths of the study time period may also produce discrepancy in the results. We examined the sensitivity of detrending by T_CMIP to the detrending time period. The blue line with squares in Fig. 2a shows the detrended spatial-mean series over 1950–2016, which is detrended by using T_CMIP over 1950–2016. In order to evaluate the impact of the detrending period, we also derive the post-1950 segment of the detrended 1920–2016 series in Fig. S1a, which is detrended by T_CMIP over longer time period (1920–2016). By comparing these two detrended series, we can find that they are almost perfectly coincidence with correlation of 0.9998 at p = 0.01. Therefore, for the spatial-mean part of the detrended ${P}_{t}$ (Eq. (10)), the study time period, using 1920–2016 or 1950–2016, as well as the data choice between CRU and Dai’s scPDSI, only has tiny influence on the extraction of the FGW-related climate modes. Considering the relatively higher spatial resolution of CRU data and the time coverage (from 1950) of the atmospheric reanalysis data, this study chooses the CRU scPDSI over 1950–2016 as the main aridity data.

According to Eq. (1) and (5–7, 9–11), the projection time series ${P}_{t}$ and its main components are plotted in Fig. S2. In this figure, the components of the ${P}_{t}$ are calculated independently according to Eqs. (6, 10, 11). In the practical calculation, the sum of the calculated ${P}_{t, mean}$, ${\stackrel{-}{P}}_{t,var}$ and ${P}_{t,var}^{^{\prime}}$ exactly equals to the ${P}_{t}$ (Eq. (1)), which verifies the correctness of the decomposition by Eq. (5). This is also the case for the decomposition for the detrended ${P}_{t}$ (Eq. (9)). Furthermore, as shown by Fig. S2, ICV-induced spatially-varying component ${P}_{t,var}^{^{\prime}}$ dominates the variance of ${P}_{t,var}$ with extremely high correlation with the detrended ${P}_{t}$ up to 0.94 passing 99% significance test. For the spatial-mean ICV-induced component, ${\stackrel{-}{P}}_{t,var}$, which only explains a small part of variation of ${P}_{t,var}$, the correlation between ${\stackrel{-}{P}}_{t,var}$ and ${P}_{t,var}$ is 0.48 at p = 0.01. So, the variation of ${P}_{t,var}$ is mainly determined by ${P}_{t,var}^{^{\prime}}$.

We also examine the influence of the study time periods on ${P}_{t}$ by comparing two series ${P}_{t},$ one constructed based on the 1950–2016 CRU scPDSI (solid purple line in Fig. S2) and the other also over 1950–2016 but derived from the ${P}_{t}$ constructed based on 1920–2016 CRU scPDSI (thick black line in Fig. S2). The correlation between these two ${P}_{t}$ is up to 0.98. We also calculate the monthly ${P}_{t}$ and find similar correlation of 0.97 between them (figure not shown). The detrended series of the two ${P}_{t}$ also have high correlation of 0.9 for yearly series and 0.88 for monthly series. These high similarities suggest that the time periods discussed here have tiny impact on ${P}_{t}$ construction.

Figure S3 further evaluates the influence of study period and data on regression patterns on ${T}_{CMIP}$. Figure S3a displays the regression pattern of the CA scPDSI on ${T}_{CMIP}$ during 1920–2016. The pattern in Fig. S3a bears a striking assemblance to the regression pattern during 1950–2016 shown in Fig. 2d. Due to the paucity of Dai’s scPDSI in CA before 1950 (Fig. S3b), we only compare the 1950–2016 forced pattern of Dai’s scPDSI (Fig. S3c) with that of CRU scPDSI (Fig. 2d). Although Dai’s scPDSI (Fig. S3c) has relatively lower (2.5° × 2.5°) spatial resolution than CRU data, the regression patterns obtained from the two data are still qualitatively similar. They are both characterized by the wetting trend in southeast CA and drying trend in southwest CA. Because the comparison between the spatial patterns in Fig. S3c and Fig. 2a can only provide a qualitative evaluation, Fig. S4 further shows the amplitude indices of the two forced patterns obtained from CRU and Dai’s data. In Fig. S4, the correlations of ${P}_{t,var}$, ${\stackrel{-}{P}}_{t,var}$ and ${P}_{t,var}^{^{\prime}}$ based on the two datasets are 0.61, 0.94 and 0.65, which all pass p = 0.01 significance tests. Considering the phase difference of ICV existing between different data or ‘experiments’, those are very high correlations for the ICV-related series.

The above similarities of ${P}_{t}$ between the two scPDSI data are also reflected by the climate modes extracted based on them (Fig. S5 and S6). As mentioned above, the time series ${P}_{t,var}$ was used to extract the forced pattern-related climate modes. Therefore, the atmospheric and oceanic variability associated with the two components of ${P}_{t,var}$ (Eq. (9)) reflects the forced pattern-related modes from the aspects of spatial-mean and spatially-varying forced variations of the CA scPDSI. Figure S5 and S6, respectively, demonstrate the correlations of 500 hPa geopotential height (Z500) and SST with monthly ${\stackrel{-}{P}}_{t,var}$ and ${P}_{t,var}^{^{\prime}}$ series. For the ${P}_{t}$ series extracted from the CRU data and Dai’s data, the associated circulation (Fig.S5) and SST variation (Fig. S6) having the similar patterns suggests the decomposition of ${P}_{t,var}$ is insensitive to dataset.

The ${\stackrel{-}{P}}_{t,var}$-related Z500 patterns (Fig. S5a, b) and SST patterns (Fig. S6a, b) show high similarities. The Z500 patterns (Fig. S5a, b) feature a blocking-like circulation system over Europe with two negative anomalies over North Atlantic and CA. The SST patterns related to ${\stackrel{-}{P}}_{t,var}$ (Fig. S6a, b) show almost all signatures of the main oceanic modes, such as IPO, Indian Ocean warming and Pan-Atlantic mode. The two scPDSI datasets examined here can be seen as two experiments, which have consistent FGW response with the out-of-phase ICV-induced response superimposing on it. For the spatial mean series of the CA scPDSI, high correlation (r = 0.9 at p = 0.01) exists between CRU scPDSI and Dai’s scPDSI. After removing the FGW-related part, the detrended spatial means (${\stackrel{-}{S}}_{var}\left(t\right)$) in Eq. (8)) extracted from the two scPDSI datasets also have high correlation of 0.93 (p = 0.01), which means the interdecadal-to-interannual modes exerts consistent influence on the spatial-mean variation of the CA scPDSI. Returning to definition of ${\stackrel{-}{P}}_{t,var}$ (Eq. (9)), ${\stackrel{-}{P}}_{t,var}$ is the product of a time-invariant coefficient ($\frac{{\sum }_{\lambda }{\sum }_{\theta }{S}_{FGW}\left(\lambda ,\theta \right)cos\theta }{{\sum }_{\lambda }{\sum }_{\theta }{S}_{FGW}^{2}\left(\lambda ,\theta \right)cos\theta }$) and ${\stackrel{-}{S}}_{var}\left(t\right)$. That means the high consistence of ${\stackrel{-}{S}}_{var}\left(t\right)$ between the two datasets will necessarily lead to the same correlation between ${\stackrel{-}{P}}_{t,var}$ series obtained from the two datasets.

For the other ${P}_{t,var}$ component, ${P}_{t,var}^{^{\prime}}$, although the CRU scPDSI and Dai’s scPDSI also show significant correlation of 0.64 (p = 0.01), the ${P}_{t,var}^{^{\prime}}$-related Z500 (Fig. S5c, d) and SST pattern (Fig. S6c, d) show relatively larger difference than ${\stackrel{-}{P}}_{t,var}$-related fields (Fig. S5a, b and Fig. S6a, b). Aside from the amplitude difference, the spatial structures of the ${P}_{t,var}^{^{\prime}}$-related circulation and SST patterns are generally similar between the two scPDSI datasets. For the ${P}_{t,var}^{^{\prime}}$-related Z500 patterns (Fig. S5c, d), both of the two datasets show midlatitude wave trains with similar spatial structure in generally same phase except that the Z500 field related to ${P}_{t,var}^{^{\prime}}$ from Dai’s scPDSI (Fig.S5d) in higher amplitude than CRU scPDSI (Fig. S5c). It is also the case with SST (Fig. S6c, d) except that the ${P}_{t,var}^{^{\prime}}$-related SST pattern from Dai’s scPDSI (Fig. S6d) has relatively weaker correlation coefficients in Indian Ocean and Atlantic. In general, the consistence between the two datasets on the ${P}_{t,var}$-related atmospheric and oceanic modes supports the main conclusions of this work.

On the other hand, the time period of EOF analysis may also influence the results based on it. In particualr, too short time period may not well isolated the decdal/interdecadal climate mode. For the sensitivity of SST mode to analysis time period, we further performed EOF analysis based on the monthly SST over 1920–2016 in the same regions as Fig. 5, i.e., tropical Atlantic (ATL_TP), extra-tropical North Atlantic (NAT_ET), Pacific (PAC) and tropical Indian Ocean PC2 (IO_TP). It should be noted that the SST field has been detrended by T_CMIP before the EOF analysis and the detrending time period for SST field is set to 1920–2015 for both the 1950–2016 and 1920–2016 EOF analysis. This long-period (97-year) detrending can avoid aliasing the FGW-induced variability with the IPO- and AMO-induced variations.

Similar to the previous sensitive analysis, the PCs of the SST EOF analysis over 1920–2016 data are truncated to retain the post-1950 segments, which are then used to make comparisons with the PCs obtained directly from the EOF analysis over 1950–2016 (Fig. 5). Correlative analysis between these two categories of SST PCs shows the correlations of 0.97 for ATL_TP,1, 0.98 for NAT_ET,1, 0.999 for PAC₁, and 0.999 for IO_TP,2, which suggest nearly perfect coincidence between the temporal variations of the two sets of EOFs extracteds over two different time periods.

The above analysis suggests the SST EOF analysis executed in this study is also insensitive to the time-period choice between 1920–2016 and 1950–2016. This is partly because the variability linearly forced by FGW has been removed before the EOF analysis. The other reason may be the spatial extent of the EOF analysis on SST is too regional to well isolate the signal of the AMO with period up to 80 years. In fact, we also execute the EOF analysis over 1920–2016 in larger spatial extent in Atlantic, such as North Atlantic (0–70˚N) and Pan-Atlantic (20˚S-70˚N and 40˚S-70˚N). The AMO signal could be well isolated in these larger-extent analysis, but there is no significant correlation found between the AMO mode and ${P}_{t,var}$. As a result of it, we only present the more regional analysis that can extract ${P}_{t,var}$-related SST modes, as those shown in Fig. 5. From the above analysis, the EOF analysis executed in this study does not intend to isolate AMO signal, and the resultant mode is therefore not strongly impacted by the length of the analyzed period.