1. Introduction
On the global scale, droughts and concurrent temperature anomalies (especially warmer than normal) have occurred frequently in recent years (Di Capua et al. 2021; Erfanian et al. 2017; Hoerling et al. 2014; Liu and Zhou 2021; Yiou et al. 2020; Zhang et al. 2021a) and are projected to increase in the future under global surface warming (AghaKouchak et al. 2014; Wang and Chen 2014; Zscheischler et al. 2018). Physical mechanisms behind droughts have been investigated based on the direct or remote influences of four large-scale triggers:
-
Atmospheric circulations and patterns. For example, propagating Rossby waves originating in the west Pacific lead to upper-level enhanced anticyclones during the 1988 Great Plains drought (Chen and Newman 1998), while the descending branch of the Hadley circulation is responsible for amplified subsidence and associated long-term precipitation deficits during the 2011 eastern China spring–summer drought (Jin et al. 2013).
-
Land–atmosphere coupling. Both local and downstream effects of land–atmosphere coupling can be observed in different geographic regions and seasons, e.g., the 2010 western Russia summer hot drought (Di Capua et al. 2021; Hauser et al. 2016; Liu et al. 2020), the 2017 persistent northeastern China drought (Zeng and Yuan 2021; Zeng et al. 2019), and the late spring 2011 Texas drought (Fernando et al. 2016). Intermediate linkage of land–air feedback to droughts occurs through local anticyclonic circulation anomalies or an enhanced midtropospheric high.
-
Tropical oceanic thermal conditions (Erfanian et al. 2017; Xu et al. 2020; Zhang et al. 2021a). As spatially concurrent extreme events, the recent 2019 record-breaking summer–autumn droughts over pan-Australia and eastern China can be attributed to the super positive Indian Ocean dipole and strong central Pacific El Niño events (Xu et al. 2020; Zhang et al. 2021a). The enhanced descending motion directly responsible for these two droughts is essentially modulated by the oceanic thermal condition.
-
Marine heatwaves adjacent to coastal areas (Rodrigues et al. 2019; Shi et al. 2021). The 2013/14 austral summer drought in eastern South America is typically influenced by the westward extension of anticyclonic anomalies over concurrent extreme marine heatwaves in the adjacent western South Atlantic (Coelho et al. 2016; Rodrigues et al. 2019).
Linking these large-scale triggers to these droughts, local-scale atmospheric dynamics (e.g., vertical motion, horizontal anticyclones/cyclones) are common bridges across most drought-inducing climatic diagnoses. Persistent anticyclonic abnormalities favor near-surface warming, especially over the northern midlatitudes (Barriopedro et al. 2011; Schubert et al. 2014; Wehrli et al. 2019), through decreased cloud cover, increased shortwave radiation, and adiabatic heating triggered by intense dynamic subsidence (Zschenderlein et al. 2019). Meanwhile, persistent enhanced vertical subsidence usually leads to long-term precipitation deficits and drought occurrence, as expected. Besides considerable routine qualitative analysis of typical dynamic fields (e.g., 500-hPa vertical motion and 200-/850-hPa divergent/convergent motion) in drought-related climatic diagnoses (Wang and He 2015; Wang et al. 2017), some quantitative drought analyses based on atmospheric dynamic variables have been conducted (Liu and Zhou 2021; Liu et al. 2017). Regarding the recent 2019 hot autumn drought in eastern China, we found that the evolution of atmospheric dynamics was highly correlated with the development of droughts and heatwaves (Liu and Zhou 2021).
The dynamically based drought-inducing roles mentioned above motivate us to reconstruct droughts using atmospheric dynamic anomalies quantitatively. On the one hand, the attempt at reconstruction might check the possibility of developing reliable tools for drought monitoring and early warning from dynamic perspectives. Similar efforts have recently been made to advance seasonal-to-subseasonal (S2S) forecasts of surface extreme events connecting upper atmospheric circulation patterns (White et al. 2022). On the other hand, quantitative reconstruction could also help quantify atmospheric dynamic contributions at specific pressure levels (e.g., vertical motion at 400 hPa) and may help discover some potential dynamic roles previously ignored. In addition, feasible dynamically based reconstruction performance might provide new insight from dynamic perspectives, as previous studies have focused on water vapor–related drought mechanisms in the hydrological community (Herrera-Estrada et al. 2019; Ramakrishna et al. 2017; Zhang et al. 2021b).
The focus of this study is dynamically based hydrometeorological reconstruction under drought situations and associated quantitative analysis of predictor contributions. On the grid scale, the predictands are hydrometeorological elements (i.e., precipitation, near-surface air temperature, surface soil moisture, and actual evaporation) at the drought peak time. The predictors are the atmospheric dynamic variables (i.e., vertical velocity, relative vorticity, and horizontal divergence) in the troposphere. The tool for statistically based model construction is the so-called XGBoost (extreme gradient boosting; Chen and Guestrin 2016) ensemble learning technique, and predictor contributions are quantified from interpretable perspectives.
Accordingly, this paper is divided into the following sections. First, we employ 12 representative severe drought events for comprehensive hydrometeorological reconstruction. In this regard, we adopt two different reconstruction schemes (i.e., statistically preexisting dynamic–hydrometeorological relationships and interannual variability). Also, we assess reconstruction performance based on associated predictor contributions and temporal evolution of atmospheric dynamics. Subsequently, we implement global-scale quick drought reconstruction with interannual variability, which eventually highlights dynamic drought-inducing characteristics from interpretable perspectives.
2. Dataset and methods
a. Seasonal-scale standardized anomalies of atmospheric and hydrometeorological variables
On the global scale during 1980–2020, ERA-5 1-hourly 0.5° × 0.5° total precipitation and 6-hourly 0.5° × 0.5° 2-m air temperature (Hersbach et al. 2020), European Space Agency Climate Change Initiative (ESA-CCI) v5.3 daily 0.25° × 0.25° surface soil moisture (Dorigo et al. 2017; Gruber et al. 2017), and Global Land Evaporation Amsterdam Model (GLEAM) v3.5a daily 0.25° × 0.25° actual evaporation (Martens et al. 2017) are comprehensively employed and transformed to a daily time scale and a spatial resolution of 1° × 1° during 1981–2020. Subsequently, global gridded precipitation, near-surface air temperature, and actual evaporation datasets are computed using a daily running 90-day-mean window, whose values are located at the last day of the time window. For example, the 90-day-mean value on 20 March 1999, is temporally averaged over the period from 21 December 1998 to 20 March 1999. At the same time, surface soil moisture maintains the daily scale because of long-term memory. Finally, the global-scale gridded standardized precipitation anomaly index (SPAI), standardized temperature anomaly index (STI), standardized anomalies of daily surface soil moisture (SMsurf_SA), and standardized anomalies of actual evaporation (E_SA) are produced, which simply measure anomalies in units of standard deviation (Schumacher et al. 2019; Zscheischler et al. 2014) within the climatological 1981–2020 period.
Similarly, global-scale ERA5-based 27-level 6-hourly 0.5° × 0.5° relative vorticity (RV), horizontal divergence (DIV), and vertical velocity (OMEGA) from 100 to 1000 hPa during 1981–2020 are finally transformed to daily running 90-day-mean standardized anomaly-based variables with a spatial resolution of 1° × 1°. The 27 pressure levels correspond to three divisions (100–225 hPa with an interval of 25 hPa, 250–750 hPa with a break of 50 hPa, and 775–1000 hPa with a gap of 25 hPa).
b. Regional drought detection and associated spatiotemporal characteristics
We detect global severe drought events using an unsupervised machine learning method and subsequently identify the drought peak time and associated drought centroids.
1) Detection of global severe drought events using the DBSCAN algorithm
The so-called DBSCAN (density-based spatial clustering of applications with noise) algorithm is employed for drought detection based on three-dimensional point clustering; the algorithm parameters are provided in the Text S1 in the online supplemental material. The input data are included in a three-dimensional (longitude–latitude–time) 1981–2020 daily running seasonal-scale (90-day mean in practice) 1° × 1° SPAI dataset with a one-day time-step interval. The main steps are as follows:
-
The DBSCAN algorithm is performed over three-dimensional grid cells with SPAI no more than −1.5 for event identification globally. Each detected drought event is essentially a set of three-dimensional clustered points.
-
Regarding events preliminarily detected in step 1, only long-lifetime events with significant influence on continental areas are selected. Specifically, the lifetime needs to be beyond 60 days, and at the same time, the accumulated duration with continental coverage over 2 × 105 km2 needs to be over 30 days.
-
For every clustered grid cell at each time step of a given event selected in step 2, grids with moderate-grade SPAI values (from −1.5 to −1) within a three-grid-unit search distance are also merged; Text S2 provides relevant details.
Finally, we detect regional severe drought events globally, and the associated global-scale spatial distribution is shown in Fig. S21 of the online supplemental material. The dataset is stored on the linkage (see the ERA5 file on Google drive: https://drive.google.com/drive/folders/1LWLJTwCc1c3BWac97XhQE8czrZcU6NWK?usp=sharing).
2) Drought peak time and associated drought centroids
Regarding a given regional drought event, drought peak time is when the continental drought coverage reaches a maximum value during the lifetime, which is a unique critical reference time throughout the present study. For all the involved grid cells at the peak time, the geographic coordinates (longitude, latitude) of the centroid are computed with the SPAI-based weight considered. Only those detected drought events with centroids between 60°N and 50°S during 1981–2020 are of concern in this study, and their total number is 393.
c. XGBoost-based drought reconstruction schemes
XGBoost (Chen and Guestrin 2016) is a highly efficient decision-tree-based ensemble learning technique, recently applied in seasonal precipitation forecasting (Gibson et al. 2021), in driver identification for wildfires (Wang et al. 2021), and in crop yield prediction (Folberth et al. 2019; Shahhosseini et al. 2021). We employ it for drought reconstruction and quantitatively assess associated predictor contributions. The XGBoost model is implemented based on the Python-based Scikit-Learn API (see the web page https://xgboost.readthedocs.io/en/latest/python/python_api.html#module-xgboost.sklearn). We specify the regression model and the squared loss-based goal as the learning task and objective in this study.
On the grid scale, the input predictors are the standardized anomalies of 90-day-mean atmospheric dynamic variables (i.e., vertical velocity, relative vorticity, and horizontal divergence) at 24 pressure levels from 100 to 925 hPa, with a total of 72 predictors. Accordingly, the output predictands are standardized hydrometeorological elements (i.e., SPAI, STI, SMsurf_SA, and E_SA). For a given grid, samples for model training are extracted from grids within a 1.5-grid radius to make reconstructed spatial distributions smooth.
At the model training stage, boosting learning rate (i.e., learning_rate), maximum tree depth for base learners (i.e., max_depth), and numbers of gradient boosted trees (equivalent to numbers of boosting rounds, i.e., n_estimators) are selected for hyperparameter optimization. The optimization approach is the so-called grid search technique, which adopts the fivefold cross-validation scheme. In addition, reconstruction performance is evaluated utilizing the three traditional coefficients [i.e., anomaly correlation coefficient (ACC), root-mean-square error (RMSE), and coefficient of determination (R2)], whose details are provided in Text S3.
Twelve well-recognized extreme global drought events (Table 1) are typical examples of drought reconstruction and our understanding of dynamic roles. The reconstruction target is the spatial distribution of hydrometeorological anomalies at the drought peak time. The XGBoost-based concurrent dynamic–hydrometeorological relationship is modeled using two different schemes. Finally, we can obtain the gridscale hydrometeorological distributions at the drought peak time while inputting 72-predictor dynamic standardized anomalies (SAs) into the calibrated concurrent relationship mentioned above. Details of these two schemes are illustrated as follows.
Characteristics of the 12 severe drought events employed for comprehensive hydrometeorological reconstruction. STI (standardized temperature index) is the regionally averaged value over grid cells under drought situations at the drought peak time. The latitudinal position herein is the centroid-based latitude of all grid cells under drought situations at the peak time. A, L, T, and M in the attribution column correspond to “atmospheric circulations and patterns,” “land–atmosphere coupling,” “tropical oceanic thermal conditions,” and “marine heatwaves adjacent to coastal areas,” respectively.
1) Preexisting statistical dynamic–hydrometeorological relationships
Considering that seasonal-scale prolonged extreme drought events usually correspond to persistent and steady atmospheric dynamic anomalies, we hypothesize that dynamic–hydrometeorological relationships could potentially exist before drought peak times. We can calibrate the XGBoost-based concurrent dynamic–hydrometeorological relationships using data within a given time window of 30 days. In the present study, six experiments are conducted, and associated time windows are described using the days ahead of the drought peak time, i.e., [−30, 0], [−60, −30], [−90, −60], [−120, −90], [−150, −120], [−180, −150]. The hyperparameter arrays for max_depth, n_estimators, and learning_rate for optimization are [4, 6, 8, 10], [80, 120, 160, 200, 240], and [0.05, 0.1, 0.15, 0.2], to achieve good performance and save computation time.
2) Interannual variability
Regarding hydrometeorological reconstruction of a given event at the peak time (e.g., 26 April 2011), the samples for XGBoost-based model training are derived from the contemporary time (e.g., every 26 April during 1981–2020 except 2011). In practice, datasets from 5 days preceding and following the peak time are also employed for training to enlarge the sample size. Associated parameter arrays for learning_rate, max_depth, and n_estimators are [0.1], [2, 4, 6, 8,10], and [80, 120, 160, 200, 240] for hyperparameter optimization, respectively.
d. Quick reconstruction of regional droughts globally
Regional droughts with centroids between 60°N and 50°S are employed for quick hydrometeorological reconstruction. The focus is to roughly estimate dynamic roles in global-scale droughts based on reconstruction performance and variable contributions. The immediate reconstruction uses the interannual variability-based scheme mentioned above. However, the predictors, predictands, and hyperparameters change slightly. Four groups of input 100–1000-hPa predictors are separately employed for drought reconstruction, which are the “OMEGA-only” (27 predictors), “RV-only” (27 predictors), “DIV-only” (27 predictors), and “ALL (OMEGA+RV+DIV)” (81 predictors) groups. We choose the two essential variables (i.e., SPAI and STI) as the predictands, and learning_rate, max_depth, and n_estimators are set as 0.1, 6, and 160.
e. Feature importance–based variable contributions
Predictor contributions in drought reconstruction are measured using so-called feature importance (FI), since each predictor acts as one feature in the field of machine learning. We compute the gridscale XGBoost-based FI via the classic “gain” approach, which calculates the fractional contribution. Regarding a given grid cell, the total FI over all predictors is naturally equal to 1, and each FI value varies within 0–1. Further, we introduce the ratios of grid cells with FI greater than a critical value for intercomparison of unified forms, as the spatial extent of different drought events at the peak time varies. The workflow of generating FI-based diagrams is shown in Fig. 1 and is illustrated as follows.
Illustration of computing feature importance (FI)-based predictor contributions regarding drought reconstruction globally.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
1) Preexisting statistical dynamic–hydrometeorological relationships
A three-dimensional FI-based matrix is obtained based on one experiment [e.g., using data during the (−30, 0) time window]. Further, a self-defined R1 (the ratio of grid cells with FI greater than 0.05 over all grid cells) is calculated for each predictor of a given experiment. Eventually, a corresponding R1-based diagram (Fig. 1a) can be achieved. Higher R1 values (darker green) suggest more contributions from a given predictor to this reconstruction experiment.
2) Interannual variability
Based on the three-dimensional FI-based matrix of a given hydrometeorological reconstruction, R2 (the ratio of grid cells with a given FI range {e.g., [0, 0.05)} over all grid cells) is calculated; the FI ranges herein are tenfold, with an increment of 0.05 {i.e., [0, 0.05), [0.05, 0.1), [0.1, 0.15), …, [0.4, 0.45), and [0.45, 1]}. The sum of R2 over all of the FI range regarding a given predictor (e.g., OMEGA at 500 hPa) is naturally equal to 100%. Finally, we can obtain an R2-based diagram (Fig. 1b). Higher R2 values over the high FI range correspond to more contributions regarding a given predictor.
3) Quick reconstruction of regional droughts globally
Similar to that for interannual variability mentioned above, we compute R3 (ratios with FI greater than 0.1) for each predictor regarding a given drought event. Finally, we use R3 and associated centroid-based latitudes to generate an R3-based latitude-predictor diagram (Fig. 1c).
3. Typical drought reconstruction and predictor contributions
Two different schemes are adopted herein to reconstruct typically severe drought events and analyze associated predictor contributions. Two well-known droughts in Europe and South America are reproduced based on interannual variability. We try to illustrate the potential influences of latitudinal differences on predictor contributions. On the other hand, we reconstruct two recent North American droughts utilizing preexisting dynamic–hydrometeorological relationships. In this regard, we attempt to highlight two key points: 1) the feasibility of reconstruction using preceding statistical relationships and 2) how the asynchronous variation of different atmospheric dynamics can affect reconstruction performance.
a. Interannual variability
Hydrometeorological reconstruction of two typical droughts performs relatively well, e.g., the locations of extreme centers under drought situations (panels a2, c2, and e2) and the spatial patterns of temperature anomalies (b2) in Fig. 2. Together with similar reconstruction performance for other typically severe droughts worldwide (Figs. S1 and S2), this suggests the feasibility of dynamically based drought reconstruction on the interannual scale.
“ALL (OMEGA+RV+DIV)” scheme-based hydrometeorological reconstruction of the interannual variability and associated predictor contributions regarding the peak time of two regional droughts (the 2010 western Russia summer hot drought and the 2015/2016 South American drought).
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
Feature importance–based predictor contributions quantitatively reveal the specific roles of atmospheric dynamics. According to the relative performance of predictor contributions (Fig. 2, Figs. S3 and S4), as expected, the four hydrometeorological elements can be roughly divided into two groups (i.e., precipitation and soil moisture, and near-surface air temperature and actual evaporation). Across these 12 severe drought events with various types of large-scale drought-inducing triggers, midtropospheric vertical motion is dominant in a lack of precipitation and soil moisture. In contrast, the anticyclonic/cyclonic circulations without an inherent vertical location are mainly responsible for near-surface temperature anomalies. In addition, the predictor contributions regarding actual evaporation reconstruction are similar but not identical to those of temperature reconstruction, possibly due to the combined effects of both near-surface temperature and soil moisture.
More importantly, dynamic roles are seemingly modulated by latitudinal and geographic differences. While reconstructing drought situations in the 2010 European hot summer, anomalies of midlevel vertical motion and low-level anticyclonic/cyclonic circulations in the troposphere are crucial (Figs. 2a3 and 2c3). In comparison, the 2015/16 South American drought reconstruction is affected mainly by midlevel vertical motion but seldom influenced by anticyclonic/cyclonic circulations in the troposphere (Figs. 2e3). Regarding reconstruction of near-surface air temperature anomalies, relative vorticities contribute to the two specific droughts to a different degree, as expected (Figs. 2b3 and 2f3). However, vertical motion has no evident impact on near-surface temperature anomalies (Figs. 2b3). These contrasting features regarding predictor contributions also appear in the reconstruction of other drought events over Europe and tropical South America, e.g., the 2018 northern European drought (Figs. S3a1,a2), the 2014 Brazilian drought (Figs. S4c1,c2), and the 2010 Amazon drought (Figs. S4d1,d2).
“ALL (OMEGA+RV+DIV)” scheme-based hydrometeorological reconstruction using the (a1)–(a28) preexisting dynamic–hydrometeorological relationship and (b1)–(b4) associated predictor contributions regarding the peak time of the 2011 Texas drought.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
b. Preexisting dynamic–hydrometeorological relationships
Statistical dynamic–hydrometeorological relationships may be steady and persistent for several months before drought peak time. Based on this hypothesis, we utilize preexisting statistical relationships to reconstruct two recent North American severe droughts (Fig. 3 and Figs. 4a5–a28) and find relatively good performance. For example, both the spatial patterns and intensities of actual evaporation SAs can be roughly captured two months before the peak time (Fig. 3 and Fig. 4a16).
As in Fig. 3, but for the peak time of the 2012 Great Plains drought.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
Height–time evolution of the standardized anomalies (SAs) of atmospheric dynamics and hydrometeorological variables for the (a)–(d) 2011 Texas drought and the (e)–(h) 2012 Great Plains drought. SAs over all the grid cells under drought situations at the drought peak time are the targets used for computing area-averaged intensities and coverage proportions. In (a)–(c) and (e)–(g), shading corresponds to SA-based area-averaged intensities. At the same time, red (blue) contours represent coverage proportions with SAs beyond +1 (below −1) exceeding 30%, 50%, and 70%. Similarly, curves in (d) and (h) correspond to area-averaged SPAI and STI evolution. Terms such as “Peak-Nd” [gray lines in (d) and (h)] represent the nth day preceding the peak time, with an interval of 30 days. Thick yellow lines correspond to the drought peak time.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
Good reconstruction performance can be captured with different lead times regarding precipitation deficit and near-surface warming. The 2011 Texas drought situation (Fig. 3a1) can be continuously captured from the latest 1-month period (Fig. 3a5) to the 5-month-lead period (Fig. 3a25). At the same time, the earliest time associated with good reconstruction performance of near-surface warming seems to extend to the 1-month-lead period (Fig. 3a10). Similarly, the warmer-than-normal conditions and drought situation in the 2012 Great Plains drought can be detected based on the 3-month-lead (Fig. 4a18) and 1-month-lead (Fig. 4a9) periods, respectively.
Regarding a specific drought event, the reconstruction performance of drought and near-surface temperature anomalies seems asynchronous, which could be physically understood based on the temporally asynchronous evolution of atmospheric dynamics. In terms of the 2011 Texas drought, enhanced vertical subsidence appears 5 months ahead of the peak time (Fig. 5c), in association with the occurrence of the persistent drought condition (the black curve in Fig. 5d). Meanwhile, the amplified anticyclonic anomalies responsible for the near-surface warming (Fig. 5b) and the gray curve in Fig. 5d) occur 2 months preceding the peak time. In comparison, the anticyclonic anomalies appear before the obvious vertical subsidence during the 2012 Great Plains drought development (Figs. 5f,g) and are responsible for the earlier good reconstruction performance of near-surface warming rather than that of the drought situation. In addition, almost all dynamic predictors are involved in the statistically based reconstruction (Fig. 3 and Figs. 4b1–b4) from interpretable perspectives. In this situation, predictor contributions seemingly vary with rolling preceding time windows.
Evolutions of ACC and RMSE as a function of different preceding time windows are displayed in Fig. 6 to reflect the hydrometeorological reconstruction performance of the 12 typical drought events. Reconstruction performance gradually improves with decreasing lead times of the preceding time windows as expected. Also, it is interesting that preexisting statistical relationship–based reconstruction of STI and E_SA performs better than that of SPAI and SMsurf_SA.
Model performance of “ALL (OMEGA+RV+DIV)” scheme-based hydrometeorological reconstruction using preexisting dynamic–hydrometeorological relationships. Boxplots are employed for comprehensive metrics of the performance of these 12 typical drought events (Fig. 3, Fig. 4, and Figs. S5–S14) utilizing (a) ACC and (b) RMSE. For each boxplot, the short line on the top, the upper box boundary, the short line inside the box, the lower box boundary, the short line on the bottom, and the gray lozenge correspond to the maximum, 75th percentile, 50th percentile (median), 25th percentile, minimum, and outlier values of ACC- or RMSE-based results.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
4. Global-scale estimation of dynamic roles in drought reconstruction
a. Evaluation of reconstruction schemes and predictor contributions
We can preliminarily understand dynamic roles in drought reconstruction using model evaluation based on four different schemes. Regarding the reconstruction schemes based on single-type atmospheric dynamics, the OMEGA-only method performs the best in drought reconstruction (yellow curves in Figs. 7a0 and 7a1), and the RV-only scheme overall simulates concurrent temperature anomalies relatively well (blue curves in Figs. 7b0 and 7b1). This is because vertical subsidence suppresses precipitation formation, and horizontal cyclones/anticyclones may induce near-surface temperature anomalies. With all three types of atmospheric dynamics considered (red curves in Fig. 7), the “All” scheme-based performance of drought reconstruction is almost identical to that of the OMEGA-only scheme. In contrast, the “All” scheme-based reconstruction of temperature anomalies is greatly improved over the RV-only method, especially over the 0.3–0.4 range of ACC. This result indicates that on the global scale, dynamic subsidence is highly dominant in drought formation; however, the influencing factors of concurrent temperature anomalies are not limited to anticyclones/cyclones.
Scheme evaluations regarding the global-scale reconstruction of drought and concurrent temperature anomalies at the peak time of global-scale detected severe droughts. R2 and ACC [scatterplots (A) and (B)] coefficients and associated cumulative frequencies starting with the best scores [(a0), (a1), (b0), and (b1)] are employed for scheme evaluation. Each colored scatter point in (A) and (B) corresponds to the metrics of the reconstruction performance of a given event with a given scheme implemented.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
To comprehensively understand predictor contributions concerning different types of atmospheric dynamics, latitudinal-predictor characteristics and event-based frequencies based on the four schemes are investigated in Fig. 8. In terms of drought reconstruction (Figs. 8A and 8a), midtropospheric vertical motion is dominant across the globe, as expected, and the low-level cyclonic/anticyclonic anomalies also contribute to drought development in the middle and high latitudes (especially north of 40°N; region I in Fig. 8A). Since drought events associated with region I are well distributed throughout the year (Fig. S19A), there are two seasonally dependent mechanisms responsible for the features of region I. The enhanced low-level anticyclones suppress precipitation formation by deflecting storm systems and descending airflow in the summer half-year, especially over Europe (Hauser et al. 2016). The other mechanism is possibly the lack of landfalling extratropical cyclones in the winter half-year, as these are important sources of extreme precipitation over Europe and North America (Hawcroft et al. 2018; Owen et al. 2021).
Latitudinal evolutions and frequencies of predictor contributions based on the “ALL (OMEGA+RV+DIV),” “OMEGA-only,” “RV-only,” and “DIV-only” schemes regarding the global-scale XGBoost-based reconstruction. The upper [(A)–(D) and (a)–(d)] and lower [(E)–(H) and (e)–(h)] panels correspond to drought and concurrent temperature anomalies at the peak time of global-scale detected severe droughts, respectively. Panels (A)–(H) display predictor contributions, and each predictor group on the predictor axis follows the 100–1000 hPa ascending 27-pressure-level based order. Meanwhile, the latitudinal axis of (A)–(H) corresponds to the centroid position of grid cells under drought situations at the drought peak time. Scatter points in (A)–(H) represent R3 values, while (a)–(h) display global event frequencies with R3 of a given predictor beyond 10%. Light red shading in (A) and (E) highlights special signal areas for illustration. A single type of dynamic variable is displayed for each diagram in (B)–(D) and (F)–(H), while the other two irrelevant types are marked with gray blocks.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
On the other hand, regarding the reconstruction of concurrent temperature anomalies globally at the drought peak times (region II in Fig. 8E), 100–250-hPa cyclonic/anticyclonic anomalies are significant for reconstruction north of 30°N. The highlighted upper-level signals in region II correspond to upper-tropospheric large-amplitude Rossby-wave packets and associated persistent anticyclones/cyclones, in relation to the near-surface temperature extremes in the northern midlatitudes (Fragkoulidis et al. 2018; Schubert et al. 2014; Wehrli et al. 2019; Zschenderlein et al. 2019). One possible physical linkage with near-surface warming is that amplified upper-tropospheric anticyclonic anomalies may trigger intensified dynamic subsidence and associated intense adiabatic heating (Zschenderlein et al. 2019). In addition, 250–500-hPa vertical motion contributes significantly to the reconstruction of near-surface temperature anomalies over tropical areas between 10°N and 20°S (region III in Fig. 8E). Physically, the upper-tropospheric vertical motion leads to fewer clouds, more downward shortwave radiation, and near-surface warming.
Predictor contributions based on single-type reconstruction schemes based on interannual variability are also investigated (Figs. 8B–D and 8F–H). Regarding drought situations, midtropospheric vertical motion, mid and lower-tropospheric relative vorticity, and a combination of upper- and lower-tropospheric horizontal divergence can contribute obviously to drought reconstruction. However, when reconstructing near-surface temperature anomalies, the upper-tropospheric relative vorticity is the absolutely dominant single-type variable.
b. Understanding geographic distributions of predictor contributions
We highlight the potentially close dynamic–hydrometeorological relationships in some particular latitudinal ranges (regions I, II, and III in Figs. 8A and 8E). However, associated geographic distributions and seasonality remain unknown. This section generates correlation maps among these standardized seasonal-scale variables in different seasons (Fig. 9 and Fig. S18) to help understand reconstruction performance (Fig. 8) and associated dynamic–hydrometeorological relationships on the global scale.
(a),(c),(e),(g) Global-scale dynamic–hydrometeorological correlation maps and (b),(d),(f),(h) associated latitudinal distributions of coverage in boreal winter (DJF, December–February) and boreal summer (JJA, June–August). Pearson correlation coefficients (r) are computed among standardized seasonal-scale variables during 1981–2020, namely, SPAI, STI, 200-/850-hPa relative vorticity (i.e., rv200 and rv850), and 300-/500-hPa vertical velocity (i.e., omega300 and omega500). Note that only significant r values (values > 0.5 or values < −0.5) with a two-tailed p value below 0.01 are displayed in (a), (c), (e), and (g) using shading (brown-green and blue-red bars) and markers (crosses and squares). Panels (b), (d), (f), and (h) display corresponding significant r-based coverage evolution, maximum coverage, and associated location.
Citation: Journal of Hydrometeorology 23, 9; 10.1175/JHM-D-22-0006.1
1) Precipitation anomalies
Vertical motion at 500 hPa (hereafter simplified to omega500, positive direction downward) and relative vorticity at 850 hPa (hereafter simplified to rv850) are chosen to be correlated with SPAI. Consistent with the dominant roles of omega500 in drought reconstruction (Fig. 8A), the significant omega500–SPAI correlation is almost negative across the globe, and the associated latitudinal evolution is well distributed through all seasons (Figs. 9b,d and Figs. S18b and d). Significant rv850–SPAI correlations are positive (negative) in the Northern (Southern) Hemisphere due to hemispheric differences in the Coriolis force. Significant positive correlation-based coverages peak at around 52°N without a noticeable seasonality difference (green curves in Fig. 9 and Fig. S18). This characteristic partly illustrates low-level anticyclonic/cyclonic contributions to drought reconstruction in the middle to high latitudes (Fig. 8A). In particular, Europe and central Asia are occupied by a large area of highly positive rv850–SPAI correlations almost throughout the year.
2) Near-surface air temperature anomalies
In the middle to high latitudes in both hemispheres, relative vorticity at 200 hPa (hereafter simplified to rv200) shows significant correlations with STI through all seasons (panels f and h of both Fig. 9 and Fig. S18). A large area of significant negative correlation appears in the Northern Hemisphere and peaks north of 40°N; it is partly responsible for the dominant role of the upper-level anticyclones/cyclones in reconstructing mid- to high-latitudinal temperature anomalies (region II in Fig. 8E). In particular, the significant negative rv200–STI correlation coefficients cover almost all of Europe in boreal summer (Fig. 9g), indicating that upper-level anticyclonic anomalies are potentially responsible for the warm-season European heatwaves (Fig. 2b3 and Fig. S3a2). In the subtropics and tropics between 30°N and 30°S through almost all seasons, upper-level vertical motion, rather than upper-level relative vorticity, is positively correlated with near-surface air temperature. In the austral summer, a large area of highly positive omega300–STI correlations appears over northeastern South America, northern Australia, and South Africa (thick squares in Fig. 9e), where enhanced upper-level dynamic subsidence possibly induces near-surface warming. In boreal summer (Fig. 9g), positive omega300–STI correlation-based coverage peaks at around 27°N but becomes weak. In any case, the results herein could aid understanding of the noticeable upper-tropospheric omega-based predictor contributions in the reconstruction of near-surface air temperature anomalies concurrent with droughts between 10°N and 20°S (region III in Fig. 8E).
5. Conclusions and discussion
Global-scale severe drought events and potential drivers are critical topics in the hydrometeorological community. Local atmospheric dynamic anomalies (e.g., anticyclonic circulations and vertical subsidence) are well recognized across different large-scale triggers (e.g., atmospheric teleconnections, tropical thermal conditions, and marine heatwaves). We reconstructed hydrometeorological variations (e.g., soil moisture and actual evaporation) under drought situations utilizing local dynamic variability and quantified the associated specific predictor contributions. To achieve this, we adopted two different reconstruction schemes (i.e., statistically preexisting dynamic–hydrometeorological relationships and interannual variability). Regarding drought and concurrent near-surface temperature anomalies reconstructed with preexisting dynamic–hydrometeorological relationships, good reconstruction performance can be captured with the same or different lead times, depending on whether the evolution of dynamic anomalies (e.g., vertical motion and relative vorticity) is temporally asynchronous. On the other hand, reconstruction on the interannual scale performs relatively well, also suggesting feasibility regardless of seasonality, geographic differences, and even drought-inducing mechanisms. More importantly, from interpretable perspectives, global-scale analysis of dynamic contributions helps discover unexpected dynamic drought-inducing roles and associated latitudinal modulation. For example, low-level cyclonic/anticyclonic anomalies also contribute to drought development in the northern middle and high latitudes, while upper-level vertical subsidence contributes significantly to tropical near-surface temperature anomalies. These aforementioned achievements (i.e., feasible variable groups for different hydrometeorological reconstruction, latitudinal laws of dynamically based predictor contributions, and seasonality and geographic differences) could provide prior information for the potential construction of dynamically based drought prediction models across the globe.
Besides the potential advances mentioned above, reconstruction performance and the associated influencing factors also need further exploration.
a. Spatiotemporal characteristics of extreme drought events
The drought peak time for global reconstruction is relatively well distributed throughout the year and across the globe (Table 1 and Figs. S19a and S19A). The standardized forms adopted herein could weaken seasonality and geographic differences to clarify the dynamic roles and help achieve good reconstruction performance. Even so, reconstruction performance may be affected by drought migration patterns, time scales, and mechanisms. In terms of migration patterns, the 2019 Yangtze River hot autumn drought propagates southward from northern China (Liu and Zhou 2021) instead of undergoing routine local development, which may be responsible for the unsatisfactory reconstruction performance of interannual variability at the drought peak time (Fig. S1c). Regarding drought time scales, the 2012 Great Plains drought is essentially a so-called flash drought with rapid onset and intensification (DeAngelis et al. 2020; Pendergrass et al. 2020), despite being captured on the seasonal scale. In addition, the 2012 Great Plains drought results from an upper-level atmospheric ridge that persists for 2–3 weeks over the United States, which is partly intensified by soil dryness via land–atmosphere feedback (DeAngelis et al. 2020) during drought development. Therefore, the current dynamically based reconstruction performance might be further improved by considering land surface thermal conditions.
b. Temporally asynchronous variation of atmospheric dynamic anomalies
Dynamically asynchronous variation (Fig. 5) is partly responsible for different reconstruction performance utilizing preexisting dynamic–hydrometeorological relationships (Figs. 3 and 4). The 2012 Great Plains drought is not unique since a similar situation with RV SAs appearing earlier than OMEGA SAs also occurs in the 2015/16 tropical South American severe drought (Figs. S12 and S15d). Another recent topic of interest is that the skillful prediction of near-surface temperature seems to partly help the precipitation forecast during the 2012 Great Plains drought (Kam et al. 2021). In this regard, the dynamically asynchronous variation in our study might be a potential physical explanation. In addition, dynamically synchronous cases also exist, e.g., the 2019 Yangtze River autumn hot drought (Figs. S15b and S8) and the 2014 southeastern Brazil austral summer drought (Figs. S16b and S11).
c. Implications of predictor contributions regarding the two reconstruction schemes
Predictor contributions (Fig. 2, Figs. S3 and S4) display no evident contributions of horizontal divergence to reconstruction on the interannual variability since vertical motion is closely related to horizontal divergence/convergence but is more dominant in drought mechanisms. Still, reconstruction based on preexisting statistical relationships highlights prominent divergence-related contributions to different degrees (Figs. 3 and 4 and Figs. S5–S14). In this regard, atmospheric dynamic situations are comprehensively described from a statistical perspective. Even so, some seemingly confusing issues merit further exploration. For example, reconstruction of near-surface warming two months ahead of the peak time in the 2018 northern European drought (Figs. S6a6 and S6a10) is considerably affected by upper- and midtropospheric vertical motion (the omega part in Fig. S6b2). The main predictor contribution seemingly does not coincide with the persistent anticyclonic anomalies in the troposphere (Fig. S15a2).
d. All types of temperature anomalies concurrent with severe drought events
Although compound droughts and heatwaves (warmer than normal, STI > +1) are a recent hot topic, temperature anomalies during drought development globally can also be normal (−1 < STI < +1) and colder than normal (STI < −1), as shown in Fig. S19B. To our knowledge, the attribution of different types of temperature anomalies has seldom been investigated. Accordingly, diagnosis of dynamic anomalies and associated decomposition of local temperature variation (e.g., advection, adiabatic, and diabatic terms) may help reveal potential mechanisms. The motivation is that RV SAs and near-surface temperature during the 2011 Yangtze River drought are within normal ranges (Figs. S17a2 and S17a4), while most other cases in Table 1 correspond to near-surface warming and obvious upper-level anticyclonic/cyclonic anomalies (Figs. S15–S17).
e. Illustration of overfitting-related hyperparameters
It is noted that the max_depth parameter adjusted in this study is one regularization parameter related to overfitting. Increasing the max_depth value will make the model more complex and perform better, but overfitting is also more likely. Accordingly, the so-called gamma parameter of the XGBoost model could be introduced and enlarged to make the model more conservative. Practically, in the present study, the max_depth parameter is set relatively small (the 2–10 range, corresponding to shallow trees), and therefore the gamma parameter remains zero by default. In this regard, we also achieve relatively good performance concerning global drought events over moderate-sized datasets without too much time or memory consumption.
f. Influence of surface pressure on training data
Low surface pressure (SP) due to high altitude can affect the validation of reanalyzed atmospheric dynamic data in specific grid cells. For convenience, we discard training data with corresponding SP greater than 925 hPa across the 12 severe drought events. Even then, some grid cells with SP less than 925 hPa are still involved (gray squares in Fig. S20), which leads to the use of artificial reanalysis datasets. We mention this point for caution in further studies; however, it eventually becomes a trade-off between rigor and convenience. On the other hand, due to considerable variation in surface pressures regarding global-scale drought events and associated seasonality, a large number of drought events with surface pressure greater than 925 hPa also exist. Therefore, we employed 100–1000-hPa pressure levels for global-scale quick reconstruction (section 2d). In terms of predictor contributions concerning global-scale quick reconstruction (Figs. 8A,E), the lack of prominent signals in vertical locations greater than 925 hPa partly suggests that near-surface dynamic roles may be negligible noise in interannual variability.
Acknowledgments.
This work is fully supported by the International Cooperation and Exchange Programme of the National Natural Science Foundation of China (42120104001). The authors are grateful for the detailed and well-advised comments from the two anonymous referees.
Data availability statement.
The ERA-5 near-surface hydrometeorological data are available at https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=form, while the associated pressure level–based atmospheric dynamic data can be obtained at https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-pressure-levels?tab=form. Both the ESA-CCI v5.3 surface soil moisture and the GLEAM v3.5a actual evaporation dataset can be downloaded publicly following the guidelines at https://www.gleam.eu/#home. The global-scale detected severe drought events have been deposited on the Google drive (see the ERA5 file: https://drive.google.com/drive/folders/1LWLJTwCc1c3BWac97XhQE8czrZcU6NWK?usp=sharing).
REFERENCES
AghaKouchak, A., L. Y. Cheng, O. Mazdiyasni, and A. Farahmand, 2014: Global warming and changes in risk of concurrent climate extremes: Insights from the 2014 California drought. Geophys. Res. Lett., 41, 8847–8852, https://doi.org/10.1002/2014GL062308.
Ainembabaz, J. H., J. Rusike, and B. Keizire, 2018: The 2015–16 El Niño-induced drought crisis in Southern Africa: Lessons from historical data and policy implications. AGRA, 31 pp., https://agra.org/wp-content/uploads/2020/08/AGRA-drought-crisis-in-Southern-Africa_v2-.pdf.
Barriopedro, D., E. M. Fischer, J. Luterbacher, R. Trigo, and R. Garcia-Herrera, 2011: The hot summer of 2010: Redrawing the temperature record map of Europe. Science, 332, 220–224, https://doi.org/10.1126/science.1201224.
Chen, P., and M. Newman, 1998: Rossby wave propagation and the rapid development of upper-level anomalous anticyclones during the 1988 US drought. J. Climate, 11, 2491–2504, https://doi.org/10.1175/1520-0442(1998)011<2491:RWPATR>2.0.CO;2.
Chen, T., and C. Guestrin, 2016: Xgboost: A scalable tree boosting system. Proc. 22nd ACM Sigkdd Int. Conf. on Knowledge Discovery and Data Mining, San Francisco, CA, J. Association for Computing Machinery, 785–794, https://doi.org/10.1145/2939672.2939785.
Coelho, C. A. S., and Coauthors, 2016: The 2014 southeast Brazil austral summer drought: Regional scale mechanisms and teleconnections. Climate Dyn., 46, 3737–3752, https://doi.org/10.1007/s00382-015-2800-1.
DeAngelis, A. M., H. L. Wang, R. D. Koster, S. D. Schubert, Y. H. Chang, and J. Marshak, 2020: Prediction skill of the 2012 US Great Plains flash drought in Subseasonal Experiment (SubX) models. J. Climate, 33, 6229–6253, https://doi.org/10.1175/JCLI-D-19-0863.1.
Di Capua, G., S. Sparrow, K. Kornhuber, E. Rousi, S. Osprey, D. Wallom, B. van den Hurk, and D. Coumou, 2021: Drivers behind the summer 2010 wave train leading to Russian heatwave and Pakistan flooding. NPJ Climate Atmos. Sci., 4, 55, https://doi.org/10.1038/s41612-021-00211-9.
Dorigo, W., and Coauthors, 2017: ESA CCI soil moisture for improved Earth system understanding: State-of-the art and future directions. Remote Sens. Environ., 203, 185–215, https://doi.org/10.1016/j.rse.2017.07.001.
Erfanian, A., G. L. Wang, and L. Fomenko, 2017: Unprecedented drought over tropical South America in 2016: Significantly under-predicted by tropical SST. Sci. Rep., 7, 5811, https://doi.org/10.1038/s41598-017-05373-2.
Fernando, D. N., and Coauthors, 2016: What caused the spring intensification and winter demise of the 2011 drought over Texas? Climate Dyn., 47, 3077–3090, https://doi.org/10.1007/s00382-016-3014-x.
Folberth, C., A. Baklanov, J. Balkovič, R. Skalský, N. Khabarov, and M. Obersteiner, 2019: Spatio-temporal downscaling of gridded crop model yield estimates based on machine learning. Agric. For. Meteor., 264, 1–15, https://doi.org/10.1016/j.agrformet.2018.09.021.
Fragkoulidis, G., V. Wirth, P. Bossmann, and A. H. Fink, 2018: Linking northern hemisphere temperature extremes to Rossby wave packets. Quart. J. Roy. Meteor. Soc., 144, 553–566, https://doi.org/10.1002/qj.3228.
Gibson, P. B., W. E. Chapman, A. Altinok, L. Delle Monache, M. J. DeFlorio, and D. E. Waliser, 2021: Training machine learning models on climate model output yields skillful interpretable seasonal precipitation forecasts. Commun. Earth Environ., 2, 159, https://doi.org/10.1038/s43247-021-00225-4.
Gruber, A., W. A. Dorigo, W. Crow, and W. Wagner, 2017: Triple collocation-based merging of satellite soil moisture retrievals. IEEE Trans. Geosci. Remote Sens., 55, 6780–6792, https://doi.org/10.1109/TGRS.2017.2734070.
Hauser, M., R. Orth, and S. I. Seneviratne, 2016: Role of soil moisture versus recent climate change for the 2010 heat wave in western Russia. Geophys. Res. Lett., 43, 2819–2826, https://doi.org/10.1002/2016GL068036.
Hawcroft, M., E. Walsh, K. Hodges, and G. Zappa, 2018: Significantly increased extreme precipitation expected in Europe and North America from extratropical cyclones. Environ. Res. Lett., 13, 124006, https://doi.org/10.1088/1748-9326/aaed59.
Herrera-Estrada, J. E., J. A. Martinez, F. Dominguez, K. L. Findell, E. F. Wood, and J. Sheffield, 2019: Reduced moisture transport linked to drought propagation across North America. Geophys. Res. Lett., 46, 5243–5253, https://doi.org/10.1029/2019GL082475.
Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146, 1999–2049, https://doi.org/10.1002/qj.3803.
Hoerling, M., J. Eischeid, A. Kumar, R. Leung, A. Mariotti, K. Mo, S. Schubert, and R. Seager, 2014: Causes and predictability of the 2012 Great Plains drought. Bull. Amer. Meteor. Soc., 95, 269–282, https://doi.org/10.1175/BAMS-D-13-00055.1.
Jimenez-Munoz, J. C., C. Mattar, J. Barichivich, A. S.-Artigas, K. Takahashi, Y. Malhi, J. A. Sobrino, and G. van der Schrier, 2016: Record-breaking warming and extreme drought in the Amazon rainforest during the course of El Niño 2015-2016. Sci. Rep., 6, 33130, https://doi.org/10.1038/srep33130.
Jin, D. C., Z. Y. Guan, and W. Y. Tang, 2013: The extreme drought event during winter–spring of 2011 in East China: Combined influences of teleconnection in midhigh latitudes and thermal forcing in maritime continent region. J. Climate, 26, 8210–8222, https://doi.org/10.1175/JCLI-D-12-00652.1.
Kam, J., S. Kim, and J. K. Roundy, 2021: Did a skillful prediction of near-surface temperatures help or hinder forecasting of the 2012 US drought? Environ. Res. Lett., 16, 034044, https://doi.org/10.1088/1748-9326/abe1f6.
Lau, W. K. M., and K. M. Kim, 2012: The 2010 Pakistan flood and Russian heat wave: Teleconnection of hydrometeorological extremes. J. Hydrometeor., 13, 392–403, https://doi.org/10.1175/JHM-D-11-016.1.
Li, M. Y., Y. Yao, I. Simmonds, D. H. Luo, L. H. Zhong, and X. D. Chen, 2020: Collaborative impact of the NAO and atmospheric blocking on European heatwaves, with a focus on the hot summer of 2018. Environ. Res. Lett., 15, 114003, https://doi.org/10.1088/1748-9326/aba6ad.
Liu, X. B., B. He, L. L. Guo, L. Huang, and D. L. Chen, 2020: Similarities and differences in the mechanisms causing the European summer heatwaves in 2003, 2010, and 2018. Earth’s Future, 8, e2019EF001386, https://doi.org/10.1029/2019EF001386.
Liu, Z. C., and W. Zhou, 2021: The 2019 autumn hot drought over the middle-lower reaches of the Yangtze River in China: Early propagation, process evolution, and concurrence. J. Geophys. Res. Atmos., 126, e2020JD033742, https://doi.org/10.1029/2020JD033742.
Liu, Z. C., G. H. Lu, H. He, Z. Y. Wu, and J. He, 2017: Understanding atmospheric anomalies associated with seasonal pluvial-drought processes using southwest China as an example. J. Geophys. Res. Atmos., 122, 12 210–12 225, https://doi.org/10.1002/2017JD026867.
Marengo, J. A., J. Tomasella, L. M. Alves, W. R. Soares, and D. A. Rodriguez, 2011: The drought of 2010 in the context of historical droughts in the Amazon region. Geophys. Res. Lett., 38, L12703, https://doi.org/10.1029/2011GL047436.
Martens, B., and Coauthors, 2017: GLEAM v3: Satellite-based land evaporation and root-zone soil moisture. Geosci. Model Dev., 10, 1903–1925, https://doi.org/10.5194/gmd-10-1903-2017.
Namias, J., 1991: Spring and summer 1988 drought over the contiguous United States—Causes and prediction. J. Climate, 4, 54–65, https://doi.org/10.1175/1520-0442(1991)004<0054:SASDOT>2.0.CO;2.
Owen, L. E., J. L. Catto, D. B. Stephenson, and N. J. Dunstone, 2021: Compound precipitation and wind extremes over Europe and their relationship to extratropical cyclones. Wea. Climate Extremes, 33, 100342, https://doi.org/10.1016/j.wace.2021.100342.
Pendergrass, A. G., and Coauthors, 2020: Flash droughts present a new challenge for subseasonal-to-seasonal prediction. Nat. Climate Change, 10, 191–199, https://doi.org/10.1038/s41558-020-0709-0.
Ramakrishna, S. S. V. S., V. B. Rao, B. R. S. Rao, D. H. Prasad, N. N. Rao, and R. Panda, 2017: A study of 2014 record drought in India with CFSv2 model: Role of water vapor transport. Climate Dyn., 49, 297–312, https://doi.org/10.1007/s00382-016-3343-9.
Rodrigues, R. R., A. S. Taschetto, A. Sen Gupta, and G. R. Foltz, 2019: Common cause for severe droughts in South America and marine heatwaves in the South Atlantic. Nat. Geosci., 12, 620–626, https://doi.org/10.1038/s41561-019-0393-8.
Schubert, S. D., H. L. Wang, R. D. Koster, M. J. Suarez, and P. Y. Groisman, 2014: Northern Eurasian heat waves and droughts. J. Climate, 27, 3169–3207, https://doi.org/10.1175/JCLI-D-13-00360.1.
Schumacher, D. L., J. Keune, C. C. van Heerwaarden, J. V. G. de Arellano, A. J. Teuling, and D. G. Miralles, 2019: Amplification of mega-heatwaves through heat torrents fuelled by upwind drought. Nat. Geosci., 12, 712–717, https://doi.org/10.1038/s41561-019-0431-6.
Shahhosseini, M., G. Hu, I. Huber, and S. V. Archontoulis, 2021: Coupling machine learning and crop modeling improves crop yield prediction in the US Corn Belt. Sci. Rep., 11, 1606, https://doi.org/10.1038/s41598-020-80820-1.
Shi, H., M. Garcia-Reyes, M. G. Jacox, R. R. Rykaczewski, B. A. Black, S. J. Bograd, and W. J. Sydeman, 2021: Co-occurrence of California drought and northeast Pacific marine heatwaves under climate change. Geophys. Res. Lett., 48, e2021GL092765, https://doi.org/10.1029/2021GL092765.
Wang, H. J., and S. P. He, 2015: The north China/northeastern Asia severe summer drought in 2014. J. Climate, 28, 6667–6681, https://doi.org/10.1175/JCLI-D-15-0202.1.
Wang, H. L., S. Schubert, R. Koster, Y. G. Ham, and M. Suarez, 2014: On the role of SST forcing in the 2011 and 2012 extreme U.S. heat and drought: A study in contrasts. J. Hydrometeor., 15, 1255–1273, https://doi.org/10.1175/JHM-D-13-069.1.
Wang, L., and W. Chen, 2014: A CMIP5 multimodel projection of future temperature, precipitation, and climatological drought in China. Int. J. Climatol., 34, 2059–2078, https://doi.org/10.1002/joc.3822.
Wang, S. S., X. Yuan, and Y. H. Li, 2017: Does a strong El Niño imply a higher predictability of extreme drought? Sci. Rep., 7, 40741, https://doi.org/10.1038/srep40741.
Wang, S. S.-C., Y. Qian, L. R. Leung, and Y. Zhang, 2021: Identifying key drivers of wildfires in the contiguous US using machine learning and game theory interpretation. Earth’s Future, 9, e2020EF001910, https://doi.org/10.1029/2020EF001910.
Wehrli, K., B. P. Guillod, M. Hauser, M. Leclair, and S. I. Seneviratne, 2019: Identifying key driving processes of major recent heat waves. J. Geophys. Res. Atmos., 124, 11 746–11 765, https://doi.org/10.1029/2019JD030635.
White, R. H., K. Kornhuber, O. Martius, and V. Wirth, 2022: From atmospheric waves to heatwaves: A waveguide perspective for understanding and predicting concurrent, persistent and extreme extratropical weather. Bull. Amer. Meteor. Soc., 103, E923–E935, https://doi.org/10.1175/BAMS-D-21-0170.1.
Xu, K., H. Y. Miao, B. Q. Liu, C. Y. Tam, and W. Q. Wang, 2020: Aggravation of record-breaking drought over the mid-to-lower reaches of the Yangtze River in the post-monsoon season of 2019 by anomalous Indo-Pacific oceanic conditions. Geophys. Res. Lett., 47, e2020GL090847, https://doi.org/10.1029/2020GL090847.
Yiou, P., and Coauthors, 2020: Analyses of the northern European summer heatwave of 2018. Bull. Amer. Meteor. Soc., 101, S35–S40, https://doi.org/10.1175/BAMS-D-19-0170.1.
Zeng, D. W., and X. Yuan, 2021: Modeling the influence of upstream land–atmosphere coupling on the 2017 persistent drought over northeast China. J. Climate, 34, 4971–4988, https://doi.org/10.1175/JCLI-D-20-0650.1.
Zeng, D. W., X. Yuan, and J. K. Roundy, 2019: Effect of teleconnected land–atmosphere coupling on Northeast China persistent drought in spring–summer of 2017. J. Climate, 32, 7403–7420, https://doi.org/10.1175/JCLI-D-19-0175.1.
Zhang, W. J., W. Mao, F. Jiang, M. F. Stuecker, F. F. Jin, and L. Qi, 2021a: Tropical indo-pacific compounding thermal conditions drive the 2019 Australian extreme drought. Geophys. Res. Lett., 48, e2020GL090323, https://doi.org/10.1029/2020GL090323.
Zhang, Y., K. Liu, Y. H. Li, W. Shen, Y. L. Ren, D. W. Zeng, and S. Sha, 2021b: Water vapor changes affect cross-seasonal strong drought events in the eastern region of northwest China. Front. Earth Sci., 9, 609321, https://doi.org/10.3389/feart.2021.609321.
Zscheischler, J., and Coauthors, 2014: Impact of large-scale climate extremes on biospheric carbon fluxes: An intercomparison based on MsTMIP data. Global Biogeochem. Cycles, 28, 585–600, https://doi.org/10.1002/2014GB004826.
Zscheischler, J., and Coauthors, 2018: Future climate risk from compound events. Nat. Climate Change, 8, 469–477, https://doi.org/10.1038/s41558-018-0156-3.
Zschenderlein, P., A. H. Fink, S. Pfahl, and H. Wernli, 2019: Processes determining heat waves across different European climates. Quart. J. Roy. Meteor. Soc., 145, 2973–2989, https://doi.org/10.1002/qj.3599.