Spatiotemporal Evolution and Nowcasting of the 2022 Yangtze River Mega-Flash Drought

Abstract

Flash droughts challenge early warnings due to their rapid onset, which requires a proper drought index and skillful nowcasting system. A few studies have assessed the nowcast skill for flash droughts using a one-dimensional index, but whether the models can capture their spatiotemporal evolution remains unclear. In this study, a three-dimensional meteorological flash drought index based on the percentile of 15-day moving average precipitation minus evapotranspiration (P-ET) is developed. The index is then used to investigate the spatiotemporal evolution of a mega-flash drought that occurred in the Yangtze River basin during the summer of 2022. The results show that the mega-flash drought started at the beginning of July in the upper reaches of the river and expanded to the middle and lower reaches at the beginning of August due to the spread of the high-pressure system. The evolution is well captured by the proposed three-dimensional index. The spatial correlations between the China Meteorological Administration global medium-range ensemble forecast system (CMA-GFS)’s nowcast and reanalysis ranged from 0.58 to 0.85, and the hit rate and equitable threat score are 0.54 and 0.26, respectively. This study shows that the CMA-GFS nowcast of the P-ET index roughly captured the drought’s evolution, which can be used for flash drought early warnings and water resource management.

1. Introduction

Flash droughts are the new normal in the warming world, challenging drought forecasting and early warning systems, which are usually built upon slow-evolving droughts [1]. With the long-term increasing trend in global mean frequency with respect to flash droughts, mega-flash droughts with severe impacts occurred frequently around the world [2,3,4]. In 2012, a mega-flash drought occurred in the central U.S., caused an economic loss of USD 30 billion [5], and raised extensive concerns in climate, hydrology, agriculture, and ecology communities. After 10 years, another mega-flash drought occurred in the Yangtze River basin in southern China, which is the most severe drought in the region in the recent 60 years [6,7]. The drought developed from the upper reaches of the river to the middle and lower reaches within a month [7] and resulted in widespread water deficit conditions across the Yangtze River basin.

The 2022 Yangtze River mega-flash drought caused a number of societal issues, including crop reduction, wildfires, water and energy supply shortages, and heatwave-related human health issues [8,9]. Moreover, the lakes and reservoirs dried up extensively, resulting in ecological degradations that may require decades to recover [10]. Besides its extensive impact similarly to other droughts, the 2022 drought is characterized as a mega-flash drought due to its rapid onset. However, exactly how fast the drought evolves requires an objective drought index for identification and measurement. In fact, there are hundreds of drought indices [11], from meteorological droughts to agricultural, hydrological, and even socioeconomic droughts, depending on the aspects or sectors concerned. All droughts originate from meteorological droughts, including flash droughts. So far, soil moisture, evapotranspiration, or vegetation-based flash drought indices have been widely used [12,13,14], but the meteorological flash drought index is usually ignored [15]. One of the reasons is that most studies are purpose-oriented, e.g., oriented toward the impacts of flash drought on agriculture and the ecosystem, while the meteorological drought index is not directly related to these impacts [14]. Another important reason is that meteorological droughts usually refer to seasonal, interannual, or even decadal droughts, for which precipitation (P) or precipitation surplus (P minus evapotranspiration (ET) or potential ET) can be good indicators after accumulation processes. While for flash droughts that usually occur over sub-seasonal time scales, the usages of P or P-ET have not been well explored [16]. Pentad mean (5-day mean) soil moisture is used to identify soil flash drought, with certain thresholds and a requirement for a decline rate in soil moisture [1]. For instance, soil moisture that drops from the 40th percentile to the 20th percentile within 20 days is regarded as a flash drought event if the minimum drought duration is met (e.g., 20 days). While for the P-ET percentile index, pentad mean values might fluctuate too much to reflect a smooth drought evolution process due to the large variability of P. Similarly to matching the standardized precipitation minus evapotranspiration index (SPEI) with soil moisture [17], the P-ET percentile index should be smoothed to reflect water-deficit conditions at a longer time scale. Although flash droughts have a shorter time scale than seasonal droughts, a 5-day non-rainfall condition might not cause a large impact if the soil is wet. Therefore, the antecedent P-ET should also be used to quantify accumulated dry conditions that may eventually develop into drought conditions. In other words, flash droughts are unique because of this rapid onset, but a rapid onset is not the only criterion. Flash droughts should last for a period of time to cause an impact; otherwise, the “dry flash” will not affect society. Therefore, a proper smoothing time scale is required for applying the P-ET flash drought index.

Besides rapid onset, another unique feature of the 2022 mega-flash drought is its wide coverage from the upper-to-lower reaches of the river [7]. The middle and lower reaches of the Yangtze River basin usually suffer from summer drought after the Meiyu season, where the land surface condition can switch from wet to dry conditions in a short time and is accompanied by intensive solar radiation [18]. However, the widespread drought across the entire Yangtze River basin is unprecedented. To characterize the spatial propagation of drought, a specific drought index is needed. Most flash drought indices are one-dimensional; i.e., droughts are identified based on the selected drought index for each grid point independently [19]. This method only accounts for temporal evolution while neglecting spatial evolution. Therefore, three-dimensional drought indices are needed for flash droughts, especially for the 2022 Yangtze River mega-flash drought. A few studies started to investigate the identification of three-dimensional flash droughts [20,21], but the specific identification of a mega-flash drought from a meteorological perspective is overlooked. Investigating the spatiotemporal evolution of the 2022 Yangtze River mega-flash drought with a three-dimensional P-ET drought index would provide information for the causes, nowcasting, and early warning. It will also provide guidelines for water resource management under extreme drought conditions.

Given that droughts usually occur at seasonal to interannual time scales, most previous studies focused on seasonal-to-decadal meteorological drought predictions [22,23], with skillful predictions under certain oceanic anomaly conditions, including the El Niño–Southern Oscillation (ENSO). In contrast, for flash drought forecasting at sub-seasonal time scales, there are only a few assessment studies with numerical weather prediction (NWP) models. For instance, the global ensemble forecast system model was used to predict flash droughts in the United States with limited skill [24]. The European Center for Medium-Range Weather Forecasts (ECMWF) model was used to assess the flash drought forecasting in China, where the results suggest that the hit rate is less than 25% within the first week [25]. Whether the NWP models provide timely nowcasting information for flash droughts needs further investigation. The China Meteorological Administration (CMA) global medium-range ensemble forecast system (CMA-GFS) [26,27] is a new generation NWP model, which has been in operation since 2018. Evaluating the performance of the CMA-GFS in nowcasting weather extremes including flash droughts provides implications for service-oriented model research and development.

In this paper, a three-dimensional meteorological flash drought index is developed and used to investigate the spatiotemporal evolution of the 2022 Yangtze River mega-flash drought. Consequently, the nowcasting capability of the CMA-GFS during the 2022 mega-flash drought is evaluated.

2. Materials and Methods

2.1. ERA5 Reanalysis and CMA-GFS Nowcast Data

In this study, P, ET, soil moisture, and 500 hPa geopotential height from the fifth-generation reanalysis (ERA5) [28] are used as the observation for investigating the evolution of the 2022 mega-flash drought. ERA5 is the latest reanalysis released by the ECMWF, which assimilates a large amount of historical observation based on state-of-the-art modelling and data assimilation systems. It has a good performance over the globe and East Asia. Here, 62 years (1961–2022) ERA5 data with a spatial resolution of 0.25-degree are used to calculate the percentiles or anomalies during the 2022 summer drought period. The hourly ERA5 data are aggregated into pentad (5 days) mean data. The land surface model of ERA5 has four soil layers with depths of 7 cm, 21 cm, 72 cm, and 189 cm, so the top 1 m soil moisture is the accumulated value in the first three soil layers.

The CMA-GFS is the operational national NWP system in China, which includes both an atmospheric model and 4D-VAR assimilation component [26,29]. The P and ET operational nowcast products used in this study are based on the CMA-GFSv3.3 with a spatial resolution of 0.25 degree. The nowcasts start every day and extend to 7 days. Here, the 5-day nowcasts that started from 2 June, 7 June, …, 31 August in 2021 and 2022 are used. The nowcasts of P and ET in the summer of 2022 are bias-corrected through subtracting the results of 2021 from the results of 2022 and adding the increment to the ERA5 results of 2021. This is similar to the interannual increment method [30].

2.2. Three-Dimensional Meteorological Flash Drought Index

In order to investigate the evolution of meteorological flash drought, precipitation minus evapotranspiration (P-ET) index is used. Given that P-ET has different time scales as compared with top 1 m soil moisture that has been widely used to identify flash drought events [12], the 15-day moving average P-ET and 10-day moving average P-ET are calculated and converted into percentiles for each pentad (5 days). For instance, the 15-day moving average P-ET for the pentad 2–6 June is the P-ET averaged during 23 May–6 June. Figure 1 shows the correlation between percentiles for the pentad mean of the top 1 m soil moisture and percentiles for the 10-day/15-day moving average P-ET during the summers of 1961–2022. It is found that soil moisture correlates well with P-ET for both time scales, where the correlations are statistically significant (p < 0.05) for all grid cells (Figure 1). However, the 15-day moving average P-ET (Figure 1b) has higher correlation with soil moisture than that for the 10-day moving average P-ET (Figure 1a), with regional mean correlation increasing from 0.47 to 0.57. Therefore, the 15-day moving average P-ET is used for meteorological drought identifications in this study, which is also close to the response time scale of the top 1 m soil moisture to the precipitation.

The framework of three-dimensional flash drought events is divided into two steps:

(1)

Extraction of one-dimensional flash drought events

To identify one-dimensional flash drought events, 15-day moving average P-ET data are aggregated for each pentad, and they are converted to percentiles according to the climatology for each pentad during 1961–2022. A one-dimensional flash drought event [1] starts as (a) the pentad mean P-ET percentile decreases from above 40% to 20%, with an average decline rate of at least 5% in percentile for each pentad. (b) Once the P-ET percentile rises above 20% again, the flash drought terminates, and (c) a flash drought event should last for at least 4 pentads (20 days).

(2)

Identification of three-dimensional flash drought events

Based on the one-dimensional flash drought event, the spatiotemporal connectivity between the flash drought grids is considered to identify three-dimensional flash drought events. The identification is conducted as follows: (a) construction of spatially connected flash drought patches in one time step; (b) determination of the spatiotemporal connectivity of drought patches in two consecutive time steps; and (c) examination of the minimum duration, where a three-dimensional flash drought event should last for at least 20 days.

For (a), this study first marks the grids where a one-dimensional flash drought event occurred. Then, the Contiguous Drought Area (CDA) analysis method [31] is used to identify drought patches. For a marked grid Z, the 3 × 3 area around Z is examined for other marked grids. If all the adjacent grids are not marked, then the grid Z is discarded. If there is at least one marked grid in adjacent grids, then the grid Z is retained. The consecutive marked grids are defined as flash drought patches.

For (b), the connectivity of the flash drought patches is determined by checking the overlapping area between the flash drought patches of two consecutive time steps. If two patches do overlap, then two adjacent patches belong to the same event, otherwise they are considered to belong to two separate flash drought events. The overlapping area of two flash drought patches should be at least 50% of the smaller patches.

Due to the fact that one-dimensional index characterizes the temporal evolution process of drought characteristics based on a specific spatial scale (fixed grid or regional average), the spatial dynamic propagation characteristics of flash droughts cannot be captured [21]. In contrast, the three-dimensional index considers the evolution of flash drought in space (two-dimensional) and time (one-dimensional), which can facilitate the understanding of the origin of the flash drought. For instance, the three-dimensional index can help to identify both local causes (e.g., land–atmospheric coupling [32]) as well as remote causes (e.g., ocean–atmosphere or land–atmosphere teleconnections [33,34]).

2.3. Evaluation Metrics

In order to evaluate the CMA-GFS nowcasts, the correlation (r) and normalized root-mean-square-error (NRMSE) are calculated as follows:

$$r=\frac{Cov\left(X_{ERA5}, X_{CMA-GFS}\right)}{{\sigma }_{ERA5}·{\sigma }_{CMA-GFS}} ,$$

(1)

where $Cov\left(X_{ERA5}, X_{CMA-GFS}\right)$ is the covariance between CMA-GFS nowcasts ($X_{CMA-GFS}$) and ERA5 reanalysis data ($X_{ERA5}$), and ${\sigma }_{ERA5}$ and ${\sigma }_{CMA-GFS}$ are the standard deviation of ERA5 reanalysis and CMA-GFS nowcasts, respectively.

$$NRMSE=\frac{\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{\left(X_{ERA5,i}-X_{CMA-GFS,i}\right)}^2}{n}}}{{\sigma }_{ERA5}} ,$$

(2)

where n is the number of nowcasts. NRMSE represents the deviation of the nowcast from the true value (ERA5 reanalysis), with a value less than 1 indicating a useful nowcast. The lower the NRMSE, the higher the nowcast accuracy.

The nowcast skill for three-dimensional flash droughts can be evaluated using a $2\times 2$ contingency table. A is the number of observed flash drought events (calculated by ERA5 reanalysis) predicted by the CMA-GFS, B is the false alarm number of flash drought events, C is the number of observed flash drought events not predicted by the CMA-GFS, D is the observed and predicted nondrought events, and N is the sum of A, B, C, and D. The hit rate (HIT) measures the probability of observed flash droughts that are predicted, the false alarm ratio (FAR) measures the probability of predicted flash droughts that do not occur in the observation (false alarm), and the equitable threat score (ETS) is a balanced score between hits and false alarms. These metrics can be calculated as follows:

$$HIT=\frac{A}{A+C} ,$$

(3)

$$FAR=\frac{B}{A+B} ,$$

(4)

$$ETS=\frac{A-A_{Ref}}{A+B+C-A_{Ref}} ,$$

(5)

$$A_{Ref}=\frac{\left(A+B\right)\left(A+C\right)}{N} .$$

(6)

3. Results and Discussion

Figure 2 shows the climatological characteristics (i.e., frequency, mean duration and mean intensity) of meteorological flash droughts over the Yangtze River basin during 2 June–4 September of 1961–2022. Frequency is the number of drought events during the summers of 1961–2022. Mean duration is the average number of pentads each drought last for. Mean intensity is calculated by accumulating water deficits (40% minus P-ET percentile for each pentad) during all drought events, and then by dividing the accumulation with the pentads and events. The results show that the spatial patterns of drought characteristics from the one-dimensional index (Figure 2a–c) are similar to those from the three-dimensional index (Figure 2d–f). The frequency of three-dimensional flash droughts is lower than that of one-dimensional droughts, which is because a few moderate droughts with limited spatial coverage are removed during the identification of three-dimensional droughts. The mean duration of three-dimensional flash droughts is slightly shorter than that of the one-dimensional flash droughts (Figure 2b,e) due to more rigorous definition of the former, and the intensity of the former is slightly greater than that of the latter (Figure 2c,f). For three-dimensional flash droughts, they have higher chance to occur over the headwater region and the Yangtze Delta in the downstream area (Figure 2d), but the drought duration is longer over the middle and lower reaches (Figure 2e). The intensity has less spatial heterogeneity (Figure 2f) than frequency. On average, the frequency, mean duration, and intensity are 17.2 events, 4.92 pentads, and 27.57%/pentad/event for one-dimensional flash droughts, and 15.5 events, 4.86 pentads, and 28.21%/pentad/event for three-dimensional flash droughts, respectively (

Table 1. The average frequency (events), mean duration (pentads/event), and mean intensity (%/pentad/event) for one-dimensional flash droughts and three-dimensional flash droughts over the Yangtze River basin during summers of 1961–2022.

Drought Type	Frequency (Events)	Mean Duration (Pentads/Event)	Mean Intensity (%/Pentad/Event)
One-dimensional flash droughts	17.2	4.92	27.57
Three-dimensional flash droughts	15.5	4.86	28.21

).

With the three-dimensional P-ET index, the spatiotemporal evolution of the 2022 mega-flash drought over the Yangtze River is shown in Figure 3. The drought started in the upper reaches of the river at the beginning of July (Figure 3c), and developed into severe drought conditions during 7–11 July (Figure 3d). The drought over the upper reaches alleviated by the end of July (Figure 3h), but the drought over the middle to lower reaches started at the beginning of August (Figure 3i), right behind the drought over the upper reaches. The flash drought over the middle and lower reaches was very extensive and intensive, with the P-ET percentile lower than 5% for half of the month (Figure 3k–m). And the drought did not recover in September for the lower reaches (Figure 3o). The spatial expansion is very fast for both the upper reaches and the middle and lower reaches, with severe water deficiency over a large area.

Figure 4 shows the anomalies of 500 hPa geopotential height during the evolution of the 2022 mega-flash drought. The propagation of the flash drought from upstream to downstream areas is associated with the movement of the high-pressure anomaly centre. The onset of the upstream flash drought in 2–6 July (Figure 4c) is caused by the strong geopotential height anomaly to the west of the upstream area (Figure 4c). As the anomaly centre of geopotential height moved eastward, the upstream flash drought developed quickly (Figure 3d–g and Figure 4d–g). The subtropical high was weakened in the northeast area during 27–31 July (Figure 4h), but it was strengthened quickly during 1–5 August (Figure 4i), which resulted in a rapid onset of a flash drought over the middle and lower reaches of the Yangtze River basin (Figure 3i). From 6 to 30 August, the positive anomaly of the subtropical high persisted over the basin (Figure 4j–n), further intensifying the drought conditions (Figure 3j–n).

In order to evaluate the capability of the CMA-GFS in nowcasting the 2022 mega-flash drought, the nowcast skill of P-ET was evaluated firstly. Given that the CMA-GFS only provides 7-day weather nowcasts, the nowcasts in the first 5 days are merged with ERA5 reanalysis in the antecedent 10 days, which is similar to the procedure of 6-month standardized precipitation index (SPI6)-based seasonal drought forecasting [22]. Figure 5 shows the correlation and NRMSE between the merged P-ET pentad series and ERA5 during the summer of 2022. The CMA-GFS model has a higher correlation over the west and southeast parts than that over the middle and outlet of the Yangtze River basin (Figure 5a). Although the correlations are statistically significant over most grid cells (Figure 5a), the NRMSE shows that it is less skillful than climatological forecast over the middle and outlet parts, where the NRMSE is great than 1 (Figure 5b), which means the RMSE is greater than the standard deviation. Except for these regions, the NRMSE values are less than 1, which means skillful nowcasts over the remaining areas. Overall, the nowcasting skill is acceptable because the regional mean correlation and NRMSE are 0.8 and 0.85, respectively.

Figure 6 shows the three-dimensional flash drought evolution based on the nowcasting data during the summer of 2022. The nowcasting performed well in reproducing the spatiotemporal evolution of the mega-flash drought, with spatial correlations ranging from 0.58 to 0.85 during 2 July–30 August (Figure 6) compared with the reanalysis results (Figure 3). However, the drought intensity was overestimated. To diagnose the nowcasting results, the summer mean values of P, ET, and P-ET from the ERA5/CMA-GFS merged results and the ERA5 reanalysis were calculated. It is found that the P nowcasting was only underestimated by 3%, but the ET nowcasting was overestimated by 32.7% (Figure 7a,b,d,e;

Table 2. The P (mm), ET (mm), and P-ET (mm) from ERA5 reanalysis and CMA-GFS nowcasts averaged over the Yangtze River basin in the summer of 2022.

Data Type	P (mm)	ET (mm)	P-ET (mm)
ERA5 reanalysis	72.80	52.02	20.97
CMA-GFS nowcasts	70.61	69.02	1.60

). As a result, the P-ET was underestimated substantially (Table 2), which ultimately caused an overestimation of the drought condition. Note that the P and ET products from the CMA-GFS had already been bias-corrected by using the differences in nowcasts between the summers of 2021 and 2022. Therefore, there are errors in the summer ET nowcasts over the Yangtze River basin. This might be partly caused by un-initialized soil moisture, where the climatological soil moisture that is used in the initialization of the CMA-GFS is higher than the actual conditions during the summer drought period. Other reasons including the deficiencies in the parameterizations of ET should also be explored in the future.

To have a quantitative assessment of the flash drought nowcasting, the HIT, FAR, and ETS are calculated. Figure 8 shows that HIT is very high over specific regions, the corresponding FAR is low, and the ETS is even higher than 0.8. However, there are also some regions with very low HIT, high FAR, and low ETS, including most parts of middle and lower reaches. On average, the HIT is 0.54, and the FAR is 0.51, which suggests that only half of the flash droughts over the Yangtze River during the summer of 2022 can be nowcast and that half of the drought nowcasts are false alarms. The basin mean ETS is 0.26, which is much higher than the national mean results in a previous study [25]. The nowcast system has low HIT over parts of the middle and lower reaches, and the missing values of FAR over these regions suggest that no drought condition is nowcast. Whether the system underestimates the variability needs further investigation, but more nowcast data is also needed to provide enough flash drought samples for fully assessing the nowcast system.

This study shows that the proposed three-dimensional index captures the spatiotemporal evolution of the 2022 mega-flash drought over the Yangtze River, which can be used for meteorological flash drought monitoring. The method is also useful for high-resolution drought monitoring at smaller watersheds, where meteorological droughts are more relevant to hydrological droughts due to the dominant rainfall–runoff processes. The CMA-GFS has potential in nowcasting flash droughts, but further improvement in modelling the terrestrial water cycle in needed. This is especially important for nowcasting weather extremes, including droughts and heatwaves. The land-surface processes have long been overlooked in the NWP because of the slow evolution at the synoptic scale. However, the land-data assimilation and land-process modelling should receive more attentions in the NWP. This is especially important due to faster onset of droughts under global warming [1], more frequent heatwaves, and their compounding effects on wildfires, ecological degradations, and the drying up of surface water bodies even with groundwater recharge [35,36]. In addition, combining numerical prediction with machine learning could also improve the skill [37].

4. Conclusions

This study developed a three-dimensional meteorological flash drought index based on the percentile of 15-day moving average precipitation minus evapotranspiration (P-ET). The results show that the spatial patterns of drought characteristics are similar between one-dimensional and three-dimensional flash droughts, while the former has higher frequency than the latter. With the three-dimensional flash drought identification, it is found that the 2022 mega-flash drought started at the beginning of July in the upper reaches, and it expanded to the middle and lower reaches at the beginning of August. The flash drought over upper reaches recovered in August, but the flash drought over the middle and lower reaches did not recover in the summer. The rapid onset of the mega-flash drought is associated with strong anomalies of high-pressure system from upper-to-lower reaches across the Yangtze River basin. The mean correlation and normalized root-mean-square-error averaged over the Yangtze River basin are 0.8 and 0.85 for P-ET nowcasted by the CMA-GFS, respectively. The spatial correlations for P-ET percentiles range from 0.58 to 0.85 during 2 July–30 August, which suggests that the CMA-GFS roughly captured the evolution of the 2022 mega-flash drought. However, the mean drought intensity was overestimated due to the over-predicted ET from the CMA-GFS. The evaluation of nowcasting flash drought event shows that the hit rate, false alarm ratio, and equitable threat score are 0.54, 0.51, and 0.26, respectively. This study shows promising nowcasts of the spatiotemporal evolution of the 2022 mega-flash drought using the CMA-GFS.