The Unprecedented 2023 North China Heatwaves and Their S2S Predictability

This study unravels the characteristics, mechanisms, and predictability of four consecutive record‐breaking heatwaves hitting North China in June and July 2023. The first three heatwaves primarily influenced the northern part of North China and were accompanied by consistent anticyclonic anomalies in the upper troposphere. The anomalous anticyclone was caused by the British–Baikal corridor teleconnection along the polar front jet, particularly during the second heatwave. In contrast, the fourth heatwave was induced by a distinct low‐pressure system, attributed to the Silk Road pattern along the subtropical jet. The presence of this low‐pressure system and its interaction with atmospheric rivers and local topography led to the foehn wind, further contributing to the rise in surface temperatures. Sub‐seasonal to seasonal models can effectively predict the occurrence of all heatwaves 2–5 days in advance despite underestimating the intensity. However, models exhibit limitations in providing reliable predictions when the lead time exceeds 2 weeks.


Introduction
Heatwaves are extreme weather events characterized by exceptionally high temperatures, which can profoundly impact agriculture, ecosystem, and human society (Fontana et al., 2015;Luber & McGeehin, 2008;White et al., 2023).In recent decades, heatwaves have escalated in terms of intensity, frequency, and duration (Perkins-Kirkpatrick & Lewis, 2020; Qian et al., 2024;Rahmstorf & Coumou, 2011;Xie et al., 2023).The increased occurrence of intense heatwaves surpasses our ability to adapt, leading to severe casualties on a global scale.Notable examples include the 2010 Russia heatwave, which caused 55,000 deaths (Katsafados et al., 2014), the 2019 Europe heatwave that claimed the lives of thousands (Rustemeyer & Howells, 2021;Xu, Wang, Liu, et al., 2020), and the 2021 North America heatwave leading to deaths in the thousands (Bartusek et al., 2022).From a physical process perspective, heatwaves can be classified into those primarily driven by dynamic processes, such as adiabatic heating and horizontal heat transport, and those primarily driven by thermodynamic processes, such as the influence of soil wetness and greenhouse gases.From an atmospheric circulation perspective, both high-pressure anticyclones and low-pressure cyclones could contribute to the occurrence of heatwaves.The role of high-pressure anticyclones in heatwaves has been extensively discussed in previous research, such as increasing incoming solar radiation, inducing adiabatic warming, among others (Lin et al., 2022;Xu, Wang, Liu, et al., 2020).In contrast, the role of low-pressure cyclone in heatwaves receives less attentions.Recent studies found that the low-pressure system can also intensify temperature through temperature advection (Jiménez-Esteve & Domeisen, 2022;Keikhosravi, 2019).The diverse mechanisms underlying heatwaves highlight the need to enhance our understanding and improve the predictive skill of these phenomena.In June-July 2023, North China was struck by a series of severe heatwaves.Historical temperature records were either reached or surpassed in 22 stations, including Dagang in Tianjin and Tangjiakou in Beijing.The Beijing Nanjiao Observatory recorded temperatures exceeding 40°C for three consecutive days from June 22nd to 24th, marking the first occurrence of such consecutive high temperatures in recorded history.Since 1949, Beijing has witnessed a total of 11 days with temperatures exceeding 40°C, with 5 of them occurring in 2023.A rapid attribution study suggested that even considering the current state of climate change, the temperatures of this event were still unusual (Qian et al., 2024).Since North China is a densely populated region and serves as an important economic and political center in China, there is significant scientific and social value in investigating the mechanisms underlying the 2023 North China heatwaves.
Predicting extreme heatwaves with a sufficiently long lead time is crucial for implementing effective disaster mitigation measures.As an intermediate field between weather forecasting and climate prediction, Sub-seasonal to Seasonal (S2S) prediction aims to provide predictions that bridge the gap between weather and climate, covering time ranges from weeks to months (Vitart et al., 2017;Vitart & Robertson, 2018).As S2S models can be used as effective tools to predict extreme weather with relatively long lead times, S2S prediction has the potential to improve emergency warnings and preparedness.Previous studies assessing the performance of S2S models in predicting extreme heat events indicate that S2S models generally have a noteworthy capability to predict the occurrence up to 1 week in advance (Gibson et al., 2020;Lin et al., 2022;Mitropoulos et al., 2023;Qi & Yang, 2019;White et al., 2023).Nevertheless, the accuracy of S2S models in predicting the occurrence of heatwaves can vary and depend on the particular heatwaves being studied.As poor prediction skills will lead to insufficient disaster mitigation actions, it is crucial to investigate the ability of S2S models to predict heatwaves and explore the reasons behind their predictive performance.In this study, we aim to analyze the characteristics and circulation patterns associated with the extreme heatwaves in North China during 2023.We also evaluate the performance of S2S systems in predicting these heatwaves and explore potential factors that explain the prediction performance.

Data and Methods
In this study, we utilized the 6-hourly Japanese 55-year Reanalysis (JRA-55) data (Kobayashi et al., 2015), which is one of the best reanalysis in reproducing surface air temperature over mainland China (Zhang et al., 2021).The dataset has a horizontal resolution of 1.25°× 1.25°and consists of 37 vertical levels.It covers a time span of 66 years, ranging from January 1958 to July 2023.Variables including 2-m temperature (T2m), geopotential height, zonal (U) and meridional (V) wind, vertical velocity (ω), air temperature (T), soil wetness (SW), and specific humidity (q) were used in this study, and the 6-hourly mean variables were aggregated to compute the daily mean.Daily Climate Prediction Center (CPC) global precipitation data, with a horizontal resolution of 0.5°× 0.5°, was also used.The daily climatology is computed by averaging each field over the 30 years  for each calendar day.Daily anomalies are derived by removing the daily climatology from the original fields.A heatwave event is identified when the T2m anomaly values exceed 95th percentiles within a 5-day window centered around the calendar day of interest for at least two consecutive days (Anderson & Bell, 2011;Ma et al., 2015).
The S2S prediction data (Vitart et al., 2017) from 10 operational centers were used to evaluate the sub-seasonal predictability of the heatwaves.The analysis involved assessing real-time and reforecast variables, including T2m, q, 500 hPa, and 200 hPa geopotential height (Z500 and Z200), and U, V. Detailed information regarding ensemble number and initialization frequency of each model was listed in Table S1 in Supporting Information S1.Forecast results initialized with lead time of 2-5, 6-9, 13-16, and 21-23 days, relative to the starting day of each heatwave, were examined.Models output is interpolated onto JRA-55 grids to facilitate the analysis in this study.Climatology in models was defined using the dates from the reforecast data that approximately matched the initialization frequency in the real-time data.As a result, there may be variations in climatology among different models and reanalysis datasets.However, these discrepancies are not expected to impact our overall conclusion significantly.1a) from June 1st to July 31st for each year from 1958 to 2022 are depicted by gray lines, and the time series for the year 2023 is indicated by the blue line.The dashed gray lines above and below represent the 95th and 5th percentiles of T2ma on each calendar day, respectively.There were four consecutive heatwaves during June and July.These heatwaves are highlighted with gray shading, while the specific data points exceeding the threshold are marked with red dots.It is noted that the controlling circulation system on July 8th is similar to that of July 5th-10th (Figure S1 in Supporting Information S1).Thus, July 5th-10th is considered as a whole heatwave event in the following analysis.For simplicity, the four heatwaves are denoted as I (6.14-6.17),II (6.21-6.24),III (6.30-7.2),and IV (7.5-7.10),respectively.The highest temperature record occurred on June 22nd, with its T2ma amplitude reaching +6.5°C, which is unprecedented in history.Additionally, the peaks in heatwaves I and III also surpass corresponding historical records.Figures 1c-1f give the specific spatial distribution of T2ma during the four sub-events.Although it is obvious that all events showed significant positive anomalies across a large area in North China, a careful examination reveals that hightemperature anomalies were primarily concentrated in the northern part of North China for heatwaves I, II, and III, while anomalies for heatwave IV shifted toward the southern part of North China.The observed spatial differences in temperature anomalies imply that the synoptic circulation patterns contributing to heatwaves I-III may differ from that influencing heatwave IV.

Local Characteristics of Heatwaves
The occurrence of unprecedented heatwaves was usually closely linked to favorable circulation conditions above (Suarez-Gutierrez et al., 2020;Xu, Wang, Huang, & Chen, 2021).Figures 1g-1j show the synoptic circulation at 200 hPa during each heatwave.During heatwaves I, II, and III, North China was primarily controlled by highpressure systems.Anticyclonic anomalies can cause wind convergence in the upper level and thus lead to large-scale subsidence in situ (Holton, 1973).In contrast, heatwave IV was associated with a low-pressure system over the northern part of North China.The circulation configuration of the low-pressure system and a ridge in the upstream induced negative vorticity over North China, where heatwave IV occurred (Figure S2 in Supporting Information S1).As indicated by the geopotential tendency equation, the increased negative vorticity advection with height also excites downward motions in order to maintain the wind-pressure balance (Holton, 1973).
To further investigate the relationship between the circulation anomalies mentioned above and the occurrence of heatwaves, Figures 1k-1n illustrate the height-longitude section of temperature anomalies and wind anomalies over North China, highlighted with the green box in Figures 1g-1j.During heatwaves I-III, the wind speed was relatively low in the near-surface and was primarily characterized by westward and downward motion (Figures 1k-1m).Correspondingly, the warm temperature anomalies also covered a broad region throughout North China.In contrast, during heatwave IV, the wind speed was strong near-surface and was primarily characterized by eastward and downward motion.The wind aligned with the topographic profile and formed the foehn wind, highlighting the crucial role of mountains in forcing the wind patterns (Figure 1n).The warm temperature anomalies that were distributed precisely along the topography further confirm this hypothesis.The critical role of topography in the occurrence of heatwaves has also been emphasized in many previous studies (e.g., González-Herrero et al., 2022;Ma et al., 2014;Tran et al., 2023;Yoon et al., 2018).In addition to the critical role of circulation, extremely dry soil can also be observed prior to and during heatwaves (Figure S3 in Supporting Information S1).The local dry land conditions will reduce latent heat flux and thus create favorable thermodynamical conditions to further exacerbate the intensity of heatwaves.
The role of atmospheric rivers in contributing to heatwaves was highlighted by recent studies (e.g., Lin et al., 2022;Mo et al., 2022).Therefore, the effect of moisture depicted by vertically integrated vapor transport (IVT) (Text S1) is also investigated in this study.As shown by Figure S4 in Supporting Information S1, little water vapor was transported to North China, indicating the negligible role of IVT in heatwaves I-III (Figures S4a-S4c in Supporting Information S1).In contrast, during heatwave IV, an atmospheric river associated with the East Asian summer monsoon persisted in the south of North China (Figure S4d in Supporting Information S1).The anomalous cyclonic circulation in North China resulted in the transportation of water vapor from the atmospheric river toward the northern areas of Mongolia (Figure S4d in Supporting Information S1).The water vapor then converged on the windward side of the mountain range (Figure S5a in Supporting Information S1) and released latent heat due to the condensation.In this process, liquid water was dropped as precipitation (Figure S5b in Supporting Information S1), leaving a warm and dry (high potential temperature) air parcel in the north of Mongolia.The anomalous cyclone further transported sensible heat (Text S1) to North China (Figure S6a in Supporting Information S1), also leading to increased temperature anomalies over China.Notably, the release of latent heat in the lower troposphere can potentially amplify the low-pressure system, leading to positive feedback between the cyclonic circulation and water vapor transport (Figure S6c in Supporting Information S1).In contrast to the significant latent heating caused by water vapor transport, the contribution of the greenhouse effect of water vapor to the heatwave is limited.As shown by Figure S7 in Supporting Information S1, the integrated water vapor (IWV) (Text S1) anomalies were negative during all four heatwaves, excluding the considerable role of the greenhouse effect in the heatwave.

Linkage to the Upstream Wave Train Across Eurasia
The circulation anomalies observed over North China are not confined to the local region but can be related to the upstream wave trains across Eurasia.Figures 2a and 2b show wave trains at 200 hPa during composited I, II, III (hereinafter called C3) and IV, respectively.The wave train in C3 originated from Eastern Europe, crosses north Eurasia, and then propagated into Northeast Asia, ultimately leading to the formation of a positive height anomaly over North China.This wave train (Figure S8-S10 in Supporting Information S1) is equivalent barotropic and shares striking similarities to the recently identified British-Baikal Corridor (BBC) pattern (Xu et al., 2019;Xu, Wang, Chen, et al., 2020).The BBC pattern is the dominant waveguide teleconnection pattern along the polar front jet and has been identified as the most important large-scale circulation pattern driving the heatwaves over northern Eurasia (Kim & Seo, 2023;Li et al., 2021;Wang & Xu, 2024;Xu, Wang, Vallis, et al., 2021).The unprecedented 2019 heatwave in central Europe is a prominent example of the BBC pattern in forming extreme heat events (Xu, Wang, Liu, et al., 2020).In 2023, the BBC pattern remained in the negative phase after June 6th (Figure S11 in Supporting Information S1), which set a favorable condition for the occurrence of heatwaves over North China.In particular, corresponding to heatwave II, which was characterized by the highest T2ma in June-July 2023, the BBC pattern index (Text S2) also reached an extremely negative value during June 21st-24th (Figure S11 in Supporting Information S1).The notably amplified activity of BBC pattern partly explains why in June 2023 North China heatwaves were so extraordinarily extreme in history.
During heatwave IV, the wave pattern was shifted to lower latitudes and predominantly propagated along the subtropical jet.This wave train is similar to the previously identified Silk Road pattern (SRP; Lu et al., 2002;Enomoto et al., 2003;Kosaka et al., 2009;Wang et al., 2017), which is the dominant waveguide pattern along the subtropical jet.The low-pressure system in heatwave IV over North China coincides with the last action center of SRP over Eurasia (see Figure 4e in Wang et al., 2017).The critical role of the Silk Road pattern in causing the heatwaves in North China has also been emphasized in previous studies such as Thompson et al. (2019), Li et al. (2021) and Wang et al. (2021).In 2023, the SRP index (Text S2) showed the third lowest value in history during July 5th-10th (Figure S12 in Supporting Information S1), which was concurrent with the extremely high temperature during heatwave IV.This indicates the important role of SPR in the occurrence of heatwave IV.

Temperature Prediction
The predicted T2ma within North China (blue box in Figure 1a) with lead time of 2-5 days, 6-9 days, 13-16 days, and 21-23 days are illustrated in Figure 3.During heatwave C3 (Figure 3a), all models and members predicted significantly above-normal temperatures 2-5 days before the heatwaves, although the predicted amplitudes were slightly smaller (by 1-2°C) than the observed values.When examining forecasts with a lead time of 6-9 days, nearly all the models predicted a weaker T2ma than the values predicted at the lead time of 2-5 days.With lead times of 13-16 days and 21-23 days, the models reported a more diverse and weaker T2ma prediction, and more models (e.g., JMA, HMCR, CMA) failed to predict the correct sign of T2ma.In general, models demonstrate the capability to detect the occurrence of heatwave C3 with a lead time of 2-5 days in advance.However, they struggle to accurately predict the intensity of C3, and the predictive skills deteriorate rapidly beyond a 2-week lead time.This result agrees with Lin et al. (2022), who found that the S2S models also lack predictive skills for the intensity of the 2021 western North American heatwave with a lead time of one week.
Similar to the prediction for heatwave C3, all models indicate significantly above-normal but weaker temperature anomalies when predicting heatwave IV with a lead time of 2-5 days (Figure 3b).When forecasts are initialized 6-9 days in advance, models such as IAP, UKMO, and ECCC show insignificant ensemble mean T2ma, with some members even reporting negative T2ma values.Predicted T2ma with lead times of 13-16 and 21-23 days exhibit larger inter-model spread, and most models struggle to predict statistically positive T2ma values 21-23 days in advance.In contrast to heatwave C3, the models' performance for predicting heatwave IV has a large uncertainty and does not show significant improvement as the forecast lead time shortens.For instance, in models CMA, IAP, UKMO, ECCC, and particularly KMA, the predicted ensemble mean T2ma in the lead time of 13-16 days is closer to the observed values than 2-5 or 6-9 days.This discrepancy appears unexpected since typically forecast with initiation day closer to the occurrence of extreme event usually shows more accurate prediction skills.In short, the forecasting performance for T2ma during heatwave IV is comparable to that of heatwave C3 within a lead time of 2-5 days but becomes less predictable as the lead time extends beyond this range.

Local and Upstream Circulation Prediction
The prediction skill of heatwaves is closely tied to the accuracy of atmospheric circulation predictions in the troposphere.As such, Figure 4 illustrates the prediction skill of Za200 (geopotential height anomaly at 200 hPa) in situ and across Eurasia for each model.
When predicting heatwave C3 with a lead time of 2-5 days, most models show a close alignment between the predicted Za200 values and the observed values (red dashed line), demonstrating a higher prediction skill for local atmospheric circulation in the upper troposphere than surface temperature (Figure 4).The amplitudes of the Za200 values in the models are generally lower than the observed values, which could be an essential factor contributing to the lower-than-observed values of T2ma prediction.In addition, models with higher prediction skills for Za200 tend to exhibit better prediction skills for T2ma, indicating the importance of atmospheric circulation in determining the prediction skill of T2ma.However, specific models, such as NCEP and KMA models, predict stronger circulation anomalies but still underestimate the intensity of T2ma.This suggests that other factors, such as the imperfect representation of the downward influence of atmospheric circulation, moisture, and land processes in the models, also contribute to the weaker intensity of T2ma predictions.When examining a lead time of 6-9 days, the predicted Za200 widely deviates from the observation compared to that of 2-5 days lead time.ISAC and UKMO even forecast insignificant Za200, indicating an increased uncertainty in Za200 prediction during this stage.For lead times exceeding two weeks, the performance of the models gradually deteriorates in predicting Za200, and several models even forecast negative Za200 values.Throughout the four leading periods, there are indications that the height of bars correlates with their darkness, implying that the prediction skill of wave trains across Eurasia influences the prediction skill of local atmospheric circulation.
For heatwave IV, when considering a lead time of 2-5 days, most models show Za200 values that are close to the observations, except for IAP (JMA), which predicts a stronger (weaker) low-pressure system (Figure 4b).As the lead time extends to 6-9 days, the models generally predict a stronger low-pressure system than the observations.In line with this, the models also exhibit realistic predictions of water vapor transport within a lead time of 2-5 days, but the strength of water vapor transport diminishes significantly beyond this timeframe (Figure S13 in Supporting Information S1).The prediction of stronger circulation anomalies combined with nearly insignificant changes in T2ma bias between the 6-9 days and 2-5 days lead times raises a point of interest.This suggests that the bias in T2ma at this stage may not have a linear relationship with the prediction accuracy of the intensity of low-pressure systems.Instead, it could be attributed to the subtle configuration between the low-pressure system, topography, and the atmospheric river.This speculation can be further supported by the fact that even though almost all models report statistically insignificant positive or negative circulation anomalies with a considerably larger Za200 spread among models at the 13-16 days and 21-23 days stages, the T2ma bias remains comparable to the 6-9 days and even 2-5 days lead times.The prediction performance of regional Za200 is closely related to the varying skills in predicting upstream wave trains across Eurasia, as indicated by the darkness of the bars.Again, this highlights the crucial role of the wave train pattern in capturing local Za200 in North China.
As discussed in Section 3.2, the Eurasian wave trains, such as the BBC and SRP, play a crucial role in influencing local circulation and thereby surface temperatures over North China.It is essential to investigate how well the BBC pattern and SRP are reproduced in S2S models and how they can be related to the predictability of heatwaves.For the BBC pattern, we focus on the S2S prediction during heatwave II, as this event shows the largest T2ma amplitude and lowest BBC index.Figure S14 in Supporting Information S1 shows that almost all models are capable of predicting an above-normal BBC index with a lead time of at least 6-9 days.However, the figure consistently shows a lack of discernible prediction skills beyond 2 weeks.This changing performance in predicting the BBC pattern explains the close alignment between the predicted Za200 and the observed values within a lead time of 2-5 days, but the biases show a significant increase when the lead time extends beyond two weeks (Figure 4a).For the SRP, the models generally tend to underestimate the intensity of the SRP index (Figure S15 in Supporting Information S1).The models cannot accurately predict the intensity of the SRP index in the lead time of 2-5 days, which contrasts with the BBC pattern that the models can generally predict its activity with high accuracy with a lead time of 6-9 days.Beyond 2 weeks, nearly all models show zero prediction skills for the SRP index.The lower predictability of SRP compared to the BBC pattern can partially explain why the models show lower prediction skills for heatwaves IV compared to heatwave C3.

Summary and Conclusions
This study examines the synoptic conditions and S2S predictability for the four record-breaking heatwaves that occurred in North China in June and July 2023.Heatwaves I-III predominantly influence the northern region of North China and are characterized by a similar anomalous high-pressure system over North China.The highpressure system is strongly linked to the BBC pattern along the polar front jet, exhibiting unprecedented intensity during this stage.In contrast, heatwave IV is primarily influenced by an anomalous cyclone over North China.The low-pressure system, interacting with the atmospheric river and local topography, leads to the formation of the foehn wind and further contributes to the rise of surface temperatures.The low-pressure system is closely linked to the SRP along the subtropical jet, whose index reaches the third-lowest value in history.It is believed that the extremely intensified activities of quasi-stationary Rossby wave trains along jet streams, coupled with favorable water vapor transport and land surface conditions, led to this record-breaking heatwave.
In general, S2S models can predict the occurrence of heatwaves within a lead time of 2-5 days but underestimate the intensity of heatwaves.The bias of T2ma and the inter-model spread increases as the lead time extends, and S2S models generally do not exhibit significant prediction skills beyond 2 weeks.The limited prediction skill beyond 2 weeks can be attributed to the models' limited ability to predict the BBC pattern and SRP along jet streams.When comparing the prediction skill for different heatwaves, S2S models exhibit a higher and more consistent prediction skill for heatwaves I-III compared to heatwave IV.This is mainly attributed to the models' stronger prediction skill for the BBC pattern compared to the SRP and the increased complexity of the circulationwater vapor-topography interaction process for heatwave IV.
breaking heatwaves in June and July 2023, consisting of four sequential synoptic-scale events • The first three and the last heatwaves are controlled by distinct local circulations induced by atmospheric wave trains across Eurasia • S2S models capture the heatwave occurrences 2-5 days in advance, but predictive skill notably diminishes beyond 2 weeks Supporting Information: Supporting Information may be found in the online version of this article.

Figure
Figure1ashows the T2m anomalies (T2ma) averaged between June and July 2023.A notably positive anomaly is evident in the North China region (114°-121°E, 35°-43°N), indicating a significant occurrence of exceptionally

Figure 1 .
Figure 1.(a) Observed June and July 2023 averaged T2ma (shading, unit: °C), based on JRA-55 reanalysis.(b) Time series of averaged T2ma (unit: °C) over 114°-121°E , 35°-43°N from June 1st to July 31st for each year.Gray and blue lines denote the historical years 1958-2022 and the year 2023, respectively.The identified four heatwave events are highlighted by red dots and gray background shading.The dashed lines above and below represent the 95th and 5th percentiles, respectively.(c)-(f) T2ma (shading, unit: °C) averaged during each heatwave.(g)-(j) Geopotential height anomalies (shading, unit: gpm) and wind anomalies (vector, unit: m/s) at 200 hPa during each heatwave.(k)-(n) Height-longitude section of temperature anomalies (shading, units: °C) and wind anomalies (vector, units: m/s) over the green boxes in (g)-(j) for each heatwave.The vertical component of wind in (k)-(n) has been scaled by a factor of 100 for visual convenience.Topography is denoted as black shading.

Figure 2 .
Figure 2. Z200 anomalies (shading, units: gpm) and wave activity flux (vector, units: m 2 /s 2 , defined by Takaya & Nakamura, 2001) averaged during heatwaves (a) C3 and (b) IV.The vertical cross-sections of geopotential height anomalies along the white lines are shown in Figure S10 in Supporting Information S1.

Figure 3 .
Figure 3. Boxplot of S2S models predicted T2ma.The upper and lower edges of the box denote the first quartile (Q1) and the third quartile (Q3) of the data, respectively, and the median is drawn with a solid line inside the box.Inter-quartile range (IQR) is defined as Q3 Q1.The whiskers extend from the lowest datum above Q1 1.5IQR to the highest datum below Q3 + 1.5IQR.Ensemble members lying outside the whiskers are drawn with hollow dots.Dots filled in orange (black) are ensemble mean, which is significant (insignificant) at the 95% confidence level by the two-tailed Student's t-test.The red dashed lines denote T2ma in observation.(a) and (b) represent heatwaves C3 and IV, respectively.

Figure 4 .
Figure 4. Boxplot is the same as Figure 3, but for Za200.The red dashed lines denote T2ma in observation.The green lines indicate the prediction bias of T2ma.The height of the bar under each panel denotes the absolute value of the Za200 ensemble mean minus observation.The filled colors of bars indicate RMSE between observation and models over 60°W-140°E, 15°-90°N.