For each YC, the background temperature is calculated using three steps. Firstly, at each latitude band and pressure level, the daily zonal mean temperature (T‾d) is calculated by averaging the temperature profiles at the ascending and descending nodes, respectively. This largely removes the gravity waves, non-migrating tides, and long-period planetary waves. Here, each latitude band has a width of 10° with centers offset by 5° from 80° S to 80° N. Secondly, linear regression is performed on T‾d at each node and is formulated as follows: 1T‾d=T‾d0+ktUT+T‾res.

Here, T‾d0 is the mean temperature in each YC, tUT is the universal time (in days), and k represents the linear variation in T‾d in each YC. After removing T‾d0 and the linear variation (ktUT) from T‾d, we get a residual temperature T‾res for each YC. Thirdly, tidal fitting is performed on the T‾res of both nodes and is formulated as follows: 2T‾res=T‾bk+∑n=13ancos⁡(nωtLT-φn).

Here, ω=2π/24 is the rotation frequency of Earth (in radians per hour); tLT is the local time (in hours); and an and φn are the respective amplitudes and phases of migrating diurnal (n=1), semidiurnal (n=2), and terdiurnal (n=3) tides. Now, T‾bk excludes atmospheric waves and is regarded as the mean temperature.

Figure  shows that the center date of each YC shifts forward about 1 month from 2002 to 2023. This forward drift propagates the seasonal variation in the temperature into T‾bk. This could further distort the long-term trend calculated from T‾bk and can be removed with the help of MSIS2.0. This is because MSIS2.0 has assimilated the SABER temperature profiles during 2002–2016. The climatological temperature of MSIS2.0 coincides with that of SABER within the uncertainties of ∼ 3 K in the MLT region (Emmert et al., 2021). The detailed procedure for removing seasonal variations is described below.

Firstly, we calculate the mean temperature of MSIS2.0. The temperature profiles (at 15 longitudes and 24 LTs each day) are calculated from MSIS2.0 under conditions of lower solar activity (F10.7 = 50 SFU, where SFU denotes solar flux units) and geomagnetic quiet time (Ap = 4 nT) throughout 1 calendar year. Therefore, solar and geomagnetic activities do not influence the seasonal variation or trend in the mean temperature. Then, the daily zonal mean is calculated for the temperature profiles of each day. This removes tides and long-period planetary waves. The daily zonal mean temperature in each YC is averaged to get the mean temperature (T‾MSISyear, where the superscript denotes the YC in that year). Figure a1 and a2 show the T‾MSISyear at 70° N in YC3 and at 70° S in YC6, respectively, during 2002–2023.

Secondly, we calculate the seasonal variations in each YC. The seasonal variations (ΔT‾MSISyear) caused by the forward drift in each YC in different years are quantified by the difference between the T‾MSISyear of that year and the reference year (i.e., T‾MSIS2002). For example, the difference between 2003 and 2002 is calculated as ΔT‾MSIS2003=T‾MSIS2003-T‾MSIS2002. More specifically, as T‾MSISyear does not include the year-to-year variations in temperature but depends on the temporal span of the YC only, ΔT‾MSIS2003 in YC3 represents the seasonal variation from 20 to 19 June. Figure b1 and b2 show ΔT‾MSISyear at 70° N in YC3 and at 70° S in YC6, respectively, during 2002–2023. It is evident that the forward drift in the YC induces temperature variations of ± 20 K at 70° N/S from 2002 to 2023 and should be removed before we determine the long-term trends in SABER temperature.

Finally, we correct the mean temperature. The corrected mean temperature (T‾bcrtyear, shown in Fig. d1 and d2) is obtained by removing ΔT‾MSISyear from T‾bkyear. This removes the seasonal variation caused by the forward drift in the YC from 2002 to 2023. Moreover, T‾bcrtyear retains the long-term trend in the mean temperature. We note that, after removing ΔT‾MSISyear, the T‾bcrtyear covered by each YC can be represented by its center date and half-span in the reference year (Table ). Table  also lists the approximate season related to each YC.

To calculate accurate trends in the MLT region, multiyear variations should be removed properly. The multiyear variations in temperature in the MLT region could be the solar cycle, with a period of about 11 years (Beig et al., 2008; Tapping, 2013; Forbes et al., 2014; Gan et al., 2017; Qian et al., 2019), and the influences from below, such as the El Niño–Southern Oscillation (ENSO) with varying cycles of around 2–7 years (Domeisen et al., 2019; Li et al., 2013, 2016; Randel et al., 2009). The solar cycle can be represented by the solar radiation flux at 10.7 cm (i.e., F10.7, with units of SFU = 10-22 Wm-2Hz-1) (Tapping, 2013). ENSO is represented by the multivariate ENSO index (MEI) (Domeisen et al., 2019). The multiple linear regression (MLR) method is effective to separate the long-term trend in temperature from the variations caused by the solar cycle, ENSO, and the quasi-biennial oscillation (QBO). The MLR equation is formulated as follows: 3Y(t)=c0+c1t+c2F10.7(t)+c3ENSO(t)+ε(t).

Here, Y represents the mean temperature at year t from 2002 to 2023; c0 represents a mean state of Y; c1 is the long-term trend in Y; and c2 and c3 represent the contributions from the solar cycle and ENSO, respectively. The terms for F10.7 and ENSO are included in Eq. () for the purpose of correctly determining the long-term trend, but they are not considered further in this work. Here, we note that both the trends (linear variations) and quasi-periodic variations represent the natural variations in the predictors. These natural variations might influence the trends and variations in temperature. Thus, MLR is applied to characterize the contributions from the natural variations in predictors. The resulting trends in the temperature then exclude the trends inhibited in the predictors. This is the trend studied in this work. Otherwise, if these predictors are detrended, their residuals are used in the MLR. Thus, the resulting trends in temperature may include the trends inhibited in the predictors.

The statistical significance of the regression coefficients is measured using a Student t test and the variance–covariance matrix in Eq. (). Specifically, in Eq. (), the sampling points are 22 and the predictor variables are 4. This results in 19 degrees of freedom. Consequently, the critical value is ∼ 2.0 based on a Student t test at a confidence level of 95 % (Kutner et al., 2005). This signifies that, with reference to a 95 % confidence level, the magnitude of the regression coefficient should be at least 2.1 times greater than the standard deviation.

The mesopause temperature (T‾msp) is defined as the minimum of the mean temperature. The pressure level at which the minimum temperature occurs is defined as the mesopause altitude (zmsp). Figure d1 and d2 show the mesopause altitudes calculated from T‾bkyear (black crosses) and T‾bcrtyear (red dots), respectively. We see that the mesopause altitudes calculated from T‾bkyear and T‾bcrtyearare nearly identical in the first several years but exhibit discrepancies over the later several years. This implies that the seasonal variation caused by the forward drift in the YC affects the mesopause altitudes to some extent. Moreover, the mesopause altitudes exhibit larger variabilities in the southern summer polar region (YC6) compared with the northern summer polar region (YC3). Figure  shows the date–latitude distributions of the mesopause temperature (T‾msp) and altitude (zmsp) calculated from T‾bcrtyear. We note that zmsp is initially defined for the pressure level (Fig. d). To compare our work with previous studies, zmsp is interpolated onto the geometric heights in Fig. .

Previous SABER studies have often discarded high latitudes, possibly due to insufficient LT coverage that induces uncertainties in the mean temperature estimation. A major advantage of binning the SABER temperature based on YC is that an accurate mean temperature can be obtained. Thus, the latitudinal variations in T‾msp and zmsp at high latitudes can be thoroughly studied. Firstly, we focus on the YCs in northern summer and winter (i.e., YC3 and YC6, respectively), as the summer mesopause at high latitudes is more sensitive to the summer-to-winter circulation (Dunkerton, 1978; Qian et al., 2017). In YC3 (YC6), T‾msp and zmsp generally decrease from 50° S to 80° N (50° N to 80° S). We note that T‾msp has local minima around the Equator throughout the 22 years in YC3 and YC6 and is the coldest at the highest latitudes of the summer hemisphere. The zmsp is the lowest at 40–60° N/S throughout the 22 years. Besides the latitudinal variations, T‾msp and zmsp also exhibit multiyear variations. For example, T‾msp is colder around the Equator during the solar minima (i.e., 2007–2008 and 2019–2021) in YC3 and YC6. In YC6, the lower zmsp at southern higher latitudes might be related to the warm phase of ENSO during 2002–2005 and 2016–2019.

In YC2 and YC5, the latitudinal variations in T‾msp and zmsp are almost hemispherically symmetrical. The T‾msp is the coldest around the Equator and the warmest at the highest latitudes. The zmsp is the lowest at lower latitudes and the highest at the highest latitudes. In YC1, T‾msp and zmsp share similar latitudinal variations in winter (YC6); the difference is that T‾msp is warmer in YC1 than in YC6, while zmsp is higher in YC1 than in YC6. In YC4, T‾msp and zmsp share similar latitudinal variations in summer (YC3); the difference is that T‾msp is warmer in YC4 than in YC3, while zmsp is higher in YC4 than in YC3. In YC1–YC2 and YC4–YC5, multiyear variations in T‾msp exhibit clear a dependence on the solar cycle. At lower latitudes, T‾msp values are colder during the solar minima (i.e., 2006–2010 and 2017–2021). At high latitudes, T‾msp values are warmer during the solar maxima (i.e., 2002–2005, 2012–2014, and after 2021). However, it seems that the multiyear variations in zmsp are not as obvious as those in T‾msp. These multiyear variations are considered in Eq. () to separate the long-term trend in T‾msp correctly, but they are not considered further in this work.

Trends in the corrected mean temperature and their significance for each YC are shown in Fig. . These trends are generally larger at high latitudes than those at lower latitudes within the six YCs. Moreover, the trends show both hemispheric symmetry and asymmetry approximately in the high-latitude MLT region.

First, we describe the hemispheric symmetry in the trends. In YC1 and YC4 and above 10-3 hPa, the cooling trends are ≥ 2 K per decade at latitudes higher than 40° N (YC1) and 40° S (YC4), respectively. Around 10-4 hPa, the cooling trends reach their peaks of ≥ 6 K per decade. In addition, there are also warming trends of ≥ 2 K per decade at latitudes higher than 30° S (YC1) and 30° N (YC4), respectively. Above the mesopause, there are cooling trends of ≥ 2 K per decade observed within the latitude range of 20–50° S for YC5 and 20–50° S for YC2. Additionally, in the region just below 10-3 hPa, there are warming trends of ≥ 2 K per decade at latitudes of 50–80° N for YC5 and 50–80° S for YC2. In YC3 and YC6, the cooling trends of ≥ 2 K per decade shift upward from the mesopause at 80° N (YC3) and 80° S (YC6) to 10-4 hPa at 50° S (YC3) and 50° N (YC6). There are also cooling trends of ≥ 6 K per decade at the high latitudes of the summer hemisphere. Meanwhile, the coldest trends are ≥ 10 K per decade just below 10-4 hPa and at 80° N/S. Although the cooling trends in the MLT region have been reported extensively at lower and middle latitudes (Beig et al., 2003; Laštovička, 2023), the extreme cooling trends at high latitudes and above the summer mesopause have not yet been reported. We note that the systematic error in the SABER operational processing is unknown. Its impacts on the credibility of the trends derived here will be discussed in Sect. .

Next, we describe the hemispheric asymmetry in the trends. In YC1 and YC4, the cooling trends of ≥ 2 K per decade in YC1 extend to a wider latitude range (20° N–80° S) than those in YC4 (30° S–80° S) above 10-3 hPa. The insignificant warming trends of ≥ 2 K per decade can be seen in the stratosphere at latitudes higher than 60° N in YC1 but at 45–60° S in YC4. In YC5 and YC2, the cooling trends of ≥ 2 K per decade can be seen around the stratopause at 30–50° S (YC5) but below the stratopause at 30–50° N (YC2). In YC3 and YC6, the significant warming trends of ≥ 2 K per decade in YC6 are stronger than those in YC3 around 0.1 hPa. In addition, the warming trends near the summer mesopause are significant in YC6 but insignificant in YC3. The simulation results in Qian et al. (2019) also demonstrated warming trends in the southern summer MLT region. Specifically, they showed significant warming trends below ∼ 95 km and cooling trends above ∼ 95 km at latitudes exceeding 45° S between November and February. In contrast, there were insignificant or warming trends at latitudes exceeding 45° N during June and July. Qian et al. (2019) attributed the warming trend in the summer mesosphere to the changing meridional circulation.

Taking advantage of the continuous long-term (22 years or equivalently two solar cycles) measurements and binning the YCs at 50° S–80° N or 80° S–50° N, the robust mean states of the mesopause temperature (T‾msp) and height (zmsp) as well as their trends and the responses of T‾msp to the solar cycle, ENSO, and QBO are quantified using MLR. Here, we focus on the mean states and trends in the mesopause temperature and altitude.

Figure a and b show the mean T‾msp and zmsp over 22 years for the six YCs. In YC1–YC2 and YC4–YC5, the mean T‾msp is in the range of 172–183 K. However, the mean T‾msp at latitudes higher than 40° N is warmer in YC1 compared with YC5, and the mean T‾msp at latitudes higher than 40° S is warmer in YC2 compared with YC4. The mean zmsp is mainly in the range of ∼ 96–102 km, but it is higher than ∼ 85 km at 40–50° N (YC1) and 40–50° N (YC4). In YC3, the mean T‾msp decreases sharply with latitude, from ∼ 180 K at 30° N to ∼ 125 K at 80° N. The mean zmsp in YC3 reaches a minimum of ∼ 85 km at 60° N. In YC6, the mean T‾msp decreases sharply with latitude, from ∼ 180 K at 35° S to ∼ 135 K at 80° S. The mean zmsp in YC6 reaches a minimum of ∼ 86 km at ∼ 50° S. The mean T‾msp (zmsp) in the northern summer polar region is ∼ 5–11 K (∼ 1 km) colder (lower) than that in the corresponding southern region. The hemispheric asymmetries of the summer mesopause temperature and altitude coincide with Xu et al. (2007), who used the SABER temperature data during 2002–2006 and showed that the mean T‾msp in the summer polar region of the NH is ∼ 5–10 K colder than its counterpart in the SH. A recent study by Wang et al. (2022), who used the SABER temperature data during 2002–2020, showed that the mean T‾msp in the summer polar region of the NH is ∼ 10 K colder than its counterpart in the SH. Moreover, the transition latitudes of the mean T‾msp (zmsp) from higher temperature (height) are 30° N in YC3 and 40° S in YC6. This coincides well with values reported by Xu et al. (2007) and Wang et al. (2022). These hemispheric asymmetries of the mean T‾msp and zmsp as well as the transition latitudes could be caused by the hemispheric asymmetry of solar radiation and gravity wave forcing (Xu et al., 2007).

Figure c shows that trends in T‾msp in YC1 and YC4 display extreme cooling (≥ 2 K per decade) at latitudes higher than 55° N/S. However, at 40° S–40° N, trends in T‾msp in YC1 show cooling, with magnitudes of ∼ 0–2 K per decade, whereas they show warming in YC4, with magnitudes of ∼ 0–1 K per decade. In YC2 and YC5, trends in T‾msp show either cooling or warming, depending on the specific latitudes and months being considered. At southern latitudes, trends in T‾msp show cooling, with magnitudes of ≥ 1 K per decade in YC2. Trends in T‾msp in YC5 change sharply from 2.0 K per decade at 45° N to -3 K per decade at 80° N. In YC3 and YC6, trends in T‾msp mainly show cooling, except for the insignificant warming trends in YC6 and at latitudes higher than 40° S. Although trends in T‾msp show warming at some latitudes in certain YCs, the all-YC mean trends in T‾msp (blue line in Fig. c) show cooling, with magnitudes of 0.3–1 K per decade at 50° S–50° N. At latitudes higher than 55° S, the insignificant cooling trends are ≤1.5 K per decade. In contrast, at latitudes higher than 55° N, the significant cooling trends are ≥ 1.5 K per decade.

To facilitate a comparison with previously reported annual and global-mean trends in the MLT region, we present the mean trends in the corrected mean temperature at 50° S–50° N and at 55–80° S or 55–80° N for the six YCs (Fig. ). The mean trends at 50° S–50° N for each YC show cooling, with magnitudes of ∼ 0.5–1 K per decade at 10–10-3 hPa. The exception is the warming trend of 0.2 K per decade around 10-2 hPa in YC1 and of 0.1 K per decade around 4 × 10-3 hPa in YC3. Above 5 × 10-3 hPa, the cooling trends increase sharply with altitude and reach ∼ 2 K per decade in YC5 and ∼ 3 K per decade in YC2 at 10-4 hPa. Compared with the situation in YC2 and YC5, the cooling trends increase more sharply with altitude in YC3 and YC6. Their magnitudes change nearly identically and are from ∼ 0.5 K per decade at 2 × 10-3 hPa to ≥ 5 K per decade at 10-4 hPa. When the mean trends at 50° S–50° N across all YCs are further averaged, we obtain an annual mean trend (blue line in Fig. a). The annual mean trend shows cooling, with magnitudes of ∼ 0.5–0.8 K per decade that vary slightly with altitude at 10 × 100–5 × 10-4 hPa.

The altitude variation and the magnitude of the annual mean trend are similar to previous results (Garcia et al., 2019; Mlynczak et al., 2022; Zhao et al., 2021). Figure 3 of Garcia et al. (2019) revealed that the global-mean (52° S–52° N) SABER temperature trends display cooling, with magnitudes of ∼ 0.5–0.9 K per decade at 10–5 × 10-4 hPa during 2002–2018. These magnitudes of these cooling trends are slightly smaller than those derived from WACCM. Table 1 of Mlynczak et al. (2022) demonstrated that the global-mean (55° S–55° N) SABER temperature also displays cooling trends, with magnitudes of ∼ 0.51–0.63 K per decade at 1–10-3 hPa. Similarly, Fig. 4 of Zhao et al. (2021) revealed that the global-mean (50° S–50° N) SABER temperature trends show cooling, with magnitudes of ∼ 0.5–0.9 K per decade at 30–105 km. At 10-4 hPa, the extreme cooling trend of 2.6 K per decade in Table 1 of Mlynczak et al. (2022) is slightly smaller than the 2.8 K per decade derived here, although within 2 times the standard deviation (blue line in Fig. a). By further examining the trends across the six YCs (Figs.  and a), it becomes evident that the extreme cooling trend is mainly attributed to the middle latitudes of the summer hemisphere (i.e., YC3 and YC6), although also partially to other months. As suggested by Mlynczak et al. (2022), the extreme cooling trend at 10-4 hPa is due to a decrease in solar irradiance that is not captured by the F10.7 index.

We note that these trends are derived from the SABER temperature. The systematic error in the SABER temperature influences the credibility of these derived trends. According to the rigorous analysis of the systematic error, the trends derived here are reliable only if their magnitudes are larger than the systematic trend uncertainty. The annual and global-mean trends that show cooling with magnitudes of 2–4 K per decade around 10-4 hPa are unreliable, as these values are in the range of the systematic trend uncertainty of 22.7 K per decade at 6.3 × 10-5 hPa and 4.5 K per decade at 2.8 × 10-4 hPa. At pressure levels lower than 10-3 hPa, the annual and global-mean trends that show cooling with magnitudes of ∼ 0.5–1 K per decade are unreliable, as these values are in the range of the systematic trend uncertainty of 3.6 K per decade around 10-3 hPa and 1.3 K per decade below 6.6 × 10-3 hPa.

These detailed comparisons showed that the trends at the pressure levels reported by Garcia et al. (2019) and Mlynczak et al. (2022) directly support the altitude variations and magnitudes of the trends derived here. Although the trends reported by Zhao et al. (2021) are for the geometric height, their altitude variations and magnitudes also agree with the trends derived here. However, these trends are unreliable, as their magnitudes are in the range of the systematic trend uncertainties. We note that the method of binning SABER observations based on the YC provides an opportunity to study the trends at latitudes higher than 50° N/S in certain months.

At latitudes higher than 50° N/S, the altitude variations in the mean trends in the six YCs (Fig. b) are seasonally symmetrical above approximately 1 hPa. The magnitudes of trends are mainly in the range of -2 to 2 K per decade below a height of 10-3 hPa. These trends are larger than the systematic trend uncertainties of 1.3 K per decade and, thus, are reliable below 6.6 × 10-3 hPa. Moreover, these trends are in the range of the systematic trend uncertainties of 3.6 K per decade and, thus, are unreliable around 10-3 hPa. An interesting feature is the warming trends of 1–2.5 K per decade at 10-2–10-3 hPa in April, August, October, and December. The altitudes of peaks in the warming trends vary from 4 × 10-3 hPa to 10-3 hPa in different months. Focusing on the latitude band of 64–70° N in June and 64–70° S in December, Bailey et al. (2021) merged the temperature data from HALO and SABER (total length of 29 years) and from HALOE and SOFIE (total length of 22 years). Their analysis revealed warming trends of 1–2 K per decade near 5 × 10-3 hPa (∼ 85 km) at 64–70° N in June and 64–70° S in December, as illustrated in Fig. 7 of their paper. The results simulated by WACCM-X showed significant warming trends at ∼ 80–95 km at latitudes higher than 45° S from November to February and close to zero or warming trends at latitudes higher than 45° N from June to July (Qian et al., 2019). The warming trends in December derived here coincide with those reported by Bailey et al. (2021) and Qian et al. (2019). The weak warming trend at 2 × 10-3 hPa in June coincides with those in Qian et al. (2019) but is much smaller than the 1–2 K per decade reported by Bailey et al. (2021). In April and October, the warming trends are hemispherically symmetrical at 10-2–10-3 hPa and reach a peak of ≥ 2 K per decade at 3 × 10-3 hPa. It should be noted that the warming trends of 1–2.5 K per decade at 10-2–10-3 hPa are in the range of the systematic trend uncertainties of 1.3 K per decade at 6.6 × 10-3 hPa and 3.6 K per decade around 10-3 hPa; thus, they are unreliable in the sense of systematic trend uncertainty. Above 10-3 hPa, the trends transit from warming to cooling.

We can see extreme cooling trends of ≥ 6 K per decade above ∼ 10-3 hPa in YC3 and YC6 as well as around 10-4 hPa in YC1 and YC4. Due to the systematic trend uncertainty, these trends are reliable around 10-3 hPa but unreliable around 10-4 hPa. These cooling trends are comparable with the global average mesosphere temperature of 6.8–8.4 K per decade derived by Mlynczak et al. (2022) after doubling the CO2 in the MLT region. However, it takes decades to double CO2. Thus, a purely radiative effect due to increasing CO2 cannot support the extreme cooling trends derived here. Mlynczak et al. (2022) proposed that the F10.7 is not a suitable proxy to indicate the effects of solar radiation on the lower thermosphere. However, the solar irradiance in the Schumann–Runge band (175–200 nm) might be responsible for the colder trend. Even so, the extreme cooling trends of ∼ 10 K per decade are still larger than those reported by Mlynczak et al. (2022). Other possible reasons for these extreme cooling trends in the high-latitude MLT region can be attributed to the dynamic feedback in the polar MLT region.

Besides the purely radiative effect on the cooling trends in the MLT region (i.e., Garcia et al., 2019; Mlynczak et al., 2022), the dynamic feedback might be another cause of the cooling trends. Based on the simplified transformed Eulerian mean (TEM) thermodynamic equation, the temperature change (ΔT) caused by dynamics can be written as follows (Eqs. 3 and 4 of Yu et al., 2023): 4ΔT=-α-1w∗S+v∗∂T‾a∂φ.

Here, α is the Newtonian cooling coefficient; w∗ and v∗ are the residual vertical and meridional velocity, respectively; S and T‾ are the static stability and zonal mean temperature, respectively; and a and φ are the Earth's radius and latitude, respectively. From Eq. (), we propose that the extreme cooling trends in the high latitudes of the summer hemispheres (YC3 and YC6) might result from the changing summer-to-winter circulation and gravity wave forcing in the MLT region. The circulation is upwelling (positive w∗) in the summer hemisphere and causes a cold summer mesosphere through adiabatic cooling. Conversely, in the winter hemisphere, the circulation is downwelling (negative w∗), leading to a warm winter mesosphere through adiabatic warming (Garcia and Solomon, 1985). A necessary condition for the extreme cooling trends at summer high latitudes is the stronger upwelling and, thus, the increasing gravity wave body force in the summer hemispheres. Previous studies have shown that the gravity wave potential energy (GWPE) in the MLT region exhibits significant positive trends at southern high latitudes in January and at northern high latitudes in July (Fig. 5 of Liu et al., 2017). The positive trends in the GWPE might enhance the strength of upwelling and, thus, result in extreme cooling trends in the high latitudes of the summer hemispheres. It should be noted that the dynamic feedback in the MLT region is only analyzed qualitatively, as quantitative analysis should be performed through model simulations. Thus, one can elucidate the physics behind the strong cooling trend in the polar MLT region.

The trends in T‾msp derived in this study are significant and mainly negative at 50° S–50° N across most YCs. The averaged trend in T‾msp for the six YCs is -0.64 ± 0.22 K per decade over 50° S–50° N. When the average is calculated over 80° S–80° N, the trend in T‾msp for the six YCs is -1.03 ± 0.40 K per decade. The cooling trend in T‾msp derived here also coincides with the -0.5 ± 0.21 K per decade trend in the mesosphere (Garcia et al., 2019), although only within 50° S–50° N. Compared with the trend derived from sodium lidar observations during nighttime only around 40° N, the trends in T‾msp from SABER are about -0.1, 0.0, -0.2, -0.8, 0.6, and -1.9 K per decade for the six YCs and have an annual mean of -0.4 K per decade. This is less than the significant cooling trend of 2.3–2.5 K per decade during 1990–2018 but is consistent with the insignificant cooling trend of 0.2–1 K per decade during 2000–2018 (Yuan et al., 2019). The comparisons of the trends in T‾msp between our results and those from satellites and ground-based observations exhibit general consistencies in the sense of the annual mean or global mean. However, the zmsp is mainly above 95 km (6.5 × 10-4 hPa), where the systematic trend uncertainties are larger than 3.8 K per decade and are larger than the trends in T‾msp. Thus, the trends in T‾msp derived here are mainly unreliable in the sense of rigorous systematic error analysis.

A notable feature is the warming trends in T‾msp with magnitudes of 0–2 K per decade at latitudes higher than 40° S in YC6. This warming trend is insignificant at a 95 % confidence level. If we change the temporal interval from 2002–2023 to 2002–2019, the trends in T‾msp are cooling with magnitudes of 1–2 K per decade. Here, we note that the year 2020 is just after the SABER temperature data were revised (version 2.08, from 15 December 2019) (Mlynczak et al., 2023). In this work, we use the SABER temperature data of versions 2.07 (before 15 December 2019) and 2.08 (after 15 December 2019). According to Mlynczak et al. (2023), the newly released data are free from algorithm instability. A recent study by Yu et al. (2023) showed that the Hunga Tonga–Hunga Ha'apai (HTHH) volcanic eruption on 15 January 2022 induced temperature anomalies of ± 10 K globally in the stratosphere and mesosphere in August. The anomalies disappeared after September 2022. This indicates that this volcanic eruption may have influenced the mesosphere temperature through circulations and waves. From the mesopause temperature of YC6 (shown in Fig. ), we see that a warmer mesopause occurred after 2020 before the HTHH volcanic eruption. On the other hand, there is no significant variation in the mesopause temperature in YC3 throughout the 22 years. Thus, the largest difference in YC6 may not have been caused by algorithm instability or the HTHH volcanic eruption but may rather have been a realistic result. As shown in Figs. d and b and reported by Wang et al. (2022), the annual variability in zmsp is ∼ 5 km in the southern high latitudes (YC6) but is relatively stable in the northern high latitudes (YC3). The large annual variability in zmsp induces a large variability in the T‾msp (indicated by the large standard deviations in the right subpanel of Fig. b). This, in turn, contributes to the large variability in the trends in T‾msp at southern high latitudes. Another possible reason is that the warming trends of 0–2 K per decade are unreliable due to the large systematic trend uncertainties in this height range.