4.1. Time-Series Comparison
Figure 1 presents retrograde and prograde seasonal terms of GAO and HAM computed from different GRACE solutions and the hydrological LSDM. The χ
P and χ
R parts reveal similar amplitudes within each GRACE solution; however, for GAO and the LSDM-based HAM, χ
R terms exhibited visibly stronger amplitudes than those observed for χ
P terms. With the new GRACE solutions, only the JPL and GFZ series revealed a reduction in amplitudes, whereas very little amplitude change was detected for other GRACE data. This was also revealed by the standard deviation (STD) values (see
Table A1 in the
Appendix A). Notably, a reduction of STD for both χ
P and χ
R was observed for JPL and GFZ, whereas we noted an increase in this parameter for other solutions.
Figure 2 compares the mean χ
P and χ
R values with ranges between minimum and maximum for the RL05 and RL06 solutions. Updating some background models and processing algorithms in the GRACE RL06 data resulted in increased compliance of HAM between individual solutions (indicated by reduced range), especially for the χ
R term. However, there was still no full agreement between GAO and mean HAM obtained from the GRACE observations. In particular, the χ
R part for the GRACE-based mean HAM data clearly underestimated seasonal variations of both GAO and LSDM-based HAM.
The χ
R and χ
P parts of non-seasonal oscillations in GAO and HAM are shown in
Figure 3. The χ
R circular term in non-seasonal variation appeared to be stronger than the χ
P term for most of the old GRACE HAM series, as indicated by the STD values presented in
Table A1. However, this was not apparent for GAO and HAM computed from the new GRACE solutions and the LSDM. The comparison of the mean χ
P and χ
R non-seasonal changes with ranges between minimum and maximum (
Figure 4) showed that with the new GRACE RL06 solutions, different estimations of HAM were more similar; however, visible discrepancies were still present. Nevertheless, the HAM from the mean of all new GRACE solutions seemed to be more consistent with GAO and LSDM-based HAM than the HAM from any single GRACE solution.
As shown in
Figure 3, the non-seasonal oscillations in GAO and HAM were characterized by both long-term and short-term oscillations. The main contributors to long-term non-seasonal variations in HAM are groundwater changes [
57] and mass loss of ice sheets and glaciers caused mainly by the warming climate [
57,
58,
59]. Other contributors include core-mantle coupling [
60] and the flattening of the inner core and its tilt angle with respect to the outer core and mantle [
61,
62]. At shorter timescales (a few years or less), the main contributors to PM changes are atmosphere and land hydrosphere [
4,
63]. Keeping this in mind, we now decompose GAO and HAM series into long-term and short-term variations with periods of <730 days (
Figure 5 and
Figure 6) and >730 days (
Figure 7 and
Figure 8), respectively.
The comparison of
Figure 1 and
Figure 2 with
Figure 5 and
Figure 6 and values in
Table A1 shows that the seasonal variations (
Figure 1 and
Figure 2) appeared to have weaker amplitudes than non-seasonal short ones (
Figure 5 and
Figure 6). However, previous research [
4,
63] emphasized that the land hydrosphere had the highest impact on PM excitation at seasonal time scales. Similar to the non-seasonal variations (
Figure 3 and
Figure 4), for shorter non-seasonal periods obtained from old GRACE data, the χ
R produced higher amplitudes than χ
P, which was especially evident for JPL RL05 and ITSG-Grace2016. With the new GRACE solutions, these characteristics were less apparent. We observed a decrease in amplitudes and STD in the new HAM series and noted that this change was most evident for χ
R terms of JPL- and ITSG-based excitation functions. Most of the short-term non-seasonal variations computed from old GRACE solutions had amplitudes comparable or larger than the amplitude variability observed in GAO, especially in the χ
R part, whereas both the new solutions and the LSDM rather underestimated GAO amplitudes (
Figure 5 and
Figure 6). The comparison of the mean χ
P and χ
R short-term non-seasonal changes with ranges between minimum and maximum (
Figure 6) shows that results from the new GRACE solutions were more consistent in the χ
R part, but visible differences between particular solutions remained for the χ
P part, despite decreased amplitudes.
The χ
P and χ
R parts of long-term non-seasonal GAO and HAM are presented in
Figure 7 and
Figure 8. In general, comparison of new GRACE data with old data revealed that the amplitudes of longer oscillations changed less than those for shorter periods. The magnitude of HAM was affected only slightly for the JPL and GFZ solutions. Notably, HAM series computed from the LSDM revealed an overestimation in the amplitudes of observed PM excitation, whereas they visibly underestimated them in the case of shorter variations (
Figure 5). Regardless of whether old or new GRACE data were used, for the χ
R term, GRACE-based HAM series were characterized by higher STD and bigger amplitudes than GAO. For the χ
P part of the oscillations, the STDs of HAM were more consistent with the STDs of the reference series. The small amplitude change obtained after updating GRACE models from RL05 to RL06 resulted in a small increase in consistency between different HAM estimations (
Figure 8). Different GRACE solutions were revealed to be more consistent for the χ
R term than for the χ
P term.
We now extend our assessment of variability of time series shown in
Figure 1,
Figure 2,
Figure 3,
Figure 4,
Figure 5,
Figure 6,
Figure 7 and
Figure 8 by introducing a more detailed analysis of their STDs. The STD values for each oscillation are given in
Table A1 in the
Appendix A. To compare the STD of different HAM with STD of reference GAO series, we computed percent error in STD as follows:
where positive results indicate higher STD for HAM series and negative results indicate higher STD for GAO (
Table 1). For seasonal variations, almost all GRACE solutions underestimated the STD of HAM as each value of STD error is negative (except χ
R for JPL RL05), and a higher disagreement with GAO was observed for the χ
R term. This result corresponds with
Figure 1, which reveals that seasonal amplitudes of χ
P for GAO were visibly stronger than for GRACE-derived HAM. Conversely, in the long-term non-seasonal spectral band, GRACE-based HAM series overestimated the STD of the reference data. For non-seasonal and non-seasonal short-term variations, the results are mixed and depend on the solution considered. In general, the highest STD agreement between HAM and GAO was obtained in the non-seasonal spectral band, whereas the lowest was found for seasonal oscillations. Taking into consideration absolute values of STD error, we can generally conclude that, with the new GRACE RL06 data, a percent error of STD has decreased for long-term non-seasonal changes, but increased for short-term oscillations.
Finally, to quantify the change of STD in HAM from new GRACE solutions compared with the old solutions, we computed the percentage of STD change as follows:
where positive values indicate an increase of STD and negative values indicate a decrease of STD (
Table 2). Besides seasonal changes for CSR, CNES/GRGS, and ITSG solutions, all HAM from new GRACE solutions had a decreased STD for all oscillations, and the highest change was observed for JPL and GFZ solutions. These results correspond with findings from our previous work [
64], in which we compared equatorial components of HAM derived from different GRACE solutions. In that study, we noted that JPL data revealed the highest STD and amplitude change in RL06 compared with RL05 of all evaluated solutions for both seasonal and non-seasonal HAM variation.
4.2. Agreement between HAM and GAO
We now analyse the agreement between different HAM series and GAO by computing correlation coefficients, relative explained variance, and standard deviation of differences between HAM and GAO, for each oscillation separately (
Figure 9,
Figure 10,
Figure 11 and
Figure 12 and
Table A2,
Table A3,
Table A4 and
Table A5 in the
Appendix A). The correlation plots were supplemented with information about the critical value of the correlation coefficient and the standard error of the difference between two correlation coefficients. The critical value of the correlation coefficient can be determined based on interpretation of autocorrelation function and statistical tables for Student’s
t-test [
8]. The autocorrelation function of the time series shows how rapidly the series changes and does not consider its previous values [
65,
66]. It shows the length of the time lag after which an evaluated series becomes decorrelated, meaning that the correlation between one series and the same series shifted with a lag is zero. Usually, the decorrelation time is assumed using one of the following four methods: (1) it is assumed to be the time required for autocorrelation function drop to the first zero crossing, (2) it is assumed to be the time required for autocorrelation function drop to 1/e, (3) it is assumed to be double the time required for autocorrelation function drop to 1/2, or (4) it is assumed to be double the time required for autocorrelation function drop to 1/e [
66]. In this paper, we first determined decorrelation time using method (2) and then computed a number of independent points by dividing the number of series points by the decorrelation time. Finally, the critical value of correlation coefficient for computed number of independent points and assumed significance level (here we assumed 95%) was read from the Student’s
t-test statistical tables. The standard error of a difference between two correlation coefficients was computed as
, where
is a number of independent points.
Relative explained variance (Var
exp) is commonly used for estimating the discrepancy between a model (evaluated series, here HAM) and actual data (reference series, here GAO). It is the part of the total variance of reference data that is explained by evaluated data. The percentage of GAO variance explained by HAM was computed here as follows:
where Var
(GAO), Var
(HAM), and Var
(GAO-HAM) are variance of GAO (reference series), variance of HAM (evaluated series), and variance of a difference between GAO and HAM, respectively. The higher the value of Var
exp, the stronger the association between the evaluated and reference series. The optimal Var
exp value is 100%, which means in our case that HAM explains the full variance of GAO. For this case, the differences between reference and evaluated series are the same for all points of the time series. In other words, the variance of these differences is equal to zero. As the variance of differences
between GAO or HAM increases, the Var
exp decreases. A similar method of quality assessment of the time series is to compute standard deviation of differences (STD
diff) between reference and evaluated data. The lower the STD
diff values, the better the evaluated series (optimal value of STD
diff is 0). However, because the computation of both Var
exp and STD
diff is based on STD or variance of differences between GAO and HAM, these parameters show the same characteristics of assessed data, and can lead to the similar conclusions. Therefore, we focused here only on detailed Var
exp analysis. The values of STD
diff are given in
Table A2,
Table A3,
Table A4 and
Table A5 (
Appendix A)
We also look closer into the magnitude of improvement in correlation and variance agreement between GRACE-based HAM and GAO after releasing new GRACE solutions. To quantify the level of increase or decrease of these parameters in each new solution, we computed the percentage change of these parameters (Corr change and Var
exp change). Using these parameters, we examined how much correlation coefficients and Var
exp for RL06 were improved compared with correlation coefficients and Var
exp for RL05. The computations were performed for each pair of GRACE new and old solutions: CSR RL06 versus CSR RL05, JPL RL05 versus JPL RL06, GFZ RL06 versus GFZ RL05, CNES/GRGS RL04 versus CNES/GRGS RL03, and ITSG-Grace2018 versus ITSG-Grace2016. We used the following equations (
Table 3 and
Table 4):
where
means absolute values,
is a mean correlation coefficient for all old GRACE solutions,
is a mean relative explained variance for all old GRACE solutions, positive results indicate improvement, and negative results indicate deterioration.
Table 5 and
Table 6 were supplemented with values for the mean GRACE correlations and variances.
Figure 9 shows that, in the seasonal part of the spectrum, the CSR RL06 solution provided the highest correlation of HAM with GAO (0.87) for the χ
R term, whereas the best result for the χ
P part was obtained for CNES/GRGS RL03 and CNES/GRGS RL04 (0.74 and 0.73, respectively). For both χ
R and χ
P terms, HAM from JPL solutions (both RL05 and RL06) provided the worst agreement with reference data, with correlations far below the required level for statistical significance. Very low correlation coefficients for JPL data were a result of phase differences between the two sinusoids representing seasonal variations for GAO and JPL-based HAM. We found an increase of correlation coefficients with GAO for HAM from new GRACE data compared with the older ones (except ITSG for the χ
R term and JPL for the χ
P term). For the χ
R term, the biggest correlation improvement was detected for the JPL solution, whereas for the χ
P term, correlation improvement was highest for ITSG (
Table 3). Similar to the correlation results, the highest relative explained variance was obtained for CSR RL06 in χ
R (51%) and for CNES/GRGS RL03 and CNES/GRGS RL04 in χ
P (49% and 52%, respectively). The highest variance improvement was detected for the JPL solution in the χ
R part and for the ITSG solution in χ
P part of the seasonal variation (
Table 4). It should be noted that the HAM function obtained from LSDM revealed a very good agreement with reference GAO series, but only in the χ
R part (correlation coefficient of 0.74 and relative explained variance of 54%).
In the non-seasonal spectral band, for χ
R, the correlation and variance agreement improved notably in all new solutions except for GFZ; however, for χ
P, a visible correlation increase was observed only in CSR (
Figure 10,
Table 3). The best correlation agreement with GAO for both χ
P and χ
R terms was detected for CSR RL06 (0.66 and 0.68 for χ
R and χ
P, respectively) and ITSG-Grace2018 (0.64 and 0.59 for χ
R and χ
P, respectively). The comparison of relative explained variances provided similar conclusions, with the best results for CSR RL06 (42% and 44% for χ
R and χ
P, respectively) and ITSG-Gace2018 (40% and 28% for χ
R and χ
P, respectively). However, despite some improvement in the results using the new GRACE RL06 data, the variance agreement was still unsatisfactory as none of the values exceeded 45% and many negative variances occurred. In the χ
P part of the hydrological excitation, LSDM-based HAM provided results comparable with those obtained for CSR RL06 and ITSG-Grace2018 (correlation coefficient of 0.64 and relative explained variance of 29%).
For short-term non-seasonal variation in HAM, only CSR RL06 and ITSG-Grace2018 provided correlation coefficients visibly above the statistical significance level for both χ
R and χ
P circular terms (
Figure 11). These GRACE models are the only ones to show a visible improvement in HAM correlation with GAO compared with the previous releases (
Table 3). Notably, HAM from the GFZ RL05 solution was characterized by the best correlation agreement in the χ
P part of the hydrological excitation (0.65); however, in the χ
R part, this consistency was poor. HAM computed using the new GFZ and JPL solutions was revealed to decrease correlation with GAO compared with the older GRACE data, and this was visible especially in χ
P (
Table 3). Similar findings were shown from an analysis of relative explained variances—in χ
R, the best results were provided by CSR RL06 and ITSG-Grace2018 (40% and 31%, respectively), whereas in χ
P, the highest variances were obtained for HAM derived from GFZ RL05, CSR RL06, and ITSG-Grace2018 (25%, 29%, and 20%, respectively). Apart from χ
P for JPL and GFZ data, all new solutions revealed an improvement in variance agreement with GAO compared with the old solutions.
The data presented in
Figure 12 suggest that long-term changes in the hydrological part of PM excitation are much better determined by GRACE observations than the shorter period variations. This is unsurprising as short oscillations are more diverse than long ones, which could affect the magnitude of correlation coefficients between HAM and GAO. Moreover, the fact that GRACE solutions are provided in only monthly intervals with occasional gaps in data might also have contributed. Similar conclusions were drawn from our previous work [
10], where we focused on analysis of the equatorial components (χ
1, χ
2) of PM excitation. As shown in
Figure 12, for the χ
R part, almost all new solutions provided correlation agreement between HAM and GAO at the level of 0.80 or more (except GFZ RL06), with the best results for JPL RL06, CNES/GRGS RL04, and ITSG-Grace2018 (correlation coefficients equal to 0.87, 0.86, and 0.85, respectively). Similarly, for χ
P, correlation coefficients exceeded 0.70 for all HAM functions and the highest value was obtained for CSR RL06 (0.81). However, a notable correlation improvement in new GRACE data was detected only for CSR and JPL solutions in χ
R, and for CSR and CNES/GRGS in χ
P (
Table 3). In terms of relative explained variance, the most satisfactory results were for JPL RL06 (69%) and ITSG-Grace2018 (64%) for χ
R, and CSR RL06 for χ
P (57%). Notably, for the χ
R part of excitation, we observed visible variance improvement in HAM functions obtained from the GRACE RL06 solutions compared with the RL05 solutions (
Table 4). Similar to the seasonal χ
R and non-seasonal χ
P terms, long-term χ
P oscillations were very well modelled by the LSDM.
The more detailed analysis of values given in
Table 3 allows us to conclude that CSR is the only solution for which there was correlation improvement between HAM and GAO in RL06 compared with RL05 for both χ
P and χ
R and for all oscillations. The biggest correlation improvement was detected for the JPL solution in the χ
R seasonal term (186% improvement) and for the ITSG solution in the χ
P seasonal term (83% improvement). Our previous work, which evaluated χ
1 and χ
2 [
64], also showed that, for the seasonal part of HAM, the JPL solution was distinguished by the greater improvement in consistency with GAO than other solutions. This was mainly because new the JPL data were smoother than the older data. Notably, we also observed a decrease in correlation, which mostly affected the GFZ and JPL solutions and was highest for the χ
P part of the non-seasonal short-term variations (maximum decrease for JPL—103% and for GFZ—99%). With the release of new GRACE solutions, HAM from the CNES/GRGS series produced the lowest correlation change, which did not exceed ±9% (except for the χ
P term in the non-seasonal long-term variation). Taking into account the mean correlation change, the seasonal correlations showed the greatest improvement (23% improvement for χ
R and 20% improvement for χ
P), whereas the smallest change was observed for the long-term non-seasonal spectral band. We noted a correlation decrease for the χ
P part of the non-seasonal short-term variation (32% decrease) and for the χ
P part of non-seasonal variation (6% decrease).
Table 4 shows that only HAM derived from the CSR solution improved variance agreement with GAO for both χ
P and χ
R and for all oscillations. However, in contrast to the correlation changes, there were fewer cases in which there was a notable decrease in variance agreement between HAM and GAO (only for χ
P in non-seasonal short-term variation for JPL and GFZ, and for χ
P in non-seasonal variation for JPL). The highest variance improvement for both χ
P and χ
R was observed for CSR in the non-seasonal spectral band and for JPL in seasonal spectral band. Taking into consideration the mean variance change, apart from the χ
P term of short-period changes, we detected a notable variance improvement for all oscillations, which exceeded 100%. These findings reveal that the mean variance improvement was several times greater than the mean correlation improvement.
At this point, it should be mentioned that our validation of χ
P and χ
R terms in GRACE-based HAM was based on correlation coefficients with GAO and relative explained variances, but there are other metrics that can be helpful in such an evaluation. The use of coefficient of determination (R
2) values from a linear regression analysis is a common method for such interpretation of the results. The R
2 value shows quality of the model’s fit to the data and ranges between 0 and 1 (with 1 being the best value). R
2 is often used in validation of hydrological models [
67], but it can be also used in assessment of other data types. Therefore, we computed R
2 between GAO and different HAM and showed the results in
Figure 13 (for seasonal and non-seasonal changes) and
Figure 14 (for non-seasonal short-term and non-seasonal long-term changes). However, the analysis of R
2 led us to the similar conclusions as a comparison of correlation coefficients. In particular, for seasonal variations, the highest R
2 values were obtained for CSR RL06 (for χ
R), GFZ RL06 (for χ
R), and CNES/GRGS RL04 (for χ
P); for non-seasonal and non-seasonal short-term variations, the highest R
2 values were observed for CSR RL06 (for χ
R and χ
P), ITSG-Grace2018 (for χ
R), and GFZ RL05 (for χ
P); for non-seasonal long-term changes, the highest R
2 values were obtained for JPL RL06 (for χ
R), CNES/GRGS RL03 and RL04 (for χ
R), ITSG-Grace2018 (for χ
R), CSR RL06 (for χ
P), GFZ RL05 (for χ
P), and LSDM (for χ
P).
In contrast to the previous studies, which demonstrated good results for χ
2 and clearly worse results for χ
1 [
3,
7,
9,
10,
11,
12,
13,
14,
23,
27,
63,
68], it was difficult to conclude whether the χ
P term or the χ
R term is better modelled by GRACE as results depended on the solution and oscillation considered. Moreover, in terms of correlation and variance agreement with GAO, there was no noticeable difference between χ
P and χ
R. To make the results more general and readable, we next computed the mean correlation and variance from all old GRACE solutions, and then the mean correlation and variance from all new GRACE solutions, for different variations separately (
Table 5 and
Table 6). The tables are supplemented with corresponding values for HAM from LSDM. In terms of the correlations, the differences in results between the χ
P and χ
R terms was small and did not exceed 0.1, and for new GRACE RL06 solutions, they were even smaller (
Table 5). Similarly, the variance explained values obtained for these new GRACE data reveal that HAM agreed better in the χ
R than in the χ
P term only for non-seasonal short-term variations (11% and −7%, respectively) (
Table 6). For other oscillations in HAM determined from GRACE RL06 data, the variance results were almost identical for both terms as the variance differences between χ
P and χ
R did not exceed one percentage point. For GRACE RL05 solutions, these discrepancies were slightly higher, but it remains unclear which term is better modelled by GRACE RL05 data. For HAM obtained from the LSDM, the discrepancies in results between χ
P and χ
R were more evident—both correlation coefficients and relative explained variances were higher for the χ
R term in the seasonal spectral band, whereas for all non-seasonal variations, these parameters were higher for the χ
P term. For LSDM, the maximum correlation difference between χ
P and χ
R reached 0.63 (for seasonal changes), whereas the maximum variance difference was equal to 129 percentage points (for non-seasonal long-term changes).
Table 5 and
Table 6 also indicate that, in general, the highest improvement in correlation with the new GRACE RL06 data was obtained for seasonal variation. However, in the short-term spectral band, the correlation with GAO dropped for the χ
P term, which contributed to a slight correlation decrease in this term for non-separated non-seasonal change (short-term plus long-term). In terms of relative explained variances, the agreement with GAO was improved in almost all considered spectral bands (except χ
P in non-seasonal short-term variations), which might be a result of amplitude changes in HAM computed from the new GRACE data. Nevertheless, it should be kept in mind that such conclusions are general and are based on the mean of GRACE solutions. The results for various solutions differed from each other (see
Figure 9,
Figure 10,
Figure 11 and
Figure 12).