the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Reconstruction of Arctic sea ice thickness and its impact on sea ice forecasting in the melting season
Abstract. Generally, the sea ice prediction skills can be improved via assimilating available observations of the sea ice concentration (SIC) and the sea ice thickness (SIT) into a numerical forecast model to update the initial fields of the model. However, due to the lack of SIT satellite observations in the melting season, only SIC fields in the forecast model can be directly updated, which will bring about the dynamical mismatch between SIC and SIT to affect the model prediction accuracy. In order to solve this problem, a statistically based bivariate regression model of SIT, named as BRMT, is tentatively established based on the grid reanalysis data of SIC and SIT, to reconstruct the daily Arctic sea ice thickness data. Both BRMT-constructed SIT and several popular reanalysis datasets are compared to each other and validated based on available SIT observations in situ. Results show that BRMT can effectively reproduce the spatial and temporal changes of ice thickness in the melting season, and BRMT-constructed SIT is more accurate in capturing the change trend of ice thickness over a period of time, also the reconstructed SIT of one-year ice and multi-year ice types in the central Arctic and E Greenland Sea are closer to the observations. Further, as SIT from BRMT and SIC from satellite remote sensing are jointly assimilated into the ice-sea coupled numerical model, the prediction accuracy of SIC and SIT in the Arctic melting season is significantly improved, especially the SIC in the marginal ice zone and SIT in the central Arctic.
- Preprint
(2594 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on tc-2022-92', Anonymous Referee #1, 08 Jul 2022
Introduction:
The current study presents a statistical method that uses SIC and SIT from a reanalysis dataset to construct a historical SIT dataset. The idea behind this is that the SIT of the source reanalysis dataset is not accurate in the melt season as no SIT measurements in the melt season are available to feed into that reanalysis dataset, and that incorporating statistical relationships between SIC and SIT leads to an improved SIT dataset. Detailed comparisons with in situ observations and other often-used SIT datasets show that the newly developed SIT performs well. In addition, assimilation runs are performed in which only SIC or both SIC and the newly constructed SIT dataset are assimilated, which are then used to initialize 7-day forecasts. The skill of forecasts initialized from the assimilation runs in which both SIC and the new SIT dataset are assimilated is shown to be higher than forecasts initialized from assimilation runs in which only SIC was assimilated. The analysis shown is detailed and interesting, but there are several major issues that the authors have to address before I can recommend publication.
Major comments:
- Although I understand that the authors are not native English speakers, the English is poor, which makes it difficult to follow the text. I suggest the authors improve the language by consulting with an English native speaker.
- Perhaps it is due to my lack of expertise in the area, or the poor English (or a combination of these factors), but I do not fully understand the statistical model that is used to construct the SIT as described in section 3.1 and in figure 2. Other readers may have similar problems and therefore this should be improved. In particular, I do not understand the ‘linear regression for each grid point’. What is particularly confusing is that the authors write that (l. 145) ‘the linear regression process is carried out at each grid point … for each year.’ And (l. 147) ‘the corresponding SIC-SIT regression ... can be obtained for each year’. This description suggest that the linear regression is done spatially for each year (i.e. regression of SIT at a location with SIC at all other locations in a fixed year), but other text later in the paper suggests the linear regression is done at a specific grid point over the time dimension (i.e. regression of SIT at a location with SIC at that same location over time). Figure 2 also suggests that the regression is done spatially for each year, but I don’t think that is what the authors mean. Please clarify.
- The abstract should be improved, as I initially did not understand the method that the authors are introducing. I understood that the aim of the authors is to construct a historical SIT dataset, but they aim to do that based on gridded SIC and It should be explained more clearly how the SIT that is the input of the BRMT method differs from the SIT output. Also, the abstract contains several statements about improved performance, without specifying the baseline:
- 17: ‘BRMT-constructed SIT is more accurate’: more accurate than what?
- 19: ‘closer to observations’: closer than what?
- 21: ‘significantly improved’: compared to what?
The baseline should be specified. Finally, some more details on the forecasting experiments should be included in the abstract. In particular, it would be helpful to note that these pertain to 7-day forecasts (to contrast with seasonal forecasts that run for up to a year and in the context of which SIT initialization is often discussed).
Minor comments (note: there are many more grammatical errors that I don’t list below, see main comment #1):
- Some key references are omitted and should be included: Dirkson et al 2015 (https://doi.org/10.1002/2015GL063930.1) develops 3 statistical methods to generate a SIT datasets, and Dirkson et al 2017 (https://doi.org/10.1175/JCLI-D-16-0437.1) shows that one of the 3 statistical methods leads to improved seasonal forecasts of SIC. The authors should include these references.
- The authors try to highlight the importance of the newly SIT dataset by comparing forecasts initialized from assimilation run in which it is used with forecasts initialized from assimilation runs in which it is not used. While this is interesting and worth reporting, it only highlights the importance of initializing SIT versus not initializing SIT. To investigate whether or not the newly developed SIT dataset provides additional value compared to other SIT datasets (e.g. that from the reanalysis dataset that it was derived from) in the context of forecasts, an additional set of forecasts would have to be presented in which an alternative SIT dataset is used for creating the initial conditions.
- 63,66: is à was
- 255: will not à does not?
- Figure 5 (bottom map): what do the colors represent?
- Figure 8: I suggest to use a non-linear scale for the Normalized standard deviation as a) the most interesting data is where normalized standard deviation is close to 1, and because most points are located there
- 365: ‘In part with small deviation evaluation criteria value’: not clear. In the following lines (including the quoted numbers for correlation coefficient), do you only use data points with ‘small’ standard deviation, and if yes, what is the cut-off value for ‘small’?
- 477 ‘significantly smaller’: this is a bit hard to see from Fig. 13 as it is hard to compare panel a with panel b. Perhaps it would make sense to add a 3rd panel showing the difference between panel a and b? Also: it is not clear what the authors mean with ‘a long-term stable effect’ in l. 479
- Figure 14: it is hard to see what the authors refer to, as all the figures are so similar. Perhaps adding a contour line would help, but as it stands the current figure 14 does not add much to the paper. Figure 15 is much more informative
- 530 ‘are largest’: except for Exp_Ctrl
- 533: ‘Variation law’: not sure what is meant with that
Citation: https://doi.org/10.5194/tc-2022-92-RC1 - AC1: 'Reply on RC1', Lu Yang, 28 Aug 2022
-
RC2: 'Comment on tc-2022-92', Anonymous Referee #2, 21 Jul 2022
In this manuscript the authors present a statistically reconstructed dataset of Arctic summer sea ice thickness (SIT), which they create using SIT & SIC from the CMEMS Arctic reanalysis system, TOPAZ. The resulting SIT reconstruction, BRMT, is then compared with several model products, and thickness estimated by BGEP Eulerian moorings and IMB Lagrangian buoys. Finally they BRMT dataset is assimilated into a short forecasting experiment for the period September 2011.
There are some interesting ideas and concepts here that are worthy of publication in The Cryosphere. However, there are some major issues that will need to be addressed before the manuscript can be published.
Major comments/concerns
- The quality of English language used throughout the manuscript leaves a lot to be desired. As well as reducing the overall readability of the paper, there are also several cases in which it is hard to understand what is being said (or why).
- One of the main motivations listed for this study, and the BRMT dataset, is that summer satellite SIT data is not available. This issue is mentioned several times and the authors go so far as to state that it is “impossible”. However, the authors do not take into account the fact that summer satellite SIT has been being developed now for many years. The first dataset of summer Arctic radar freeboard (10-years) was published at the beginning of this year (Dawson et al., 2022). (NB. conversion from radar freeboard to SIT has also been done but that paper is still in press.)
I’m not necessarily saying that the existence of these new datasets invalidates the motivation for this study, but it should definitely be referenced and included in the discussion. How does/would the BRMT compare with the Dawson et al. (2022) freeboard? - The paper manages to be both too long and detailed, and too short and vague at the same time. By this I mean that, despite the paper being rather long, sufficient details are not provided of the methods used to create BRMT. Given the length of the paper, I recommend that the authors consider dropping Section 5 and focussing properly on the BRMT reconstructed dataset – both creation and evaluation.
- The “retrospective forecast experiments” in Section 5 are really only cursory, with less than one month of forecasts performed for only one particular year. To properly assess the impact of assimilating the BRMT SIT a much more comprehensive assessment would be needed – including the whole summer period on at least 2 different years. Many would also question the fact that BRMT uses information from the future (i.e., 2012-2018) in the reconstruction. This makes it unusable for real-world forecasting situations. How much skill would be lost if you were to only use past data?
- I struggle with the concept and motivation for the reconstructed SIT dataset. BRMT essentially uses the relationship between SIC and SIT in the TOPAZ reanalysis. Much of the motivation for using TOPAZ in this way is not included and so I am left with so many questions in my head: So why not just use TOPAZ? What extra is BRMT bringing to the table? Why do you trust the relationships in TOPAZ so much? How much difference would using a different reanalysis make to the SIT reconstruction? I’m also concerned that there is some horrible kind of circularity in the analysis here, whereby desired traits – such as the relationship between SIC & SIT – are included in the design of the system and then used as part of the evaluation.
- The comparison of the model-based and reconstructed SIT observations with the in-situ observations is either not performed carefully enough or not described adequately. I am not convinced that these comparisons are being performed or interpreted correctly for the following reasons:
- The authors do not specify anywhere how they define Sea Ice Thickness (SIT). Is it the "floe thickness" (i.e., the average thickness of all the sea ice floes present in the grid-cell) or the "grid-box-mean thickness" (i.e., sea ice volume per unit grid-cell area)? The former is certainly what your point/in-situ observations (BGEP/IMB) are measuring. However, the latter is the prognostic used in the sea ice continuum models formulation and so very likely what you are using from the model-based products (TOPAZ/PIOMAS/GIOMAS/etc.). Obviously for SIC=1 these definitions are the same but not for SIC<1. Another consideration is that if using the "grid-box-mean thickness" definition, SIT will likely be much more correlated with SIC than using the "flow thickness" definition.
- There is no discussion of how much one would expect agreement between the model and the in-situ observations. In particular the BGEP data are point observations (of ice draft converted to thickness) and are being compared with the modelled thickness in a large grid-cell. Even for the case SIC=1, when the SIT definition issue above is not present, a direct comparison is not obvious. It might be that spatial and temporal averaging of the BGEP data makes it comparable to the mean of the model ITD but that is not discussed. The same is true for the IMBs which will only ever model a single floe. Furthermore there is no discussion of the sampling issues one would expect in the IMB dataset. The IMBs are of course Lagrangian in nature and are permanently attached to the same ice floe. So one would expect changes in thickness to relatively slow, given that they are purely driven by thermodynamics. Meanwhile the dynamical nature of the Eulerian model could have huge changes in thickness from one time-step to the next. Finally, there are known sampling biases in the IMB dataset that should be discussed. IMBs are normally deployed just before the freeze-up in ice that has survived the summer. Mid-thickness floes are normally chosen – avoiding thick floes for practical reasons and thin floes to limit the chance of losing expensive equipment too quickly.
Minor comments/concerns
- In many cases results seem somewhat overstated. In particular the performance of BRMT in the East Greenland region, described as ‘outstanding’, is based upon comparison with only 2 IMB buoys – although the Lagrangian trajectory will include several individual measurements, they will only be of 2 individual ice floes! (NB. the same is true for the forecast results, which are based on only a few forecasts performed in a single year, but this is already mentioned above.)
- Too little information is provided about the model/reanalysis products being used. In particular, what observations are assimilated and what surface forcing is being for the reanalyses. This applies for both the reanalyses datasets in Section 2 and the MITgcm model in Section 3.
- RMSE of SIC is not a very good metric for sea ice forecasts because of the errors in the passive microwave satellite observations. This is particularly true in the summer when the SSMIS cannot distinguish surface melting/ponds from open water. The SSMIS accuracy is also lower in areas of low concentration or thin ice. These points are the motivation for people using sea ice “extent” to compare with satellites.
- In particular, I would drop the MIZ analysis in 5.2.1/Figure 15 because the SSMIS satellite is likely not able to resolve that. If you redid this analysis using AMSRE2 observations (which is higher resolution and more able to resolve thin/low concentration ice) then the results could be quite different.
- Some of the figures do not bring any useful information and so could either be removed or reformulated. For example, the data in Fig 13 can be understood easily from Fig 12b. Similarly, Fig 14, for which all the panels in look the same, could be improved by changing the model fields on the lower rows to be model-obs differences.
Typos and technical comments
I attach and annotated version of the original pdf with technical comments.
I do not highlight all instances where the English language needs to be improved only the cases where the language is unclear to the point that the scientific understanding is inhibited.
- AC2: 'Reply on RC2', Lu Yang, 28 Aug 2022
Status: closed
-
RC1: 'Comment on tc-2022-92', Anonymous Referee #1, 08 Jul 2022
Introduction:
The current study presents a statistical method that uses SIC and SIT from a reanalysis dataset to construct a historical SIT dataset. The idea behind this is that the SIT of the source reanalysis dataset is not accurate in the melt season as no SIT measurements in the melt season are available to feed into that reanalysis dataset, and that incorporating statistical relationships between SIC and SIT leads to an improved SIT dataset. Detailed comparisons with in situ observations and other often-used SIT datasets show that the newly developed SIT performs well. In addition, assimilation runs are performed in which only SIC or both SIC and the newly constructed SIT dataset are assimilated, which are then used to initialize 7-day forecasts. The skill of forecasts initialized from the assimilation runs in which both SIC and the new SIT dataset are assimilated is shown to be higher than forecasts initialized from assimilation runs in which only SIC was assimilated. The analysis shown is detailed and interesting, but there are several major issues that the authors have to address before I can recommend publication.
Major comments:
- Although I understand that the authors are not native English speakers, the English is poor, which makes it difficult to follow the text. I suggest the authors improve the language by consulting with an English native speaker.
- Perhaps it is due to my lack of expertise in the area, or the poor English (or a combination of these factors), but I do not fully understand the statistical model that is used to construct the SIT as described in section 3.1 and in figure 2. Other readers may have similar problems and therefore this should be improved. In particular, I do not understand the ‘linear regression for each grid point’. What is particularly confusing is that the authors write that (l. 145) ‘the linear regression process is carried out at each grid point … for each year.’ And (l. 147) ‘the corresponding SIC-SIT regression ... can be obtained for each year’. This description suggest that the linear regression is done spatially for each year (i.e. regression of SIT at a location with SIC at all other locations in a fixed year), but other text later in the paper suggests the linear regression is done at a specific grid point over the time dimension (i.e. regression of SIT at a location with SIC at that same location over time). Figure 2 also suggests that the regression is done spatially for each year, but I don’t think that is what the authors mean. Please clarify.
- The abstract should be improved, as I initially did not understand the method that the authors are introducing. I understood that the aim of the authors is to construct a historical SIT dataset, but they aim to do that based on gridded SIC and It should be explained more clearly how the SIT that is the input of the BRMT method differs from the SIT output. Also, the abstract contains several statements about improved performance, without specifying the baseline:
- 17: ‘BRMT-constructed SIT is more accurate’: more accurate than what?
- 19: ‘closer to observations’: closer than what?
- 21: ‘significantly improved’: compared to what?
The baseline should be specified. Finally, some more details on the forecasting experiments should be included in the abstract. In particular, it would be helpful to note that these pertain to 7-day forecasts (to contrast with seasonal forecasts that run for up to a year and in the context of which SIT initialization is often discussed).
Minor comments (note: there are many more grammatical errors that I don’t list below, see main comment #1):
- Some key references are omitted and should be included: Dirkson et al 2015 (https://doi.org/10.1002/2015GL063930.1) develops 3 statistical methods to generate a SIT datasets, and Dirkson et al 2017 (https://doi.org/10.1175/JCLI-D-16-0437.1) shows that one of the 3 statistical methods leads to improved seasonal forecasts of SIC. The authors should include these references.
- The authors try to highlight the importance of the newly SIT dataset by comparing forecasts initialized from assimilation run in which it is used with forecasts initialized from assimilation runs in which it is not used. While this is interesting and worth reporting, it only highlights the importance of initializing SIT versus not initializing SIT. To investigate whether or not the newly developed SIT dataset provides additional value compared to other SIT datasets (e.g. that from the reanalysis dataset that it was derived from) in the context of forecasts, an additional set of forecasts would have to be presented in which an alternative SIT dataset is used for creating the initial conditions.
- 63,66: is à was
- 255: will not à does not?
- Figure 5 (bottom map): what do the colors represent?
- Figure 8: I suggest to use a non-linear scale for the Normalized standard deviation as a) the most interesting data is where normalized standard deviation is close to 1, and because most points are located there
- 365: ‘In part with small deviation evaluation criteria value’: not clear. In the following lines (including the quoted numbers for correlation coefficient), do you only use data points with ‘small’ standard deviation, and if yes, what is the cut-off value for ‘small’?
- 477 ‘significantly smaller’: this is a bit hard to see from Fig. 13 as it is hard to compare panel a with panel b. Perhaps it would make sense to add a 3rd panel showing the difference between panel a and b? Also: it is not clear what the authors mean with ‘a long-term stable effect’ in l. 479
- Figure 14: it is hard to see what the authors refer to, as all the figures are so similar. Perhaps adding a contour line would help, but as it stands the current figure 14 does not add much to the paper. Figure 15 is much more informative
- 530 ‘are largest’: except for Exp_Ctrl
- 533: ‘Variation law’: not sure what is meant with that
Citation: https://doi.org/10.5194/tc-2022-92-RC1 - AC1: 'Reply on RC1', Lu Yang, 28 Aug 2022
-
RC2: 'Comment on tc-2022-92', Anonymous Referee #2, 21 Jul 2022
In this manuscript the authors present a statistically reconstructed dataset of Arctic summer sea ice thickness (SIT), which they create using SIT & SIC from the CMEMS Arctic reanalysis system, TOPAZ. The resulting SIT reconstruction, BRMT, is then compared with several model products, and thickness estimated by BGEP Eulerian moorings and IMB Lagrangian buoys. Finally they BRMT dataset is assimilated into a short forecasting experiment for the period September 2011.
There are some interesting ideas and concepts here that are worthy of publication in The Cryosphere. However, there are some major issues that will need to be addressed before the manuscript can be published.
Major comments/concerns
- The quality of English language used throughout the manuscript leaves a lot to be desired. As well as reducing the overall readability of the paper, there are also several cases in which it is hard to understand what is being said (or why).
- One of the main motivations listed for this study, and the BRMT dataset, is that summer satellite SIT data is not available. This issue is mentioned several times and the authors go so far as to state that it is “impossible”. However, the authors do not take into account the fact that summer satellite SIT has been being developed now for many years. The first dataset of summer Arctic radar freeboard (10-years) was published at the beginning of this year (Dawson et al., 2022). (NB. conversion from radar freeboard to SIT has also been done but that paper is still in press.)
I’m not necessarily saying that the existence of these new datasets invalidates the motivation for this study, but it should definitely be referenced and included in the discussion. How does/would the BRMT compare with the Dawson et al. (2022) freeboard? - The paper manages to be both too long and detailed, and too short and vague at the same time. By this I mean that, despite the paper being rather long, sufficient details are not provided of the methods used to create BRMT. Given the length of the paper, I recommend that the authors consider dropping Section 5 and focussing properly on the BRMT reconstructed dataset – both creation and evaluation.
- The “retrospective forecast experiments” in Section 5 are really only cursory, with less than one month of forecasts performed for only one particular year. To properly assess the impact of assimilating the BRMT SIT a much more comprehensive assessment would be needed – including the whole summer period on at least 2 different years. Many would also question the fact that BRMT uses information from the future (i.e., 2012-2018) in the reconstruction. This makes it unusable for real-world forecasting situations. How much skill would be lost if you were to only use past data?
- I struggle with the concept and motivation for the reconstructed SIT dataset. BRMT essentially uses the relationship between SIC and SIT in the TOPAZ reanalysis. Much of the motivation for using TOPAZ in this way is not included and so I am left with so many questions in my head: So why not just use TOPAZ? What extra is BRMT bringing to the table? Why do you trust the relationships in TOPAZ so much? How much difference would using a different reanalysis make to the SIT reconstruction? I’m also concerned that there is some horrible kind of circularity in the analysis here, whereby desired traits – such as the relationship between SIC & SIT – are included in the design of the system and then used as part of the evaluation.
- The comparison of the model-based and reconstructed SIT observations with the in-situ observations is either not performed carefully enough or not described adequately. I am not convinced that these comparisons are being performed or interpreted correctly for the following reasons:
- The authors do not specify anywhere how they define Sea Ice Thickness (SIT). Is it the "floe thickness" (i.e., the average thickness of all the sea ice floes present in the grid-cell) or the "grid-box-mean thickness" (i.e., sea ice volume per unit grid-cell area)? The former is certainly what your point/in-situ observations (BGEP/IMB) are measuring. However, the latter is the prognostic used in the sea ice continuum models formulation and so very likely what you are using from the model-based products (TOPAZ/PIOMAS/GIOMAS/etc.). Obviously for SIC=1 these definitions are the same but not for SIC<1. Another consideration is that if using the "grid-box-mean thickness" definition, SIT will likely be much more correlated with SIC than using the "flow thickness" definition.
- There is no discussion of how much one would expect agreement between the model and the in-situ observations. In particular the BGEP data are point observations (of ice draft converted to thickness) and are being compared with the modelled thickness in a large grid-cell. Even for the case SIC=1, when the SIT definition issue above is not present, a direct comparison is not obvious. It might be that spatial and temporal averaging of the BGEP data makes it comparable to the mean of the model ITD but that is not discussed. The same is true for the IMBs which will only ever model a single floe. Furthermore there is no discussion of the sampling issues one would expect in the IMB dataset. The IMBs are of course Lagrangian in nature and are permanently attached to the same ice floe. So one would expect changes in thickness to relatively slow, given that they are purely driven by thermodynamics. Meanwhile the dynamical nature of the Eulerian model could have huge changes in thickness from one time-step to the next. Finally, there are known sampling biases in the IMB dataset that should be discussed. IMBs are normally deployed just before the freeze-up in ice that has survived the summer. Mid-thickness floes are normally chosen – avoiding thick floes for practical reasons and thin floes to limit the chance of losing expensive equipment too quickly.
Minor comments/concerns
- In many cases results seem somewhat overstated. In particular the performance of BRMT in the East Greenland region, described as ‘outstanding’, is based upon comparison with only 2 IMB buoys – although the Lagrangian trajectory will include several individual measurements, they will only be of 2 individual ice floes! (NB. the same is true for the forecast results, which are based on only a few forecasts performed in a single year, but this is already mentioned above.)
- Too little information is provided about the model/reanalysis products being used. In particular, what observations are assimilated and what surface forcing is being for the reanalyses. This applies for both the reanalyses datasets in Section 2 and the MITgcm model in Section 3.
- RMSE of SIC is not a very good metric for sea ice forecasts because of the errors in the passive microwave satellite observations. This is particularly true in the summer when the SSMIS cannot distinguish surface melting/ponds from open water. The SSMIS accuracy is also lower in areas of low concentration or thin ice. These points are the motivation for people using sea ice “extent” to compare with satellites.
- In particular, I would drop the MIZ analysis in 5.2.1/Figure 15 because the SSMIS satellite is likely not able to resolve that. If you redid this analysis using AMSRE2 observations (which is higher resolution and more able to resolve thin/low concentration ice) then the results could be quite different.
- Some of the figures do not bring any useful information and so could either be removed or reformulated. For example, the data in Fig 13 can be understood easily from Fig 12b. Similarly, Fig 14, for which all the panels in look the same, could be improved by changing the model fields on the lower rows to be model-obs differences.
Typos and technical comments
I attach and annotated version of the original pdf with technical comments.
I do not highlight all instances where the English language needs to be improved only the cases where the language is unclear to the point that the scientific understanding is inhibited.
- AC2: 'Reply on RC2', Lu Yang, 28 Aug 2022
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
772 | 247 | 47 | 1,066 | 45 | 43 |
- HTML: 772
- PDF: 247
- XML: 47
- Total: 1,066
- BibTeX: 45
- EndNote: 43
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1