Characterizing the California Current System through Sea Surface Temperature and Salinity

Round 1

Reviewer 1 Report

This paper is devoted to a study of surface water mass characterization off the Pacific coast of the United States (California Current system). The methodology is developed using saildrone data from 2018 and 2019, and the application is first on collocated SST and SSS satellite data and the on satellite gridded fields illustrating two contrasting summers (2015 and 2021).

I have some concerns on what are the scales and structures that are targeted. For example, the saildrone data are aggregated to daily resolution which I assume is done to target rather large spatial scales (this should be > 100 km ? what is it exactly during this survey?), whereas the satellite data-based gridded SST and SSS products have spatial scales that are smaller (albeit two-dimensional, instead of one-dimensional), but at least for SSS temporal scales that are much longer. Then, of course, there are the systematic biases as well as random errors in the SST and SSS products, and the fact that there are very few SMAP data retained close to a coast (50 km) (is that the JPL V5.0 L3 dataset, or is it the RSS SMAP version used?); Thus if I understand correctly, in the product used, there might be some extrapolation for cells very close to shore which are plotted and used (or not?). Also, the average SSS field is adjusted to a climatology (here, probably a HYCOM model reanalysis) which can have some bias pattern. In some ways, I would expect the clustering to capture a mix of large-scale patterns, as well as transient meso-scale features, such as the upwelling cells or the Columbia river plume.

However, the use of the clusters then to contrast interannual situations, such as the examples of the 5-days periods during the summers 2015 and 2021, seems not to me that helpful. I don’t understand what information this brings compared to a simpler analysis of the fields and of the T-S distributions. Also, I don’t see why the clusters as defined above would be particularly relevant (maybe the ones associated to Columbia river plume and transient upwelling cells?). What new information do we get here? (for that the discussion does not help much, as it seems to a large extent to describe the water mass distribution changes that one sees from the SST, SSS maps (together with a SST/SSS scatter plot)

I also have two other major comments/questions with this paper:

A significant bias in SSS was identified for April-September 2015 in the SMAP data, and an adjusted version has since been produced (5.3). Maybe this period should be discarded from the paper discussion of T-S, or at least some comment made on whether it has any impact on the results presented. For that, the choice of having selected five days at the end of July 2015 (l. 163-164) is a bit unfortunate. On the other hand, the range of SSS values and interannual anomalies compared to this bias may be large enough that the impact of this bias may not be important (this is worth checking).

Clustering methodology (starting line 113). I wonder there is a need of respective normalization of T and S for the clustering algorithm. I noticed that the references 22, 23 deal only with temperature and do not mix two different variables… when this is the case, I would expect the choice of a specific Eucledian norm, possibly after normalizing each variable by it rms before combing them. Also, l. 125, how are the centroids defined in a two-dimensional state (what norm?)?  I read afterwards that the saildrones data are ‘reclassified’ to nearest centroid, which also implies that a norm with respective weight for T and S is defined. An other choice (as in figure 2 that density is plotted) could have been to weight SST and SS with their respective expansion coefficients, so that it gives some idea of respective contribution to surface density distribution.  What is a full covariance type (mentioned on line 122) and how the extreme values as discussed later after line 125 (and Fig. A1)?. Based on the discussion in Figure A1, it seems that these are the extremes in each property which are used to define the bounds 0 and 1. What is the advantage of that over considering rms variability/variance? (compared to A1, eq. 1 is strange or areSSS and SST normalized in it?)

I am not sure of the interest of section 3.3.1 (and figure 5) with the application of the clusters to the seasonal climatology. To a large extent, these are not the scales (in time and space) that are selected in the classification of the saildrone data, so the results are somewhat obvious.

Minor comments:

On figure 4, the near coastal band not surprisingly shows a larger scatter. This is in the region with few SMAP SSS retirevals and with larger errors due to land-sea contamination. It would help to have the data in these scatter plots normalized by the error estimates. Also, notice that the ‘mean’ SMAP field in the product used is likely to be a numerical model run, and thus the negative coastal bias for SSS mentioned on line 245 could be a model bias.

In section 3.3.2, what motivates chosing 5-day averages? Is it to remove some of the SST day-to-day variability in MUR SST (and thus, why is 5-days chosen?).

l. 74: ‘of multiple water masses…’

l. 99: what is 0.6° resolution (I guess in latitude, not longitude). This is slightly ambiguous (it is not the footprint, and not the slightly smoothed version). Is it JPL or RSS version that was used? If it is the JPL version, it has an approximate spatial resolution of 60 km.

l. 196-197.  ‘The two other clusters…’. What is the number of points in those (proportion of the total number of points). (after, this can be compared with what happens in the classification of the Remote sensing collocated data). Based on what they are, I am wondering whether a 4 clusters choice (instead of 6) would not have done as well a job, at least for the saildrone data.

l. 213: ‘Few navy clustered data…’ ?

On Fig. 2 and Fig. 3, I am rather surprised to see lots of collocated data within 50 km of the coast, where there is no SMAP L2 SSS data. Thus, these SSS data near the coast seem actually interpolated in the L3 products from offshore data.

l. 264: ‘colored…’

l. 452-453. I am quite puzzled: for me ‘higher SST values in the remote sensed data than in situ data’ unambiguously means that the SST Biases of the MUR SST product are positive, not negative. This needs to be clarified and be consistent throughout the paper.

Ref. 20: Fore, A.G.

Ref 22: Guillaume M.  should be Mazé G.

There have been other published efforts to combine T and S to describe meso-scale variability from satellite SSS data (SMAP or SMOS, or combined products on weekly or monthly time-scales), and the literature cited is not complete on that (but I am not sure that this is important to mention them in this paper).

Author Response

This paper is devoted to a study of surface water mass characterization off the Pacific coast of the United States (California Current system). The methodology is developed using saildrone data from 2018 and 2019, and the application is first on collocated SST and SSS satellite data and the on satellite gridded fields illustrating two contrasting summers (2015 and 2021).

I have some concerns on what are the scales and structures that are targeted. For example, the saildrone data are aggregated to daily resolution which I assume is done to target rather large spatial scales (this should be > 100 km ? what is it exactly during this survey?), whereas the satellite data-based gridded SST and SSS products have spatial scales that are smaller (albeit two-dimensional, instead of one-dimensional), but at least for SSS temporal scales that are much longer. 

The goal of aggregating Saildrone data into daily means is to match satellite data, not to specifically target any spatial structure. Besides, Saildrone tracks are not uniform, therefore the spatial coverage of the daily means could vary. 

Then, of course, there are the systematic biases as well as random errors in the SST and SSS products, and the fact that there are very few SMAP data retained close to a coast (50 km) (is that the JPL V5.0 L3 dataset, or is it the RSS SMAP version used?); Thus if I understand correctly, in the product used, there might be some extrapolation for cells very close to shore which are plotted and used (or not?). Also, the average SSS field is adjusted to a climatology (here, probably a HYCOM model reanalysis) which can have some bias pattern. In some ways, I would expect the clustering to capture a mix of large-scale patterns, as well as transient meso-scale features, such as the upwelling cells or the Columbia river plume.

We are using the JPL V5.0 L3 dataset, which has been specified in the main text and not only in the Data section.

The reviewer is correct in that the L3 JPL SMAP product models higher resolution data, including the coastal area. 

However, the use of the clusters then to contrast interannual situations, such as the examples of the 5-days periods during the summers 2015 and 2021, seems not to me that helpful. I don’t understand what information this brings compared to a simpler analysis of the fields and of the T-S distributions. Also, I don’t see why the clusters as defined above would be particularly relevant (maybe the ones associated to Columbia river plume and transient upwelling cells?). What new information do we get here? (for that the discussion does not help much, as it seems to a large extent to describe the water mass distribution changes that one sees from the SST, SSS maps (together with a SST/SSS scatter plot)

The reviewer is not wrong in that the results could have been the same if we analyzed temperature and salinity separately. However, the goal of the paper is not a ‘novel’ description of the California Current, but to present a method to use remote sensing data, in coastal areas, despite the limitations due to biases and uncertainties in SSS data due to land proximity. In this paper we focus on an area with previous descriptions for the exact reason that it is well sampled. We use it as a case study to test the method and its related uncertainty, but argue that this method and data could be used in other areas with limited in situ data - which is the advantage of satellite data. As such, we aim to describe well-known processes, because the goal is to demonstrate that the method can identify conditions associated with those processes giving confidence to its use in other regions, or in changes in such conditions.

I also have two other major comments/questions with this paper:

A significant bias in SSS was identified for April-September 2015 in the SMAP data, and an adjusted version has since been produced (5.3). Maybe this period should be discarded from the paper discussion of T-S, or at least some comment made on whether it has any impact on the results presented. For that, the choice of having selected five days at the end of July 2015 (l. 163-164) is a bit unfortunate. On the other hand, the range of SSS values and interannual anomalies compared to this bias may be large enough that the impact of this bias may not be important (this is worth checking).

We are not aware of this bias during April-September 2015. The authors would appreciate a reference on the issue, as in the data site there is no warning of problems with the data other than July 7th 2015, nor in known issue pages. 

We are also not aware that a new version (5.3) has been released. This new product is not available at the SMAP NASA site, an important requirement for us to use the data, therefore, we will, for the time being continue using version 5.0.

We appreciate the reviewer might be an expert on this data and aware of this issue and releases, however, this information is not accessible or available for end users (or us at the moment).

Clustering methodology (starting line 113). I wonder there is a need of respective normalization of T and S for the clustering algorithm. I noticed that the references 22, 23 deal only with temperature and do not mix two different variables… when this is the case, I would expect the choice of a specific Eucledian norm, possibly after normalizing each variable by it rms before combing them. Also, l. 125, how are the centroids defined in a two-dimensional state (what norm?)?  I read afterwards that the saildrones data are ‘reclassified’ to nearest centroid, which also implies that a norm with respective weight for T and S is defined. An other choice (as in figure 2 that density is plotted) could have been to weight SST and SS with their respective expansion coefficients, so that it gives some idea of respective contribution to surface density distribution.  What is a full covariance type (mentioned on line 122) and how the extreme values as discussed later after line 125 (and Fig. A1)?. Based on the discussion in Figure A1, it seems that these are the extremes in each property which are used to define the bounds 0 and 1. What is the advantage of that over considering rms variability/variance? (compared to A1, eq. 1 is strange or areSSS and SST normalized in it?)

For the clustering analysis, there was a need to normalize the T and S values for one not to dominate the clustering given their big difference in magnitude ranges. We calculate a normalization by scaling the range of values from 0 to 1 and use that normalized T-S space to do all the clustering and classification analysis. We agree that there are many ways to this normalization, but we have chosen one that is simple, as it still retains the linear change in variables, and also clearly represents the magnitude and variability we can see in the data, to avoid confusion or misinterpretation. 

We have clarified in the text about the full covariance type, since it is meant to be a full covariance matrix type that is one of the clustering parameters. In this case it means that the two variables (or axes) are allowed to be correlated. 

I am not sure of the interest of section 3.3.1 (and figure 5) with the application of the clusters to the seasonal climatology. To a large extent, these are not the scales (in time and space) that are selected in the classification of the saildrone data, so the results are somewhat obvious.

Indeed the scale at which the climatology is calculated is different from the saildrone data. The purpose of the climatology is to show that the several years of satellite data shows a similar pattern than the two years of saildrone data. Furthermore, doing it in daily scales produces much noise, while a climatological mean is a common practice in earth science.

Minor comments:

On figure 4, the near coastal band not surprisingly shows a larger scatter. This is in the region with few SMAP SSS retirevals and with larger errors due to land-sea contamination. It would help to have the data in these scatter plots normalized by the error estimates. Also, notice that the ‘mean’ SMAP field in the product used is likely to be a numerical model run, and thus the negative coastal bias for SSS mentioned on line 245 could be a model bias.

Indeed, the coastal errors are not unexpected, as discussed in the manuscript. We have added a few sentences to specify the error that could arise, as the reviewer points out, from the modeling of L3 the L3 product. However, we are not clear what would be the advantage of normalizing the plots by error estimates. It is not the purpose of this manuscript to evaluate or validate the coastal SSS data, and the purpose of figure 4 is to illustrate the magnitude of the differences in the coastal regions. An in-depth analysis, validation and discussion of the coastal biases are found in other studies and out of the scope of this manuscript.

In section 3.3.2, what motivates chosing 5-day averages? Is it to remove some of the SST day-to-day variability in MUR SST (and thus, why is 5-days chosen?).

The following explanatory text was added to the manuscript: A 5-day period was chosen to illustrate interannual variability; a compromise between smoothing the day-to-day variability and still capturing synoptic events like upwelling.

1. 74: ‘of multiple water masses…’

Done

1. 99: what is 0.6° resolution (I guess in latitude, not longitude). This is slightly ambiguous (it is not the footprint, and not the slightly smoothed version). Is it JPL or RSS version that was used? If it is the JPL version, it has an approximate spatial resolution of 60 km.

Thank you for pointing this out. It is an error. The resolution of L3 JPL SSS is 0.25degrees, in both latitude and longitude. We have now corrected this in the manuscript.

197. 196-197.  ‘The two other clusters…’. What is the number of points in those (proportion of the total number of points). (after, this can be compared with what happens in the classification of the Remote sensing collocated data). Based on what they are, I am wondering whether a 4 clusters choice (instead of 6) would not have done as well a job, at least for the saildrone data.

While the two clusters mentioned could be combined into one cluster or, as the reviewer suggested, eliminate them in favor of only four clusters, we didn’t choose the number of clusters by hand. The clustering algorithm used here selects the optimal number of clusters to use to divide the data. This led to those two clusters having a similar amount of data point that closer clusters, as seen from Figure 2, and representing a significant range of SSS and SST conditions. Furthermore, a visual inspection of different numbers of clusters showed that 6 clusters indeed make more sense to describe the tilted V shaped observed, supporting the algorithm selection. 

1. 213: ‘Few navy clustered data…’ ?

Rephrased to ‘Fewer data belonging to the navy cluster…’

On Fig. 2 and Fig. 3, I am rather surprised to see lots of collocated data within 50 km of the coast, where there is no SMAP L2 SSS data. Thus, these SSS data near the coast seem actually interpolated in the L3 products from offshore data.

Indeed, the L3 data has data near the coast, although it is the one with the largest errors. The details of the modeling that lead from L2 to L3 is in the SMAP user guide [20], but we consider its discussion beyond the scope of this manuscript. 

1. 264: ‘colored…’

Done

453. 452-453. I am quite puzzled: for me ‘higher SST values in the remote sensed data than in situ data’ unambiguously means that the SST Biases of the MUR SST product are positive, not negative. This needs to be clarified and be consistent throughout the paper.

In the specified line, and throughout the manuscript we have changed the word ‘bias’ when we mean ‘difference’, and specify the direction of the difference when needed to avoid this confusion.

Ref. 20: Fore, A.G.

The document cited does not include a G. We leave this revision to the editorial staff, as the citation is done correctly.

Ref 22: Guillaume M.  should be Mazé G.

Thank you for this correction.

There have been other published efforts to combine T and S to describe meso-scale variability from satellite SSS data (SMAP or SMOS, or combined products on weekly or monthly time-scales), and the literature cited is not complete on that (but I am not sure that this is important to mention them in this paper).

We aim to provide relevant citations to the study, and R. Sabia and team have written several or those. If the reviewer considers there is another study that is relevant to this analysis and we are missing it, we will appreciate letting us know.

Reviewer 2 Report

This manuscript use a classic oceanographic SST-SSS diagram framework to describe variability in surface ocean conditions, identifying distinctive water masses。SST and SSS data derived from satellit remote sensing MURSST V4.1 L4 and SMAP V5.0 L3 dataset, were aggregated into domains using a cluster algorithm over a T-S diagram, to describe surface characteristics of the CCS, validating them with in situ data from uncrewed Saildrone vessels.

The analysis results in this manuscript demostrate the capbility of remote sensed SST and SSS data  identify anomalous conditions in the CCS and describe this variability in a larger context than only to limited in situ data.

The significance of content are clear, the method is reasonable, and the results are sufficient to support the conclusions. 

There are  several minor questions to be modified or explained with more details.

1. This paper said “present a new methodology”，but MURSST V4.1 L4 and SMAP V5.0 L3 dataset have been available for years，cluster algorithm and SST-SSS diagram framework are both classic technology，therefore, the innovation of the method proposed in this paper needs to be explained more clearly.

2.  line 18, "SST and SST data" is wrong

3.  line 138, "Figure A1" is not found

4.  line 308, "a couple of turquoise points... is likely wrong", but a close examination of the SSS data in Figure 7C and the SST data in Figure 7D shows that these two turquoise points are correct. 

5. line 323, "at high temporal resolution", but the temporal resolution of SMAP SSS L3 data is low.

Author Response

This manuscript use a classic oceanographic SST-SSS diagram framework to describe variability in surface ocean conditions, identifying distinctive water masses。SST and SSS data derived from satellite remote sensing MURSST V4.1 L4 and SMAP V5.0 L3 dataset, were aggregated into domains using a cluster algorithm over a T-S diagram, to describe surface characteristics of the CCS, validating them with in situ data from uncrewed Saildrone vessels.

The analysis results in this manuscript demostrate the capbility of remote sensed SST and SSS data  identify anomalous conditions in the CCS and describe this variability in a larger context than only to limited in situ data.

The significance of content are clear, the method is reasonable, and the results are sufficient to support the conclusions. 

We appreciate the reviewer’s encouraging words.

 There are  several minor questions to be modified or explained with more details.

1. This paper said “present a new methodology”，but MURSST V4.1 L4 and SMAP V5.0 L3 dataset have been available for years，cluster algorithm and SST-SSS diagram framework are both classic technology, therefore, the innovation of the method proposed in this paper needs to be explained more clearly.

We have rewritten the first paragraph of our discussion and modified the abstract to address this point. We explain in more detail what we mean by new methodology (combining TS diagrams for surface data with clustering technique) and for which purpose (bypass the large uncertainties of SSS data in coastal areas).

2. line 18, "SST and SST data" is wrong

Done.

3. line 138, "Figure A1" is not found

Apologies. Figure A1 was submitted separately, but it doesn’t seem to be added to the final file. It is now included in the main file to avoid this problem again.

4. line 308, "a couple of turquoise points... is likely wrong", but a close examination of the SSS data in Figure 7C and the SST data in Figure 7D shows that these two turquoise points are correct. 

They are indeed classified correctly. We reworded as: 

‘Note that there are a couple of turquoise points in southern California, but the T-S diagram (Figure 7B) shows that they are better (visually) aggregated with the orange cluster.’

5. line 323, "at high temporal resolution", but the temporal resolution of SMAP SSS L3 data is low.

We removed this comment because we indeed do not demonstrate this point in the study. Our intention was to refer to the capacity of SSS data to describe variability in daily resolution, vs. other data with extensive spatial coverage like field campaigns that only acquire data every few months, for example. We removed this comment to not add confusion or superfluous text.

Reviewer 3 Report

The comments are given in pdf file attached.

Comments for author File: Comments.pdf

these are in the pdf file attached

Author Response

Effort and knowledge is utilised to use high number of sea surface in- situ and satellite data that could have been analysed with usual oceanographic methods, probably resulting with the same type of classification as given in the table 1.

The results of the method have shown an average description of oceanographic conditions and processes, and better description could be done using vertical temperature and salinity profiles (surely some vertical profiles must have been measured), TS diagrams and a few appropriate satellite images. The development of oceanographic processes could be better explained by temporal analysis at some particular places. Comparison of the two years summer season could be better shown by comparing a few characteristic SST and SSS satellite images.

I do not see the advantage of this kind of statistical method in comparison to usual oceanographic analysis. For example, what is the purpose of detecting areas under anthropogenic pressure with this method, when such locations are known in advance. Or, upwelling could be recognised from satellite images or from temporal evolution of in-situ data at some places.

If the authors would be able to detect or confirm some rare or unknown process in the ocean, then this method would be especially worth of using it.

Although I do not like this approach very much, the method may be useful when the number of data from unexplored area become huge.

Anyhow, advantages of this method should be much better documented.

The authors have not pointed to different possibilities for using this method but have mentioned the management purpose, although without examples.

The reviewer is not wrong in that the results could have been the same if we analyzed temperature and salinity at depth to describe this region. However we disagree in two points:

i) We are presenting a new set of data with different coverage, i.e., surface, that is not similar to other in situ or at-depth instruments. That similar description of conditions is achieved is a confirmation that surface conditions reflect subsurface ones, and it is a replication of previous results (based only on a few months of data decades ago). We disagree that only rare or unknown processes are worth describing and that research should be limited to follow classic methods and data. Understanding how remote sensing is connected to vertical structure is a major area of research; it provides both spatial and temporal sampling for monitoring key changes in upwelling, a vital component in studying climate impacts in coastal areas.  

ii) More importantly, the goal of the paper is not a ‘novel’ description of the California Current, but to present a method to use remote sensing data, in coastal areas, despite the limitations due to biases and uncertainties in SSS data due to land proximity. In this paper we focus on an area with previous descriptions for the exact reason that it is well sampled. We use it as a case study to test the method and its related uncertainty, but argue that this method and data could be used in other areas with limited in situ data - which is the advantage of satellite data. As such, we aim to describe well-known processes, because the goal is to demonstrate that the method can identify conditions associated with those processes giving confidence to its use in other regions, or in changes in such conditions.

We presented this description and argued that given the readily available satellite data, it can be used in near-real time for management decisions, however, we do not have examples of such. We added some potential uses, instead.

References are not consistent, and should be checked. For example the same author is once mentioned as Fore, A.G. and another time as Fore, A. (in refs 19 and 20). Hristova (ref 30) is without a name.

We have checked the references. References 19 and 20 are named exactly as in the manuscript submitted, one as Fore A.G. and the other as Fore, A. We have corrected reference 30 that was indeed missing a name by mistake.

Conclusions are drastically short. Main results must be clearly mentioned in Conclusions. If there is a particular value of the method in this oceanographic context it must be mentioned in Conclusions. 

The conclusions have been rewritten. They are still short, as we aim to avoid repeating points already made in discussion, but have revised the text to expand on the value of the method. 

High temporal resolution is mentioned few times, at the end even in Conclusions. Why nothing is shown on a temporal scale?

We removed the phrasing of high temporal resolution in some locations because we indeed do not demonstrate this point in the study. Our intention was to refer to the capacity of SSS data to describe variability in daily resolution, vs. other data with extensive spatial coverage like field campaigns that only acquire data every few months for example. We remove this comment to not add confusion or superfluous text.

Some errors:

Line 18 correct twice mentioned SST, one should be SSS

Done 

Lines 22-23, line 68, line 71, line 131, line 138... and elsewhere throughout the whole text, change "remote sensed" to "remote sensing", or to "remotely sensed"

Done

Line 66 change "ecologically and societally relevant processes" to "anthropogenically influenced processes"

We mean processes that are relevant for society and the ecosystem (river outflow, land run-off, and coastal upwelling, among others), not processes that are influenced by humans, although some of these can be. Therefore, the current wording is adequate.

Line 82 is it "uncrewed" or "unmanned"?

Uncrewed is an accepted non-gender-specific word to describe this vessel. While unmanned is the term still used in industry, and it is technically meant to be genderless, it is by many (including the authors) considered gender-specific and gender-biased. As such, we chose not to use it. 

Line 103 is it "array" not "xarray" and delete unnecessary bracket after

It is “xarray”, but the work package was added afterwards to avoid confusion.

The extra parenthesis was deleted.

Line 149 " Classification on" change to "Classification of"

Done

Line 175 is it "vortex" or "vertex" or just V shape.

It is a description of the data distribution, which is a tilted V-shape. Vortex or vertex do not apply in this case.

Line 264 "coloreded" change to "coloured" or "colored" 

Done

Reviewer 4 Report

 The manuscript by García-Reyes et al presents a new and interesting approach to T-S concepts and satellite data to identify different water types.  The analysis is clearly described and well presented.  I have a few overarching concerns relating to some specific wording and clarification of implications for future application, but believe the work will be a nice reference after addressing these concerns.  It is my hope that these comments will improve the ultimate impact of the paper.

First, I have a concern related to the broad usage of the term “bias” (particularly lines 20, 204, 24, 442, 473 but also elsewhere) where I feel “uncertainty” or something similar would be more appropriate.  To me, a measurement error consists of a bias (something more systematic) and a random uncertainty.  Bias to me refers to a more systematic tendency.  While residual cloud (SST) and proximity to land (SSS) can indeed lead to biases, I believe the impact of resolution (e.g. 205) is more random and is a source of uncertainty.  As another specific, on lines 455-6, I don’t see how regridding would cause a true “bias”.  If that is truly meant, please clarify how that is possible/expected.

Closely following on resolution, I would like to see a bit more discussion/clarification regarding the interpolation (line 153) in the regridding of the SST product to the SSS grid.  By traditional interpolation, I think of just trying to sample the inherent higher resolution of the SST data at the coarser grid points of the SSS product.  Is this regridding truly a “match” as said on line 456?  This approach tends to preserve the resolution difference of the SST and SSS data.  Arguably, averaging the SST data over the SSS grid would provide more of a “match.”  I do not advocate redoing the analysis with the average, but I would like to see some discussion of what impact this difference would have.  Ideally, it would be very nice to show just what impact the change to averaging would have on the classification error (Table 2) and the collocated T-S analysis (Fig. 1E).  On a related note, near line 462, why is this a resolution issue with MUR?  To me, the issue is with the SSS resolution and the need to regrid to that coarser resolution.

My final overarching comment is that I would really like to see some expanded discussion about some of the additional uncertainties and limitations of applying the approach in the future.  In particular, frequent mention is made of the desire to provide information at high temporal resolution (e.g. line 474), but I think more must be discussed about the tradeoffs and limitations with this, particularly as highlighted by the issue of the “blue” classification type and the time-scale of the corresponding feature.  Could the authors speak a bit more to the tradeoffs between training requirements, additional factors potentially influencing the classifications, and the ultimately applicable time scales?  Some more specifics:  Can more guidance be provided on the training required – presumably enough to capture all the possible regimes?  Would separate seasonal analyses be required?  Is the resolution/accuracy of remotely sensed data sufficient?  What would have happened had the Saildrone data only been available in the year of the large marine heat wave?  On line 458 it is said that coastal data could be excluded, but what other challenges be introduced by doing so?  What other factors could lead to uncertainties, particularly those related to more dynamic air-sea-land interactions as opposed to broader water properties?  Line 33d speaks to “many processes” but perhaps more specifics could be suggested.  For example, no mention is made of precipitation, but could an intense precipitation event alter the salinity characteristics enough to influence the classification?  How would diurnal warming be a factor?  (Diurnal warming is mentioned on line 247, but I’m not sure it is referenced in the most appropriate context – see specifics below. 

I understand that it could be difficult to be completely definitive on this matter, but at a minimum it could be clarified what factors would need to be studied in more detail before applying.  Ultimately, what are the characteristics that must be achieved for successful application of the technique?  The last sentence of the conclusions touches on these, but I think a bit more would be valuable.

Other specific minor comments:

Line 48:  dynamics -> dynamical

Line 102:  How were data matched to L2 data when MUR is already a L4 project?  Please clarify.

Lines 137-8:  recommend clarifying this is the “collocated” data

Line 156:  Clarify these are the points matching the Saildrone data – to distinguish from all the data considered next.

Line 190:  I fully understand the point, but while “clearly separated” there are still some exceptions like the turquoise points off of SF Bay.  These are points potentially impacted by other factors like alluded to above.

Fig. 4:  Units could be added to the axes.  Also, tan -> than in the caption.  For D, it could also be explicitly clarified that the symbol color is related to the distance from shore as in the other panels.

Lines 242-3:  For the coastal issues, is this all proximity to land?  Could not resolution also be a factor in the coastal region where things can also be quite variable?

Line 247:  I’m a little unclear how diurnal warming could be a large factor here given that the Saildrone data were averaged over daily time scales and the MUR data is a “foundation” product where observations with larger expected diurnal warming were (at least in earlier versions) were excluded from product development.

Line 264:  coloreded - > colored?

Author Response

The manuscript by García-Reyes et al presents a new and interesting approach to T-S concepts and satellite data to identify different water types.  The analysis is clearly described and well presented.  I have a few overarching concerns relating to some specific wording and clarification of implications for future application, but believe the work will be a nice reference after addressing these concerns.  It is my hope that these comments will improve the ultimate impact of the paper.

We thank the reviewer for their encouraging words, and appreciate their constructive feedback.

First, I have a concern related to the broad usage of the term “bias” (particularly lines 20, 204, 24, 442, 473 but also elsewhere) where I feel “uncertainty” or something similar would be more appropriate.  To me, a measurement error consists of a bias (something more systematic) and a random uncertainty.  Bias to me refers to a more systematic tendency.  While residual cloud (SST) and proximity to land (SSS) can indeed lead to biases, I believe the impact of resolution (e.g. 205) is more random and is a source of uncertainty.  As another specific, on lines 455-6, I don’t see how regridding would cause a true “bias”.  If that is truly meant, please clarify how that is possible/expected.

We agree with the reviewer that the word bias generates confusion, as we are utilizing it to describe errors and/or uncertainties in many cases. We have revised the text to reflect when we mean error/uncertainty to differentiate when we actually mean bias.

Closely following on resolution, I would like to see a bit more discussion/clarification regarding the interpolation (line 153) in the regridding of the SST product to the SSS grid.  By traditional interpolation, I think of just trying to sample the inherent higher resolution of the SST data at the coarser grid points of the SSS product.  Is this regridding truly a “match” as said on line 456?  This approach tends to preserve the resolution difference of the SST and SSS data.  Arguably, averaging the SST data over the SSS grid would provide more of a “match.”  I do not advocate redoing the analysis with the average, but I would like to see some discussion of what impact this difference would have.  Ideally, it would be very nice to show just what impact the change to averaging would have on the classification error (Table 2) and the collocated T-S analysis (Fig. 1E).  O

Thanks for pointing that out, the word ‘interpolation’ in L153 was an error, indeed it was meant averaged, not interpolation. 

On a related note, near line 462, why is this a resolution issue with MUR?  To me, the issue is with the SSS resolution and the need to regrid to that coarser resolution.

This statement referred, as you pointed out, to the re-gridded MURSST data. We have clarified that in the text now.

My final overarching comment is that I would really like to see some expanded discussion about some of the additional uncertainties and limitations of applying the approach in the future.  In particular, frequent mention is made of the desire to provide information at high temporal resolution (e.g. line 474), but I think more must be discussed about the tradeoffs and limitations with this, particularly as highlighted by the issue of the “blue” classification type and the time-scale of the corresponding feature.  Could the authors speak a bit more to the tradeoffs between training requirements, additional factors potentially influencing the classifications, and the ultimately applicable time scales? 

 Some more specifics:  Can more guidance be provided on the training required – presumably enough to capture all the possible regimes?  Would separate seasonal analyses be required?  Is the resolution/accuracy of remotely sensed data sufficient?  

What other factors could lead to uncertainties, particularly those related to more dynamic air-sea-land interactions as opposed to broader water properties?  

Ultimately, what are the characteristics that must be achieved for successful application of the technique?  The last sentence of the conclusions touches on these, but I think a bit more would be valuable.

Thank you for this comment. We have revised the discussion and added a final paragraph to include most points raised here. In addition, we revised the conclusions to further address these points. We agree they have improved the manuscript. 

We have left out a discussion on high temporal resolution because our intention was to refer to the capacity of SSS data to describe variability in daily resolution, vs other data with extensive spatial coverage like field campaigns that only acquire data every few months for example. We removed this comment to not add confusion or superfluous text.

What would have happened had the Saildrone data only been available in the year of the large marine heat wave?  

This is a concern with all descriptions of data; we acknowledge this by selecting months where we do have two years of data, and enough latitudinal coverage for the classification to not be skewed by one year or one short-lived condition. We revised the description of the Saildrone data to clarify this point.

On line 458 it is said that coastal data could be excluded, but what other challenges be introduced by doing so?  

We expanded on this statement, as those points should be excluded as outliers in the description of the classified data, as in any other analysis. We did not exclude them to be able to identify them as outliers.

Line 33d speaks to “many processes” but perhaps more specifics could be suggested.  For example, no mention is made of precipitation, but could an intense precipitation event alter the salinity characteristics enough to influence the classification?  

We do not mention precipitation explicitly here because this is not a process only of coastal areas, however, we do mention river outflow and land runoff which vary with precipitation. Furthermore, during the summer in this region, precipitation is not common and therefore not a factor to consider. In the winter, we expect lower salinity due to precipitation to generate a number of points with said characteristics that would influence, but also determine the clustering classification. We include a brief comment on this in the discussion when talking about other seasons.

I understand that it could be difficult to be completely definitive on this matter, but at a minimum it could be clarified what factors would need to be studied in more detail before applying.  

We revised this section to include what would be a factor to study in more detail. 

Other specific minor comments:

Line 48:  dynamics -> dynamical

Done

Line 102:  How were data matched to L2 data when MUR is already a L4 project?  Please clarify.

Thanks for pointing that out, that L2 referred to the SSS data. Now it is clarified in the text.

Lines 137-8:  recommend clarifying this is the “collocated” data

Done

Line 156:  Clarify these are the points matching the Saildrone data – to distinguish from all the data considered next.

Thanks for pointing this out. Done.

Line 190:  I fully understand the point, but while “clearly separated” there are still some exceptions like the turquoise points off of SF Bay.  These are points potentially impacted by other factors like alluded to above.

The paragraph has been reworded to include these exceptions. 

Fig. 4:  Units could be added to the axes.  Also, tan -> than in the caption.  For D, it could also be explicitly clarified that the symbol color is related to the distance from shore as in the other panels.

Done

Lines 242-3:  For the coastal issues, is this all proximity to land?  Could not resolution also be a factor in the coastal region where things can also be quite variable?

Yes, the land contamination is due to its resolution, as a large footprint (low resolution) would likely to contain land (or similar) features near the coast. This has been clarified in the text.

Line 247:  I’m a little unclear how diurnal warming could be a large factor here given that the Saildrone data were averaged over daily time scales and the MUR data is a “foundation” product where observations with larger expected diurnal warming were (at least in earlier versions) were excluded from product development.

This argument has been corrected and updated. Indeed, diurnal warming would not be a problem in the MURSST data. 

Line 264:  coloreded - > colored?

Done

Round 2

Reviewer 1 Report

please, see attached document.

Comments for author File: Comments.pdf

Mostly fine

Author Response

We appreciate the reviewer’s comments, and the clarification on those not adequately addressed in the previous version.

Please see the attached file for the detailed response.

Author Response File: Author Response.pdf