Overview

The authors discuss that changes in phytoplankton biomass, proxied by chlorophyll a concentration (Chla), partitioned between different phytoplankton functional types (PFTs) provides not only direct ecological information but also indirect information about environmental change due to niche differences between the functional types. PFTs in optical applications are groupings of phytoplankton separable by optical signatures of pigment composition (and possibly ancillary environmental information), that however contain a reasonable degree of taxonomical and ecological information. Such partitioning of bulk Chla can in principle be applied to remote sensing, and has been demonstrated globally for open ocean areas. However, due to the limited lifespan of orbital remote sensing missions, long term analysis based on remote sensing requires merging of data from multiple sensors, including harmonization to account for differences in sensor specification, calibration, and performance. Nonetheless, the official PFT product released by the Copernicus Marine Services (CMEMS) includes appropriate merging only for data between 2002 and Apr-2016, using data based only from the Sentinel-3 mission from May-2016. This resulted in a continuity issue in the time series of PFT products from CMEMS. The authors attempted a previous harmonization method they had developed for a study in a specific ocean region, but despite it achieving a good performance on the global average, spatially resolved data over the globe showed large discrepancies between the datasets on regional scales. Therefore, the authors developed a new harmonization method that also takes as input the spatial information, in the form of geographic coordinates, in order to harmonize the data on a spatially resolved level, and the global average as a consequence.  This allowed the authors to evaluate a 2 decade long time series of PFT abundance at the global average and over selected regions. The authors then describe the temporal patterns observed and provide some comment concerning the use of this information.

The study represents further progress in an relevant research topic, in which the consortium of authors includes recognized experts and leader in the field. It provides a method that potentially improves the official PFT CMS products, and provides new information on temporal and spatial changes of optical/ecological phytoplankton groups. The written and visual presentation is mostly clear, with exceptions noted in the detailed review below.

While in general my perspective is positive, I do have concerns related to the methodology and results that can impact the interpretation in the study and I believe should be addressed for a publication. My comments are divided in three sections: Major comments and Minor comments, have comments that would be expected to be addressed by the authors for a publication, and Suggestions, which contain suggestion for improvement that need not be addressed by the authors.

Major comments

1. Source data for PFT products and the need for harmonization at the PFT product level. Section 2.1 provides the description of the database used for the PFT products, indicating that GlobColour merged product is used from Jul-2002 to Apr-2016 and OLCI data used from May-2016 to present. The justification to not use the GlobColour merged product for the whole dataset period (Jul-2002 to present) is that VIIRS/SNPP data has been identified to contain biases possibly caused by sensor degradation. I have little experience with GlobColour or the CMEMS PFT product, so it is possible I’m missing something, but I do not find that justification to be sufficient, specially considering that among the authors are the producers of the GlobColour merged dataset. For example, is it the case that GlobColour cannot exclude VIIRS/SNPP data since 2017 from the merged product calculation due to some specific reason? Or is a solution being sought but not implemented in the time frame of this study? This is relevant because if the radiometric data, base for the PFT calculation, is not harmonized between sensors, then the PFT product will have to be harmonized, which is the topic of this study. And since Xi et al. (2023) have already dealt with such harmonization, the problem seems to be known for a few years. To be clear, I believe it is fair that the study took this course, but I’m missing more discussion on the why it was so. For example, regardless of the situation with GlobColour, as discussed above, this course could be justified if PFT derived from OLCI/Sentinel-3 alone would be superior (due to higher spatial resolution and more bands) but that does not seem to be the case as PFT data derived from OLCI/Sentinel-3 data is corrected to match the PFT data derived from the merged dataset. The authors could argue that the way forward is merging of the derived product (e.g., PFT) and not of the base radiometric data as the configurations of the new sensors become more diverse, and so the future of initiatives such as GlobColour is the merging of derived products. Again, I may be missing something, but it seems to me a clearer justification is necessary.
2. Overlapping period for model calibration and partitioning of data for model calibration and validation. Section 2.1 describes that though the official PFT CMEMS product uses the merged GlobColour product only from 2002 to Apr-2016, this was extended by the authors to Apr-2017 in order to provide a year of overlap with PFT from OLCI/Sentinel-3 (May-2016 onwards) for model calibration and validation. It is possible that at least a year of overlap is necessary to fully take in account the variability caused by seasonality, but having a single year is challenging as that specific year may present conditions that are specific to that year only. Considering the argument for not using the merged product beyond 2017 discussed in the previous point, I understand that one year is all data that it is available, however, the potential impacts of this single overlap year should be discussed. A closely related issue is the data splitting between calibration and validation sets described in Section 2.2. A random subset can produce “data leakages” (information in the validation set also present in the calibration set) that produce overly optimistic performance statistics (e.g., Meyer et al., 2018). In the case of this study, the model application (May-2016 to present) is beyond the time domain of the model calibration (May-2016 to Apr-2017), so in my view an ideal validation strategy should demonstrate that the model can well extrapolate in time, which is not possible when the data is randomly split as the calibration and validation sets cover the same time period. I note that this is specially necessary because while before (regression method of Xi et al., 2023) the correction was a matter of relating two variables over a large spatial domain, now the relation depends explicitly on geographical location, for which dynamics might be changing in time. This is somewhat a conundrum considering again the data availability for the GlobColour merged product as it is (with potentially bias data from VIIRS/SNPP from 2017) and the potential need to include all seasons. Nonetheless, I believe those aspects to be very relevant and that they should be considered in the discussion.
3. Reference PFT from in situ data. Though Section 2.3 provides information about the different datasets, their sampling and partitioning, no information is given on how PFT Chla is calculated from in situ measurements. The authors have addressed this important issue in previous publications, but a summary of the method should also be included here, with citations to their previous work where the methodology is discussed in greater detail. In addition, any changes to methods or to the base data analysis methods (instruments, protocols, operators, etc) for calculation of PFT from in situ measurements should be noted, and if it is the case, discussed if they may impact the validation in Figure 3b (dataset 2, Section 3.2).
4. Calculation of spatial averages. The methods do not specify how the regional and global averages were calculated, in particular if the area contribution of each pixel was compensated for the area distortion of a regular lat/lon grid. It is possible that this was taken into account, but since it is not mentioned in the manuscript I ask the authors to confirm. This is mostly relevant for the global average, but also applicable to the regional averages. In my understanding, the average should contain a weight proportional to the cosine of the latitude to make the contributions of each pixel to the average proportional to their contributions to the surface ocean area. This would likely change considerably the global trend in diatom Chla, as the increase seem to be driven by increases at high latitudes.
5. Temporal pattern shift after the overlapping period. Despite the efforts in data harmonization, the limitations of the cal/val procedure discussed before and the validation analysis for high latitude data ask for additional support for the temporal patterns observed after the overlapping period. For example, the Chla of diatoms changes considerably from 2017 onwards (Figure 4). A corresponding change in behaviour is observed for green algae and prokaryotes. Maybe the authors can provide independent evidence sourced from the literature (specially in situ studies, if any is available) for the temporal / regional patterns they observe in given PFT, in order to provide support that the changes are not potentially influenced by the correction procedure. The two external supports provided by the authors (CMEMS trend product and the study of Van Ostend et al., 2023) concern only bulk Chla, that is, without information about increasing or decreasing abundances of specific groups. This is also relevant specially considering that validation in Figure 3b suggests poor performance of the retrievals against in situ data for the Arctic. This performance results should be considered in the discussion of the global average and Arctic average trends - currently the discussion states only, and in my opinion incorrectly, that no overestimation was observed for the validation with Arctic data in dataset 2.
6. PFT anomaly in 2023. I’m not convinced this analysis add information to the study. I would argue that in principle a given year anomaly is relevant in a specific study about that year phenomena, but in the context of global change, a given year anomaly would seem relevant to me only if the anomaly is beyond a given range of variability of the data along the 2 decade of the climatology (e.g., 2 standard deviations).
7. Relative composition. While the study provides the analysis in terms of Chla as a proxy for biomass of each PFT, one could expect that new information is available when normalizing the data to evaluate changes in relative composition. For example, even though diatoms might be increasing in the a given environment, other groups might be increasing faster and so the relative contribution of diatoms be decreasing. I believe such analysis would add more to the current study than the 2023 anomaly analysis commented above.  However, such analysis is not necessary and the study is sufficient without it.

Minor comments

1. Abstract: The OLCI acronym is not defined when used in the abstract.
2. Line 16: "the merged sensor-derived PFT" - it not clear to me what the authors are referring to. The main text makes clear this is specifically the GlobColour merged dataset, one of the inputs to the CMEMS PFT product. While understanding the limitations of space in the abstract, the authors are encouraged to modify the abstract to make it more specific.
3. Line 45 and other locations: The study is described as merging PFT products from different sensors, but I understood that what is being merged is the PFT products from two datasets: PFT from GlobColour merged sensors and PFT from OLCI/Sentinel-3. In that sense, phrasing it as merging between sensors can cause confusion.
4. Line 68: It seems the potential bias is identified specifically in the VIIRS/SNPP, and not on VIIRS/NOAA-20. If that is correct, please add the mission specification to the text.
5. Line 86: Please provide a citation (Xi et al., 2021? or “this study”, if it is the case) for the statement that PFT from OLCI/Sentinel-3 carries higher uncertainty than PFT data from merged GlobColour. A similar validation analysis to that presented in Figure 3 Section 3.2 could be provided for the OLCI data before harmonization to the GlobColour merged sensor data as supplementary material.
6. Line 101: It is not clear to me what the PFT data transformation achieves. It is stated that the log-transform and scaling/shifting achieve a normal distribution in the rage [0, 1], but a normal distribution is not restricted to a finite range and shifting/scaling does not change the data distribution type, just the magnitude of its specific parameters (e.g., mean and standard deviation). Authors are encouraged to rephrase or provide further details. In addition, they are encouraged to provide more information on the transformation used to achieve latitude and longitude data in normal distribution and in the range [-1, 1].
7. Line 123: The acronym CMEMS is used for the first time here, but not defined. Note that Copernicus Marine Service is mentioned for the first time in Line 42, where the acronym could be defined.
8. Line 175: The authors evaluate that only diatoms, dinoflagellates and prokaryotes show some relation between sensor-derived and in situ PFT data in the validation analysis presented in Figure 3b. However, in my perspective the Arctic dataset shows no relation for any group except perhaps prokaryotes. This suggests that the Arctic dataset (or region) is the most problematic (not necessarily polar regions or high latitudes in general), and where future work could focus.
9. Line 269: First use of the acronym OMI, but it is not defined in the manuscript.
10. Table 1: This is likely planned to be updated upon acceptance, but currently the documentation for the third row, concerning the new in situ PFT data states only “data to be submitted to PANGEA” and adds a citation that is not in the reference list. Considering that PANGEA offers the possibility of data publication with moratorium specifically to address situations of peer-review, I recommend that the data submission process, which might take time, be initiated and the information on the manuscript updated to reflect final citations/references and links. This point is also applicable to the data availability statement in lines 118-119 and 298-299.
11. Figure 1: The reference data set is presented in the Y-axis in subfigures (c) and (d). While this is unconventional, it is not per se an issue, however some statistics (regression slope, R2 and MARD) are not symmetrical and do depend on which variable is taken as reference. The exception is model II regression for slope when using major axis estimation (cf. Legendre and Legendre, 2012), but that is not indicated.
12. Figure 3: The plots present values of “MDPD” though the acronym is not defined in the main text or figure caption. I note that in the main text MARD is defined, so even if it is a typo, the acronym for percentage still need to be formally defined.
13. Figure 4: The indication of the regions in the map is not very legible. Though I believe there is no ambiguity where the regions are and what they cover, I think the visual presentation can be improved. The authors could consider using semitransparent filled boxes with a colour as in the legend by the side of the figure (removing the diagonal lines and names, and the grey shading of the polar areas).

Suggestions

1. Chlorophyll a is presented both with italicized (abstract) and roman (main body) "a". I suggest using the italicized form, as that is the form used in reference documents in the field of photosynthetic pigment analysis (e.g., Roy et al., 2011).
2. Communities vs Assemblages vs Ensembles vs Guild. Some terms in ecological literature have been used so loosely over the years such as to lose meaning. Currently, my opinion is to follow Fauth et al. (1996), who provides a classification and definition based on an intersection of three major areas: Phylogeny, Geography, and Resources. In their nomenclature, an ensemble is a "phylogenetically bounded group of species that use a similar set of resources within a community". Alternatively, assemblages are "phylogenetically related groups within a community". And "local guild" (though in this case, "local" seems inappropriate) is a group of "species that share a common resource and occur in the same community". While not a perfect system, at least it is a structured one. Admittedly, I have used "phytoplankton assemblage" in my own work, though "ensemble" in their nomenclature or "guild" (considering phylogeny in algae is a loose link, and that phytoplankton includes cyanos) seem superior descriptors. I encourage the authors to consider this nomenclature system.
3. "Composition structure" appears several times in the text. Consider using just "composition".
4. Line 33: The start of the phrase, “Dedicated to marine biogeochemistry, ”, is unclear to me and seems unnecessary. Maybe just use a connector like “Therefore,”?
5. Line 39: Chla was already defined in line 28.
6. Line 50: “correction skill” -> “correction procedure”.
7. Line 52: The corrected PFT time series” -> “The harmonized PFT time series”.
8. Line 54: “For a longer goal” -> “Considering that ocean colour missions are planned to be continued into the next decade and beyond, ”.
9. Line 59: "The PFT datasets with per-pixel uncertainty (product ref. no. 1 in Table 1) have been generated by the EOF-PFT approach adapted based on the version proposed by Xi et al. (2021). The updated algorithms (...)" -> "The PFT datasets with per-pixel uncertainty (product ref. no. 1 in Table 1) are produced with a modified version of the EOF-PFT approach proposed by Xi et al. (2021). The modified algorithms (...)".
10. Line 135: “cannot be assured” -> “could not be assured”?
11. Line 151: “The slope of the corrected dataset is close to one” -> “The slope of the  regression when using the corrected dataset is close to one”.
12. Line 289: I’m not sure if “white ocean” and “green ocean” are well established jargon, but if the jargon is not necessary to facilitate communication, I suggest describing the intended concepts directly.
13. Equation 1: As it is, it seems to me a bit lost in the manuscript and it could be generalized. Maybe include after the call to it in line 143, “relative difference (RD in %) between the OLCI-derived and merged sensor-derived diatom Chla. RDs were calculated as: RD_i = (Chla_i^OLCI - Chla_i^Merged) / Chla_i^Merged, where i is the ith PFT.”
14. All figures that include a global map: For visual consistency, I suggest that all global maps are presented in the same projection. I recommend the projection of Figure 6 to also be used in Figures 1, 2 and 5, or the Mollweide projection (equal area) to be use for all global maps.
15. Figure 3: A recommendation is to not extend the regression lines beyond the range of the X-axis data, as this is beyond the domain of the data used to calculate the regression.

References

Fauth, J. E.; Bernardo, J.; Camara, M.; Resetarits, W. J.; Van Buskirk, J.; McCollum, S. A. 1996. Simplifying the Jargon of Community Ecology: A Conceptual Approach. The American Naturalist, 147(2), 282–286. doi: http://www.jstor.org/stable/2463205

Ho, Tin Kam. 1995. Random Decision Forests. Proceedings of the 3rd International Conference on Document Analysis and Recognition, Montreal, QC, 14–16 August 1995. pp. 278–282.

Legendre, P.; Legendre, L. 2012. Numerical Ecology. 3rd edition. Elsevier, Amsterdam.

Meyer, H.; Reudenbach, C.; Hengl, T.; Katurji, M.; Nauss, T. 2018. Improving performance of spatio-temporal machine learning models using forward feature selection and target-oriented validation. Environmental Modelling & Software 101, 1-9. doi:  https://doi.org/10.1016/j.envsoft.2017.12.001

Roy, S.; Llewellyn, C. A.; Egeland, E. S.; Johnsen, G. (editors). Phytoplankton Pigments: Characterization, Chemotaxonomy, and Applications in Oceanography. Cambridge University Press, Cambridge, UK, 2011.

Xi, H.; Bretagnon, M.; Losa, S. N.; Brotas, V.; Gomes, M.; Peeken, I.; Alvarado, L. M. A.; Mangin, A.; Bracher, A. 2023. Satellite monitoring of surface phytoplankton functional types in the Atlantic Ocean over 20 years (2002-2021). Contribution to the 7th edition of Copernicus Marine Service Ocean State Report. State of the Planet 1-osr7, 5. doi: https://doi.org/10.5194/sp-1-osr7-5-2023, 2023.

This paper introduces a machine learning-based correction method to harmonize global phytoplankton functional type (PFT) data obtained from various ocean color sensors, addressing the discrepancies caused by differences in sensor characteristics. The authors propose the use of a random forest-based ensemble learning method (MLBE) for this task. By correcting the OLCI-derived PFT data to match the merged sensor-derived PFT data, this method ensures more consistent and reliable global PFT observations. The study demonstrates the utility of this correction for analyzing long-term trends in PFTs, revealing significant changes in the biomass of diatoms and dinoflagellates, while showing more stable trends for haptophytes and prokaryotes. Additionally, it examines anomalies in PFTs, noting significant increases in diatom and dinoflagellate Chla concentrations, particularly in higher latitudes and coastal regions.

Overall, the paper is clear in its objectives, and the methodologies employed are robust. The innovative use of machine learning to calibrate and harmonize PFT data from different sensors is a particularly valuable contribution. This approach significantly improves the accuracy and reliability of the resulting PFT time series, showcasing the potential of machine learning in enhancing ocean color application. Given the importance and novelty of this research, I recommend the publication of this paper. Before final acceptance, I have a few suggestions (listed below) that I hope the authors will consider to further refine and enhance the quality of the work.

1. During model training, the authors randomly partitioned the dataset into a training set (70%) and 100 test sets (30%), achieving good validation results. However, the inclusion of latitude and longitude as input features may raise concerns about potential data leakage and shortcut learning, as spatial dependencies could result in overly optimistic estimates of model accuracy (https://doi.org/10.1038/s41559-023-02162-1). To mitigate this risk and improve the rigor of the validation, I suggest the authors incorporate temporal partitioning in addition to random partitioning. By dividing the dataset based on time and ensuring that the training and test sets are strictly independent in terms of temporal coverage, the MLBE model's ability to generalize across both time and space can be more rigorously assessed.
2. When calculating the relative difference (RD in %) using PFT data, is the RD calculation performed on the log-transformed data or the raw PFT data? It is recommended to clarify this in the paper.
3. Similarly, when calculating the PFT anomaly, is the calculation performed on the log-transformed data or the raw PFT data?
4. When calculating the average, it is important to clarify whether the authors computed a weighted average based on factors such as latitude, or if they simply calculated the unweighted mean of all pixels. Given that PFT variations are primarily observed in high-latitude regions, using a latitude-weighted average would be more reasonable.
5. The MLBE model demonstrates excellent performance, but it would be valuable to explore whether it can be applied to data from other sensors. Given the potential for future expansion, I recommend that the authors include a brief discussion in the paper about potential future work, particularly how the correction method could be further improved or extended to incorporate data from other satellite sensors. This discussion would not only highlight the adaptability and scalability of the method to other satellite datasets but also significantly enhance the broader impact and relevance of this research. I believe that including this discussion would add considerable value to the paper.