please see attachment

A hyperspectral and multi-angular synthetic dataset for algorithm development in waters of varying trophic levels and optical complexity

Jaime Pitarch, Vittorio Ernesto Brando

Comments

This paper presents a new synthetic data set linking apparent and inherent optical properties based on a very substantial set of radiative transfer simulations that are intended to provide comprehensive representation of optical water types found in nature. The purpose is to support ocean colour algorithm development and there is specific effort made to cover a wide range of sun sensor geometries, high spectral resolution and other important features. I am generally supportive of the effort and believe that the ambition of the work is significant. However, there are a couple of areas where I feel there are issues that might be either addressed or at least acknowledged before publication goes forward.

Limitations of measured data sets: One of the key themes of the paper is an ambition to better replicate the true range of variability found in nature. This is particularly emphasised with respect to oligotrophic waters which are reasonably claimed to be relatively under-sampled. In several sections, the authors point to existing field data sets and attempt to replicate all of the observed variability. Whilst this appears sensible on first inspection, I believe there is an underlying issue that needs to be considered. Essentially this boils down to the quality of field data. Any measurement is going to be subject to uncertainty and in many (most?) cases this uncertainty will become more significant as signal levels become smaller. This has been explored in some papers e.g. ref 1. Examples from the current manuscript that I think need to be considered include Figures 6 and 9 which both show apparently very strong variations in spectral slopes that just happen to coincide with signa levels dropping to very low levels.  Is this real variability or is it the result of poor quality fits caused by limited data quality when signals are very low?  Does it make sense to reproduce this level of variability in a synthetic data set if it is actually effectively noise and therefore potentially misleading? I think this at least needs to be considered. There are also known issues with aspects of filter pad absorption measurements (Ref 2 pathlength amplification and baseline correction - the latter can also be an issue for CDOM absorption) that are not discussed but that could lead to significant discrepancies in observed data sets. These issues are effectively being baked into the training of this synthetic data set. The description of how these data were measured is lacking detail and I think there is scope to at least mention that there may be issues of this nature.

Oligotrophic under-sampling: The authors make a significant play on extending coverage of oligotrophic waters that have been historically under-sampled. Whilst this is true, it remains the case that these waters have been sampled. I am concerned that Figure 9 appears to show at least a full order of magnitude of additional CDOM (ag440) range that has never been observed, even with Ultrapath CDOM sampling. I am perfectly happy to criticise measurement quality (see above) but I am a bit concerned about the justification for effectively inventing an additional decade of variability in this parameter? It is possible that community measurements have a lower limit that inhibits resolution of lower signals, but it is also potentially true that there is a background level of dissolved organic absorption that is a natural feature. I am not convinced that this aspect of the data set is as reliable as the paper currently suggests. Again, a more careful discussion of potential merit or otherwise would be advisable I think.

Parameterisation: The paper takes considerable effort to describe and justify construction of the bio-optical model and other aspects which go into parameterising the Hydrolight runs. Inevitably there are decisions that need to be made and options discarded as a result. This is fine, but in several cases here various decisions are presented as inevitable when in fact alternative option could have been chosen. I would not ask for these decisions to be revered or for models to be reworked in addition - that would be unfair. However, I think it is possible for the authors to recognise that alternatives would be available and might also be legitimate options. For example, they have opted to use the a version of the Hydrolight input generation where they calculate backscattering from  backscattering ratios applied to scattering coefficients rather than directly inputting backscattering SIOPs. I can point to a small number of papers where there have been efforts made to directly estimate thee parameters (refs 3 and 4) and which would have provided alternative options that could be considered. Again, I would like to emphasise that I am not looking for more work to be done here, just that there is a slightly less emphatic description of what is possible and available (or not), taking into account material that is not hard to find in the literature.

Validation: The synthetic data set produces hyperspectral remote sensing reflectance spectra that may be of great value for algorithm development. However, it is unclear how representative the simulated spectra actually are? The discussion of the outputs very rapidly branches off into cluster analysis and consideration of geometric effects, but there is no real analysis of how representative the spectra are of natural distributions. I would like to see a comparison with existing measured data sets to get a sense of where there are overlaps and divergences that may or may not be of interest when considering value as an allegedly global data set. I would emphasise that I have no trouble with the quality of the simulated reflectance spectra per se – Hydrolight will produce essentially the right reflectance spectrum for whatever conditions you tell it to work with. However, the value of this synthetic data set is very much in its ability to cover the range of naturally occurring variation and I would like to see harder evidence that it does this e.g. for turbid coastal waters as well as more open coastal and oceanic conditions.

Final comment: I have pointed to four references that are all from my own work. I am very uncomfortable doing this and I am NOT looking for these to be specifically referred to. They do, however, represent the basis for where my opinions have been shaped on these matters and where I believe we might have some philosophical differences that are not, however, insurmountable. I would be more comfortable with a slightly less emphatic version of the paper that provides the reader with clear explanations of the decisions that were taken, but that notes that alternative options could have been taken in at least some cases. I genuinely think the authors need to carefully consider the rationale for reproducing all of the observed variability, including measurement uncertainties, some of which are very significant indeed. Ultimately I would be unlikely to use this synthetic data set as I would struggle to accept some of the decisions that have gone into producing it, but I can imagine it being welcomed by a significant part of the community more or less as is. As with all of these things, the expression caveat emptor pertains. I hope that these comments will help to encourage a slightly less emphatic description of the data set and encourage potential users to be mindful of where the limitations might still be found.

References

1. McKee, D., R. Röttgers, G. Neukermans, V. Sanjuan Calzado, C. Trees, M. Ampolo-Rella, C. Neil and A. Cunningham Impact of measurement uncertainties on determination of chlorophyll-specific absorption coefficient for marine phytoplankton. J. Geophys. Res. Oceans: 119, 9013–9025, doi:10.1002/2014JC009909, 2014.
2. Lefering, I., R. Röttgers, R. Weeks, D. Connor, C. Utschig, K. Heymann, and D. McKee Improved determination of particulate absorption from combined filter pad and PSICAM measurements, Opt. Express 24, 24805-24823, 2016.
3. Bengil, F., D. McKee, S. T. Beşiktepe, V. S. Calzado, and C. Trees A bio-optical model for integration into ecosystem models for the Ligurian Sea, Prog. Oceanography 149, 1-15, 2016.
4. Lo Prejato M., McKee D. (2023) Optical Constituent Concentrations and Uncertainties Obtained for Case 1 and 2 Waters From a Spectral Deconvolution Model Applied to In Situ IOPs and Radiometry. Earth and Space Science, 10 (12), art. no. e2022EA002815. DOI: 10.1029/2022EA002815