Review of "CRITER 1.0: A coarse reconstruction with iterative refinement network for sparse spatio-temporal satellite data"

by Matjaž Zupancic Muc, Vitjan Zavrtanik, Alexander Barth, Aida Alvera-Azcarate, Matjaž Licer, and Matej Kristan

This manuscript is a description of a novel machine learning technique for gap filling of SST analysis in the presence of possibly significant missing (satellite) observational data.    The reconstruction method uses high resolution, multi-sensor, binned where observations exist, L3S SST products from Copernicus Marine Service, and then uses a two stage approach machine learning technique to both fill the true missing data as well as further missing data removed from the L3S product to be used for training and validation.  The analysis is then validated against this removed data, showing improvements over other methods -- primarily DINEOF of Alvera-Azcárate et al., (2005).  

Firstly, I am not an expert on machine learning techniques, and therefore will offer limited comment of the techniques involved, but rather a potential user of improved SST analysis, and therefore offer comments more aligned with that perspective.   This is one of my major points of commentary on the present manuscript:  The paper as a whole is a rather technical description of the proposed method -- and rightly so.  However, I believe some additional commentary on potential users of the system, and what benefits it might offer them should be addressed in the introduction.  As it stands now, this is only addressed very briefly and casually in literally the first 4 lines of the introduction, after which the manuscript pivots to solely detailing the technical details.   My additional major comment would be to better describe some of the terminology in the manuscript.  The meaning of seemingly simple terminology, such as that used for variance, as well as deleted and visible regions, is likely inherently obvious to the authors, however, the interpretation of these terms by the reader could lead to some confusion.  Some more detailed descriptions with regards to the used terminology may be necessary, even if this seems painfully obvious to the authors.  

Major Comments:

1)  Not enough motivating background information in the introduction.    Other than the first 4 lines of the introduction, no motivating information is provided as to why improved high-resolution SST reconstructions are necessary.   While everyone would presumably like the best possible SST reconstruction, what applications would best benefit, and how might they benefit?   Although more directed towards satellite capabilities than gap filling techniques, a review article such as "Observational Needs of  Sea Surface Temperature" (https://doi.org/10.3389/fmars.2019.00420) would seem a good starting point for building motivation.    Other articles exploring the use of improving the resolution of SST boundary conditions for numerical weather prediction could also prove useful.  A quick search yielded me these two possibilities (10.1175/JCLI-3275.1, 10.5194/hess-24-269-2020).  Presumably a more detailed background search would yield more.  

 - Given that high resolution global NWP systems -- ECMWF's IFS is 9km (1/12^o) -- better high resolution global SST products are also required.  The SST reconstructions pursued in this manuscript are all regional (Mediterranean, Adriatic, North Atlantic).  It is not mentioned whether it would be practical to scale the proposed technique to global domains, such as gap filling the Copernicus Marine Service 1/10^o ODYSSEY L3 product.  

2) Some terminology used in the manuscript, while seemingly obvious, on further contemplation the meaning and interpretation is not so obvious.  

 - Uncertainty/Variance (σ²):  The uncertainty or variance outcome from the machine learning training process is introduced and summarized with the generic statement leading off section 3 in the opening 3 lines (ll. 82-84).  This statement represents the only description of how this quantity, which plays a large role in the analysis of the techniques performance and skill  over the rest of the manuscript.  If possible a more detailed description of how this term is output or diagnosed from the machine learning process would be warranted.  From a naive aspect, I would assume this variance, or uncertainty is the range of SST values that would lead to the same best fit outcome in the training process, but obviously, not enough information is given to confirm this.  Furthermore, as detailed in the paper on "Observational Needs of Sea Surface Temperature" given above, and the outcome of many workshops on the needs required of SST observations and analysis, there is a strong need for estimates of uncertainty to accompany estimates of SST.  The estimate of variance/uncertainty outcome from  this technique seems well posed to fulfill this requirement -- if its definition is an adequate measure of this.  

 - The definition of variance becomes further confused with the introduction of scaled error (error divided by variance, l. 251, 3rd line of S5.3).  While the authors again use symbol σ for the scaled variance, or more precisely, σ_ε, this is well identified.  The confusion (for me) was then when scaled variance , σ_ε << 1 was compared with an idealized reconstruction where σ_ε = 1, this is casually referred to as an overestimate of the variance (ll. 261-262).  It took me more than a few moments to eventually realize this was the scaled variance, with the actual variance being a divisor to this scaled variance -- and therefore scaled variance ,  σ_ε < 1, does indeed represent an overestimation of actual variance.   At the risk of insulting some all knowing readers, but lifting up some of the slower to comprehend readers, please somehow remind the readers that this is the scaled variance which is divided by the actual variance -- and therefore the statement does actually makes sense.   

- Visible and Deleted regions: The definition of deleted regions seem relatively obvious:  The regions where SST observations have artificially been removed from the L3 product.  However, the definition of visible, sometimes referred to as observed, regions seems less definite:  Is it the fully observed region in the L3 SST before deletion, or the observed region in the L3 SST after removal of the deleted regions?   Please provide a precise definition of deleted and visible regions.

Typographic and style comments:

Section 4 Results (l. 168) is empty?

Figures 4-6, B1-B3:  Limits on σ and rmse.  The colour scale limits on  σ and rmse seem to be all automatically generated.  This is a hindrance to both comparing between techniques (CRITER/DINCAE2) and comparing over-dispersive and under-dispersive regions (σ vs rmse).  Although I realize this will often lead to regions of colour saturation, I would strive (at least for individual scenarios) to have the colour scale range identical between CRITER/DINCAE2 results and between σ and rmse (preferably with the zero value always represented).    This would likely enhance your ability to discuss the results in the text, and by setting the scales for σ and rmse identically, it would then allow you to connect the results in Section 5.3 with the earlier results -- for instance, you would easily be able to identify regions where DINCAE2 has insufficient variance compared to RMSE, and vice versa for CRITER).  

l. 263 :  in order of 10^−2◦ or lower.    Not sure what is meant by to the power of -2o?  Possible typographical error.  "In order of" is more conventionally referred to as "Of the order of"

Further Comments:

Comments outside scope of manuscript that may be worthy of at least some discussion.

1) As already mentioned, I do not know much about technical details of machine learning techniques.  But what I do know is that the techniques are relatively agnostic to the physical relationship between the inputs and outputs.   In this study, one inputs binned temperature observations and outputs the full field temperatures.  The input L3S products used in the study have undergone a variety of processing (radiance algorithms, bias corrections) to produce a binned multi-sensor temperature.  Would this technique be generalizable to the underlying radiances, complicated by requiring different trainings for each instrument?   The advantage might, however, be better instrument bias corrections and a further reduction in error?

2) As mentioned in Major Comment #1:  Is this scalable to a global analysis?  

3) It could be interesting to apply a (spatial) spectral analysis on the results and underlying inputs, which admittedly would likely require large cloud free areas, at least for analyzing the spectral characteristics of the inputs.  Do the wavelength characteristics of the CRITER and DINCAE2 results differ, and how do they compare to the original wavelength characteristics of the binned SST L3S products:  Are certain wavelengths removed and/or enhanced.   

The paper by Matjaž Zupančič Muc et al. (2025) presents a novel ML-based architecture, which includes the combination of a Vision Transformer model with a U-Net, to reconstruct satellite-derived sea surface temperature values not measured by infrared sensors, mostly due to cloud coverage.

While the technical parts concerning the networks involved are mostly well explained and exhaustive, I think there is a lack of attention when dealing with the datasets involved and the presentation of the results.

I suggest reconsider publication after addressing some majors concerns explained in details in the pdf attached.