Assessment of Machine Learning-Driven Retrievals of Arctic Sea Ice Thickness from L-Band Radiometry Remote Sensing

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The research is generally clear and engaging. The selection of machine learning methods is logical and appropriate for the task. The structure of the data description is somewhat unconventional, as some elements are presented within the methodology section, but this does not hinder understanding. Overall, I found the paper interesting and well-organized, though I have several comments and suggestions for improvement:

Some abbreviations are not introduced before being used in the abstract (e.g., ESA, BGEP, IRO2). These should be defined upon first mention.

Line 80: It would be helpful to add a sentence such as: “The remainder of the paper is organized as follows: Section 2 …” to improve the logical flow and guide the reader.

The authors should consider clearly stating the key contributions of the paper, preferably in the introduction.

The model uses only four input variables: SMOS brightness temperature, sea ice temperature, sea ice salinity, and snow presence. Is this input set truly sufficient to achieve reliable predictions? Additionally, what is the impact of each variable in the Random Forest model's decision-making? A feature importance analysis would be beneficial.

Section 3.2: If I understand correctly, the authors use a 3×3 grid as input to estimate ice thickness at the central pixel. What is the motivation behind choosing a 3×3 window? Have the authors experimented with different input sizes, such as 5×5, etc.? Moreover, have they considered using Vision Transformers (ViT) as an alternative to CNNs for this task?

Figure 2: It would be clearer to annotate the images using letters (e.g., a, b, c) and remove textual descriptions from within the figure itself. Descriptions should instead be included in the figure caption.

The authors are addressing a regression task in this study. However, in pixel-based approaches such as RF and LSTM, transforming the problem into a classification task could offer benefits, including the application of data balancing techniques and the use of a broader set of evaluation metrics.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript evaluates multiple machine learning methods, including Random Forest (RF), Convolutional Neural Network (CNN), and Long Short-Term Memory network (LSTM) in retrieving Arctic Sea Ice Thickness,  and compares  them with the existing ESA operational products. The study demonstrates some value in the field and may be useful for  applying machine learning methods in remote sensing data processing.   Before the manuscript can be accepted for publication, there are some major issues the authors may want to address.

 

Issues:

1）The limitation of the training dataset fundamentally restricts the performance of ML models, which is entirely determined by the accuracy of the physical model. The authors should clearly state this limitation in the article and emphasize that the study’s value lies in "evaluating the ability of different ML architectures to learn and approximate this physical model", while the validation comparison demonstrates how this approximation performs in real-world scenario.

 

2）The results (page 10) and Figure 4 clearly show that the LSTM model predicts "negative sea ice thickness", which is physically impossible. This indicates that the model cannot generalize in some cases or has learned incorrect relationships. However, in the discussion section, the poor performance of LSTM is only attributed to spatio-temporal resolution mismatch, with  no in-depth explanation or discussion of this critical phenomenon of negative values. Why do negative values occur? Does this imply fundamental flaws in the model structure or training method? This phenomenon seriously undermines the credibility of the LSTM method in this study. 

 

3）The end of the discussion section mentions that the 2D-CNNLSTM method is suboptimal, but only in a single sentence. Since this study uses the CNN model to process spatial features and the LSTM model to process temporal information, the 2D-CNNLSTM as a spatio-temporal model should be introduced in detail—for example, whether it first uses CNN to extract spatial features and then inputs feature sequences into LSTM, or adopts a more integrated ConvLSTM structure. It is necessary to explore why this theoretically most powerful model considering both spatial and temporal information performs poorly. Adding detailed descriptions and comparative experiments of this model can support the manuscript’s arguments. 

Comments on the Quality of English Language

The writing of the manuscript needs further improvement, especially, the format (paragraph indenting) and the using of the word.

 

Here are only a few I noticed.

Line 30,  “on 2015” . Should  “in 2015” be used?

Line 80,  Why does indenting change in this paragraph?

Line 86， used in situ observation.  “in situ” means on the spot. Is this the meaning of the sentence.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have done a good job. I have no further recommendations.

Author Response

On behalf of all the authors, I would like to thank the reviewer very much for revising our manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have provided a thorough and satisfactory response to all the concerns raised in my initial review. They have embraced the core suggestions and implemented extensive revisions that have substantially improved the overall quality of the manuscript. The revised version demonstrates enhanced scientific rigor, clarity of argumentation, and improved readability. I recommend that the manuscript be accepted for publication.

Author Response

On behalf of all the authors, I would like to thank the reviewer very much for revising our manuscript.