A Non-Uniform Grid Graph Convolutional Network for Sea Surface Temperature Prediction
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript presents a study on sea surface temperature (SST) prediction using a Graph Convolutional Network (NGGCN) model based on obtained SST data. The use of Non-uniform Grids to capture sharp SST spatial gradients is particularly noteworthy. The Non-uniform GCN has been demonstrated to outperform models that utilize uniform grids. This approach appears to be a reasonable and effective method for improving SST prediction accuracy. In this respect, I concur with the major conclusions drawn in the manuscript.
However, I have some concerns regarding the limitations of the predictions made by the model. Specifically, the predictions are based solely on prior SST data and do not account for external factors such as heat fluxes between the air and sea, ocean current advections, and other related anomalies. These factors can significantly influence SST variations. Therefore, I recommend that the authors consider incorporating these external forcing factors as additional input features or at least discuss their potential impacts on SST predictions.
Overall, the English writing is adequate.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsDifferent to previous results, this paper tried to perform GCN based on non-uniform grids. From the results, this method shows some advances to predict SST variations. Generally, this paper was written well. However, I think some modifications are needed before its publication.
Major:
1. There is no explanation in the article on the training process of deep learning and the selection of optimizer hyperparameters. Necessary explanations need to be given so that readers can follow and reproduce the results of the paper.
2. The author first uses the GCN module to extract the spatial features of the non-uniform SST grid points, then uses the FC module to decode the feature extraction, and then uses the GRU module to decode the temporal features. Why is FC decoding needed in the middle? Can the encoded spatial features be directly processed by GRU?
3. Why you carry out spatial feature extraction firstly and then temporal feature extraction secondly? Can the process be reversed? That is, perform temporal feature extraction firstly and then perform spatial feature extraction secondly.
4. The author adopts a cyclic prediction method, using the previous 7 days to predict the future day, using the previous 30 days to predict the future 7 days, and using the previous 120 days to predict the future 30 days. How to determine these previous/leading days (7, 30, 120 days)? Have you done a sensitivity test?
5. Is it enough to only consider MSE in the loss function? Can other meteorological statistical indicators be included in the loss function?
Minor:
1. Page 6 says that the coarse grid resolution of 0.25 is uniform grid points, but page 7 also says that the obtained effective SST data is divided into coarse grid point data. Then a question is: The uniform grid point data is fine grid point data? The coarse grid point data is uniform grid point data?
2. Page 3: Coupled Model Interpolation (wrong word) Project (CMIP)
The Coupled Model Interpolation Project (CMIP) is a typical physical model … This is wrong, CMIP is not ONLY a model. Re-write it.
3. Figure 4. The colors are too dark. Change colorbar to replot it.
4. Figure 11. What is ground truth?
Comments on the Quality of English LanguageEnglish writing is generally fine.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsReview report
The manuscript "A Non-uniform Grid Graph Convolutional Network for Sea Surface Temperature Prediction" by Lou et al. introduces a non-uniform grid for training deep learning networks, based on the dynamic characteristics of non-uniform changes in sea surface temperature. The study is well-designed, the main argumentation is reasonable, and the conclusions are credible. The non-uniform grid deep learning network method proposed in this paper has significant utility and reference value for geoscience data processing. However, the structure of the paper requires some minor adjustments, and a few issues need optimization. Overall, the article meets the high standards of this journal but still requires some minor revisions.
Hence, I recommend a minor revision of this manuscript.
minor comments:
1. The second section of this article, which provides an overview of the three SST forecasting methods, could be incorporated into the introduction section without the need for a separate section. This integration would allow for a more seamless and logical flow of the introduction's narrative.
2. The English in the manuscript could be further polished. It is recommended that the text be revised to correct grammatical errors, appropriately rephrase certain expressions, and avoid repetition of words and meanings to improve the narrative quality of the article.
3. The paper mentions using SST gradients to determine appropriate grids for training deep learning networks to predict SST variations. However, in reality, the SST gradients in many ocean areas exhibit significant spatial differences throughout the year, especially during different seasons. Please discuss the impact of these variations on the application of this forecasting method.
4. The method's sensitivity to data resolution should be analyzed, examining how prediction accuracy may be affected by variations in input data resolution.
5. The title for section 5 should be “ Validation and Evaluation” instead of “Discussion”.
6. The figure captions in the manuscript require careful revision (Figs 2,5,6,7,10,11). Generally, figure titles should contain sufficient information to convey the essence of the figure's content without the need to refer to the text, allowing readers to understand the main message of the figure promptly.
7. Adjust the color scale range of Figure 6 to ensure that the gradient distribution depicted in Figure 6 matches the spatial resolution of the grid shown in Figure 7.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe comments are answered well.
