Quantification of Shortwave Surface Albedo Feedback Using a Neural Network Approach
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper explains Neural Network’s ability to reproduce the complex nonlinear relationship between radiation flux and different atmospheric variables, such as surface albedo, cloud optical depth and their coupling effects. It is a good try to use the Neural Network to solve the nonlinear problem. The paper is well written.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsI accepted in present form.
1. The study addressed to test whether NNs can be used to evaluate shortwave radiative feedbacks and to assess 4 their performance
2. I consider the topic original. It addresses a specific gap in the field.
3. Using Neural Network (NN) models is the main thing in the article to evaluate shortwave radiative feedbacks.
4. What further controls should be considered?
5. The conclusions consistent with the evidence and arguments presented . They address the main question posed.
6. The references appropriate.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe main concerns are:
1. The rationale for the specific neural network architecture (e.g. number of layers, nodes per layer) is not clearly explained or justified. More details on the model selection and optimization process could strengthen this.
2. There is limited discussion around the model's generalizability beyond the specific case studies presented. Testing on a more diverse set of data would help establish broader applicability.
3. Some concepts, such as the bivariate feedback analysis, are introduced rapidly without much background. More explanation for readers less familiar with these concepts would be helpful.
4. The presentation of results relies heavily on visual inspection of figures. Including some quantitative evaluation metrics could make the assessment more objective.
Based on the concerns some major comments for improvement are:
1. Elaborate on the neural network architecture selection, optimization and validation to justify the final model configuration.
2. Use a more diverse dataset with varying conditions for testing generalizability. Assess performance across range of test cases.
3. Add more background and context when introducing key concepts to aid reader comprehension.
4. Supplement visual assessments with quantitative metrics like RMSE, R2 etc. to evaluate model performance. Statistical significance testing could also lend more weight to the conclusions.
5. Better highlight the advancements made by the neural network approach compared to existing linear models for feedback analysis. Emphasize the benefits.
6. Discuss any challenges encountered with the neural network methodology and potential limitations or drawbacks.
7. Suggest next steps for how to build on this work - further applications, improvements to model, integration with other models etc.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThis paper is a very nice discussion of the use of neural networks to better understand the interacting effects of various surface albedos and cloud covers on short wave radiation levels for the Earth. This is an important topic since it is a step towards a rational integration of non-linear processes where previous work has been unable to adequately incorporate them.
The paper reads like a cut down version of a higher degree report. It is thus, necessarily, clear and gives detailed explanations. This is very good. The arguments for, and the discussion of, neural networks cover quite well known issues but they are usefully and compellingly included with pertinent examples from the field. These descriptions of neural network issues are useful but, are also well known and some editing could be worthwhile.
This work represents a worthwhile step towards properly incorporating non-linear multi-parameter effects in climate modelling and the data presented in Figures 4-7 demonstrate its potential.
The paper is potentially publishable as it stands, representing good progress but also demonstrating a clear need for further application of the technique.
The field is full of abbreviations, mostly obvious and well-known. Still, even obvious ones should be introduced by words. This is not always the case here and is worth checking - unless I missed some. In this context, note the column labels of Figure A3. Some sentences are incomplete (line 95, line 178). In line 65, we seem to have 8 even years between1990 and 2020, this does not seem likely. Lines 210-215 are important, particularly 213-215. The authors might consider expanding this latter sentence. Text on Figure 1 is too small. Figure 3(b) has text which is unclear and not helpful whilst the individual curves are not identified (although reference to 3(a) does allow one to understand). Figures 4(c), 4(d), and 5(c) (particularly 4(d)) are hard to understand with their current palettes. These figures are important and could be improved. Line 424 has an incomplete reference "[?]". Line 462 seems to be converted from another document, fix the Fig. numbering.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 5 Report
Comments and Suggestions for AuthorsThe paper proposes the Neural Network (NN) models for nonlinear radiative transfer with aim to test whether NNs can be used to evaluate shortwave radiative feedbacks and to assess their performance. The study is focused on the shortwave radiative feedback driven by surface albedo by analyzing two heuristic cases: univariate feedback perturbing albedo and bivariate feedback perturbing albedo and
cloud cover concurrently, testing the NN’s ability to capture non-linearity in albedo-flux and albedo-
cloud-flux relationships, an actually and usefully approach.
Ø Title is informative and reflects the contents and the Introduction chapter reflects a good documentation for the nonlinear climate feedback effects.
Ø Methods section presents in details three approach for diagnosing radiative feedbacks, PRP method; Kernel method and Neural network method.
Ø The model’s training encompasses the data of all 12 months of even-numbered years within the timeframe from 1990 to 2020. For validation, data from the odd years within the same time period is employed. The NN model’s performance is quantitatively evaluated using mean bias error (MBE) and root mean square error (RMSE).
Ø In order to asses the model’s performance, we compare the predicted net shortwave 348
Ø radiative flux at the top of the atmosphere (TOA) against the reanalysis data from ERA5.
Ø The obtained results are suggestive represented by figures. A comparison between NN model’s results and data from ERA5 is highlighted in Fig.4 e.
Ø The conclusions pointing out the relevance of results obtained and highlighted the accuracy of the NN model’s but also the weaknesses of this and a future work is proposed for changing the NN model to predict flux increments.
Ø Authors could suggest changes on the NN model’s for improvement as their performance at the poles as well as regions of elevated topography?
Ø The list of 30 references is consistent with good documentation in the field.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThe manuscript can be accepted in its present form.
