Application of Fast MEEMD–ConvLSTM in Sea Surface Temperature Predictions
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsA fast Multidimensional Ensemble Empirical Mode Decomposition combined Convolutional Long Short-Term Memory model was used to predict Sea Surface Temperature, as a sample in the Bay of Bengal. The results show a good accuracy. But some points need be clarified.
1, In section 3.4, paragraph 2, the reason behind the consideration of 2016 Marine heat wave isn’t included. Give a probable reason to this part of the analysis.
2, In section 2.5, paragraph 4, line 172, give some references to justify the moving window method’s capabilities in time series analysis.
3, In the Introduction section, paragraph 7, line 77, what does the author mean by ideal accuracy?
4, Some spelling format should be corrected. Such as, time format in Figure 10.
Comments on the Quality of English LanguageNot bad.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors combine sequence analysis by decomposition and machine learning model Convolutional Long Short-Term Memory to predict SST. The results show that the method is reliable and effective in spatiotemporal SST modelling. The manuscript can be accepted with some small clarifications and changes.
1, For the train-test split ratio selection In section 3.2, how do you decide in Figure 5, please provide some reference to justify the author’s explanation.
2, Authors claim to have limited the number of IMFs created to 5 in section 3.2. But in figure 6, there are 6 IMFs. Why is this? If mistaken, correct it accordingly.
3, For model loss function the author has used less popular ‘logcosh’ loss function (line 193). Explain the reason behind this and some reference to justify the use of this function.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper uses a reduced data-driven model to forecast weekly SST in the Bay of Bengal. The novelty is supposedly in a unique combination of data compression and ML-prediction methods. I had trouble following the procedure and have a suspicion that the results do not really demonstrate the prediction skill, but rather the interpolation errors (so that future information is inadvertently used in the forecasts). I'd advise authors to (1) rectify the description of methodology; and (2) to compare the performance to the benchmark forecasts using linear inverse models (LIMs), in which next week's SST vector is linearly related to the current week's SST vector. Such models are typically trained on a continuous segment of the SST time series (usually all data prior to and including the current initial condition from which the forecast is run forward). The forecast skill is quantified, for each weak, by the basin-average rms error or anomaly correlation. The authors' procedure can only be assessed (quantitatively) in a sensible way in relation to the above benchmark performance. Meanwhile, describing how the linear forecasts are trained and initialized would help the reader to better follow the multi-step procedure used by the authors (which, presumably, should follow the same set up as those of the linear model, but they are completely non-transparent in the authors' workflow figure 2).
The key question in any prediction study is: "What are the maximum lead times for which any of these models can achieve useful skill?" The authors managed to completely avoid answering this question in their description.
A side note: the paper does use an SST product derived in part from remotely sensed data, but this is very the connection to remote sensing ends... Is it really an appropriate journal to publish this work in?
Comments on the Quality of English Language
English is acceptable
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for Authors
Review of “Application of Fast MEEMD-ConvLSTM in Sea Surface Temperature Predictions” by Wanigasekara et al.
Based on the Multidimensional Ensemble Empirical Mode Decomposition (MEEMD) and Convolutional Long Short-Term Memory (ConvLSTM), the authors build a model to predict the sea surface temperature (SST) in the Bay of Bengal. Results indicate that the MEEMD method is a reliable and effective tool in modelling and predicting the spatiotemporal SST in the Bay of Bengal. The content of this study falls in the scope of Remote Sensing. The followings are my detailed comments.
1, Why do the authors use the weekly mean SST rather than daily mean SST? Can the method developed in this study catch the variation of SST in a period shorter than one week?
2, I cannot understand why changing the train-test split ratio significantly change the validation loss. Can the author analyze the corresponding physical reason?
3, The words and numbers are too small to be recognized in Figures 5, 7, 8 and 10. Moreover, these figures are not very clear.
4, Add the unit of variable shown in Figures 4, 7, 8, 9 and 10.
5, Brackets, including round brackets (e.g. Lines 20 and 22) and square brackets (e.g. the citation number of references) should be placed at the end of the sentence but before the full stop.
6, There are some typos in the paper. Please carefully check in the revision.
Comments on the Quality of English Language
Minor editing of English language is required.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsI still find the paper to be poorly written and articulated and have the remaining doubts about the results. A particular area of concern is the way the time-lagged data is split. From the authors' description it's still not clear whether the training and validation subsets do not overlap within the maximum time lag. Re: LIM model - the comparison is not fair as no regularization has been used in fitting the model, which is bound to negatively impact its predictive skill. A standard way is to fit such models in the (reduced) phase space of leading EOFs, which would limit the number of model parameters to be estimated; the model selection can also be fine-tuned using other regularization methods, such as lasso or partial least squares.
Comments on the Quality of English LanguageEnglish ok
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsI think that the authors have addressed my concerns.
Author Response
Thank you for your kind response. Your comments and suggestions have been of much help to increase the quality of the manuscript.
