BMAE-Net: A Data-Driven Weather Prediction Network for Smart Agriculture

Round 1

Reviewer 1 Report

In this study, a multi-head attention encoder-decoder neural network optimized via Bayesian inference strategy (BMAE-Net) is proposed here to accurately predict time-series changes of climate factors. It has some research value, However, the following problems still exist and need to be improved.

1. The abstract feels empty and lacks specific research conclusions.

2. The paper lacks a discussion section, without a comparative analysis with the results of others, so that it is not at all clear whether your research is innovative in the field.

3. Conclusive speculations inside the introduction, which should contain references from others, such as lines 38-40 et al.

We hope that the authors will continue to improve the quality of their papers before further publication.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper is well written and the subject is of great interest for applications in agronomy. In order to provide a better theoretical frame, I recommend to mention in the introduction the existence of deterministic methods based on PDEs (partial differential equations) for weather prediction and say a few words about how they compare with respect to deep learning models. I suggest to cite https://doi.org/10.1098/rsta.2020.0097 or equivalent papers.

Also, it would be very interesting to know if your method sounds equally promising for other meteorological parameters like humidity or rain, as they are extremely important inputs - along with temperature - for risk models in agronomy.

An other issue it that the best performance of your method is provided with Case 1 (24 h in the past day to predict 24 h in the next day), while longer period predictions (like Case 3, which exhibits worse performance) are desirable. Could you elaborate on that?

Then, you say that the introduction of Bayesian optimization "increases the computational cost and requires more training time". You also provide details about the hardware used for your numerical experiment but, surprisingly, do not provide any information about typical computer times for training the models and performing inference. Please, let the reader know, as this is an important point when deciding whether or not to adopt a proposed methodology.

Last, are data and software available on open source basis for non commercial use? 

-----------------------------------------------------------------------------------------

In the sequel I provide a list of minor changes and suggestions

line 54: "humanity" ?

line 75: "Standard DNN" ---> ""standard deep meural networks (DNN)"

line 79: "it can be seen that the followings" ---> "it can be seen the followings" ?

line 79: while statements 1 and 2 (lines 80-88) are generally true, I don't see how they turn out from your above analysis. Consider to replace with: "It is well known that", or something similar

line 103: "Attention(BMA)" ---> "Attention (BMA)"

lines 113-117: use Arabic instead of Roman numerals, as is for the section numbering 

line 141: "processing(NLP)" ---> "processing (NLP)"

line 142: "RNN" ---> "Recurrent neural networks (RNN)"

line 182: "by Vaswani" ---> "by Vaswani et al."

line 195: "drinking" ?

line 252: "unit(Bayesian-GRU)" ---> "unit (Bayesian-GRU)"

eq. 12: "AttentionWight" ---> "AttentionWeight ?

line 296: could you add a short explanation on how parallelism is exploited here, both at the algorithm level (what is exactly computed in parallel?) and hardware level (which sort of devices do you use? Multicore CPUs? GPUs? ...)

line 316: formalism here is not coherent with eq. (14) 

line 323:  "r is the prediction length" ---> "tau is the prediction length" ?

eq. (30) (SMAPE): why there is no absolute value for both terms at the denominator?

line 427: "Intel Core i7-6800 processor CPU 3.4GH" ---> "Intel Core i7-6800K 3.4 GHz CPU"

line 428: Please check if the Nvidia GTX1080Ti actually has 16 GB memory

Along section 4.2: semicolons should not be followed by capital letters 

line 484: "Section 4.3" ---> "Section 4.2" ?

line 524: "Bayesian" ---> "Bayesian optimization" ?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This article study BMAE-Net in comparison with the existing models to predict time-series changes of climate factors accurately. The introduction is logical and the manuscript is well-written. The issue of the existing models was discussed and the solution was proposed based on the research topic. I believe this research is valuable as a piece of new knowledge. However, there were some remarks:

-Lines 75 and 76. When the abbreviations are mentioned for the first time in the text they should be described.

-Lines 85-88. Please, provide precise time information on how long existing models can not forecast. According to the experiment result, we can see that BMAE-Net is more precise than others. However, the information on how long BMAE-Net can precisely forecast was not given. 

-Lines 110-112, I suppose this sentence requires citation though the reference was mentioned above.

-Please provide references to formulas all over the text where necessary. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Dear Authors,

your work is well-written and clear in most aspects. The organization of the equations and the text motivates the readers to follow your ideas.

There are a few paragraphs that need attention and the conclusion must be rewritten. Conclusion, in my point of view, is part of the discussion because you presented results and discuss them.

In line 69 you say "data is complex", and it is not clear what you mean. Complexity is something defined and discussed in several Barabasi papers, among other authors, and definitely, your data do not fit in any definition of complexity I know.  You also suggested in lines 82-83 that deep neural network are "noise safe". It is something that must be clarified. Sensors have an error in their readings more than noise properly. There is noise in the environment that may change the reading, but your text suggests sensor accuracy than noise.

Not sure you want to say inefficiency or efficiency in line 147.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 5 Report

My conclusion is as follows: major review. The article is interesting, but: 1. The title is inadequate to the content. The agricultural application aspect of the model is not emphasized at all, and its usefulness was tested for the period when there is no vegetation in the fields. The data for testing came from large urban agglomerations, not from agricultural areas.

2. The range of tests of the developed model was too narrow and the interpretation of the obtained results too superficial.

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 5 Report

Just like last time, I have no objections to the description of the BMAE-Net model used in the reviewed article. Minor editorial notes have been placed next to the text in the attached PDF file.

My criticisms concern the lack of important details of the description of the datasets used (the issue of missing data, outliers/noise).

My previous doubts about the location of sites in urban areas, the choice of December data for testing the model (a period of no vegetation in the northern hemisphere), and the insufficient interpretation of the results presented in Figure 9 still remain unaddressed in the text. The issues are included in the review response (file "agronomy-2213079-reply to reviewer5.pdf") and not in the revised version of the article. This information should go to the readers of the article, not just the reviewer.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx