Application of a Fusion Attention Mechanism-Based Model Combining Bidirectional Gated Recurrent Units and Recurrent Neural Networks in Soil Nutrient Content Estimation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a model based on a fusion attention mechanism that combines bidirectional gated recurrent units (BiGRU) and recurrent neural networks to estimate soil nutrient content, specially soil organic matter (OM), nitrogen (N), phosphorus (P) and potassium (K).

The ideas are not very new form the point of view of the intelligence artificial area, i.e., there exist some many research papers using different AI, machine learning  or deep learning techniques to calculate models, but the application can be quite OK. But the paper must be rewritten again.

First, there are numerous references in the bibliography that have nothing to do with the area of this paper. You must change them and put only references with any relationship with this paper.

On the other hand, there are a very little number of references with the machine or deep-learning techniques used to model a process, and in concrete to predict the soil nutrients. You must put more references in this area.

You must rewrite the section 4, the model overview. Why do you explain the bagging method? How are you using this method in your proposal? You have not explained at all what is the model proposed in the paper.

Why do you say that used GRU neurons (Bidirectional gated recurrent units) If you put in the Figure 4 LSTM units, not GRU units that are different, and no BiGRU or BiLSTM units, as you put in the title of the paper. Also, you have not explained what a recurrent network or a recurrent unit (GRU or LSTM) is and how do you use it in your proposal.

You say that uses fully connected layers after the CNN but the activation function of all the units in those layers cannot be a softmax function, because you use a BiGRU unit, the softmax function is only in the output layer. So, you must explain so much your proposed model.

You do not have talk about the attention mechanism, what is this technique and how can you use it in your model?

You have put a flow diagram or explain step by step of all the steps to use you model, you have 298 samples, these samples are converted into inputs to the model, but what are really the inputs to the network? How many data do you have to train the model?

You say that you use cross-validation to train the network to avoid the overfitting, but if you have a little data, you must use a k-folder cross validation, with k> 2, if not the results are not useful, especially if you have little data to train a so high neural network. Because how many parameters have you to train in the proposed network?

How you decide the structure of your model?

The CNN used to compare the results; what architecture has? How do you have chosen its architecture to compare with your results?

Why the order of the fractional differentiation is different before and after the EPO correction? You must explain so much the Figure 5

Author Response

Response to reviewer 1

 

Dear Professor:

Thank you for your careful review of our papers and recognition of our research, and we will learn your attitude towards academics. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have made serious changes to the questions you have raised. Below, we grouped actions taken in response to your comments, organized under the major headings supplied, and numbered them. We attempted to be succinct while fully explaining our actions. We focus on explaining your problem and we take it seriously and solve the comments of the other reviewers.

Given your patience waiting for our first revision, we wanted to make every effort to return this revision as promptly as possible.  We were able to make this revision our top priority. We have devoted most of our working (and nonworking) days to the revision. As a result, we are able to return the paper earlier than we estimated.

Your and the reviewers’ comments have again stimulated changes we feel further improved the paper.

Best regards,

Lixin Zhang.

 

Reviewer#1, Concern # 1: First, there are numerous references in the bibliography that have nothing to do with the area of this paper. You must change them and put only references with any relationship with this paper. On the other hand, there are a very little number of references with the machine or deep-learning techniques used to model a process, and in concrete to predict the soil nutrients. You must put more references in this area.

Author response and action: Thank you for reviewing our paper and providing valuable feedback. We have made significant revisions in response to your comments and have also revised the references as per your request. Firstly, we have removed references from the bibliography that are not relevant to the scope of our paper, retaining only those that are directly related to our research to ensure the relevance and accuracy of the references. Secondly, you pointed out the limited number of references related to the use of machine learning and deep learning techniques in modeling processes, particularly in predicting soil nutrient levels. We have taken your advice into consideration and added more references in this area to comprehensively support our research methods and findings. We appreciate your guidance and suggestions, and these improvements will contribute to enhancing the quality and credibility of our paper. (On  page 1-3 of the new manuscript)

 

Reviewer#1, Concern # 2: You must rewrite the section 4, the model overview. Why do you explain the bagging method? How are you using this method in your proposal? You have not explained at all what is the model proposed in the paper.

Author response and action: Thank you for your review and feedback. We have rewritten the "Model Overview" section of the paper, as per your suggestions, and removed the description of the Bagging method to provide a clearer introduction to the model proposed in this paper. The new "Model Overview" section will provide a more detailed explanation of our proposed model, allowing readers to better understand its core principles and functionality. We have taken your recommendations into consideration and actively addressed them to ensure the paper is more clear and accurate.(On page 6-12 of the new manuscript)

 

Reviewer#1, Concern # 3: Why do you say that used GRU neurons (Bidirectional gated recurrent units) If you put in the Figure 4 LSTM units, not GRU units that are different, and no BiGRU or BiLSTM units, as you put in the title of the paper. Also, you have not explained what a recurrent network or a recurrent unit (GRU or LSTM) is and how do you use it in your proposal.

Author response and action: Thank you for highlighting the discrepancy in our paper regarding the mention of using GRU neurons while Figure 4 depicts LSTM units. We also appreciate your observation that we lacked an explanation of recurrent networks or recurrent units (GRU or LSTM) and how they are utilized in our proposal. We have addressed these concerns by making the following modifications to the paper. We have added a detailed description of the "Att-BiGRU-RNN Model," including the "Schematic Diagram of the Att-BiGRU-RNN Model" and the "Attention Mechanism Structure Diagram." These additions serve to explain the structure of our model and the working principles of the attention mechanism.(On page 7-10 of the new manuscript)

 

Reviewer#1, Concern # 4: You say that uses fully connected layers after the CNN but the activation function of all the units in those layers cannot be a softmax function, because you use a BiGRU unit, the softmax function is only in the output layer. So, you must explain so much your proposed model.

Author response and action: Thank you for your review and detailed feedback. You pointed out that in our paper, we mentioned the use of fully connected layers after the CNN, but all units in these layers cannot have a softmax activation function, as we are using BiGRU units, and the softmax function is only used in the output layer. Therefore, you suggested that we need to provide a more detailed explanation of our proposed model. We have made the following revisions as per your suggestions. Specifically, in the new "Att-BiGRU-RNN Model" section, we explain that only the output layer uses the softmax activation function, and we provide a detailed description of the model's structure to clarify its working principles. Additionally, we have modified the description in the "Convolutional Neural Network" section to ensure clarity and remove any ambiguity.(On line 379-387 of page 10 of the new manuscript)

 

Reviewer#1, Concern # 5: You do not have talk about the attention mechanism, what is this technique and how can you use it in your model?

Author response and action: We have made the modifications as per your suggestion. Specifically, in the newly added "Att-BiGRU-RNN Model" section, we have included an "Attention Mechanism Structure Diagram" to explain the structure of the attention mechanism and its computational process within the model. This will aid readers in better understanding our model and how the attention mechanism is utilized to enhance its performance. (On page 7-10 of the new manuscript)

 

Reviewer#1, Concern # 6: You have put a flow diagram or explain step by step of all the steps to use you model, you have 298 samples, these samples are converted into inputs to the model, but what are really the inputs to the network? How many data do you have to train the model?

Author response and action: Thank you for your feedback. You mentioned that we should provide a flow diagram or step-by-step explanation of how to use our model. You also inquired about the input data for our model, especially considering we have 298 samples, and the amount of data used for training the model. We have made the necessary revisions as per your suggestions. We have added a detailed explanation of the sample partitioning method and the input data for the model. Specifically, we explain how the 298 samples are divided into a training set and a validation set, and we provide a thorough description of how the input data for the model is constructed. Additionally, we have included information about the amount of data used for training the model to provide a more comprehensive understanding. (On line 224-239 of page 5 of the new manuscript)

 

Reviewer#1, Concern # 7: You say that you use cross-validation to train the network to avoid the overfitting, but if you have a little data, you must use a k-folder cross validation, with k> 2, if not the results are not useful, especially if you have little data to train a so high neural network. Because how many parameters have you to train in the proposed network?

Author response and action: The original text mentioned the use of cross-validation to train the neural network and reduce the risk of overfitting. However, considering the limited amount of data, we believe that employing K-fold cross-validation is a more suitable approach, where the value of K should be greater than 2, to assess the model's performance more reliably. This is especially crucial when dealing with high-level neural networks that typically involve a large number of trainable parameters. Inadequate cross-validation folds can lead to insufficient stability and reliability of results, particularly in cases with limited data. Furthermore, our neural network architecture consists of three convolutional layers, two pooling layers, and three fully connected layers, which necessitates the training of a significant number of parameters. The exact number of parameters depends on the number of neurons and connections in each layer, making it one of the factors that require careful consideration. We appreciate the valuable feedback you provided. We have made modifications to the article to emphasize the importance of using K-fold cross-validation, particularly when dealing with limited data. Additionally, we will provide further clarification regarding the complexity of the neural network structure and the required number of parameters to ensure that readers have a better understanding of our approach. Once again, we thank you for your suggestions, which have been highly beneficial in improving the quality of our research. (On line 379-387 of page 10 of the new manuscript)

 

Reviewer#1, Concern # 8: How you decide the structure of your model?

Author response and action: The decision regarding our model's structure was made after careful consideration and rational selection. In addition to the Att-BiGRU-RNN model we constructed ourselves, we also conducted comparisons and validations with models used by peer researchers in the field of soil nutrient prediction. Specifically, we referenced model structures from existing literature that had already been validated and applied in the context of soil nutrient prediction. We chose these models as reference benchmarks to facilitate comparisons and validations within our research. To ensure the scientific rigor and reliability of our paper, we added citations to these models in the text to clarify their sources and foundations. This approach helps ensure that our research builds upon existing knowledge and can be compared and validated against other studies. Once again, we appreciate your question, as it contributes to further clarifying our methodology and decision-making process.

 

Reviewer#1, Concern # 9: The CNN used to compare the results; what architecture has? How do you have chosen its architecture to compare with your results?

Author response and action: The architecture of the CNN model used for result comparison was selected based on existing literature that had already been validated and applied in the field of soil nutrient prediction. We chose these models as comparison benchmarks to ensure the comparability and verifiability of our research. Specifically, we provided detailed descriptions of the structure and computational processes of the Att-BiGRU-RNN model in our paper. The structures of the other models were referenced from existing literature, where these models had been mentioned and validated for soil nutrient prediction. We added citations to these models within the text to clearly establish their sources and foundations.

 

Reviewer#1, Concern # 10: Why the order of the fractional differentiation is different before and after the EPO correction? You must explain so much the Figure 5

Author response and action: We have made significant revisions to the section titled "Results and discussion" and have also added detailed explanations in the figure (now Figure 8) to provide a clearer illustration of the differences in fractional differentiation before and after EPO correction. These modifications are aimed at ensuring that readers can better understand our methodology and the information presented in the figures. We appreciate your feedback, as it contributes to improving the comprehensibility and quality of our paper. (On page 12-14 of the new manuscript)

 

 

Reviewer 2 Report

Comments and Suggestions for Authors

The submitted article does not meet the editorial requirements. It is required to organize the entire content according to the structure recommended by the journal.

Introduction consisting of subsections is too long shorten it. Remove subsections and build only 1 Introduction and in the last paragraph state the purpose of the paper.

The purpose of the study is not clearly defined take this into account.

The second chapter should cover the methods that appear in the article's results. Please correct it. Describe the methods briefly but consider the existing literature.

Line 279. Correct the text. It should be "Table 1".

Line 282. Provide spectral range and origin for "ASD Fieldspec spectrometer"

Line 295. The drawing is not enough confirmation of the experiment performed. Confirmation of previously published literature is required.

For subsection "3.3 Spectroscopy and pretreatment" it is required to confirm the presented experiment with literature. One citation is not enough.

Line 309, 338, 449, 450. Is the rule an original solution? If not, you are required to quote at the end of the sentence before the pattern appears: "It can be expressed in matrix form as follows:" This applies to every rule in the article, which the authors did not do.

Lines: 310, 339. Do not start sentences or paragraphs with "where". Fix it everywhere.

Line 338: References to the rule are required.

The authors use methods using Matlab2019b. Will the methods used be scalable to current solutions?

In the figure, the authors present a machine learning algorithm using a training and test set. They further discuss the validation set in Tables 2 and 3. What is the difference between the test set and the validation set in the presented data. Please standardize this. Please specify the proportions of the training, test and validation sets. Did the test or validation set contain data from the training set?

The authors did not provide any supplementary or statement data. Please add this. It is required to provide the learning history for the developed models and, as part of scalability, to provide model breakdowns, e.g. with the .h5 extension for CNN in Python. Currently, in the era of machine learning and the development of artificial intelligence, the interpretability of the acquired models should be necessary to verify the structure.

Line 411. Figure 4 for CNN does not have a complete explanation. Please correct it. What does C1, C2, C3, S1, S2, S4 etc. mean? It appears in the text, but please also include it in the description for Figure 4.

In section 4.4 Convolutional neural network. references are required. The authors discuss learning algorithms and their activation functions. For proper use, the model must be presented in a supplementary data or data statement.

Why did CNN use the "softmax" algorithm and not the "adam" algorithm?

Was regularization used in the CNN during training? If not, why? The learning results presented in Table 2 may indicate a lack of model tuning, which is so important in machine learning. What did the authors do to fine-tune CNN?

It is required that the authors thoroughly improve the description of the methodology to take into account the principles of current machine learning of models, e.g. according to the ideas of most of the literature published so far: data collection, preprocessing data, building neural models, tuning model (is use: regularization, dropout, change of activation function, learning algorithm, selection of the number of epochs), development model. Please describe these machine learning methods step by step in the methodology section. Correct it.

Line: 695: Please delete or provide details for the patent.

No discussion, no comparison of results to previous literature.

Comments on the Quality of English Language

Generally, grammatical errors depend on the structure of the text prepared for individual sections. The language style is good.

Author Response

Response to reviewer 2

 

Dear Professor:

Thank you for your careful review of our papers and recognition of our research, and we will learn your attitude towards academics. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have made serious changes to the questions you have raised. Below, we grouped actions taken in response to your comments, organized under the major headings supplied, and numbered them. We attempted to be succinct while fully explaining our actions. We focus on explaining your problem and we take it seriously and solve the comments of the other reviewers.

Given your patience waiting for our first revision, we wanted to make every effort to return this revision as promptly as possible.  We were able to make this revision our top priority. We have devoted most of our working (and nonworking) days to the revision. As a result, we are able to return the paper earlier than we estimated.

Your and the reviewers’ comments have again stimulated changes we feel further improved the paper.

Best regards,

Lixin Zhang.

 

Reviewer#2, Concern # 1: The submitted article does not meet the editorial requirements. It is required to organize the entire content according to the structure recommended by the journal.

Introduction consisting of subsections is too long shorten it. Remove subsections and build only 1 Introduction and in the last paragraph state the purpose of the paper.

Author response and action: We appreciate your feedback regarding the organization of the article's introduction. We understand the need to adhere to the recommended structure of the journal. We will make the necessary revisions to the introduction by removing the subsections and consolidating them into a single Introduction section. In the final paragraph of the introduction, we will explicitly state the purpose of the paper to align with the journal's requirements. Thank you for your guidance in improving the article's structure. (On  page 1-3 of the new manuscript)

 

Reviewer#2, Concern # 2: The purpose of the study is not clearly defined take this into account.

The second chapter should cover the methods that appear in the article's results. Please correct it. Describe the methods briefly but consider the existing literature.

Author response and action: Thank you for your valuable feedback regarding the clarity of the study's purpose and the structure of the second chapter. We will ensure that the purpose of the study is explicitly defined in the paper to provide a clear understanding of the research objectives. Regarding the second chapter, we will make the necessary revisions to ensure it covers the methods that appear in the article's results. We will provide a concise yet informative description of the methods, considering the relevant existing literature. This adjustment will enhance the coherence of the paper and its alignment with the journal's requirements. Thank you for your guidance in improving these aspects of the manuscript.

 

Reviewer#2, Concern # 3: Line 279. Correct the text. It should be "Table 1".

Author response and action: Certainly, we will correct the reference to "Table 1" in the text to ensure accuracy. Thank you for pointing out the correction needed.(On line 227 of page 5 of the new manuscript)

 

Reviewer#2, Concern # 4: Line 282. Provide spectral range and origin for "ASD Fieldspec spectrometer"

Author response and action: Thank you for your feedback regarding the need for providing the spectral range and origin information for the "ASD Fieldspec spectrometer" mentioned in line 282. We will make sure to include this essential information in the manuscript to provide a comprehensive description of the equipment used. Your suggestion is appreciated and will enhance the clarity of our research.(On line 242-244 of page 6 of the new manuscript)

 

Reviewer#2, Concern # 5: Line 295. The drawing is not enough confirmation of the experiment performed. Confirmation of previously published literature is required.

Author response and action: Thank you for your valuable feedback. As mentioned in line 295, we indeed need to provide more experimental confirmation to strengthen our research. To meet this requirement, we plan to provide additional experimental confirmation by referencing previously published literature to ensure the reliability and accuracy of our research findings. We will actively work on incorporating this information and including the necessary citations in the revised version to better support our research outcomes. Once again, we appreciate your suggestions, as they will contribute to improving the quality and credibility of our research.(On line 242-258 of page 6 of the new manuscript)

 

Reviewer#2, Concern # 6: For subsection "3.3 Spectroscopy and pretreatment" it is required to confirm the presented experiment with literature. One citation is not enough.

Author response and action: Thank you for your feedback regarding subsection "3.3 Spectroscopy and pretreatment." We acknowledge your concern about the need for more literature citations to confirm the presented experiment. We will certainly address this issue by adding additional relevant citations to substantiate the experimental procedures and results presented in this subsection. Our aim is to enhance the credibility and robustness of our research by providing a more comprehensive review of the literature that supports our experimental methods and findings. Your valuable input is appreciated and will be taken into consideration during the revision process.

 

Reviewer#2, Concern # 7: Line 309, 338, 449, 450. Is the rule an original solution? If not, you are required to quote at the end of the sentence before the pattern appears: "It can be expressed in matrix form as follows:" This applies to every rule in the article, which the authors did not do.

Author response and action: We have made significant revisions to the section titled "Results and discussion" and have also added detailed explanations in the figure 8 to provide a clearer illustration of the differences in fractional differentiation before and after EPO correction. These modifications are aimed at ensuring that readers can better understand our methodology and the information presented in the figures. We appreciate your feedback, as it contributes to improving the comprehensibility and quality of our paper. (On page 12-14 of the new manuscript)

 

Reviewer#2, Concern # 8: Lines: 310, 339. Do not start sentences or paragraphs with "where". Fix it everywhere.

Author response and action: Thank you for your feedback regarding lines 310 and 339. We will make the necessary revisions to ensure that sentences or paragraphs do not start with "where" throughout the document. Your input is appreciated, and we will ensure that this issue is addressed in the revised version.

 

Reviewer#2, Concern # 9: Line 338: References to the rule are required.

Author response and action: Thank you for your comment regarding line 338. To address this, we will include references to the relevant rules in order to provide proper citations and support for the information presented. This will enhance the clarity and credibility of the document.(On line 300 of page 7 of the new manuscript)

 

Reviewer#2, Concern # 10: The authors use methods using Matlab2019b. Will the methods used be scalable to current solutions?

Author response and action: Thank you for your question. Regarding the scalability of the methods used in Matlab 2019b, it's important to note that while we employed Matlab for fractional-order differentiation in this study, it does not imply an exclusive reliance on this particular software. In fact, we have also explored the possibility of achieving the same results using alternative programming languages such as Python. Regardless of the method chosen, our primary objective is to perform fractional-order differentiation on spectral data, allowing us to subsequently feed the processed data into various models for training and analysis.We acknowledge the importance of keeping our research methods adaptable and relevant to evolving technological landscapes. Consequently, we will consider scalability as a crucial factor and choose the most appropriate tools and techniques, whether in Matlab or other platforms, to ensure the continued applicability and versatility of our research. We appreciate your inquiry, as it highlights our commitment to robust and adaptable methodologies.(On line 295-300 of page 6 of the new manuscript)

 

Reviewer#2, Concern # 11: In the figure, the authors present a machine learning algorithm using a training and test set. They further discuss the validation set in Tables 2 and 3. What is the difference between the test set and the validation set in the presented data. Please standardize this. Please specify the proportions of the training, test and validation sets. Did the test or validation set contain data from the training set?

Author response and action: In response to your query, it is crucial to clarify the distinction between the test set and validation set, as well as their relationship with the training set. In our manuscript, the test set and validation set do not contain data from the training set. To provide further clarification on this matter, we will include information regarding the allocation of the training set, test set, and validation set in the "Determination of Soil OM, N, P, and K" section. Additionally, we will specify the proportions of the training set, test set, and validation set to ensure transparency and replicability in the handling and utilization of the dataset. We appreciate your feedback, as it will contribute to enhancing the clarity of our research report in presenting the data handling procedures and results.(On line 224-239 of page 5 of the new manuscript)

 

Reviewer#2, Concern # 12: The authors did not provide any supplementary or statement data. Please add this. It is required to provide the learning history for the developed models and, as part of scalability, to provide model breakdowns, e.g. with the .h5 extension for CNN in Python. Currently, in the era of machine learning and the development of artificial intelligence, the interpretability of the acquired models should be necessary to verify the structure.

Author response and action: We have incorporated additional references into the manuscript. In addition to our self-built Att-BiGRU-RNN model, we have also compared and validated our work against models used by peer researchers in the field of soil nutrient prediction. Specifically, we have cited model structures from existing literature that have been validated and applied in the context of soil nutrient prediction. We have chosen these models as reference benchmarks to facilitate comparisons and validation within our research. To ensure the scientific rigor and reliability of our paper, we have included citations to these models in the text to clarify their sources and foundations. This approach helps ensure that our research is built upon existing knowledge and can be compared and validated against other studies.Once again, we appreciate your inquiry, as it helps further clarify our methodology and decision-making process.

 

Reviewer#2, Concern # 13: Line 411. Figure 4 for CNN does not have a complete explanation. Please correct it. What does C1, C2, C3, S1, S2, S4 etc. mean? It appears in the text, but please also include it in the description for Figure 4.

Author response and action: Thank you for your comment regarding Line 411 and Figure 4. We apologize for any confusion, and we appreciate your feedback. To address this, we will provide a more comprehensive and explicit explanation in Figure 4 to clarify the meanings of C1, C2, C3, S1, S2, S3, etc., in the context of the CNN model. This addition will enhance the clarity of the figure and ensure that readers can easily understand the components of the model.(On line 401-404 of page 10 of the new manuscript)

 

Reviewer#2, Concern # 14: In section 4.4 Convolutional neural network. references are required. The authors discuss learning algorithms and their activation functions. For proper use, the model must be presented in a supplementary data or data statement.

Author response and action: Thank you for your valuable feedback. We acknowledge the need to include references in Section 4.4 for the discussion of learning algorithms and activation functions. Additionally, we understand your request for the model to be presented in a supplementary data or data statement. To address these concerns, we will include appropriate references in Section 4.4 to support our discussion of learning algorithms and activation functions. We will also consider the best way to present the model, whether as supplementary data or in a data statement, to ensure the proper documentation and transparency of our research. Your input is highly appreciated, and we are committed to enhancing the rigor and completeness of our research report.(On line 371-375 of page 10 of the new manuscript)

 

Reviewer#2, Concern # 15: Why did CNN use the "softmax" algorithm and not the "adam" algorithm?

Author response and action: First, in addition to the Att-BiGRU-RNN model we built ourselves, we also compared and verified it with models used by peer researchers in the field of soil nutrient prediction. Therefore, in the CNN model, we borrowed other people’s model structures. In addition to the Att-BiGRU-RNN model we constructed ourselves, we also conducted comparisons and validations with models used by peer researchers in the field of soil nutrient prediction. Specifically, we referenced model structures from existing literature that had already been validated and applied in the context of soil nutrient prediction. We chose these models as reference benchmarks to facilitate comparisons and validations within our research. To ensure the scientific rigor and reliability of our paper, we added citations to these models in the text to clarify their sources and foundations . This approach helps ensure that our research builds upon existing knowledge and can be compared and validated against other studies. Once again, we appreciate your question, as it contributes to further clarifying our methodology and decision-making process.

 

Reviewer#2, Concern # 16: Was regularization used in the CNN during training? If not, why? The learning results presented in Table 2 may indicate a lack of model tuning, which is so important in machine learning. What did the authors do to fine-tune CNN?

Author response and action: We appreciate your attention to the regularization aspect of our CNN model in our study. In the revised 'Results and Discussion' section, we have explicitly mentioned the use of regularization techniques during the training of our CNN model, including the incorporation of Dropout layers to combat overfitting. These regularization techniques were employed to enhance the model's generalization performance and stability during training. We place a strong emphasis on model fine-tuning to ensure optimal performance. To further refine the CNN model, we have applied regularization techniques along with other relevant adjustments. These measures contribute to improving the model's performance and adaptability to our research's specific data and tasks. We thank you for your interest and endeavor to make our research methodology more transparent and comprehensive. If you have any further questions or suggestions, we are more than willing to hear and respond to them.

 

Reviewer#2, Concern # 17: It is required that the authors thoroughly improve the description of the methodology to take into account the principles of current machine learning of models, e.g. according to the ideas of most of the literature published so far: data collection, preprocessing data, building neural models, tuning model (is use: regularization, dropout, change of activation function, learning algorithm, selection of the number of epochs), development model. Please describe these machine learning methods step by step in the methodology section. Correct it.

Author response and action: Thank you for your suggestion, as well as for the significant revisions made to the methodology section. You have done valuable work, especially in the "Model Overview" section, where you have added detailed descriptions of the Att-BiGRU-RNN Model, including algorithm flowcharts and explanations of the principles. Furthermore, you have made extensive modifications to other sections, such as "Spectroscopy and pretreatment" and "Convolutional neural network." These improvements will greatly enhance the clarity and transparency of the methodology section in your research report, better aligning it with the current principles of machine learning models. If you need any further advice or assistance, we are here to support you. Thank you for your efforts, as they will contribute to the overall quality of your research report.(On page 6-12 of the new manuscript)

 

Reviewer#2, Concern # 18: Line: 695: Please delete or provide details for the patent.

Author response and action: Thank you for your feedback. Regarding the issue mentioned in Line 695 concerning the patent, if this portion primarily includes details related to 'Author Contributions' and 'Funding' to meet the requirements of the journal, we will consult with the editor to determine whether further details about the patent are necessary or if it is appropriate to retain this section as is. We will actively collaborate with the editor to ensure that the research report aligns with the journal's requirements. We appreciate your input, which helps us make necessary adjustments and improvements to the paper.(On line 697-705 of page 20-21 of the new manuscript)

 

Reviewer#2, Concern # 19: No discussion, no comparison of results to previous literature.

Author response and action: Thank you for your feedback. In response to the peer review comment regarding the need for more discussion and comparison of results to previous literature, we have reworked the "discussion".This revised "discussion" section now includes a more in-depth discussion and contextual comparison of the results with previous literature, addressing the peer reviewer's feedback.(On line 662-697 of page 20 of the new manuscript)

 

 

Reviewer 3 Report

Comments and Suggestions for Authors

In this paper, the authors proposed a model based on a fusion attention mechanism that combines bidirectional gated recurrent units (BiGRU) and recurrent neural networks (RNN) to estimate soil nutrient content. The proposed model integrates the fused attention mechanism with BiGRU and RNN to enhance the accuracy and effectiveness of soil nutrient prediction. The fused attention mechanism captures key features in the input data, while the BiGRU architecture captures both forward and backward contextual information, enabling the model to capture long-term dependencies in the data. The results demonstrate that the proposed Att-BiGRU-RNN model outperforms other constructed models, exhibiting higher prediction accuracy and robustness. The model shows good estimation capabilities for soil OM, N, P, and K with estimation accuracies (R2) of 0.959, 0.907, 0.921, and 0.914, respectively. The application of this model in soil nutrient estimation has the potential to optimize fertilizer management, enhance soil fertility, and ultimately improve crop yield. Further research can explore the applicability of this model in precision agriculture and sustainable soil management practices, benefiting the agricultural sector and contributing to food security and environmental sustainability.

There are some suggestions to be incorporated

1. Explain eq(1) and (2) by taking examples, it will increase the interest of the readers.

2. The authors have claimed that the proposed models capture the complex relationship between hyperspectral data and soil nutrients. But it is not shown anywhere in the proposed model(s). Kindly include it in the Sections 3 or proceeding sections.

3. The authors are encouraged to look into the articles for other model evaluation parameters and sampling approach- Estimation of soil properties from the EU spectral library using long short-term memory networks(2019) Geoderma Regional; Quantitative estimation of soil properties using hybrid features and RNN variants(2022), Chemosphere. 

4. Discuss the input to the proposed models.

Author Response

Response to reviewer 3

 

Dear Professor:

Thank you for your careful review of our papers and recognition of our research, and we will learn your attitude towards academics. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have made serious changes to the questions you have raised. Below, we grouped actions taken in response to your comments, organized under the major headings supplied, and numbered them. We attempted to be succinct while fully explaining our actions. We focus on explaining your problem and we take it seriously and solve the comments of the other reviewers.

Given your patience waiting for our first revision, we wanted to make every effort to return this revision as promptly as possible.  We were able to make this revision our top priority. We have devoted most of our working (and nonworking) days to the revision. As a result, we are able to return the paper earlier than we estimated.

Your and the reviewers’ comments have again stimulated changes we feel further improved the paper.

Best regards,

Lixin Zhang.

 

Reviewer#3, Concern # 1: Explain eq(1) and (2) by taking examples, it will increase the interest of the readers.

Author response and action: Thank you for your suggestion. To make equations (1) and (2) more engaging and reader-friendly。By integrating practical examples and citing the relevant literature, we aim to make equations (1) and (2) more relatable and engaging for our readers, helping them appreciate the practical significance of these methods in our research." This revised explanation combines practical examples with literature references to enhance reader understanding and interest in equations (1) and (2). (On line 371 and 300 of page 7 of the new manuscript)

 

Reviewer#3, Concern # 2: The authors have claimed that the proposed models capture the complex relationship between hyperspectral data and soil nutrients. But it is not shown anywhere in the proposed model(s). Kindly include it in the Sections 3 or proceeding sections.

Author response and action: Thank you for your suggestion. In response to your feedback regarding the need to demonstrate how the proposed models capture the complex relationship between hyperspectral data and soil nutrients, we have included a more detailed description of the proposed model in Section 3.3, titled "Att-BiGRU-RNN Model." This section provides a comprehensive overview of the model's structure and algorithm flow, allowing readers to gain a better understanding of how the model works and how it captures the complex relationship. We believe that this addition will address your concern and provide more insight into the model's functionality and its role in the study.(On page 6-12 of the new manuscript)

 

Reviewer#3, Concern # 3: The authors are encouraged to look into the articles for other model evaluation parameters and sampling approach- Estimation of soil properties from the EU spectral library using long short-term memory networks(2019) Geoderma Regional; Quantitative estimation of soil properties using hybrid features and RNN variants(2022), Chemosphere.

Author response and action: Thank you for your valuable input. We appreciate your suggestions for referring to specific articles for additional model evaluation parameters and sampling approaches. We have incorporated citations to the articles you mentioned. These references have greatly contributed to the enhancement and method extension of our study. They offer valuable insights into model evaluation and sampling approaches, enriching the overall quality of our research. We appreciate your guidance, which has been instrumental in improving our paper and making it more comprehensive.

 

Reviewer#3, Concern # 4: Discuss the input to the proposed models.

Author response and action: Thank you for your feedback. To address your request regarding the discussion of the input to the proposed models, we have made extensive revisions to the "Results and Discussion" and "Conclusion" sections of the paper. In these sections, we provide a more detailed and comprehensive discussion of the input data used for the proposed models, including the spectral data and any preprocessing steps that may have been employed. We also highlight the significance of the input data in the context of our study and how it impacts the model's performance and results. These revisions aim to offer a clearer and more informative discussion of the input to the proposed models, ensuring that readers have a better understanding of the data used and its relevance to the research.

 

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a model based on a fusion attention mechanism that combines bidirectional gated recurrent units (BiGRU) and recurrent neural networks to estimate soil nutrient content, specially soil organic matter (OM), nitrogen (N), phosphorus (P) and potassium (K).

You have explained your proposal so much better that in the previous version of the paper, but if not enough.

First, I suppose that your proposal in the one shown in Figure 3, in this case you must explain what kind of RNN are you using in the decoder part of the network, i.e., explain the layer RNN of Figure 3.

What is S, in equation 4?, because you explain the a’s, the k’s but not the S, i.e., the output is y=f(s), f is the softmax function, but what is S?

In Table 2, you have put the parameter of the first part of your model, the BiGRU layers, but what happen with the RNN layers and the output layer? You must put in the Table all the parameters of the whole model.

Why do you explain the CNN in section 3.4 and Figure 4? Is it another proposal yours or you have obtained it form the literature?  If is this the case, you must put the reference.

By the way, the comparison with your proposal and the results shown are with this CNN model or with a standard CNN model? If it is the case you must put the architecture of the CNN used for the comparison purposes and again explain why do you explain the CNN model of section 3.4

You have not put a flow diagram of all the steps to use your model. You must put it in the paper.

You must explain better the inputs to the network, how are them?

How many data do you have in order to train the network? Explain this in the paper. Because you have 298 samples, but you preprocess the data, with EPO and so on. So, how many data do you have at the end to train the network?

It is not enough to put in the paper a paragraph saying that you have to do a k-cross validation procedure in order to train the network, you must do it in your experiments and to say exactly what value of k have you used. It is neccesary in the esperiments of your proposal and in the CNN model, nor only in the CNN modle as you put in this version of the paper.

Author Response

Response to reviewer 1

 

Dear Professor:

We want to express our sincere gratitude for your ongoing engagement and valuable feedback throughout the review process. Your insights have been instrumental in further refining our work. In this second major revision, we have addressed the concerns and suggestions you provided in your previous review comprehensively. Below, we outline the key changes and improvements made in response to your recommendations. Your guidance has been invaluable in enhancing the quality and clarity of our research.

Best regards,

Lixin Zhang.

 

Reviewer#1, Concern # 1: First, I suppose that your proposal in the one shown in Figure 3, in this case you must explain what kind of RNN are you using in the decoder part of the network, i.e., explain the layer RNN of Figure 3.

Author response and action: Thank you for your valuable feedback. We have taken your suggestion into account and made the necessary revisions to clarify the type of RNN used in the decoder part of our network, as depicted in Figure 3 and Table 2 of our paper.

In the revised manuscript, we have provided detailed information about the RNN layers used in the decoder section of our Att-BiGRU-RNN model. Specifically, we have described the architecture of the RNN layer, including its input size, output size, the number of layers, and the activation function used. This addition provides a more comprehensive understanding of our model's components and structure, addressing your concern effectively.

 

We appreciate your input, which has helped improve the clarity and completeness of our paper.

 

Reviewer#1, Concern # 2: What is S, in equation 4?, because you explain the as, the ks but not the S, i.e., the output is y=f(s), f is the softmax function, but what is S?

Author response and action: You pointed out the lack of explanation for variable "s_t" in Equation 4. We have added a detailed explanation in the revised manuscript to provide a clearer understanding of "s_t". As mentioned in your comments, "s_t" represents the output of the decoding module after passing through the hidden layer of the decoding module. This information is now explicitly stated in the paper.

 

Reviewer#1, Concern # 3: In Table 2, you have put the parameter of the first part of your model, the BiGRU layers, but what happen with the RNN layers and the output layer? You must put in the Table all the parameters of the whole model.

Author response and action: We want to express our sincere gratitude for your invaluable feedback and insightful suggestions concerning our paper on the Att-BiGRU-RNN model for estimating soil nutrient content. Your thorough review is greatly appreciated, and we are fully dedicated to effectively addressing your concerns. In response to your comment regarding the inclusion of all parameters of the model in Table 2, we have made substantial revisions to the table. It now encompasses not only the input and output sizes but also the model parameters for each layer, providing a comprehensive overview of the entire Att-BiGRU-RNN model. This updated table can be found in "Table 2 of the revised manuscript."

We believe that this revised table offers a complete representation of the model's architecture, including the detailed parameters for each layer, aligning with your suggestion. We sincerely hope that this update meets your expectations. Once again, we would like to express our gratitude for your valuable input, which has significantly contributed to improving the quality of our paper. We are open to any further feedback you may have and remain committed to ensuring the highest standard of our research.

 

Reviewer#1, Concern # 4: Why do you explain the CNN in section 3.4 and Figure 4? Is it another proposal yours or you have obtained it form the literature?  If is this the case, you must put the reference. By the way, the comparison with your proposal and the results shown are with this CNN model or with a standard CNN model? If it is the case you must put the architecture of the CNN used for the comparison purposes and again explain why do you explain the CNN model of section 3.4.

Author response and action: Thank you for your comments and suggestions on our paper. We appreciate your feedback and have made the necessary revisions to address the issues you raised, aiming to improve the clarity of our work. Here are our responses to your questions:

1. Explanation of the CNN model in Section 3.4 and Figure 4:We apologize for the lack of clarity in our previous presentation. We have revised the text to explicitly state that we referenced an existing CNN model from the literature to construct our model. This CNN model was included for the purpose of comparison with our proposed "Att-BiGRU-RNN Model." We understand the importance of proper referencing, and we will ensure that the appropriate reference is added for the CNN model we used.
2. Comparison with the CNN model: The comparison in our paper is conducted between our proposed "Att-BiGRU-RNN Model" and the CNN model we obtained from the literature. We will provide the architecture details of the CNN model used for comparison purposes, and we will clarify in the paper that the comparison is between these two models. This comparison is essential to demonstrate the effectiveness of our proposed model in addressing the research problem at hand.

We appreciate your thorough review, and your feedback has been valuable in enhancing the quality and clarity of our paper.

 

Reviewer#1, Concern # 5: You have not put a flow diagram of all the steps to use your model. You must put it in the paper.

Author response and action: We wish to extend our sincere appreciation for your diligent review of our paper on the Att-BiGRU-RNN model for soil nutrient content estimation. Your insightful feedback has been invaluable in enhancing the quality and clarity of our research.

In response to your comment regarding the absence of a flow diagram outlining the steps to use our model, we have made significant improvements to Figure 3 to address this important aspect. The revised Figure 3 now provides a comprehensive flow diagram that elucidates all the steps involved in utilizing the Att-BiGRU-RNN model for soil nutrient content estimation. This flow diagram offers a clear visual representation of the model's application process.

Moreover, we have taken steps to enhance the comprehensibility of our model by providing explicit details in Table 2. The table now offers a comprehensive breakdown of the Att-BiGRU-RNN model's architecture, encompassing inputs, outputs, and key parameters. These additions will enable readers to gain a deeper understanding of our model's structure and functionality.

We are deeply committed to delivering research of the highest quality, and we appreciate your dedication to the peer-review process. We trust that these enhancements align with your expectations and contribute to the overall clarity of our work.

Should you have any further questions, need additional clarifications, or require any further improvements, please do not hesitate to contact us. Your expertise is a valuable asset to our research.

Once again, thank you for your invaluable insights and commitment to advancing scientific discourse.

 

Reviewer#1, Concern # 6: You must explain better the inputs to the network, how are them?

 

Author response and action: We extend our sincere appreciation for your thorough review of our paper on the Att-BiGRU-RNN model for soil nutrient content estimation. Your insightful feedback has significantly contributed to the refinement of our research.

Regarding your comment on providing a better explanation of the inputs to the network, we have taken steps to enhance the clarity of our paper in this regard. In Table 2, we have expanded the details to include the inputs and outputs of each layer within the Att-BiGRU-RNN model. This comprehensive overview allows readers to gain a deeper understanding of the data flow and transformations at each stage of the model.

Additionally, in section 4.1, titled "Removal of soil moisture factor based on EPO," we have explicitly clarified that the removal of the soil moisture factor through the Orthogonalization of External Parameters (EPO) does not alter the inputs to the Att-BiGRU-RNN model. This ensures that readers have a clear understanding of the data preprocessing step and its impact on the model's input.

We trust that these enhancements have addressed your concerns and improved the overall comprehensibility of our paper. Your expertise and insights have been instrumental in refining our research, and we are grateful for your dedication to the peer-review process.

 

Reviewer#1, Concern # 7: How many data do you have in order to train the network? Explain this in the paper. Because you have 298 samples, but you preprocess the data, with EPO and so on. So, how many data do you have at the end to train the network?

Author response and action: Thank you for your valuable feedback regarding our paper. We appreciate your thorough review. You raised a critical point regarding the dataset size used to train our network, considering the preprocessing steps, such as EPO, applied to our data. We understand the importance of providing clarity on this matter. Allow us to explain: In our study, we initially collected a dataset consisting of 298 soil samples. However, to ensure the accuracy and reliability of our analysis, we performed preprocessing steps, including the removal of the soil moisture factor based on the orthogonalization of external parameters (EPO). These preprocessing steps were essential to enhance the quality of the spectral data and reduce noise. Following the application of these preprocessing steps, our dataset size remained consistent at "2151*1" for each sample. While we started with 298 samples, the data dimensionality was maintained at "2151*1" throughout the analysis. This dataset, comprised of 2151 data points for each of the 298 samples, was used for training and testing our Att-BiGRU-RNN model. We believe that the application of EPO and other preprocessing steps significantly improved the dataset's quality and allowed us to obtain more accurate results. We are grateful for your valuable input, and we will ensure that our paper incorporates this clarification regarding the dataset size and preprocessing steps to provide a more comprehensive understanding of our research.

 

Reviewer#1, Concern # 8: It is not enough to put in the paper a paragraph saying that you have to do a k-cross validation procedure in order to train the network, you must do it in your experiments and to say exactly what value of k have you used. It is neccesary in the esperiments of your proposal and in the CNN model, nor only in the CNN modle as you put in this version of the paper.

Author response and action: We would like to express our sincere gratitude for your careful review of our paper on the Att-BiGRU-RNN model for soil nutrient content estimation. Your feedback has been immensely valuable in shaping the quality of our research.

Regarding your specific comment concerning the need for a more detailed discussion of the k-cross validation procedure, including the specific value of k, we would like to clarify our approach and rationale.

In the section mentioned, we provided a concise description of our experimental design, which includes the application of k-fold cross-validation, inspired by reference [55]. However, we opted not to delve extensively into this aspect in the text to avoid unnecessary repetition and to prevent an undue increase in the manuscript's length, which could potentially detract from the overall readability.

We highly appreciate your feedback, and upon careful consideration, we have taken significant steps to address this concern. In the revised version of our paper, we have not only expanded on the cross-validation procedure but also explicitly stated the specific value of k that we have employed for the experiments involving both our proposed Att-BiGRU-RNN Model and the CNN model. This addition ensures that our readers have a clear understanding of our experimental methodology and the rigor with which we have evaluated our models.

By revising this section comprehensively, we believe we have found a balanced approach that provides essential experimental details while maintaining the overall flow and readability of the manuscript. Additionally, we have integrated your other valuable suggestions into this revision.

We are committed to producing high-quality research, and your feedback has played a crucial role in enhancing the clarity and transparency of our work. Should you have any further questions, require additional clarifications, or seek further improvements, please do not hesitate to reach out to us. Your expertise and insights are deeply appreciated.

Once again, we sincerely thank you for your diligent review and contributions to advancing scientific discourse.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper is acceptable in current form.

Author Response

Response to reviewer 3

 

Dear Professor:

We sincerely thank you for your thoughtful review of our paper and for finding it acceptable in its current form. Your feedback has been instrumental in refining our work. In response to your valuable comments, we have carefully revised the manuscript.

We have improved the introduction by adding more background information and ensuring all relevant references are included. Additionally, we have enhanced the clarity of the research methods and results presentation. The conclusions have been thoroughly reviewed and aligned with our research findings.

With these revisions, we believe the paper has significantly benefited and is now better structured and more comprehensible. We genuinely appreciate your expertise and constructive feedback, which have played a crucial role in enhancing the quality of our research.

Thank you once again for your time and commitment to the peer review process.

Best regards,

Lixin Zhang.