Round 1

Reviewer 1 Report

Dear Authors,

I have reviewed the manuscript titled “Predicting the Feed Intake of Cattle Based on Jaw Movement 2 Using a Triaxial Accelerometer” This research is interesting and proposes new algorithms for low sample rates, which reduces the amount of data for processing. Additionally, sites for placing accelerometers on the cows' faces are compared. The authors need to explain why this work is novel because there are several devices in the market that measure feeding behavior in dairy cattle. Additionally, they need to support why a lower sample rate is needed when there is more powerful computers to handle a large amount of data efficiently. Please consider adding this to your introduction and discussion.

It is valuable that data from sensors were contrasted with visual observations and that feed intake was measured.

The written quality of this manuscript must be improved. I recommend a revision by a native speaker. Some sentences are too long and hard to read.

 

L12: replace by “Accelerometers are considered”

L14: What do you mean with management adjustment?

L17: In what stage of lactation? Add this in the abstract. Cows reduce their feeding intake during different periods of lactation.

L31: What is the different between feeding behavior and intake. Please consider this for the introduction and discussion.

L38-39: the word feeding is used 3 times in the same sentence.

L44: Not sure what management adjustment means.

L58: Machine learning methods “were” used.

L59: Please do not use good to describe performance. This is subjective.

L65: Why do you need to reduce the amount of data? Expand on this.

L66: “Main affect”

L99: In what period of lactation these cows were at?

L136: Videoed?

L109: How long was the study period? For how long a single cow was observed? Please clarify.

The sample size is very low.

L224: Feed intake rate.

L360: What is a good prediction. Please be objective.

L389: considering using time gap instead of window size.

L425: What are the units of the y-axes?

 

Figure 8 y-axis?

Author Response

Point 1: I have reviewed the manuscript titled “Predicting the Feed Intake of Cattle Based on Jaw Movement Using a Triaxial Accelerometer” This research is interesting and proposes new algorithms for low sample rates, which reduces the amount of data for processing. Additionally, sites for placing accelerometers on the cows' faces are compared. The authors need to explain why this work is novel because there are several devices in the market that measure feeding behavior in dairy cattle. Additionally, they need to support why a lower sample rate is needed when there is more powerful computers to handle a large amount of data efficiently. Please consider adding this to your introduction and discussion.

 

Response 1: Thanks for the comments and we are highly appreciate for your efforts on help improving this manuscript. We have carefully addressed all the review comments and improved the quality of this manuscript accordingly.

Cattle collect feeds into mouth and chew them through the movement of the jaws, including a series of secondary activities such as ingesting, chewing and swallowing. For sure there are several devices in the market that measure feeding behavior. But these devices can only measure the primary behaviors such as feeding, lying, walking rather than the activities of jaw movements such as ingesting, chewing, etc. The significance in measuring the secondary activities is to predict feed intakes of an induvidual cow. The objective of this manuscript is to identify the jaw movements including ingesting, chewing and to predict feed intake based on jaw movement. This totally different from the functions of market available devices. This is explained in Line 47-53 in the revision. Additionally, for sure more powerful computers to handle a large amount of data efficiently. But the significance in a lower the sample rate is to increase the battery life of wearable devices as the data sampling, pre-processing and transmission are conducted in the device. According to the literature, a main challenge with accelerometer sensing is to reduce the amount of data while maintaining the satisfactory of accuracy due to the battery life in practical use. This is explained in Line 66-72 in the revision.

 

Point 2: It is valuable that data from sensors were contrasted with visual observations and that feed intake was measured. The written quality of this manuscript must be improved. I recommend a revision by a native speaker. Some sentences are too long and hard to read.

 

Response 2: The manuscript has been polished through a professional language editing service in MDPI in the revision.

 

Point 3: L12: replace by “Accelerometers are considered”; L14: What do you mean with management adjustment? L17: In what stage of lactation? Add this in the abstract. Cows reduce their feeding intake during different periods of lactation. L31: What is the different between feeding behavior and intake. Please consider this for the introduction and discussion. L38-39: the word feeding is used 3 times in the same sentence.

 

Response 3: All the “accelerator” in the manuscripte has been replaced by “accelerometer”. “management adjustment” is substitued by “daily management” in the revision. Cows were in mid-lactating and added in the abstract. Feed intakes of an cow can be predicted and deduced from its feeding behavior, especiall from its jaw movements. This is explained in the introduction and results & discussion (jointed together according to the suggestion of third reviewer). The sentence in L38-39 is rewritten to reduct the use of “feeding”.

 

Point 4: L44: Not sure what management adjustment means. L58: Machine learning methods “were” used. L59: Please do not use good to describe performance. This is subjective. L65: Why do you need to reduce the amount of data? Expand on this. L66: “Main affect”; L99: In what period of lactation these cows were at? L136: Videoed?

 

Response 4：“management adjustment” in Line 44 is substitued by “daily management such as providing supplementary feeding, or eliminating cows with low feed conversion ratio.” further explaination.”good perfomance” in Line 59 (Line 65 in revision) is substitued by “ better perfomance” and the language has been edited through the perfessional service in MDPI. L65 (Line 72 in revision): The reduce the amount of data is to lower the sampling frequency and increase the battery life of monitoring device in practical use. As the data sampling, pre-processing and transmission are conducted in the device, and the most chanlleding in pratical use is the battery life of monitoring device. The battery life of monitoring devices can be improved 2 or 3 times when the sampling rate reduced from 20 Hz to 1 Hz. L66: this grammatical error is corrected through perfessional editing service. Cows were in mid-lactating, added in Line 107 in the revision. L136 (Line 139 in revision): “Videoed” is substitued by “recorded”.

 

 

Point 5: L109: How long was the study period? For how long a single cow was observed? Please clarify. The sample size is very low. L224: Feed intake rate. L360: What is a good prediction. Please be objective. L389: considering using time gap instead of window size. L425: What are the units of the y-axes? Figure 8 y-axis?

 

Response 5：It’s one-week period for data sampling in dairy farm and a single cow had at least 30 min record for feeding beahvior. This is provide in Line 146 in the revision. The sample size is 13 cows in this study, and the sample size is 2 to 10 cows/goats in similar studies previousely (Yukinori et al., Comput. Electron. Agr., 2013,92:54–65; Barwick et al., Comput. Electron. Agr., 2018, 145:289-297; Chelotti et al., Comput. Electron. Agr., 2018, 145:89-91; Alvarenga et al., Comput. Electron. Agr., 2019, 168: 105051; Shen et al., IPA, 2020, 7: 427–443; Arablouei et al., Comput. Electron. Agr., 2020, 183:106045).

All the “feed intaking rate” is substitued by “feed intake rate” in the revision. L360  (Line 418 in revision): “good prediction” is substitued by “better prediction”. LStack model and ETR model showed higher R² and lower RMSE and NME. This indicates a better prediction than the other twelve models. L389 (Line 374 in revision): we prefer to use window size than time gap in the manuscript. Time window is a term in machine learning whoes definition is a fixed interval of time when the data stream is processed for query and mining purposes. The time gap of time window usually described as “window size” and can be found in exiting literatures, for example, Riaboff et al. (Comput. Electron. Agr. 2019, 165, p. 104961). For figure 7 and figure 8, the caption are added for y-axis and units are added in the figure caption.

Author Response File: Author Response.docx

Reviewer 2 Report

This is an interesting paper focusing on the development of a system to be used for monitoring cattle feed activities and feed intake automatically, which is a topical and open issue in the research field. The proposed system allows to achieve good precision of the results, especially if compared to the existing technologies adopted in the sector. The paper also provides indications for the application of the systems in commercial and productive barns. Therefore, I recommend minor revisions addressing the comments in the attached file (see attached file).

Comments for author File: Comments.pdf

Author Response

Thanks for the comments and we are highly appreciate for your efforts on help improving this manuscript. All the grammatical error in the attached file has been revised and corrected through a professional language editing service. The “accelerator” has been replaced by “accelerometer” in the manuscripte. And “G”, referring to the gravitational acceleration, is substitued by “g” in the whole document. 

Author Response File: Author Response.docx

Reviewer 3 Report

The article uses accelerometers to estimate the ingesting, chewing, ingesting-chewing and other behaviors in 13 cows. 14 machine learning models were trained.

Authors must write the objectives and contributions of the article. I understand that the XGB and HMM-Viterbi algorithm were decisive for the correct interpretation of the signals obtained from the accelerometers. I suggest that the authors better describe these algorithms and how they were applied to the data.

I did not understand how the authors interpreted the accelerometer signals and converted/classified into the investigated behaviors. I suggest that the authors present a chapter in Results and Discussion detailing how the accelerometer signals were interpreted and eventually corrected. The summarized values ​​presented (0.0298G±0.7005G, 0.2703G ±0.6388G and 0.0796G±0.6561G) for each of the behaviors of interest, have a very high standard deviation, probably making unfeasible to classify these behaviors . I understand that three accelerometers were used in each animal, in the P1, P2 and P3 muscles. These results were not presented.

Results are presented in the Discussion chapter. I suggest that authors join the Results and Discussion chapters.

On page 12, section 4.1, it was a surprise to find the analysis of the observation window size effect. If these data were recorded and the results of this analysis were considered when choosing the best size of observation window to develop the ML models, this should be included in the objectives and methodology of the article.

Specifically:

Line 271-272: I think the definition of FP is wrong.

Line 301: There is a problem with the configuration of the table that makes it difficult to interpret.

Line 366-368: I think it would be interesting if the tested models were represented with the same colors in the three graphs.

Line 392: I didn't find the value of 1784 in Table 4.

 

Author Response

Point 1: The article uses accelerometers to estimate the ingesting, chewing, ingesting-chewing and other behaviors in 13 cows. 14 machine learning models were trained. Authors must write the objectives and contributions of the article. I understand that the XGB and HMM-Viterbi algorithm were decisive for the correct interpretation of the signals obtained from the accelerometers. I suggest that the authors better describe these algorithms and how they were applied to the data.

 

Response 1: Thanks for the comments and we are highly appreciate for your efforts on help improving this manuscript. We have carefully addressed all the review comments and improved the quality of this manuscript accordingly. Moreover, the manuscript has been polished through a professional editing service in MDPI to improve the language expression.

 

The objective of this study is to develop an integrated machine learning algorithm framework to identify jaw movement during feeding at a relatively low sampling frequency of triaxial accelerometer, and to predict feed intakes based on the acceleration variables of ingesting and chewing behaviors. Moreover, possible monitoring sites and pre-processing effects of window size were investigated and compared to assess its impact on jaw move-ments classification and feed intake prediction. A main challenge with accelerometer sensing is to reduce the amount of data while maintaining the satisfactory of accuracy to increase the battery life of monitoring device in practical use. Traditionally, acceleration signals were collected at around 10-30 Hz for behavior classification. Efficient classification algorithm at a low sample frequency can reduce the amount of recorded data for processing.The battery life of monitoring devices can be improved 2 or 3 times when the sampling rate reduced from 20 Hz to 1 Hz. Results of this study would contribute to automatically recognition of feeding behaviors and feed intake modelling, and provide a reference for the appropriate installation position of the wearable device equipped with triaxial accelerometers as well.

 

The basic idea of XGB is to superimpose the results of multiple weak classifiers and combine them into a strong classifier. The superposition method is to add the results of each base classifier. When generating the next base classifier, the goal is to fit the residual between the sum of the results of historical classifiers and the label.Each base classifier is a weak classifier, and the base classifier used by XGBoost used in this study was CART. The exactred features are directly input into the XGB model for modeling throgh , and the model output is the predicted behavior results.

Then, use HMM to correct results from XGB: 1) State transition matrix, observation probability matrix and initial probability of HMM model are calculated by using accelerometer data and XGB prediction results.The state transition matrix is the statistical probability of the change of cows' behavior at the current moment and at the next moment.The observed probability matrix is the statistical probability of the change from the real behavior to the predicted behavior after the behavior of cows is predicted by XGB at the current moment.The initial probability is the probability of the behavior of the accelerometer data of multiple cows at the first moment. 2) The Viterbi algorithm of HMM is used to find the original behavior sequence corresponding to the XGB prediction sequence, so as to work out the prediction result with higher accuracy and realize the correction of XGB prediction result.

 

The XGB and HMM-Viterbi algorithm are classical machine learning algrotithms. This study focuses on the specific application rather than the modification of these algorithms. More detailed information of these two algorithm can be found in “XGBoost: A Scalable Tree Boosting System (Chen et al., ACM, 2016, 785-794. http://dx.doi.org/10.1145/2939672.2939785)” and “ An inequality with applications to statistical estimation for probabilistic functions of Markov processes and to a model for ecology (Baum,et al., Bull.amer.math.stat, 1967, 37(3):360-363.)”.

 

 

Point 2: I did not understand how the authors interpreted the accelerometer signals and converted/classified into the investigated behaviors. I suggest that the authors present a chapter in Results and Discussion detailing how the accelerometer signals were interpreted and eventually corrected. The summarized values presented (0.0298G±0.7005G, 0.2703G ±0.6388G and 0.0796G±0.6561G) for each of the behaviors of interest, have a very high standard deviation, probably making unfeasible to classify these behaviors . I understand that three accelerometers were used in each animal, in the P1, P2 and P3 muscles. These results were not presented.

 

Response 2: A flowchart was added in the material & method section to explain how to interpret the accelerometer signals and converted/classified into the investigated behaviors. The high standard deviation is due to the back and forth movement of jaw during feeding. The accelerometer signals are vector and the positive vector position are shown in figure 1. Thus, samples accelerometer signals would be regularly changed in positive and negative in the same behavior. This results to a very low mean and a very high standard deviation. While in behavior classification, fifty-five features were computed from the accelerometer signals. The top eight important features were used for behavior identification using the XGB and HMM-Viterbi algorithm rather than a simple threshold. The high standard deviation won’t affect the identification of behaviors. The results of sampled accelerometer signals in different position were added in Line 313 in the revision.

 

Point 3: Results are presented in the Discussion chapter. I suggest that authors join the Results and Discussion chapters.

 

Response 3: The sections fo “Results” and “Discussion” are merged into one section of “Results and Discussion” in the revision.

 

 

Point 4: On page 12, section 4.1, it was a surprise to find the analysis of the observation window size effect. If these data were recorded and the results of this analysis were considered when choosing the best size of observation window to develop the ML models, this should be included in the objectives and methodology of the article.

 

Response 4: The window size effect is added in the objectives in introduction and methodology section.

 

 

Point 5: Line 271-272: I think the definition of FP is wrong. Line 301: There is a problem with the configuration of the table that makes it difficult to interpret. Line 366-368: I think it would be interesting if the tested models were represented with the same colors in the three graphs. Line 392: I didn't find the value of 1784 in Table 4.

 

Response 5: The definition of FP is modified according to Fogarty (Comput. Electron. Agr. , 2020, 169:105175). It refers to the number of instances where the behavior of interest was incorrectly identified as not being observed. The configuration of the table 2 is restructured to make it easier to interpret. The tested models were represented with the same colors in the three graphs of figure 6 in Line 425-427. It should be 5247 in Line 392 (Line 391 in the revision), data in text and table 4 were double checked to make sure they are correct.

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

L49: Use culling instead of eliminating. I don't think dairy cows are culled for a low-feed conversion ratio, there are other factors such as production, disease, or bad reproductive performance. Delete this phrase. 

"The reduce the amount of data is to lower the sampling frequency and increase the battery life of monitoring device in practical use" I would highlight this in the abstract as one of the impacts of this study.

I still consider the sample size very low. We have conducted behavioral studies using sensors with over 200 cows. For example, https://www.mdpi.com/1424-8220/22/1/1

Please discuss the validity of your study regarding sample size. 

Make sure to proofread the manuscript again.

 

Author Response

Point 1: L49: Use culling instead of eliminating. I don't think dairy cows are culled for a low-feed conversion ratio, there are other factors such as production, disease, or bad reproductive performance. Delete this phrase.

 

Response 1: Thanks again for taking time to review this submission. We appreciate your kind suggestions and we have amended the manuscript accordingly. This phrase is deleted.

 

Point 2: "The reduce the amount of data is to lower the sampling frequency and increase the battery life of monitoring device in practical use" I would highlight this in the abstract as one of the impacts of this study.

 

Response 2: Thanks again for the suggestion. The significance of lower sampling frequency is added in the abstract. Described in line 14-17 as below:

To address further need for commercial use, an efficient classification algorithm at a low sample frequency is needed to reduce the amount of recorded data to increase the battery life of moni-toring device, and a high-precision model needs to be developed to predict feed intake based on feeding behavior.

 

Point 3: I still consider the sample size very low. We have conducted behavioral studies using sensors with over 200 cows. For example, https://www.mdpi.com/1424-8220/22/1/1

Please discuss the validity of your study regarding sample size.

 

Response 3: Discussion about the sample size was added in the section “3.5. Comparison with similar studies” as below:

Moreover, sample size of experimental animals is essential in behavior identification and feed intake prediction. The commonly used sample size is 2 to 10 cows/goats for similar studies [10, 15, 20, 38]. The sample size is 13 cows in this study, which is consistent with the sample size commonly used in existing literature. However, it is important to expand the sample size of experimental animals to enhance and evaluate the robust of the established models or algorithm framework in practical use. This could provide a new view to study complex behavioral patterns across time and in a wider range of contexts. For example, McVey et al. used 200 cows to discovery knowledge of complex behavioral patterns through the commercial sensor platform of CowManager, finding that the tradeoffs between behaviors in time budgets can be improved to mimic the complex error structures of sensor data [39]. In the future, the developed algorithm framework is suggested to embedded in a device or sensor platform, and validate its effectiveness in practical use with a bigger sample size.

 

Point 4: Make sure to proof read the manuscript again.

 

Response 4: Thanks for the suggestion. We proof read throughout the manuscript and some optimizations have been made to enhance continuity.

 

Author Response File: Author Response.docx

Reviewer 3 Report

The authors responded and justified all points of the review. There was substantial improvement in the quality of the article.

Author Response

Point 1: The authors responded and justified all points of the review. There was substantial improvement in the quality of the article.

 

Response 1: Thank you very much for your recognition of our revised manuscript. We really appreciate your kind suggestions on help improving this manuscript. Thanks again!