Development of an Algorithm for Predicting Broiler Shipment Weight in a Smart Farm Environment
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn this paper, the authors analyzed the broiler weight data automatically measured in a smart broiler house with an intelligent system and conducted a study to predict the weight until the time of shipment.
The authors provided more details on background and motivation in the Introduction section. Please provide clear details on the contribution.
In related work, the authors reviewed six works. The authors selected 3 papers before 2010, and they gave summaries for those three works. Then the authors studied 3 recent papers. But the overall related work is not summarized. Please provide limitations of the related work and please explain how the proposed work overcomes those limitations.
Section 2.2 title is Method for estimating daily weight representative value; the authors explained the details of clustering and Kernel Density Estimation. Section 2.3 title is Method for predicting shipping weight; the authors explained ARIMA and Prophet. These sections contain the general details of clustering, ARIMA, Prophet, etc. It would be good if a proper title were applied for these sections.
Why were these models/techniques selected for the proposed work? Please justify.
In order to measure weight, the authors used a weight-measuring device developed by Emotion Co., Ltd.
What is the scale number?
What are 110 learning data sets?
What is KF in Table 1?
In section 3.1, the authors explained data collection. The following details should be given clearly.
How many parameters are used for predictions?
Explanation of parameters.
Period of data collection
Size of the dataset, etc.
In the 3.2 section, the authors mentioned three steps, namely the daily weight estimation step, the weight representative value selection step, and the shipment weight prediction step.
But these steps are not explained properly.
Please add a flow diagram.
Is it possible to compare the proposed technique with any other existing technique?
Since the experiments are conducted in real time, please add the date, duration, number of days the experiment was conducted, etc.
Novelty of the work is missing.
Author Response
Thank you for reviewing our paper.
Thank you for your unexpected comments, which helped our paper grow in a better direction. We mainly revised the overall content of the paper to properly explain the reasons for using the technology and its appropriateness, and we also greatly referenced the reviewer's comments for the description of the additional experimental technology.
Thank you again for reviewing our paper.
Comments 1: [The authors provided more details on background and motivation in the Introduction section. Please provide clear details on the contribution]
Response 1: [Thank you for pointing out the shortcomings. We have taken your feedback into consideration and recognize the need to clearly specify what our research focuses on and how it differs from existing studies. Therefore, we emphasized the challenges of utilizing broiler weight data collected through automatic measurement methods and clarified that our approach allows for the transformation of complex data into a simplified form, which can be effectively used as input for prediction models. Additionally, we highlighted that, unlike previous studies that focused on growth curve modeling for broilers, our main research topic is prediction through time series forecasting. Additional information related to this has been added to lines 49 to 63 of the main text.]
Comments 2 : [ In related work, the authors reviewed six works. The authors selected 3 papers before 2010, and they gave summaries for those three works. Then the authors studied 3 recent papers. But the overall related work is not summarized. Please provide limitations of the related work and please explain how the proposed work overcomes those limitations.]
Response 2: [Additional clarifications regarding the three papers with insufficient explanation of related research have been provided in lines 145 [14], 153 [15], and 162 [16]. The performance scores presented in these papers have been specified, along with their experimental results. Furthermore, in lines 168–175, it has been emphasized that previous studies relied on a small number of refined datasets, and it has been further clarified that the focus of this study is not on growth curve representation but on improving prediction accuracy.]
Comments 3: [Section 2.2 title is Method for estimating daily weight representative value; the authors explained the details of clustering and Kernel Density Estimation. Section 2.3 title is Method for predicting shipping weight; the authors explained ARIMA and Prophet. These sections contain the general details of clustering, ARIMA, Prophet, etc. It would be good if a proper title were applied for these sections.]
Response 3: [Thank you for your suggestions regarding the section titles. For Section 2.2, I have changed the title from 'Method for estimating daily weight representative value' to 'Clustering method and density analysis method for estimating weight representative value.' This revision clearly indicates that the section focuses on estimating the weight representative value and specifies that the techniques discussed are clustering methods and density analysis methods. For Section 2.3, I have modified the title from 'Method for predicting shipping weight' to 'Growth trend representation method and time series approach for shipping weight prediction.' Since this study uses growth trends as a key input data, I wanted to highlight that the section introduces methods dealing with growth trends. These changes can be found in Section 2.2 at line 176 and Section 2.3 at line 236.]
Comment 4 :[ Why were these models/techniques selected for the proposed work? Please justify.]
Response 4 : [I will explain the reasons for selecting the five main algorithms discussed in this paper. For K-means clustering, there is a clear distinction in that it performs distance-based partitioning, compared to the introduced KDE method. We attempted various methods to extract meaningful data from outlier data, and through experimentation, we recognized that the distance-based clustering method of K-means is an effective approach when dealing with large amounts of accumulated data, which led to its adoption. This explanation can be found in line 437 of the manuscript. Regarding the analysis of broiler chicken weight data with outliers, the suitability of the KDE method has been validated in previous studies. The underlying concept of this study is based on the observation that, despite the data being contaminated by outliers, dense clusters form at specific weight points. This is discussed in line 458 of the manuscript. The weighted moving average filter algorithm is used to correct outliers by replacing them with the average of past normal variations. The selection of the weight representative value using this algorithm is an effective way to track the simplified growth progression of broilers on a daily basis. This method is described in line 499. The reason for applying the Prophet model lies in the suitability of its growth trend representation function and its ability to handle outliers within the simplified growth trend through residual modeling. This is explained in line 543. ARIMA, although a somewhat traditional time series forecasting method, is particularly noted for its strong autoregressive pattern approach. The explanation for this can be found in line 592 of the manuscript.]
Comment 5: [ In order to measure weight, the authors used a weight-measuring device developed by Emotion Co., Ltd. What is the scale number? What is KF in Table 1?]
Response 5 : [Thank you for pointing out our mistake. Through your feedback, we realized that our paper lacked sufficient explanation regarding the data collection device. The automatic weighing scale we used for weight measurement is the Kokofarm (KF), developed by Emothon Co., Ltd. This device measures the weight of broiler chickens in grams every second and is specifically designed for poultry, with dimensions of 680mm in height, 400mm in width, and 270mm in depth. Relevant details have been provided in Figure 2 and Table 2 on line 364 of the manuscript. Additionally, through your feedback, we noticed a typo in the weight column of Figure 3, which has now been corrected to 'weight (g).' Thank you again for your valuable input.]
Comment 6 : [What are 110 learning data sets?]
Response 6 : [ Our study aims to perform predictive evaluation on a single dataset and, based on the results, select the optimal model combination comprehensively. The reference to 110 datasets indicates that we conducted predictive evaluations on 110 datasets. These are referred to as training datasets because the meaningful growth trends extracted from these 110 datasets are used as training parameters for the predictor, which directly forecasts the shipping weight. This explanation is provided in line 625. ]
Comment 7 : [How many parameters are used for predictions? , Explanation of parameters. Period of data collection, Size of the dataset, etc.]
Response 7 : [The parameter used for prediction is a single time series dataset of broiler chicken weight. The main contribution of this paper lies in refining data that includes a large amount of noise and outliers to obtain a clear weight time series dataset. To facilitate prior understanding, a brief mention of this has been added in line 68. The dataset we used contains between 7,200,000 and 10,368,000 weight data points in a single dataset. Through the reviewer’s comment, we recognized the need to specify the number of data points and have additionally explained the differences in the number of data points based on the rearing period. A general introduction and explanation of the data have been added in line 367.]
Comment 8 : [In the 3.2 section, the authors mentioned three steps, namely the daily weight estimation step, the weight representative value selection step, and the shipment weight prediction step. But these steps are not explained properly. ]
Response 8 : [Thank you for pointing out the aspects we had overlooked. The explanation of the daily weight estimation step is provided in detail in lines 413–434; therefore, additional writing was deemed unnecessary. However, for the weight representative value selection step, it was necessary to clearly explain that this process involves selecting a single representative value at daily intervals from the multiple estimated weight lists generated in the previous step. To enhance clarity, we explicitly stated that the growth trend of broiler chickens is based on the time dependency of previously selected weight representative values and that the daily weight representative value list is used to estimate and select a representative value that closely approximates the actual weight of an individual chicken. This explanation has been added in line 473.
For the shipment weight prediction step, we clarified that the selected daily weight representative values from the previous step are connected to form a growth trend time series and that this is compared with actual shipment performance indicators. Additionally, we specified that this step begins when the weight representative value exceeds 1000g, citing the lower reliability of the data before this point as the reason. This explanation has been added in line 520.]
Comment 9 : [Please add a flow diagram.]
Response 9 : [Thank you. We have realized that the previous flow diagram was overly condensed. Designing a diagram while ensuring visibility was a challenging task. Since the steps of this study follow a linear process, we hope that the revised diagram enhances clarity. Please refer to Figure 6 in line 411.]
Comment 10 : [Is it possible to compare the proposed technique with any other existing technique?]
Response 10 : [There are very few prior studies that have been conducted on datasets containing a large number of outliers, making it challenging to compare our proposed algorithmic framework with previously researched algorithms. Therefore, we divided our algorithm into three distinct stages and conducted a step-by-step comparison of comparable techniques. However, we also acknowledge the necessity of further comparisons with other methodologies.
To address this, we incorporated additional predictive methods into our experiments to enable a more diverse comparison in The shipment weight prediction step. Specifically, we introduced the Gompertz growth curve, which has been widely used in growth curve modeling and has demonstrated strong performance in representing broiler growth, as well as the Double Exponential Smoothing predictor, known for its high performance in short-term time series forecasting.
The methodologies for these additional predictive methods have been incorporated into Section 2.3. The rationale for adopting the Gompertz model and its applied parameters are detailed in lines 558–570. Furthermore, in lines 571–580, we explain why it is appropriate to compare growth curve models with time series predictors, given the time-dependent and upward growth pattern of broiler chickens. The suitability of Double Exponential Smoothing for short-term forecasting and its ability to predict future values based on past data has been described in lines 581–591.
To reflect the results of these additional implementations, much of the content from line 705 to the end has been newly revised and expanded.]
Comment 11 : [Since the experiments are conducted in real time, please add the date, duration, number of days the experiment was conducted, etc.]
Response 11 : [We apologize for any confusion. Our study utilized stored broiler weight data to construct 110 datasets, and the analysis was not conducted in real-time. We have identified and revised certain terms that may have led to misunderstandings. The term real-time was initially used to emphasize that the broiler weight data was collected in a natural and automated manner during the growth process. However, we recognize that this term may have caused misinterpretation. To clarify, the data used in this study was collected and stored from January 2023 to September 2024. We have added an explicit explanation in line 369 to prevent any further misunderstanding.]
Comment 12 : [Novelty of the work is missing.]
Response 12 : [We acknowledge this critical comment, and we must admit its validity. Our study may lack groundbreaking novelty. However, it was specifically designed to address real-world challenges encountered in smart poultry farms in the Republic of Korea.
As previously mentioned, the majority of existing studies have been conducted in controlled environments where researchers apply adjustable variables to observe growth changes, rather than in actual farming conditions. In contrast, most of the data collected from real farms is heavily distorted by outliers and noise to an extent that renders the direct application of existing research impractical. As a result, many farms struggle to determine the appropriate shipping weight, leading to penalties.
Our study directly addresses these challenges and aims to develop an algorithm that can assist both farmers and processing facilities. Moving forward, we plan to expand the number of predictive variables to a multivariate approach, further improving accuracy and providing even greater benefits to the poultry industry.]
Reviewer 2 Report
Comments and Suggestions for AuthorsThis study estimates the accurate daily weight representative value of broiler weight data using the K- means clustering method and the kernel density estimation method, and further predicts the broiler shipment weight using the Prophet predictor and the ARIMA predictor. In total, this study could help provide important weight information for understanding the growth progress of broilers and adjusting the breeding schedule. Some comments are shown as follows.
Line 32, “, [3]”: The reference should be placed before the comma.
Line 61, “with the ARIMA prediction model”: Since there are many prediction models based on time series data in reality, could you please compare your method with more than one existing predictor?
Line 64, “2.1. Previous studies on broiler weight analysis”: Why don't you compare your proposed method with the methods in these previous studies? In addition, I suggest that the authors list a table for these previous studies.
Line 81, “(MAPE) of 0.03”: The result is promising and I suggest that the authors can include it in the comparison experiments.
Line 488, “Table 4. K-means + ARIMA Predictor Experiment Results (2023.01~2023.03).”: Could you please combine tables 3, 4, 5, and 6 into a single table? It could help us visualize the comparison results of these two methods more intuitively.
References: Many references are outdated. Please cite more recent studies and related research.
Author Response
Thank you for reviewing our paper.
Our paper was able to grow significantly through the reviewer's review that contained his concerns. The lack of the test model was a part that we greatly felt, and I think it was a great revision because the test model was added so that we could analyze our research more objectively.
Thank you again for reviewing our paper.
Comment 1 : [ Line 32, “, [3]”: The reference should be placed before the comma. ]
Response 1 : [Thank you for reading our paper in detail and checking it carefully. It was completely our mistake. We will check for typos more thoroughly. The typo has been corrected.]
Comment 2 : [Line 61, “with the ARIMA prediction model”: Since there are many prediction models based on time series data in reality, could you please compare your method with more than one existing predictor? ;
Line 64, “2.1. Previous studies on broiler weight analysis”: Why don't you compare your proposed method with the methods in these previous studies?
Line 81, “(MAPE) of 0.03”: The result is promising and I suggest that the authors can include it in the comparison experiments.]
Response 2 : [We acknowledge the reviewer's comment regarding the limited number of test models and recognize the need for improvement. In response, we have decided to incorporate one additional time series forecasting model and one additional growth curve model. For the time series forecasting model, we have selected the Double Exponential Smoothing (DES) model. The rationale behind this choice is that broiler weight exhibits a dependency on past values, and despite the limitations of nonlinear modeling, DES demonstrates strong short-term forecasting performance. The details of this decision have been further elaborated in line 581 of the manuscript.
For the growth curve model, we have chosen the Gompertz growth curve model. While the reviewer suggested the Weibull model, our tests indicated that its suitability was significantly low, making it impractical for prediction. Instead, we aimed to select one model from the Logistic, Gompertz, and Von Bertalanffy models. Including all models would have reduced the clarity of the prediction results, so we selected the Gompertz model as the optimal choice. The Gompertz growth curve has been widely applied in previous studies and is particularly well-suited for capturing the early nonlinear growth of broilers, as discussed in lines 558–570 of the manuscript.
To incorporate the experimental results of the additional models, we have revised Sections 2.3.1 and 2.3.2 to include the methodologies of the Gompertz and DES models. Furthermore, the analysis of the experimental results of these models has been added to Section 4.1.3 and subsequent sections.]
Comment 3: [In addition, I suggest that the authors list a table for these previous studies.]
Response 3 : [That’s a great suggestion. Following the reviewer’s recommendation, we have created a table summarizing existing studies in line 167, <Table 1>. The main categories of this summary differentiate the data collection methods and analysis models used in this study from those of previous research, and briefly outline the results of those studies.]
Comment 4 : [Line 488, “Table 4. K-means + ARIMA Predictor Experiment Results (2023.01~2023.03).”: Could you please combine tables 3, 4, 5, and 6 into a single table? It could help us visualize the comparison results of these two methods more intuitively.]
Response 4 : [This was also a point of consideration for us, and we agree with this improvement. Without compromising visibility, we have divided and displayed four prediction models and two data analysis methods, and included the error rates below the predicted values. However, as the table size increased, we judged that displaying 15 data points would not be suitable, so we have revised it to display 10 data points. This can be found in line 723, <Table 4>.]
Comment 5 : [References: Many references are outdated. Please cite more recent studies and related research.]
Response 5: [I agree. However, we felt that replacing the papers on the main algorithms used with more recent papers wouldn't be meaningful. Therefore, we included recent studies from existing research that do not overlap too much with the content. I believe that adding the latest research is an excellent way to emphasize the necessity of our study. Four additional studies were included, which focus on modeling differences in broiler growth environments, modeling growth curves for specific species, selecting the optimal curve model by applying various growth curve models, and determining the optimal growth curve for broilers affected by diseases versus healthy chicks. Additionally, we cited several other papers. The details of the added studies are outlined in lines 108-134.]
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors answer all the questions. The authors are advised to clear all grammatical and typo mistakes