Dynamic Prediction of Chilo suppressalis Occurrence in Rice Based on Deep Learning

Round 1

Reviewer 1 Report

This is an interesting and well-conducted research on rice pest prediction based on time series data. A private dataset spanning 20 years was collected and used, with in-depth results. The manuscript is already in good shape. I recommend to the authors only to improve the quality of the figures, which is sometimes low, and to summarize the introductory part, which in my opinion is too didactic.

Author Response

Response to Reviewer 1 Comments

Point 1: 

This is an interesting and well-conducted research on rice pest prediction based on time series data. A private dataset spanning 20 years was collected and used, with in-depth results. The manuscript is already in good shape. I recommend to the authors only to improve the quality of the figures, which is sometimes low, and to summarize the introductory part, which in my opinion is too didactic.

 

Response 1: Thanks. As your suggestion, We improve the quality of the all figures.

The figures are now located in lines 87, 579, 907, 931, 954, 962, and 880 of the manuscript(in the revised version). For the introductory part, we tried our best to improve the manuscript. These changes will not influence the content and framework of the paper. And here we did not list the changes but marked them in red in the revised paper. 

Author Response File: Author Response.docx

Reviewer 2 Report

The aim of this study is to build a predictive model for the occurrence of SRSB population in Hunan Province, China. The authors propose that combining time series data of related pests and ground meteorological data can improve the accuracy of the prediction model compared to previous studies that used only ground meteorological data, which is supported by the stepwise multivariate regression model constructed in the study. The authors develop MLR, GBDT and DeepAR models based on the combination of meteorological variables, associated pests and time features. The results show that the deep learning DeepAR model achieves the best predictions.

The approach of the study seems original and the content of the manuscript is quite interesting due to the methodology and quantification tools used. The manuscript reads smoothly and is easy to understand. The objectives, scope and results of the study are clearly stated.

Here are some suggestions for the improvement:

• In the Datasets Preparation (lines 177-183) it is not clear how and why did you select the data for training and testing dataset
• Table numberings are inconsistent, there are two Tables 1 and 2.
• Table 2 (Types and units of meteorological factors) is not easy to follow.
• Reconsider Table 2 (Rice Pests, Weather and Time Datasets) -  Is there a better way to present the information?

Author Response

Response to Reviewer 2 Comments

The aim of this study is to build a predictive model for the occurrence of SRSB population in Hunan Province, China. The authors propose that combining time series data of related pests and ground meteorological data can improve the accuracy of the prediction model compared to previous studies that used only ground meteorological data, which is supported by the stepwise multivariate regression model constructed in the study. The authors develop MLR, GBDT and DeepAR models based on the combination of meteorological variables, associated pests and time features. The results show that the deep learning DeepAR model achieves the best predictions.

The approach of the study seems original and the content of the manuscript is quite interesting due to the methodology and quantification tools used. The manuscript reads smoothly and is easy to understand. The objectives, scope and results of the study are clearly stated.

Here are some suggestions for the improvement:

Point 1: In the Datasets Preparation(lines 177-183) it is not clear how and why did you selece the data for training and testing dataset.

Response 1: Thanks. We are very sorry for the trouble caused to you by our mistake. The pest’s data were obtained from five regions (Hongjiang, Yuanjiang, Dong’an, Linli, and Liling), four of which (Hongjiang, Yuanjiang, Dong’an, Linli) were collected from 2000-2020 and one (Liling) from 2010-2020. For the 2000-2020 regions, we used 2000-2018 data as the training dataset (7013 samples) and 2019-2020 data as the test dataset (566 samples); for the 2010-2020 regions, we used 2010-2019 data as the training dataset (3726 samples) and 2020 data as the test dataset (200 samples). Overall, we follow 90% of training datasets and 10% of test datasets principles. We corrected the samples size of training datasets in line 183 of the manuscript (lines 728 in the revised version).

 

Point 2: Table numberings are inconsistent, there are two Tables 1 and 2.

Response 2: Thanks. It is true as the Reviewer points out that table numbering is inconsistent, there are two tables 1 and 2. We corrected the numbering of corresponding tables in lines 163, 171, 186, 285, 296, and 318 of the manuscript (lines 624, 718, 733, 841, 852, and 890 in the revised version)

Point 3: Table 2(Types and units of meteorological factors) is not easy to follow.

Response 3: Thanks. As your suggestion, we modified the Type, Abbreviation, and unit of the meteorological factors. These modifications are in lines 596(in the revised version) in the manuscript, and the other parts of the manuscript involving meteorological factors have been revised (lines 721, 834, and 836, etc. in the revised version).

Point 4: Reconsider Table 2 (Rice Pests, Weather and Time Datasets) – Is there a better way to present the information?

Response 4: Thanks. The Table 2 number has been modified to Table 5 (Rice Pests, Weather, and Time Datasets). As per your suggestion, we adjusted the position of Weather variables and Time series of related pests and put a hyphen (-) to replace the null value so that Table 5 can display more clearly the information.

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript 'Dynamic prediction of Chilo suppressalis occurrence in rice based on deep learning' - presents statistical, machine learning and deep learning models for the prediction of destructive pests in rice production in selected farming areas in China. The authors aimed at developing an integrated pest management program for the farmers to increase rice yield. The manuscript is scientifically written and readers will find it interesting and it is likely to receiving high citations. However, there is the need for a minor revision to improve its quality to the highest level. 

1. Abstract: The headings (1) Background; (2) Methods; (3) Results; (4) Conclusions; should be checked with the Journal's Instructions for Authors if they are acceptable.
2. Keywords: Reduce the many words to a maximum of two/three. Eg. ground meteorological observation data - should be 'meteorological data'; the time series of related pests - should be 'time series analysis'.
3. Introduction: Ok!
4. Materials and methods: Provide the Latin name for rice planthopper. If there is no name available, then put a hyphen (-). Explain the 'Other' in the methodology. The 'Other' represents what? Data Collection: 19 factors have been mentioned in line 134 but in Table  2, there are 18 factors. The TABLES numbering, are not in ascending order. Please, correct the tables numbering accordingly. Model: The subsections in lines 186, 205, 216, 242, 247, 258, 261, 264 should be italicized and given a numbering for clarity in the text. What percentage of the data was used for the testing and training - Table 3?
5. Results: Ok
6. Discussion: The discussion needs to be revised to include more citations in relation to the works of other authors. There are no comparisons of the present results with previously published works. Citations are missing in the Discussion. Please, revise accordingly. 
7. Conclusions: In lines 442 to 444, the mention of table and figure should be removed. Provide the specific findings.

Author Response

Response to Reviewer 3 Comments

The manuscript 'Dynamic prediction of Chilo suppressalis occurrence in rice based on deep learning' - presents statistical, machine learning and deep learning models for the prediction of destructive pests in rice production in selected farming areas in China. The authors aimed at developing an integrated pest management program for the farmers to increase rice yield. The manuscript is scientifically written and readers will find it interesting and it is likely to receiving high citations. However, there is the need for a minor revision to improve its quality to the highest level.

 

Point 1: Abstract: The headings (1) Background; (2) Methods; (3) Results; (4) Conclusions; should be checked with the Journal's Instructions for Authors if they are acceptable.

 

Response 1: Thanks.

 

Point 2: Keywords: Reduce the many words to a maximum of two/three. Eg. ground meteorological observation data - should be 'meteorological data'; the time series of related pests - should be 'time series analysis'.

 

Response 2: Thanks. As your suggestion, we reduced the number of words for the keywords, including changing 'ground meteorological observation data' to 'meteorological data', 'the time series of related pests' to 'time series analysis', 'integrated pest management plans' to 'integrated pest management'.

 

Point 3: Introduction: Ok!

 

Response 3：Thanks.

 

Point 4: Materials and methods: Provide the Latin name for rice planthopper. If there is no name available, then put a hyphen (-). Explain the 'Other' in the methodology. The 'Other' represents what? Data Collection: 19 factors have been mentioned in line 134 but in Table 2, there are 18 factors. The TABLES numbering, are not in ascending order. Please, correct the tables numbering accordingly. Model: The subsections in lines 186, 205, 216, 242, 247, 258, 261, 264 should be italicized and given a numbering for clarity in the text. What percentage of the data was used for the testing and training - Table 3?

 

Response 4：

 

Materials and methods: Thanks. As your suggestion, we put a hyphen (-) to replace the Latin name for rice planthopper. The 'Other' represents the sum of other species captured by our light traps except the rice pests shown (Number 0-9) in Table 1, we have illustrated it at lines 590(in the revised version) of the manuscript and put hyphen (-) to replace the name and Latin name for 'Other'.

 

Data Collection: We carefully examined Table 2 and determined that had 19 factors (Number from 0-18). We corrected the numbering of corresponding tables in lines 163, 171, 186, 285, 296, and 318 of the manuscript (lines 641, 650, 665, 768, 779, and 803 in the revised version)

 

Model: We italicized and numbered the subsections in lines 186, 205, 216, 242, 247, 258, 261, 264 in the text (lines 735, 762, 773, 805, 816, and 822 in the revised version). Now the Table 3 number has been modified to Table 5, where 90% of the data were used for the training, and 10% of the data were used for the testing.

 

Point 5: Result: OK!

 

Response 5：Thanks.

 

Point 6: Discussion: The discussion needs to be revised to include more citations in relation to the works of other authors. There are no comparisons of the present results with previously published works. Citations are missing in the Discussion. Please, revise accordingly.

 

Response 6: Thanks. As your suggestion, we added the citations in relation to the works of other authors, including the application of meteorological data to build pest prediction models (line 986 in the revised version) and the effect of temperature on the SRSB (lines 897 in the revised version). The data we used came from the self-acquisition, and the deep learning method we adopted has not been used for pest prediction in the past, so it is not compared to previously published works.

 

Point 7: Conclusions: In lines 442 to 444, the mention of table and figure should be removed. Provide the specific findings.

 

Response 7: Thanks. As your suggestion, we removed the mention of table and figure in lines 442 to 444(lines 1034 to 1036 in the revised version). Correspondingly, we modify contents of lines 440 to 444(lines 1030 to 1036 in the revised version) to the

'According to the high correlation coefficient of the MLR model, the main features of the MLR model of the SRSB captured by the light trap in the research areas were selected as follows: Yuanjiang is APmean, RPH, PSB, YSB, and Season. Hongjiang is RHmin, RPH, PSB, YSB, and Season. Dong'an is TEMP, RPH, YSB, and Season.  Linli is TEMP, PSB, YSB, PLR, and Season. Liling is TEMP, PSB, YSB, PLR, and Season. The GBDT model performs better than the MLR model in four regions (Hongjiang, Yuanjiang, Dong'an, and Linli), and DeepAR performs better than MLR and GBDT in all areas. '

Author Response File: Author Response.docx