Characterization of the Pearl Millet Cultivation Environments in India: Status and Perspectives Enabled by Expanded Data Analytics and Digital Tools

Round 1

Reviewer 1 Report

A clear distinction between calibrable and optimizable parameters should be done.

A lot of calibration and optimization methods already exists methods. Please di not reinvent the wheel

Author Response

A clear distinction between calibrable and optimizable parameters should be done
Answer: Following reviewers’ suggestions, we decided to change the terminology we used. We replaced by what we used to call “optimized parameter inference” by model calibration. We follow the definition of Wallach et al. (2021) stating that calibration is the estimation of model parameters based on fitting the model to observed data. We updated the manuscript accordingly (lines 169 to 176).

A lot of calibration and optimization methods already exists methods. Please do not reinvent the wheel
Answer: We re-emphasized that crop models can be characterized by different type of parameters that generally fall under the three following categories: a) genotype parameters; b) environment parameters (e.g. soil related parameter like soil water content), and c) management parameters (Varella, 2010). In the present work, we did not calibrate the crop model for genotype parameter value. The calibration for genotype parameters was done in a previous study (for details see Garin et al. 2022) according to standard practices. In the present study, we investigated the possibility to calibrate crop models for environment and management parameters. Indeed, it is acknowledged that environment and management parameters 
represent an important part of the crop model uncertainty and that in large scale prediction exercises (e.g. at the regional level) it can be difficult to get precise information about those parameters. This incomplete knowledge can have some important consequences on the crop model prediction ability (Therond et al. 2011). Therefore, we proposed a relatively simple estimation procedure for the environment and management parameters based on large scale simulation and optimization of the correlation function between the prediction and the observation. To the extent of our knowledge, little research has been done on the specific calibration (estimation) of environment and management parameters in crop models. The work of Varella et al. (2010) is an exception that presents a Bayesian approach to estimate soil related parameters given observed yield and LAI. Compare to our approach this kind of approach is potentially more precise but it conceptually much more complicated and computationally very intensive, so it would be difficult to apply to the large-scale type of simulation covered in our study. Those elements were added to the manuscript (lines 174 to 198 and 247 to 263).

References
Garin, V. et al. (2023) New algorithm for pearl millet modelling in APSIM allowing a mechanistic simulation of tillers. bioRxiv

Therond, O. et al. (2011). Using a cropping system model at regional scale: Low-data approaches for crop management information and model calibration. Agriculture, Ecosystems & Environment,

Varella, H. et al. (2010). Soil properties estimation by inversion of a crop model and observations on crops improves the prediction of agro-environmental variables. European journal of agronomy

Wallach, D. et al. (2021). The chaos in calibrating crop models: Lessons learned from a multi-model 
calibration exercise. Environmental Modelling & Software

Author Response File: Author Response.pdf

Reviewer 2 Report

1.The paper abstract is too general, I suggest that the main conclusions such as quantitative results be given.

2.What is the temporal resolution of the introduction and input variables of the crop model in the study methodology? The article only analyzes up to 2018, suggest extending the study to the present.

3.How was the contribution of environmental variables to crop yield and quality determined?

4.The conclusion section is too much, I suggest to streamline it to give substantive and conclusive content.

The language expression is not very easy to read, it is recommended to ask English professionals for language embellishment.

Author Response

1.The paper abstract is too general, I suggest that the main conclusions such as quantitative results be given.

Answer: We revise the abstract following reviewers’ suggestions to put more emphasize on specific quantitative results and the used methodology.

2.What is the temporal resolution of the introduction and input variables of the crop model in the study methodology? The article only analyzes up to 2018, suggest extending the study to the present.

Answer: The temporal resolution of the used variables is 1998-2017. We have used the latest available data. Indeed, the yield available time series data from ICRISAT go only until 2017. We emphasize that better in the text (line 95). Beyond that year, we cannot validate the model performance with observation.

3.How was the contribution of environmental variables to crop yield and quality determined?

Answer: We have determined the influence of environmental variables on yield using linear regression. We could use generated data for a large amount of parameter combinations and environmental settings to estimate the relationship between the predicted yield and the crop model parameters, which included the weather parameters. The linear regression is a simple and efficient technique to jointly evaluate the contribution of each source of variation on the yield. We did not analyze yield quality. We extended the explanation about this part of the material and method for further clarity (lines 288 to 300)

4.The conclusion section is too much, I suggest to streamline it to give substantive and conclusive content.

Answer: We reduced the conclusion section to put more emphasize on the most important results and less on recommendations (lines 643 to 733).

The language expression is not very easy to read, it is recommended to ask English professionals for language embellishment.

Answer: We performed a complete revision of the English quality.

Author Response File: Author Response.pdf

Reviewer 3 Report

I find this work interesting but it needs major revisions before publishing. First, the authors need to include more references and choose recent publications (less than 5 years).  The first part of the manuscript concerning the building of the typology sounds scientifically. The secund part concerning  the crop model parameter influence on grain yield in the new TPEs  needs to be completely revisited. The authors have to better explain the methods they use (see the specific comments) and make effort on figures and tables. The idea to have a Rshiny application and a github repository in order to put data and script used is interesting and contribute to reproducibility of the results. The abstract must also be improve in order to give a better and complete overview of the work.

Comments for author File: Comments.pdf

Author Response

I find this work interesting but it needs major revisions before publishing. First, the authors need to include more references and choose recent publications (less than 5 years). The first part of the manuscript concerning the building of the typology sounds scientifically. The secund part concerning the crop model parameter influence on grain yield in the new TPEs needs to be completely revisited. The authors have to better explain the methods they use (see the specific comments) and make effort on figures and tables. The idea to have a Rshiny application and a github repository in order to put data and script used is interesting and contribute to reproducibility of the results. The abstract must also be improve in order to give a better and complete overview of the work.
“The authors need to include more references and choose recent publications (less than 5 years).”

Answer: We did an extra review of the literature to identify more recent work and added those references:

Rao, C. R. et al. (2019). Yield vulnerability of sorghum and pearl millet to climate change in India. Indian Journal of Agricultural Economics, 74(3), 350-362.

Casadebaig, P., et al. (2016). Assessment of the potential impacts of wheat plant traits across environments by combining crop modeling and global sensitivity analysis. PloS one, 11(1), e0146385.

Hajjarpoor, A., et al. (2021). Environmental characterization and yield gap analysis to tackle genotypeby-environment-by-management interactions and map region-specific agronomic and breeding targets in groundnut. Field Crops Research, 267, 108160.

Rahimi-Moghaddam, S., et al. (2023). Understanding wheat growth and the seasonal climatic characteristics of major drought patterns occurring in cold dryland environments from Iran. European Journal of Agronomy, 145, 126772.

Munasib, A., Roy, D., & Birol, E. (2019). Networks and low adoption of hybrid technology: the case of pearl millet in Rajasthan, India. Gates Open Res, 3(1133), 1133.

Wallach, D. et al. (2021). The chaos in calibrating crop models: Lessons learned from a multi-model calibration exercise. Environmental Modelling & Software

We also would like to re-emphasize that pearl millet cultivation is under-researched area for example, the list of connected articles provided by MDPI at the submission contained a single reference.

“The secund part concerning the crop model parameter influence on grain yield in the new TPEs needs to be completely revisited. The authors have to better explain the methods they use (see the specific comments) and make effort on figures and tables.”

Answer: We have improved the description of the methodology, especially we made an extra effort to explain better our approach for environment and management parameter calibration/estimation (lines 169 to 197 and 247 to 262). We followed all specific author suggestions. We improved the quality of tables and figures following the reviewer suggestions accordingly.

Specific comments
Abstract: to improve in particular you don’t mention the statistical approach you used

Answer: We revise the abstract following reviewers’ suggestions to put more emphasize on specific quantitative results and the used methodology.

L2: “the latest available data and methodology”: To reword with less emphasis, because model simulation and PCA have been used for a long time.

Answer: we reformulated this expression with only emphasis on data novelty:
“Here, we propose a new characterization of the pearl millet production environment using the latest available data.” (line 2)

L5: “unprecedent scale”. To reword with less emphasis, because large scale simulations are common. 

Answer: we reformulated this expression like that:
“…, and large scale crop model simulations.” (line 3)

L10: “constant increase of precipitation”. Need to investigate this point, and find references, such as Praveen, B., Talukdar, S., Shahfahad et al. Analyzing trend and forecasting of rainfall changes in India using non-parametrical and machine learning approaches. Sci Rep 10, 10342 (2020). https://doi.org/10.1038/s41598-020-67228-7

Answer:
We performed extra investigations concerning the observation of increasing rain during the kharif season, like reading the reference proposed by the reviewer (Praveen et al. 2020). This reference goes against our hypothesis of an increasing precipitation trend during the monsoon season. However, after careful reading, several elements let us think that the results of the above reference can be balanced with other elements from the literature.First, Praveen et al. (2020) only use one source of observed data that is only shortly described. For example, they only mention 34 meteorological subdivisions without specifying the number of weather station and how the data are collected. On the other hand, Katzenberger et al. (2021) include 32 prediction models with associated data source including historical observed data. Jin and Wang (2017) base their conclusions on 13 datasets including the same IMD observed data as Praveen et al. (2020). Second, Praveen et al. (2020) present a neural network prediction method without specifying which predictors are used (it seems to be the past year trend). Moreover, sentences like: “We set the parameters of artificial neural network again and again until best prediction had not achieved which we evaluated using RMSE technique” let us think that there is high chances that models were overfitted. Finally, many results of the prediction on which is based the conclusion of a decreasing or increasing rain trend look over-interpreted or even wrong. For example: (please see the attached file) Nevertheless, we still added those extra elements of literature and revised our conclusion to better emphasize the hypothetical nature of our conclusion. We also consider that, given the fact that the article 
already contains a lot of material, a detailed analysis of the Indian rain pattern and its implication for pearl millet cultivation is beyond the scope of this work (lines 524 to 532).

Jin, Q., & Wang, C. (2017). A revival of Indian summer monsoon rainfall since 2002. Nature Climate Change, 7(8), 587-594.

Katzenberger, A., Schewe, J., Pongratz, J., & Levermann, A. (2021). Robust increase of Indian monsoon rainfall and its variability under future warming in CMIP6 models. Earth System Dynamics, 12(2), 367-386.

L30: “complex agronomic systems”. Is it really? Need to add references and explain why “pearl millet agroecosystem” is complex.

Answer: Pearl millet agroeconomic system is complex because it covers a wide range of conditions in terms of weather with low precipitation to high precipitation, different types of soil from poor sandy soils from semi-desrtic area in the North Rajasthan and black vertisol in the plain of Maharastra. Pearl millet cultivation also covers a wide range of input level and socio-economic usages from rainfed landrace cultivar oriented toward subsistence and higher input cultivation oriented toward market. All those differences create many interactions at different levels of system organization that make the system rather more complex compared to others (e.g. commercial crops grown with high inputs). We added the following references to support our argumentation:

Rao, C. R., Raju, B. M. K., Rao, A. V. M. S., Reddy, D. Y., Meghana, Y. L., Swapna, N., & Chary, G. R. (2019). Yield vulnerability of sorghum and pearl millet to climate change in India. Indian Journal of Agricultural Economics, 74(3), 350-362.

We modified the text accordingly (see lines 31 to 34).

L32-34: References are old, need to add recent publications as references.

Answer: We added the following more recent references about target population of environment definition:
- Casadebaig, P., Zheng, B., Chapman, S., Huth, N., Faivre, R., & Chenu, K. (2016). Assessment of the potential impacts of wheat plant traits across environments by combining crop modeling and global sensitivity analysis. PloS one, 11(1), e0146385.
- Hajjarpoor, A., Kholová, J., Pasupuleti, J., Soltani, A., Burridge, J., Degala, S. B., ... & Vadez, V. (2021). Environmental characterization and yield gap analysis to tackle genotype-byenvironment-by-management interactions and map region-specific agronomic and breeding targets in groundnut. Field Crops Research, 267, 108160.
- Rahimi-Moghaddam, S., Deihimfard, R., Nazari, M. R., Mohammadi-Ahmadmahmoudi, E., & Chenu, K. (2023). Understanding wheat growth and the seasonal climatic characteristics of major drought patterns occurring in cold dryland environments from Iran. European Journal of Agronomy, 145, 126772.

Figure 1: What is AICRP, you mention in the legend of this figure?

Answer: We added the full acronym description in the figure caption: All India Coordinated research project.

Figure 1B: Not relevant to put this figure 1B here, because it is a result, and you don’t speak of it in the introduction.

Answer: We are aware that this figure is a result and that it is not mentioned in the introduction. However, from a didactical point of view, we find interesting to have the initial TPE (Figure 1A) and the final one 
(Figure 1B) next to each other because the reader can have a direct comparison between the initial status and the final TPE proposition. It can also save some space. However, if the reviewer judges it too confusing, we will still separate those two figures.

L60: “… across spatio-temporal scale”. To reword the sentence, because crop model are designed and used, first at plot scale. You can also mention upscaling crop model at a spatial scale. Give references.

Answer: we modified the formulation following reviewer suggestion and added two extra references “CSM reconstruct the soil-plant-atmosphere continuum of a system at the plot level. Those simulation can be scaled up across spatio-temporal scales (e.g. region) to identify TPEs with similar biophysical properties forming the TPE.” (lines 63 to 64)

L78: Figure 2 need to be improved. You mention 2 parts or on the figure, and we have from left to right, 3 parts. The labels are not consistent to the text, in particular what is “CMparameter exploration OPI”. Do the DLD data feed crop model simulation? In particular for soil parameters. Add labels on each blue arrow, explaining the different treatments of the workflow.

Answer: We modified the description of the figure to cover the three parts. We uniformize the labels with the text and added some labels on the arrows to explain the different treatments of the workflow. (lines 82 to 91).

L95: explain how you calculate the trends.
L96: same for crop and area estimation

Answer: We added some extra explanations to further specify how we estimated the variable trend “as their linear regression coefficient estimate over years”. (lines 104 to 108)

L98: You choose the majority soil type, but give some information on the number of soils by district and their importance (% area)

Answer: We are not sure to understand where reviewer 3 identified some information about different type of soil per district. In Table 3, we give some percentage about the majority soil type but at the TPE level (aggregation of several districts). So, the percentage of soil type is the relative proportion of majority soil at the TPE (zone) level. Because the crop model simulations are done for an ‘average’ point per district, we selected the majority soil type for each district. For districts with several type of soil we could have considered a weighted information given the type of soil. We admit that considering the majority type of soil was a simplification to not over complexify this specific question.

L101: “For irrigation, fertiliser and price we could also estimate the trend”. Don’t understand.
Answer: we also calculated trend of % under irrigation, fertilization, and price using the same method of calculation as the area, production, or yield (linear regression coefficient of the variable over years). For example the increase/decrease in irrigated surface over years. We added some clarification in the text (line 113-114).

L106: Crop model often require PET. You don’t mention it, can you explain?
Answer: we are not certain to understand what you mean by PET. Do you mean potential evapotranspiration? The APSIM crop model we used only requires temperature (minimum, maximum), daily precipitation and solar radiation as weather input. The potential evapotranspiration is caculated by APSIM from the input variables. We have re-emphasized the required crop model input (line 121).

L108: I cannot access Ref 30
Answer: We checked again the reference given by the package authors in R and it is the same. We tried to add the url link that connect to the CRAN webpage: https://CRAN.Rproject.org/package=apsimx. But we could not find an option in the MDPI Latex template that allows the printing of the url.

L112: I don’t understand, need to reword this sentence and explain better.
Answer: We wanted to say that we used three sources of data for the PCA to determine the TPEs: district level data, weather data, and information derived from the crop model simulation. We reformulated the sentence as such: “The agronomic district level data and the weather data were complemented by information about the system like the most probable soil depth or sowing date inferred from the crop model simulations (Table 1, see next section)” (lines 128 to 130).

L115: Add a reference for the k-mean clustering method. Why you choose the numbers 3 4 5 for partitioning the data?
Answer: We added a reference to the k-mean clustering (Hartigan and Wong 1979). Concerning the number of selected clusters (3, 4 and 5), we started from the number of clusters in the actual TPE (3) and, after the discussion with experts, extended upto 5 which appeared sensible from the practical point of view and experience of the experts working in those regions. We stopped at 5 because beyond those numbers (6, 7, 8, …), we started to have non continuous TPEs with individual districts from far away 
TPE (e.g. AE2) inserted in other TPE (e.g. B). Those insertion are potentially more due to statistical artifact than reflecting the bio-physical elements of the reality. We added some extra explanations about the selection of the cluster number. (lines 133 to 136).

L130: a modifications
Answer: Corrected

L141: Need to develop more the sources you use for each type of parameters. Example, for soil parameters, why you don ‘t use soil database. Different papers mention black soil, red soil … well described in the Indian context? Do you use survey ? …
Answer:
For the soil type we mostly used the information from the district level data (DLD) because with information at the district level it potentially constitutes the information with the highest resolution. This information was cross-referenced with the data from the ISRIC global database but it offered less resolution compared to the DLD because it contained fewer point with specific soil profiles whose representativity is difficult to assess. For the soil characteristics we selected generic soil type that were the closest to the observed type because we wanted to float parameter related to the water content and the soil depth. For those two parameters, it is impossible to get precise information at the district level for all the regions we covered in our study.Concerning the management practices like the sowing dates and the plant density, we based our choice on the available literature and on practitioners’ information. We were not able to identify a reference with a systematic description of the farmer practices over the regions cover by the study. In terms of variety usage, few systematic data collections (e.g. Asare-Marfo 2010) and information from ICRISAT breeders provided general information like the importance of landrace in the A1 zone and the higher adoption of improved varieties in other regions. This information allowed us to select generic varieties that were likely representative of the regional practice. For fertilization and irrigation, the district level data (DLD) represent the most precise source of information because it contains yearly observations at the district level. The range of those parameters were determined using the DLD. We modified the text to emphasize better our sources (lines 199-237) but we also would like to 
remember the difficulty to compile precise information about the environment and the agronomy of pearl millet cultivation in India. This is the reason why we proposed to use the optimized parameter inference 
strategy to determine the most likely parameters based on the data and the model.
Reference:
Asare-Marfo, D., Birol, E., Roy, D., et al. (2010). Investigating Farmers' Choice of Pearl Millet Varieties in India to Inform Targeted Biofortification Interventions: Modalities of Multi-stakeholder Data Collection. University of Cambridge, Environmental Economy and Policy Research Group.

L151:” we modified the generic APSIM soil profiles”. To reword the sentence and improve the explanation. I suppose it’s only parameterization ?
Answer: Here we wanted to say that we used three soil types (sandy, loam, and clay), for each of those soils, we defined three variants in terms of depth (60, 120, 180), which means at the end nine different soil that 
allowed to test the parameter soil texture (soil water content) and depth. We rewrote the sentence like that:
“For each soil type we defined three soils with depths fixed at 60, 120 or 180 cm - representing shallow, medium and deep soil profiles.” (lines 212 to 214)

L165: “three crop types”. Do you mean “varieties”?
Answer: Yes, we mean varieties. We modified the text following your suggestion.

Paragraph 2.3.2 and table 2: You don’t mention how you manage tillage operations and harvest, in the simulation, indeed these parameters are important, so explain.
Answer: No tillage operation was performed. The crop was harvested (some grain yield was recorded) only if the crop reached the “ripe” phenological stage before the end of the simulation. 

Paragraph 2.3.2 and table 2: You don’t mention how you proceed for the initialization of the model (time 0 of the simulation).
Answer: We added the following sub-section to describe better the simulations setup (lines 157 – 168):
“The CSM were run in APSIM 7.10 (Holzworth et al. 2014). The simulations were initiated on the 15 of April, two months before the hypothetical sowing date, to allow a potentential accumulation of soil water. Inside the defined sowing window, The CM was set to initiate the crop when at least 8.5 mm of rain was accumulated within five days and the soil contained a minimum of 25 mm of available water. If those conditions were not met, the crop was sown on the last day. Crop were sown at a three cm depth. No tillage operation was performed before sowing. Fertilization and irrigation were performed according to the simulation values. The crop was harvested (some grain yield was recorded) one the day it reached 
the “ripe” phenological stage before the end of the simulation (maximum four months after sowing). For each combination of district and parameter, the simulation was re-initialized after each season (year).”

L180: As far I understand what you explain in the paragraph, I think the title isn’t good. It is more “model calibration”.
L 199: “inferred optimized parameters” → calibrated parameters
Answer: Following reviewers’ suggestions, We replaced by what we used to call “optimized parameter inference” by model calibration. The title and the whole manuscript have been modified accordingly. For extra clarification about the calibration, see answers to reviewer 1. (lines 169 to 197 and 247 to 262)

L 204-212: I disagree with the fact that the method you used is similar to cross validation, in particular because you say before that there is a trend over years concerning climate series. You have either to give more arguments and references or to reword the sentence. If necessary, change the way you split 
the years in two parts, in order to perform an appropriate evaluation step.
Answer: We agree with the reviewer that the existence of time dependence between the training and the validation dataset is a special case of cross validation that would require other assumptions. Therefore, we removed the reference to cross-validation.

L209: What do you mean by “present years” and “future years” ?
Answer: The present years were 1998-2007 and the future years 2008-2017. We re-emphasis this information in the text. (line 276 to 277)

L212: Explain what is “rhô”, and what method you used for the comparison?
Answer: “rho” was the Pearson correlation; it is already defined in the previous section (lines 258-259). However, we respecified it and mentioned the reference data:
“The prediction abilities over time, and over space were also estimated by $\rho$ between the predictions and the observations.” (lines 286 to 287)

L214: I disagree that making a regression on outputs is a sensitivity analysis. You have either to give more arguments and references or to reword.
Answer: We perform a regression of the output (yield) on the inputs not on the outputs. According to us, the use of a regression is a valid method to evaluate the way the crop model behaves given variation in the parameter values (inputs). With the large amount of data we had and the large number of parameters we floated the linear regression is interesting because it allows the assessment of each parameters conditionally on the others. We extended this section to add more clarity to the material and 
methods description (lines 288 to 295).

L228: concerning weather data, you have to precise the semantic of all the variables you put in table 2: cumulative rain? Average temperature? …
Answer: We suppose that the reviewer talks about the variable definition of table 3 and not table 2, which makes no reference to weather data. We did a complete revision of the semantic of each variable from table 3.

Table 3: Need: i) to give the semantic of each variable of the table (for example , I don’t understand what is “Share btw env”), ii) to improve the presentation of the table in an harmonized way (ex sometimes variances are mentioned in () and sometimes are in sepcif lines ).
Answer: i) see previous answer, ii) We harmonize the way table three is presented (lines 324 to 325). For each variable we calculated the average value of the TPE over the period and the standard deviation in parenthesis. The trend values are the linear regression coefficient over year with standard deviation. The soil type information does not have standard deviation because it does not change. For the crop model parameters, since we obtained one value per district for the whole time series, it represents the average over districts of the TPE with standard deviation. We extended the table to add this extra information.We made a small exception for the total rain variance over seasons. We consider that this variable should be presented as such because it gives an information about rain variability, which is a crucial information in a system that can be perturbated by the erratic nature of the precipitation.

Table 3: I suggest you only put in this table the information concerning the new TPEs: AE1 A1 B G AE2. In addition, after in “Results” you comment only the elements of this typology.
Answer: Even though the suggestion of the reviewer could improve the table readability, we prefer to stick to a table with seven columns that allows the reader to get descriptive about all TPE scenarios presented with three, four or five TPE. Indeed, during the discussions we had with the practitioners and Indian pearl millet experts, it came out that it was better to keep some flexibility concerning the number of TPEs because the translation of our recommendations in reality can depend of many factors like the available means to run separate activities for extra TPE, the logistic in place like the existing testing centers, or other consideration about the breeding activities in those regions not captured by our study. Therefore, experts appreciated the possibility to have some latitude in the choice of the final picture that would be the most suitable for them. The actual version of the table would reflect more this flexibility offered to the reader.

Table 3: The rain variances are very big. Need to check.
Answer: We checked again the rain variance. The values are high because the average difference in cumulated rain between season is 137 mm which correspond to a variance of 18’769 mm2. The rain differences are particularly high between seasons in Gujarat (G) due to its proximity to the sea.

L242-L275: I suggest you have 2 paragraphs : 3.1.1 for AE1 TPE and 3.1.2 for AE2.
Answer: We split the paragraph into two paragraphs according to the reviewer suggestion. (lines 325to 373)

L300: Explain better what table 4 contains, and the method used to calculate the different indicators
Answer: We first improve the description of table 4 calculation in the ‘Crop model evaluation and parameter stability over time and space’ paragraph (lines 264 to 287). We also reinforced those explanations in the table caption (lines 435 to 436).

Paragraph 3.3: This statistical part seems not relevant. You cannot conclude on any effects with these R2. Maybe check or improve the statistical work behind.
Answer: We used the difference R squared to assess the relative importance of each parameter on the estimated yield. The difference R squared estimate the contribution to the R squared of a specific variable by calculating the difference in R squared between a full model containing all the parameters and a model without the parameter of interest. The R squared metric correct the estimated contribution by accounting for the other effects. According to us it is a valid metric that was uniformly applied to all parameter in all configuration. Every time the reference model was the same, which prevent from comparing model with different content, a general criticism to the R squared statistics. R squared statistics are well spread and widely understood and have been used in many configurations from crop model assessment (Wallach et al. 2019) to genetics (Sun et al. 2010). Nevertheless, we recalculate the parameter relative contribution using alternative metrics like the parameter relative sum of squared 
compared to the dependent variable (yield) sum of squared derived from an ANOVA. We also calculated the relative proportion of the parameter effect range on the yield range. We did not find substantial differences in the parameters’ importance, except the fact that according to the proportion of parameter range effect, the weather data (rain, temperature) and irrigation had a larger effect. We added those results to the supplementary material and updated the manuscript accordingly (lines 298 to 300 and 456to 463).
References:
Sun, G., Zhu, C., Kramer, M. H., Yang, S. S., Song, W., Piepho, H. P., & Yu, J. (2010). Variation explained in mixed-model association mapping. Heredity, 105(4), 333-340.
Wallach, D.; Makowski, D.; Jones, J.W.; Brun, F. Working with dynamic crop models: methods, tools and examples for agriculture and environment; Academic Press, 2018

Paragraph 3.4: this paragraph is not clear at all. Need to reword. Why a focus on irrigation and fertilization OPI parameters?
Answer: We rewrote the whole paragraph and shorten it. We focused on irrigation and fertilization because they are the only parameters that allow a direct comparison between the estimated ExM parameters and some observations (lines 472 to 503).

Table 4: I don’t understand the meaning of this table. Need to explain better the method used and what you put in this table.
Answer: Cf previous revision of the description of table 4.

L362: github repository doesn’t exist
Answer: We checked again the provided link to the Github repository 
(https://github.com/vincentgarin/PMapp). It is accessible. We also checked the sequence to install and run the app as well as a quick check of all the application function, without any problems devtools::install_github("vincentgarin/PMapp")library(PMapp)run_app()
According to us the problem of accessibility is due to the compilation of the url link in the MDPI latex template that does not break the link and mix it with the line numbering information, which mess up the link. We modified the text to have the full link printed correctly.

L363: errors when using the Rshiny application
Answer: We checked again the application https://agrvis.shinyapps.io/PMapp/ and did a rapid tour of all the functionalities without encountering any major errors. In some cases when no district is selected 
or in specific combination of variable, district, and year with missing values some information cannot be plotted and an error message is thrown but this is due to data availability not due to the application code.

L381: “increasing kharif rain” same remark than L10
Answer: see previous answer at the beginning.

Paragraph 4: you don’t mention the importance of soil. It is not consistent of what you present in the Results part.
Answer: Even though the influence of soil properties was already mentioned in the conclusion, we emphasized better those results in the revised version of the conclusion (e.g. line 648-649, 674 or line 
687)

Supplementary material

Add information on climate (map of rain, temperature and trends).
Answer: We added two new figures about rain and temperature pattern (average and trend) in the supplementary material

Author Response File: Author Response.pdf