Bayesian Posterior-Based Winter Wheat Yield Estimation at the Field Scale through Assimilation of Sentinel-2 Data into WOFOST Model

Round 1

Reviewer 1 Report

The manuscript entitled “Bayesian posterior based winter wheat yield estimation at the field scale through assimilation of Sentinel-2 data into WOFOST model” is an interesting research on the modelling  wheat yield. The use of NDVI to determine the LAI deserves special attention. This is an innovative approach to modelling the yield from satellite images, so far many researchers have used NDVI alone for this purpose.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This manuscript focused on field-level yield estimation with Sentinel-2 LAI and the crop growth model using a Bayesian method. This is an important try for yield estimation at fine resolution, and should be considered for potential publication in this journal. I personally recommended the minor revisions of this manuscript as some small issues needed to be handled. 

1.      In this study, the “prior” and “likelihood” are mentioned many times, but seems not to be consistent in the context. In Line187-198, the authors call “the parameter priors” are the nine single-valued variables in Table 2 and (DVS, LAI, TAGP, and TWSO) are variables for constructing the likelihood function. However, in Line16-18, Line74-76, Line105-106, Line147-149, and Figure 3, the authors never mention “the parameter priors”, but declare “county-level yield statistics, field observations” are used to construct “the prior and likelihood”. Are these "priors" the same as the "priors" mentioned in the Line187? Also, what is “prior information” in Line105? This is confusing and need to be clarified. 

2.      The parameter ranges in Table 2 are derived from references, actual measurements, or personal experience? 

3.      It's better to place the sub-picture serial number of Figure 6 in the picture frame. 

4.      I do not agree the reason explain in Line289-293. In my opinion: the "the posteriors of the model simulated variables" is a compromise between prior and likelihood (observation), according to the Bayesian framework. That is the reason why the "the posteriors of the model simulated variables" be different from observation. In addition, as far as I know "the posteriors of the model simulated variables" is more commonly called "the posterior prediction distribution".

 5.      The description of Figure 10 is not sufficient, what do the different subgraphs actually mean for the model?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

In this study, the Authors proposed a novel data assimilation framework. Firstly, they constructed the county-level prior and likelihood constraints for a process-based crop growth model based on the previous year’s statistical yield and the current year’s field observations.

This is an extensive research, with a lot of numerical analysis. 
Thematically the manuscript is well approached for the researchers and professionals and the proposed manuscript is relevant to the scope of the journal.

I found it appropriate for publication in the Remote Sensing journal, but only after some modifications and clarification from the Authors.
The title is a clear representation of the manuscript's content, but the list of keywords could be improved, by adding (or changing) one or two more terms.

The overall organization and structure of the manuscript are appropriate and well structured. The paper is relatively well written and the topic is appropriate for the journal.
The aim of the paper is well described and the discussion was well approached, its results and discussion are correlated to the cited literature data.
In the introductory part, the authors give elaboration of the overall context stating the motivation and the objectives of the work, literature review of the research pathways .
The literature review is comprehensive and properly done.

It has to be clearly stated by the authors what is their contribution that makes the research different enough in comparison to the other authors' works and they have to further elaborate the extent of novelty in their research. The novelty of the work must be more clearly demonstrated.

The significance of the Work: Given the large number of analyzed data, this is an interesting study with a possible significant impact in this area.
Statistical interpretation of the analytical data must be more properly presented.

The verification of the model should be performed with more details. 

Other Specific Comments: The work is properly presented in terms of the language. The work presented here is very interesting and well done, it is presented in a compact manner.

The results are presented in a logical sequence and the discussion and analysis of the results are properly elaborated.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

My problem with this work is that I strongly disagree with the conclusion: "We apply it in a prefecture-level area for field-level yield estimation and prove its suitability with overall high accuracy." I'm an agronomist. I worked at an agricultural cooperative and saw, heard and learned from the farmers. They would never accept an estimate with that degree of inaccuracy. Mainly if we look at the real data that resulted in the statistical values ​​and observe that the errors at each location can be huge. In figure 13, for example, we can see that for the observed value of 4100 kg/ha it was estimated at 5900 kg/ha, for an observed value of 7000 kg/ha, the simulated value was 5500 kg/ha. In other words, the model can individually miss more than a ton, more or less. This does not meet the needs, neither of managers, as a source of agricultural statistics, nor for the farmer to manage his planting area. The effort and amount of data and technology employed are enormous for a result far from "general high precision".

I really can't say how best to present all the work done, but it certainly shouldn't be based on a good quality of fit between simulated and observed data.
I appreciate the effort, but I don't think the proposed model can be of any real use. The use of a Baysean model, seemed to me to only complicate even more, both the calculations and the general presentation of the work. Statistical preciousness is not justified if the observed data are not explained by the hypothesized causes of variation. What happens is that between the spectral response value of a pixel (and any derived index such as ndvi, lai etc) and the reality of photosynthesis in wheat leaves and grain filling, there is still a huge amount of influencing factors, which are not well represented by the model. Agrometerological data, also the result of simulation models, available at scales totally different from the scale of pixels and even more different from the scale of plants, as much as they help, they are not enough to generate good results at the intended scale.

There is still the problem of availability of cloud-free images in the quantity and opportunity needed to be used in the model. There is no consideration of how to proceed in the absence of images about locations or at critical occasions. A farmer cannot wait for the satellite to pass on a cloudless day in order to make his decisions. Likewise, a manager who intends to monitor production at the pixel scale will not be able to give up the estimate in areas where images were not obtained during the crop development period.

Reviewer 2 Report

The authors present a promising approach to assimilate R/S derived parameters into a sophisticated crop growth model. This is the way to go to make full use of the large magnitude of R/S data available today. The approach of using the R7S derived data is sohisticated and uses appropriate statistical tools to allow for a clear analysis of uncertainties. Nevertheless it is a pitty that the authors are so conventional in deriving the LAI as the central R/S parameter from NDVI correlations. This results in a pretty poor result of the relation betweenNDVI and LAI in Fig.6. It reveals the "old" problem of saturation of NDVI. All in all the novel approach, which has been carried out thoroughly and with competence, justifies publication of the paper with minor revisions mainly regarding language polishing.