A Spatial–Temporal Depth-Wise Residual Network for Crop Sub-Pixel Mapping from MODIS Images

Round 1

Reviewer 1 Report

This paper presents a novel spatial-temporal deep learning-based approach for sub-pixel mapping of different crop types within mixed pixels from MODIS images. 

This paper can be improved on the following aspects. 

1. A systematic literature review about sub-pixel (or superpixel) mapping is required in the Introduction section. Authors may refer to the relevant articles from Professors Peter Atkinson, Qunming Wang, and Yuehong Chen. 

2. Model validation should be added in the Methods and Results sections. 

3. Data and Methods should be separate sections. 

4. Authors should notice a space between numbers and units. For instance, "20m" should be written as "20 m".

5. "3.1. Experiment Settings" should be moved to the Methods section instead of the Results section. 

6. Results and Discussion should be separate sections. The discussion section is used to discuss the reasons for the primary findings in the study. 

7. The proposed approaches should be compared with previous sub-pixel mapping models that Professor Peter Atkinson and others developed. 

8. In the conclusions section, the numbers of findings should be revised as "First, ..., Second, ... etc". The findings have been mentioned in the above Results and Discussion sections, so it is unnecessary to present them again. You can summarise the primary findings and contributions of the study. In addition, it is essential to highlight the study's contribution to broader research fields. 

 

 

 

 

Author Response

Dear Reviewer,

We thank and appreciate the comments and we have revised the manuscripts accordingly to address all of the concerns raised by the reviewers and the editors. We hope that the revised manuscript is considered ready for publication in revised form. Please see the attachment about the specific content.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presents a significant advancement in satellite image
processing of MODIS imagery and the authors have clearly indicated
the points at which their own work has gone beyond what is already
known. It is however important that the authors notice in their
paper that all the considerations expressed in their paper and the
applicability of their methods ultimately depend not only on the
resolution of the imagery, but on the spatial complexity of the area
also (for more information about this they may consult the book by
Fivos Papadimitriou "Spatial Complexity: Theory, Mathematical
Methods and Applications" Springer 2020). With this addition, the
paper will become ready for publication.

Author Response

Dear Reviewer,

We thank and appreciate the comments and we have revised the manuscripts accordingly to address all of the concerns raised by the reviewers and the editors. We hope that the revised manuscript is considered ready for publication in revised form. Please see the attachment about the specific content.

Author Response File: Author Response.pdf

Reviewer 3 Report

This research project is intriguing and important. However, this article needs to have its experimental design enhanced, the sections optimized, and the wording and expression polished. To make the claims solid for publication, a substantial revision is suggested. General comments -Why do you select two nearby study areas as opposed to two distant study areas like Hao et al. (2020), given that transfer learning is beneficial? Could you please explain how representative these two study areas are? -Since Landsat and Sentinel can be used to produce CDL, why did you use the coarser MODIS products to produce more refined crop mapping? Is it just because of the higher time resolution? Is your method superior to Landsat and Sentinel-2 in crops' early mapping? -I'm unable to assess your creative work because the explanations of the comparative methods and the description of the proposed method are both underdeveloped. -You used Unet and ESPCN as comparison algorithms, but these methods, proposed several years ago, are hardly called SOTA. -Why use mIoU? In the field of remote sensing classification, are Kappa coefficients and F1-score not more commonly used evaluation indicators? -The results and discussion shouldn't be combined in one part, and I didn't find enough discussion in your article. Do you compare them to related literary works? -Transfer learning has proved to be effective in numerous current studies such as Hao et al. (2020) and Zhang et al.(2022). It can be well performed by traditional machine learning such as the random forest classifier. Why was it poorly implemented in your deep learning methods?   [1] Hao, P., Di, L., Zhang, C., Guo, L., 2020. Transfer Learning for Crop classification with Cropland Data Layer data (CDL) as training samples. Sci Total Environ 733, 138869. [2] Zhang, L., Gao, L., Huang, C., Wang, N., Wang, S., Peng, M., Zhang, X., Tong, Q., 2022. Crop classification based on the spectrotemporal signature derived from vegetation indices and accumulated temperature. International Journal of Digital Earth, 1-27.   minor concerns: L16. mIoU. Please ensure that the first written abbreviation in the abstract and the main text includes the full name. L37. What is "small diverse parcel size distributions"? I cannot find any explanation from the reference [24] you pointed. L95-103 & L116-117. Please ensure that the active voice throughout the text is the same. L116. "countries"? Please verify. FIgure 1. Legend was missing. Section 2.1.3. Please add the description of the remote sensing data producing CDL. L150-151. Why did you remove other crops and non-crop covers instead of treating them into another two classes to be classified? Please verify. Figure 2. Why were there "confusing phenological periods"? Was there wrong reference data? Section 3.1. Experimental settings. L179-194. These two paragraphes are more suited for methodology than the results section. Figure 3. Please add the essential descriptions of layers, including the full names of layers. Besides, was there a add_ or concatenate_layer between BN and ReLU layers before "output"? L189-191. What do you mean by these two sentences? Did you use two training sets with different sample ratios in Sherman in 2017? What about the dataset in 2018? L235-248 and Table 2. Why did you attribute the poor transfer learning results to phenological differences? In my opinion, there were no obvious crop type adjustments from 2017 to 2018 in these two adjacent counties as shown in Figure 4. Could the failed transfer learning be attributed to the problem of your model?

L263-268. These statements are unacceptable because your experimental results and analysis are insufficient.

Author Response

Dear Reviewer,

We thank and appreciate the comments and we have revised the manuscripts accordingly to address all of the concerns raised by the reviewers and the editors. We hope that the revised manuscript is considered ready for publication in revised form. Please see the attachment about the specific content.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

All my concerns have been addressed. 

Author Response

Dear Reviewer,

Thank you very much for the suggestion. Some tense mistakes in the manuscript have been corrected. And the language has been edited by the native coauthor accordingly. We hope the reviewer found the revised manuscript clearer to read.

Reviewer 3 Report

With some revision as follows, the paper could be published: 1, There are still numerous mistakes about that active voice in your revised manuscript such as: L109 "We use", L116 "We conduct", L134 "we selected", L139 "we chose", L148 "We used", L209 "we improve", L233 "we calculated", L235 "we fixed", L241 "we do not make", L249 "we converted", L274 "we compare".... Why were some sentences in the present tense and some in the past tense? 2, The performance indicator, F1 score, can provide an accurate value for each class like You & Dong (2020). However, in your revised manuscript, it seems like an overall indicator similar to OA and Kappa. Please verify.

[1] You, N., Dong, J., 2020. Examining earliest identifiable timing of crops using all available Sentinel 1/2 imagery and Google Earth Engine. ISPRS Journal of Photogrammetry and Remote Sensing 161, 109-123.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf