Winter Wheat Mapping Method Based on Pseudo-Labels and U-Net Model for Training Sample Shortage

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study proposes a method that utilizes the random forest algorithm to establish a pseudo-labeled sample set for training a semantic segmentation model (U-Net), aimed at extracting winter wheat. The research initially extracted winter wheat using the WCI index on the Google Earth Engine (GEE) platform, and then constructed training samples (i.e., pseudo-labels) for the semantic segmentation model through an iterative process of generating random sample points, training a random forest model, and extracting winter wheat. Subsequently, the researchers trained the U-Net model using multi-temporal remote sensing images and ultimately applied the model to generate a spatial distribution map of winter wheat in Henan Province in 2022.

It is a commendable application case of U-Net. However, I have several questions:

1. What is the role of pseudo-labels in the proposed method? The authors handle this point ambiguously, dedicating extensive sections to introducing the advantages of pseudo-labels, but without experimental validation in the study. That is to say, without pseudo-labels, how does the algorithm perform?

2. U-Net is a network model with excellent image segmentation capabilities, which has been widely applied in medical imaging and remote sensing since its proposal in 2017. After several years of continuous optimization and innovation, many excellent segmentation networks have emerged, including even derivative versions of U-Net++, such as U-Net++. Why did the authors choose to use U-Net for segmentation? This straightforward application seems to lack innovation.

3. The experimental section of the paper lacks comparative experiments. It is not appropriate to compare with random forest, first, it is not a popular method for image segmentation; second, comparing deep learning with traditional machine learning is inherently unfair. The authors are recommended to use popular deep learning models for comparison.

4. Regarding the issue of insufficient training data, it seems that the authors have not conducted sufficient experimental validation during the study.

 

Revisions are recommended as follows:

1. The steps of the technical approach should correspond one-to-one with the subsections mentioned later in the text.

2. Variables should be set in italics (for example, in line 338 of the pseudo-code).

 

3. There is a lack of complete reference information, such as in reference 4, which is from 2015. Why is the DOI number used instead of the volume and issue information?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents a novel technique to construct a pseudo label for winter wheat field extraction and demonstrated that the semantic segmentation model over performs random forest model in such a large area task. The approach proposed in this paper is promising and would be beneficial to many applications. However, I still have some concerns that should be addressed before it can be published.

 

One of my biggest concerns is that would the pseudo labels generalize well for areas with crop types and climate conditions in different years, as it’s extracted based on the phenological feature and refined with RF. For now, it’s only verified for the same crop type in nearby area within the same year. It’s not reliable, at least before it’s tested for different growing seasons. Authors may refer to some papers using trusted pixels to generate training labels without field data for crop type mapping to see how the reliability of pseudo labels is reached.

 

Line 18: WCI is not explained in its first appearance. Please also check other abbreviations, such as OTSU.

 

Line 166: Should be Figure 2.

 

The usage of “study area” is quite confusing. Sometimes it refers to area A, while in some cases it refers to Henan Province. Please define and declare the study area of Henan Province and typical area A, and other study areas if exists, and try not mixing them.

 

Line 190: if not, delete this line.

 

One concern about generating pseudo-labels using random forest is the over-fitting problem based on the high overlap rate.

 

Line 230: Dates of data here are not consistent with that in Table 1. Dose the number “3” in Data Processing part of Figure 3 mean the “3” time periods in Table 1.? How are the multi-time series NDVI constructed? It’s a little confusing here. I would suggest declaring the number of sentinel 2 images collected in table 1 for each period, and the number of images used in generating each NDVI series.

 

Title in Figure 4 is not right. Please correct it.

 

Line 307: The radiometric resolution of Sentinel 2 MSI is 12 bit, which means the pixel value of sentinel 2 images ranges from 0-4095. Typically, radiometric resolution of remote sensing images is in the range of 8 to 16 bits. So, the statement here is not right. Please correct it.

 

Line 325: c in 512*512*c is not clear. I guess it’s 4 as line 204 indicates, but what it means is not clarified.

 

Line 340: It’s Table 3. Please check all the tables and figures for their order and titles.

 

Please remove the template words in Line 370-372.

 

Please point out the location of Nanyang city and Luoyang city in Figure 6 so that readers can easily track the differences authors are analyzing between the results of two models.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The idea of ​​using a combination of machine learning and deep learning techniques to increase the accuracy of the winter wheat extraction method seems to be a good one, with promising results. The study is very well designed, put into practice and explained.

However, although I understand the intention of the authors to compare machine learning algorithms with those of deep learning, thus choosing to compare the results of the U-Net model with those of the random forest model algorithm, as an alternative method of the extraction of winter wheat, I consider that it would have made more sense to compare the U-Net model with other semantic segmentation models. I suggest the authors to consider this in future research.

Author Response

Please see the attachment.

Author Response File: Author Response.docx