Round 1

Reviewer 1 Report

The manuscript entitled "Bidirectional Segmented Detection of Land Use Change based on Object-Level Multivariate Time Series", the authors shows a bidirectional segmented detection method based on Object-Level Multivariate TS to sample selection and salt-and-pepper noise of pixels. In general, the article is well-written (only be careful with the past tense in some topics), presenting the methodology and the results comparing with other methods.

Anyway, the authors should improve the introduction section showing what fails in the other methods, and why we should use the new proposed method.

In this sense, I consider the proposed article suitable for publication on Remote Sensing, after minor review.

Author Response

Response to Reviewer 1 Comments

Point 1：The manuscript entitled "Bidirectional Segmented Detection of Land Use Change based on Object-Level Multivariate Time Series", the authors shows a bidirectional segmented detection method based on Object-Level Multivariate TS to sample selection and salt-and-pepper noise of pixels. In general, the article is well-written (only be careful with the past tense in some topics), presenting the methodology and the results comparing with other methods.

Response 1：Accepted and revised. First of all, thank you for your positive comments of our manuscript.

As for the problem of the past tense in some topics, we carefully checked the manuscript and corrected the wrong tense. In addition, we have asked English professionals to polish the manuscript.

As you said, we have compared with the LandTrendr method in terms of detection time accuracy, type accuracy, etc (Line 402~426).

Point 2：Anyway, the authors should improve the introduction section showing what fails in the other methods, and why we should use the new proposed method.

Response 2：Accepted and revised. Your suggestion is very helpful to improve the manuscript. According to your suggestion, we reflect on the disadvantages of traditional methods and elaborate the advantages of new methods in the introduction. The details are as follows.

Many of the aforementioned methods use a single index to measure the similarity of TS, which highly depends on the accuracy of the selected sample points, and cannot detect change time. However, given the variety of land use types and the uncertainty of temporal nodes, it is difficult to accurately select samples for all types and times of change for the long-term study of land use change. This in turn affects the accuracy of change detection. Moreover, detecting land use changes based on pixel-level TS has stringent requirements for the registration accuracy and radiation correction of remote sensing images. The results also contain significant salt-and-pepper noise, which limits the application potential of the approach (Line 63~70).

Many studies focus on analyzing the characteristics of the TS of each segmented object to detect changes in land use. The object-level TS method has advantages over the pixel-level TS method in terms of precision. Nevertheless, there are still many challenges. First, the mean value of all pixels within each object is generally selected as the object’s representative feature in the TS. The use of the mean weakens the characteristics of the majority pixels because pixels close to the object’s edge are less representative. This in turn reduces the discriminability between the object and other objects. Second, the types and time of land use change vary in general. To ensure accuracy, samples of all the types and times of change must be selected and used in the detection of change. This is a difficult task to implement. Therefore, detection methods based on incomplete samples cannot accurately identify the type and time of complex land use change (Line 85~94).

To address the challenges outlined above and detect the types and times of land use change from Landsat images, we propose bidirectional segmented detection (BSD) based on object-level multivariate TS. The proposed method uses the median of an object because the median value of a segmented object is more representative than the commonly used mean value…We mainly solve the problem of detecting the change time and change type of land use at the same time without the changed samples (Line 95~102).

Point 3：In this sense, I consider the proposed article suitable for publication on Remote Sensing, after minor review.

Response 3：Thank you again for your positive comments.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper entitled "Bidirectional Segmented Detection of Land Use Change based on Object-Level Multivariate Time Series" is an interesting paper that tries to identify several land features, both historical and contemporary, based on the usage of time series satellite images. While the approach used is clearly described, and sometime too much detailed, my main concern would be related to the implementation of such study on a limited number of land cover types. Maybe authors could add few more detailed types and then check for the accuracy of the proposed method. Anyhow, my detailed comments are as follows:

Lines 95-96: I don’t understand how the usage of median instead of mean would reduce the salt-and-pepper noise. Lines 181-187: With the usage of such outstanding approach, maybe it would be better to select more land use types and go further into details, such as crop types. While these categories represent commonly available land cover and use classes found in previously produced map, I believe that the distinguishing between these classes was produced in a more straightforward and time saving process (e.g. Faour et al., 2018; Nasrallah et al., 2018), even using the NDVI index only. Thus, the importance of using the proposed approach would be questionable.

Faour, Ghaleb, Mario Mhawej, and Ali Nasrallah. "Global trends analysis of the main vegetation types throughout the past four decades." Applied geography 97 (2018): 184-195.

Nasrallah, Ali, Nicolas Baghdadi, Mario Mhawej, Ghaleb Faour, Talal Darwish, Hatem Belhouchette, and Salem Darwich. "A novel approach for mapping wheat areas using high resolution Sentinel-2 images." Sensors 18, no. 7 (2018): 2089.

Kindly change “pure” object to “homogenous” object throughout the manuscript. Lines 205-207: How and why boundaries were selected manually (yellow boundaries)? More discussions and clarifications should be made to Table 3 and Figure 3. This should replace Lines 213-217. Lines 218-229: Very common and not needed. The same applies for Figure 4. Lines 254-262: Repeated information. Line 291: This section should be placed in the Method section not in Results. Tables 4 and 5 are not needed. Lines 311-323: If the scale with the largest local variance (i.e. 44) with too many unneeded details was chosen, then what was the importance of using ESP? Section 4.3 should also be placed in the Method section. My main concern would be why the usage of a limited number of land uses classes. Any straightforward approach would generate same accuracies.

Author Response

Response to Reviewer 2 Comments

The paper entitled "Bidirectional Segmented Detection of Land Use Change based on Object-Level Multivariate Time Series" is an interesting paper that tries to identify several land features, both historical and contemporary, based on the usage of time series satellite images. While the approach used is clearly described, and sometime too much detailed, my main concern would be related to the implementation of such study on a limited number of land cover types.

Point 1：Maybe authors could add few more detailed types and then check for the accuracy of the proposed method.

Response 1：Accepted and revised. Your suggestion is valuable for improving the quality of manuscript. The land use types selected in this manuscript (water area, woodland, paddy field, construction land, and dry land) are generally concerned types in the study of land use change in different regions. The contribution of our research is to solve the problem of detecting the change time and change type of land use at the same time using the samples with unchanged land use type. Our method does not need the samples of land use change. It makes the sample selection easy and the land use detect stable. The recognition of crop type is different from the detection of land use change. Moreover, it needs a higher time-resolution remote sensing image series. Due to the influence of cloud and rain weather, the number and phase of available optical remote sensing images are limited and vary in different regions and years. Maybe, combining phenological information, optical and SAR multiple remote sensing image series, can solve the problem. In the future study, we would like to carry out relevant exploration.

Point 2：Lines 95-96: I don’t understand how the usage of median instead of mean would reduce the salt-and-pepper noise.

Response 2：Accepted and revised. We are sorry to misunderstand you for your careless writing. It should be “The proposed method uses the median of an object because the median value of a segmented object is more representative than the commonly used mean value.”. We corrected it (Line 95~96). The detail explanation is in the later experiments (Line 223~228).

Point 3：Lines 181-187: With the usage of such outstanding approach, maybe it would be better to select more land use types and go further into details, such as crop types. While these categories represent commonly available land cover and use classes found in previously produced map, I believe that the distinguishing between these classes was produced in a more straightforward and time saving process (e.g. Faour et al., 2018; Nasrallah et al., 2018), even using the NDVI index only. Thus, the importance of using the proposed approach would be questionable.

Faour, Ghaleb, Mario Mhawej, and Ali Nasrallah. "Global trends analysis of the main vegetation types throughout the past four decades." Applied geography 97 (2018): 184-195.

Nasrallah, Ali, Nicolas Baghdadi, Mario Mhawej, Ghaleb Faour, Talal Darwish, Hatem Belhouchette, and Salem Darwich. "A novel approach for mapping wheat areas using high resolution Sentinel-2 images." Sensors 18, no. 7 (2018): 2089.

Response 3：Accepted and revised. We carefully read the literature you provided, found and many valuable contents, and analyzed the research progress in the introduction part (Line 60~61). The article of Faour et al. uses the method of comparison after classification to detect land use change. This method is simple, but has error accumulation. The classification accuracy of every phase affects the final detection results. Nasrallah et al. extracts wheat field information based on the characteristics of NDVI time series of one year, which is difficult to be used for detecting different types of land use change.

The method we proposed can detect a variety of land use change time points and types, and has a wider range of application（Line 62）.

Point 4：Kindly change “pure” object to “homogenous” object throughout the manuscript.

Response 4：Accepted and revised. Your suggestion is valuable and more precise. We have changed the pure object to the homogenous object.

Point 5：Lines 205-207: How and why boundaries were selected manually (yellow boundaries)?

Response 5：Accepted and revised. There are mixed pixels in the segmented objects, especially near the boundaries of the objects. To ensure the selected objects homogenous and represent the majority pixels of the segmented objects, we compared with Google high-resolution image and manually sketch the boundary of homogenous objects. Certainly, as long as the selected homogenous objects are within the ground truth boundary, the results of experiments in Table 3 and Figure 3 are consistent. Experiments show that the median value of the object is more representative of the main characteristics of the object than the mean value.

Point 6：More discussions and clarifications should be made to Table 3 and Figure 3. This should replace Lines 213-217.

Response 6：Accepted and revised. Your suggestion is of great help to the improvement of our manuscript. We discussed Table 3 and Figure 3 fully and replaced 213-217. The modified content is as follows: The segmentation results show that there are many mixed pixels in the segmented objects in the construction land. We choose four mixed samples of construction land to compare the mean and median NDVI value of pixels in the objects. Meanwhile, we choose the samples of nonconstruction area (woodland land, water area, dry land, and paddy field) for comparison (Table 3). For each sample, we manually selected a homogenous object of majority pixels within the segmented boundary (indicated by a yellow line). The experimental results show that the mean and median NDVI values of nonconstruction land are almost the same because the mixed pixels in the segmented objects of nonconstruction land are fewer. For each sample of construction land, the mean NDVI values of a multiresolution segmented object and a homogenous object are quite different, whereas their median NDVI values are almost the same. To further validate the results, we selected 300 segmented objects and corresponding homogenous objects of construction land for statistics. Figure 3 shows that the median NDVI values of segmented objects are closer to the mean NDVI values of corresponding homogenous objects. (Line 209~220).

Point 7：Lines 218-229: Very common and not needed. The same applies for Figure 4. Lines 254-262: Repeated information.

Response 7：Accepted and revised. Your suggestions can make the structure and content of themanuscript clearer. We have deleted unnecessary content and duplicate information.

Point 8：Line 291: This section should be placed in the Method section not in Results. Tables 4 and 5 are not needed. Section 4.3 should also be placed in the Method section.

Response 8：Accepted and revised. Your suggestions can make the structure of the manuscript clearer. We have moved the content of 4.1 to 3.3.1 of the method part (Line 181~192), and deleted the unnecessary method part in Table 4 and Table 5. We have moved Section 4.3 to the method part (Line 288~299), while the result of precision evaluation remains in the result part (Line 345~362).

Point 8：Lines 311-323: If the scale with the largest local variance (i.e. 44) with too many unneeded details was chosen, then what was the importance of using ESP?

Response 9：Accepted and explained. For multi-resolution segmentation, numerous studies have demonstrated the importance of the scale parameter. The scale parameter controls the dimension and size of segmented objects, which may directly affect subsequent results. In numerous applied studies, land-cover extraction mainly relied on a trial-and-error approach, with segmentation scale parameters determined based on previous experience. However, this approach has been deemed inadvisable. The other researchers have since proposed methods to determine optimal segmentation scale parameters. As one such successful approach to scale optimization, Local Variance (LV) and Rates of Change of LV (ROC-LV) can be combined to determine appropriate segmentation scales. The corresponding Estimate Scale Parameter (ESP) tool was made public for optimizing scale parameters and has been successfully implemented in the eCognition software.

Point 9：My main concern would be why the usage of a limited number of land uses classes. Any straightforward approach would generate same accuracies.

Response 9：Accepted and revised. Your suggestions are of great value to the in-depth promotion of the manuscript. The land use types (water area, woodland, paddy field, construction land, and dry land) selected in this manuscript are more concerned types in the study of land use change in different regions. The main contribution of this paper is to realize the simultaneous detection of change time and change type. As the answer to the first point, the recognition of crop type is different from the detection of land use change. Moreover, it needs a higher time-resolution remote sensing image series. Maybe, combining phenological information, optical and SAR multiple remote sensing image series, can solve the problem. In the future study, we would like to carry out relevant exploration.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript “Bidirectional segmented detection of land use change based on object-level multivariate time series” solves the problem of the complexity of sample selection and the salt-and-pepper noise of pixels when detects the type and time of land use change from Landsat images by proposing a bidirectional segmented detection (BSD) method. And by combining the median of the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Built-up Index (NDBI), and the Modified Normalized Difference Water Index (MNDWI) with the BSD to identify the type and time of the land use. The study of time series of remote sensing images is a hot topic and a remote issue that is in line with human interests, and provides a good basis and reference for detection of land use change. The data used in this study is reliable and used a series of data sheets to illustrate the problem. The analysis is reasonable and logically clear. Conclusion is convincing. I think it is worth publishing in the journal.

I read this manuscript carefully and found that there are still some questions, listed below:

Proper use of proper nouns, for example: Line 23-24: “normalized difference vegetation index (NDVI), normalized difference built index (NDBI), and modified normalized difference water index (MNDWI)” should be examined carefully, especially the normalized difference built index.

 

Figure 4. The NDVI, NDBI,and MNDWI ’s full name had already appeared. Figure 5. The Numbers on the axes are too small. Figure 10, 11, 12, 13 and 16. Graph a and graph b have different latitude and longitude. In addition, they are not necessary to show zero seconds in your figure. Figure 15. Pictures representing different land use types in the three stages are better placed on the changing trend of the NDVI. Figure 16. The legend may be a little small and fuzzy.

 

 

Author Response

Response to Reviewer 3 Comments

The manuscript “Bidirectional segmented detection of land use change based on object-level multivariate time series” solves the problem of the complexity of sample selection and the salt-and-pepper noise of pixels when detects the type and time of land use change from Landsat images by proposing a bidirectional segmented detection (BSD) method. And by combining the median of the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Built-up Index (NDBI), and the Modified Normalized Difference Water Index (MNDWI) with the BSD to identify the type and time of the land use. The study of time series of remote sensing images is a hot topic and a remote issue that is in line with human interests, and provides a good basis and reference for detection of land use change. The data used in this study is reliable and used a series of data sheets to illustrate the problem. The analysis is reasonable and logically clear. Conclusion is convincing. I think it is worth publishing in the journal.

I read this manuscript carefully and found that there are still some questions, listed below:

Point 1：Proper use of proper nouns, for example: Line 23-24: “normalized difference vegetation index (NDVI), normalized difference built index (NDBI), and modified normalized difference water index (MNDWI)” should be examined carefully, especially the normalized difference built index.

Response 1：Accepted and revised. First of all, thank you for your positive comments and serious guidance. We have carefully checked the proper nouns and corrected their misuse.

Point 2：Figure 4. The NDVI, NDBI, and MNDWI ’s full name had already appeared. Figure 5. The Numbers on the axes are too small. Figure 10, 11, 12, 13 and 16. Graph a and graph b have different latitude and longitude. In addition, they are not necessary to show zero seconds in your figure. Figure 15. Pictures representing different land use types in the three stages are better placed on the changing trend of the NDVI. Figure 16. The legend may be a little small and fuzzy.

Response 2：Accepted and revised. Since Figure 4 represents NDVI and other index images, it is not of great value to the manuscript, we finally decided to delete it. The Numbers on the axes are enlarged. The longitude and latitude of all images Graph a and Graph b have been adjusted and unified, and zero seconds is not displayed. In figure14, we added two time-phase high resolution images, which are located before and after the change time point. The legend of Figure 15 are enlarged and adjusted clearly.

Author Response File: Author Response.pdf

Reviewer 4 Report

The manuscript is well very impressive and well constructed. Congratulation

The abstract must be improved with more details in results.

Several statements need reference back up eg line 37, 39, 44-48 etc revise

From the study area description, I would suggest you remove the statement on land cover its kind of overshadow the results.

Just to confirm if all images have 0% cloud cover? If not add one column showing percentage cover of each image on Table 1

Remove the highlight in the manuscript.

Just confirm minor English spell check

Author Response

Response to Reviewer 4 Comments

The manuscript is well very impressive and well constructed. Congratulation

Point 1：The abstract must be improved with more details in results.

Response 1：Accepted and revised. First of all, thank you for your positive comments on our research. We have added the details of the research results in the abstract section (Line28~31).

Point 2：Several statements need reference back up eg line 37, 39, 44-48 etc revise

Response 2：Accepted and revised. Thank you for your advice. We have provided relevant references in corresponding positions (Line 38,41,46~50, 62).

Point 3：From the study area description, I would suggest you remove the statement on land cover its kind of overshadow the results.

Response 3：Accepted and explained. Your suggestion is very helpful to our manuscript and we remove the statement on land cover.

Point 4：Just to confirm if all images have 0% cloud cover? If not add one column showing percentage cover of each image on Table 1.

Response 4：Accepted and explained. We have made a statistical analysis of the cloud cover of each time phase image，and add one column showing percentage cover of each image on Table 1 （Line 126）and Table 8 (Line 503).

Point 5：Remove the highlight in the manuscript.

Response 5：Accepted and revised.

Point 6：Just confirm minor English spell check.

Response 6：Accepted and revised. Thank you for your suggestion. We have asked English professionals to polish the manuscript.

Author Response File: Author Response.pdf

Reviewer 5 Report

This study proposes a bidirectional segmented change detection, based on object-level multivariate time series, to detect the type and time of land-use change from Landsat images.

The manuscript is well written and proposes a very interesting method to improve the performance of multitemporal classification analysis in OBIA environments.

The experiment was tested on a congruous number of images and the segmentation step was correctly executed by finding the optimal segment parameters. The application is certainly new and interesting to read.

Author Response

Response to Reviewer 5 Comments

Point 1：This study proposes a bidirectional segmented change detection, based on object-level multivariate time series, to detect the type and time of land-use change from Landsat images.

The manuscript is well written and proposes a very interesting method to improve the performance of multitemporal classification analysis in OBIA environments.

The experiment was tested on a congruous number of images and the segmentation step was correctly executed by finding the optimal segment parameters. The application is certainly new and interesting to read.

Response 1：Thank you for your positive comments on this manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I would to thank the authors for answering my concerns. Great paper!

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

I directed several revision requests to the authors of the manuscript and I can see that they have had a significant effort to improve the quality of the paper by adding more detailed information and clarify the methodology they applied. Although the paper is in much better shape, I have some doubts about the novelty and the methods used in the application of segmentation and change detection. Firstly, the time span is too short for a change detection study as I commented before, which undermines the quality of the paper. For my 5th comment, an explanation was provided with two new references, which were used in the manuscript too. Authors claimed that they used the "control variable method" of (Frohn, Reif, Lane, & Autrey, 2009). However, the authors of that paper do not provide the same method name and the given details of the parameter selection. Also, it was stated that covariance statistics suggested by (Tansey, Chambers,
Anstee, Denniss, & Lamb, 2009) was applied. When I checked the source I could not come across the covariance matrix and statistics estimation. Please check your references again.

 The authors should explain why the LV-RoC graph was changed and optimal scale values were changed. Also, clarify why the selected site was changed in Figure 7. In figures 11 and 15, the north arrow was not used, mistakenly another character was shown. Please check all your figures carefully. Lastly, please check the English grammar and typing errors.  

 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Review comments: Manuscript ID:-remotesensing-643958

General Comment

The manuscript (ID: remotesensing-643958) entitled in “Bidirectional Segmented Detection of Land Use Change based on Object-Level Multivariate Time Series” has not gone through a significant revision as compared to the earlier submission (ID: remotesensing-483840). Specially, issues from my side about discussion were not seriously considered. Still there is mix-up of discussion and results. As mentioned in previous submission, authors have two options, either to combine the two sections into one, or into two, as in the current version.

The current version of discussion can simply confuse the readers. Author(s) need to consider these points for the discussion part.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

A great deal of effort has been put by authors to revise the previous edition of the manuscript. However, the following issues still should be taken into account.

1)    A variety of methods based on dynamic time warping have been applied to remote sensing analysis. It is necessary to conduct a thorough review of the relevant literature on dynamic time warping for remote sensing analysis. The similarities and differences between the proposed method and current DTW methods should be clarified clearly to highlight your novelty.

2)    The example of selecting only one building sample is not sufficient to prove the superiority of the use of the median. I don’t suggest you to list the use of median as one of the contributions.

3) The study area is large, and thus the ground objects have great heterogeneity. Random selection of 10 samples from each of the five types of land use as samples isn’t sufficient.

4) The d is not defined in equation (6). Why i started from 3 in equation (6).

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

 All the concerns are addressed. This manuscript is suitable for publication in Remote Sensing.