Digital Mapping of Land Cover Changes Using the Fusion of SAR and MSI Satellite Data

Round 1

Reviewer 1 Report

Multisource remote sensing image fusion technology is to register the image data of the same area obtained by different types of sensors, and then use certain algorithms to organically combine the information advantages or complementarities contained in each image data. It has become an indispensable technology in the field of image processing and image information analysis. And multisource remote sensing data fusion is a research hotspot in recent decades. From the perspective of research methods in the paper, the innovation of this paper is not outstanding, it is suggested to clearly introduce the contribution in this paper. Also, the following suggestion should be discussed.

1. The fusion of multisource remote sensing data for land cover change detection has been studied for decades, the introduction of previous studies is insufficient.

2. Lines 69-76, the contributions in the paper have not been given clearly.

3. Some mapping information is missing in some figures in the paper such as north arrow.

4. In Equation (6), how to define the exact values “705” and “35”?

5. In Fig. 15, how to determine the correctly or incorrectly identified land cover changes? In another case, it should be noted that the correctly or incorrectly identified land cover changes are related to the distribution location. That is, even if the number of identified changes is close to the truth, it may not be the truly changing areas. Some indicators should be selected to evaluate the fusion of SAR MSI satellite data.

Author Response

REVIEWER_1

Comments and Suggestions for Authors

Multisource remote sensing image fusion technology is to register the image data of the same area obtained by different types of sensors, and then use certain algorithms to organically combine the information advantages or complementarities contained in each image data. It has become an indispensable technology in the field of image processing and image information analysis. And multisource remote sensing data fusion is a research hotspot in recent decades. From the perspective of research methods in the paper, the innovation of this paper is not outstanding, it is suggested to clearly introduce the contribution in this paper. Also, the following suggestion should be discussed.

1. The fusion of multisource remote sensing data for land cover change detection has been studied for decades the introduction of previous studies is insufficient.

Answer: We agree with reviewer comment. In the article we write (61-65 in the old version of article): “Information collected from different image sensors and properly processed by synthesising satellite images allows data gaps to be filled and more accurate results to be obtained [19]. Various studies on the fusion of different data can be found in the scientific literature [20-22]. Such works began in the 1980s, but some of the processes have not yet been automated”. Additionally we add more information about the scientists studies. “Some authors used optical images with different resolution (10-30 m).  The data fusion results were image with 31 spectral bands at a 30 m spatial resolution was performed and used for the identification of soil information. The modelling process produced an increase of about 10% in the prediction accuracy compared to that obtained from single-sensor images (New reference: Gasmi, A.; Gomez, C.; Chehbouni, A.; Dhiba, D.; Elf, H. Satellite Multi-Sensor Data Fusion for Soil Clay Mapping Based on the Spectral Index and Spectral Bands Approaches. Remote Sens. 2022, 14, 1103. DOI: 10.3390/rs14051103). The data fusion technics apply when scientists have optics image limitations by the cloudy time. The authors identify that the process of combining data from multiple sources to produce more accurate, consistent, and concise information than that provided by any individual data source (New reference: Munir A., Blasch E., Kwon J., Kong J., Aved A. Artificial Intelligence and Data Fusion at the Edge. IEEE Aerosp. Electron. Syst. Mag. 2021;36:62–78. doi: 10.1109/MAES.2020.3043072). The literature review by the Jayme Garcia Arnal Barbedo exclude (išskirti) Fusion Techniques and accuracy.  The Self-normalizing neural networks (CNN), Backpropagation in neural network (BPNN), Random Forest (RF), Partial least squares regression (PLSR), Dempster–Shafer and Kalman filter, Least squares support vectors machine gave the best mean accuracy of result (New Reference: Barbedo, J.G.A. Data Fusion in Agriculture: Resolving Ambiguities and Closing Data Gaps. Sensors (Basel). 2022 Mar; 22(6): 2285. DOI: 10.3390/s22062285). The authors explain that in the case of agriculture, image data fusion has been particularly effective at the orbital level, both for artificially increasing the spatial resolution of sources with revisit frequencies and for compensating cloud cover using the information present in SAR images. In these cases, the improvement can exceed 50%. A study by N. Kussul and co-authors [29] in which vegetation was classified using EO-1 and RADARSAT-2 satellite images fusion. In their study, they distinguished four classes: corn, soybeans, sunflower, and winter / spring wheat. They show that the result obtained with the data fusion is 2.5% more accurate compared to the result where only SAR images were used and 6.6% more accurate than with optical images only. Nicola Clerici and colleagues [30] used satellite imagery of Sentinel-1 and Sentinel-2 in her study. The aim of the study was to investigate the accuracy of land cover classification using different algorithms, classifying SAR and MSI images separately, and their fusion. Their result shows that regardless of the algorithm used, image fusion gives a more accurate result. They obtained a 58% more accurate result using the random forest algorithm than using SAR images alone for land cover classification and 16% more accurate than using only optical images. The same Sentinel data were used by other authors, such as from Julien Denize et al. [31] shows that SAR and MSI data fusion is 21% more accurate than SAR-only images and 7% more accurate than MSI-only images. Audrey Mercier and co-authors [32] also performed better in image fusion, which is 20% more accurate than using SAR data alone and 5% more accurate than results obtained using optical images alone. A very similar result was obtained by Armugha Khan, Himanshu Govil, Gaurav Kumar, and Rucha Dave [33], who distinguished 5 land cover classes using S-1 and S-2 satellite imagery - waters, urban areas, ravines, barren land, and crops. Their result is 25% more accurate with image fusion than SAR data alone, and 5% more accurate with optical images alone.

2. Lines 69-76, the contributions in the paper have not been given clearly.

Answer: We agree with reviewer comment. We have replaced the research objectives presented in the study with a clearer research purpose. New text: The different MSI and SAR data fusion techniques have been used for the segmentation of land cover in this study too. The purpose of this article is to choose the most appropriate data fusion technique for the identifying and managing land cover changes over time. 

3. Some mapping information is missing in some figures in the paper such as north arrow.

Answer: We checked all the figures and add the missing information.

4. In Equation (6), how to define the exact values “705” and “35”?

Answer: Thank you for the comment. The S2REP red-edge index is based on linear interpolation (following the method of Guyot & Baret, 1988) by exploiting Sentinel-2A bands 5 and 6, both positioned on the red edge slope, at 705 and 740 nm respectively. So, these numbers are used as constants in the formula. Nevertheless, we left only the most important formulas in the study and eliminate well known.

5. In Fig. 15, how to determine the correctly or incorrectly identified land cover changes? In another case, it should be noted that the correctly or incorrectly identified land cover changes are related to the distribution location. That is, even if the number of identified changes is close to the truth, it may not be the truly changing areas. Some indicators should be selected to evaluate the fusion of SAR MSI satellite data.

Answer: Thank you for comment. In the article we write “To estimate the actual change in all class pairs, 20 random objects (or all if fewer changes were identified) were selected and the result was checked against the 2018 and 2019 RGB images.” In Fig. 15 you can find the result of quality check. Talking about identified changes, seasonal or crop type changes were eliminated before quality check. Also in the conclusions, we additionally add: Changes that were falsely identified during the qualitative accuracy check of the identified changes (92.08% of all changes checked) were False Positive results and no False Negative results were observed in the analysis of the images. Although changes are incorrectly identified in some identified cases, visual inspection (especially when potential locations for potential inaccuracies are known) and manual correction would still use less time than not automating all the process. We apologize if we misunderstood your question and did not answer it.

Author Response File: Author Response.docx

Reviewer 2 Report

While this could be a useful work, it does have many shortcomings as it currently stands. Hence a rigorous review is needed to improve its readership, international significance, and novelty. I have outlined major issues below. Considering them will help authors to provide improved the readership. Abstract needs improvement. This is an incremental work, offers no new methods and insights. It is an area specific (western Lithuania) work, thus requires showing how the results could advance current knowledgebase.  Authors attempted to improve results utilising a few MSI indices but such attempts are a lot in the literature. I was wondering if corner reflection and layover or foreshortening could have impact the results.   

[1] We all know that land use is possibly the most important contributor to global environmental change including ecosystem services loss, enhancement of local warming, habitat loss, augmentation of flood risk and so on. Since this is context specific, referring to existing works related to land use change and associated impacts could strengthen its motivation, rationale etc.

https://www.nature.com/articles/nature01675; https://doi.org/10.1016/j.jenvman.2020.111885; https://www.nature.com/articles/s41586-020-03138-y; https://doi.org/10.1029/2021EF002401

[2] You have used Sentinel data but described details of pre-processing steps. As they are input to your work, you must demonstrate major steps (not too detailed as people working in this area know them well) and classified into different land use categories. What supervised method was used? Why six classes? What criteria was used to decide these six classes? Why only 2018 and 2019 data used? Could include 2016 and 2021? Using a large range could provide better results

[3] If equations are not developed by you, so refer to works that developed them, you don’t need to put them again to show that they are newly placed in this work

[4] Missing discussion section. Since you are not developing any new method, I would suggest you use existing works on similar topic to strengthen this section.  Though this way you could offer new insight to readers, above works will definitely help you to develop a well discussion part

[5] Check consistency of expressions and grammars throughout the texts. Lulc must be consistently written as “LULC” following an elaboration of the first use then acronym can be used throughout

What are the limitations of this work? Didn’t you face any issue? Putting them will help future workers. How you want to improve this work to address land use change impact of climate?

Author Response

REVIEWER_2

Comments and Suggestions for Authors

While this could be a useful work, it does have many shortcomings as it currently stands. Hence a rigorous review is needed to improve its readership, international significance, and novelty. I have outlined major issues below. Considering them will help authors to provide improved the readership. Abstract needs improvement. This is an incremental work, offers no new methods and insights. It is an area specific (western Lithuania) work, thus requires showing how the results could advance current knowledgebase.  Authors attempted to improve results utilising a few MSI indices but such attempts are a lot in the literature. I was wondering if corner reflection and layover or foreshortening could have impact the results. 

Thank you for all of your insights that have helped us supplement and enrich the article, and some of your specific observations have contributed to the clarity and quality of the study.

[1] We all know that land use is possibly the most important contributor to global environmental change including ecosystem services loss, enhancement of local warming, habitat loss, augmentation of flood risk and so on. Since this is context specific, referring to existing works related to land use change and associated impacts could strengthen its motivation, rationale etc.

https://www.nature.com/articles/nature01675; https://doi.org/10.1016/j.jenvman.2020.111885 ; https://www.nature.com/articles/s41586-020-03138-y ; https://doi.org/10.1029/2021EF002401

Answer: Thank you for recommendation. We have examined the recommend articles and add the information to the our manuscript and reference list.

Correction in the text: Land and soil are limited non-renewable resources that disappear over time. This process is particularly accelerated by the intensity of various human activities that have transformed and fragmented ecosystems of the land, as well as the functionality and performance of shared land resources. The anthropogenic influences on climate are the emission of greenhouse gases and changes in land use, such as urbanization and agriculture. Of course, soil erosion, soil organic matter decline, soil contamination and compaction, and surface temperature also contribute to this decline but human activities, such as land development, intensive land use and land abandonment, only accelerate these processes [1, Kalnay, E., Cai, M. Impact of urbanization and land-use change on climate. Nature 423, 528–531 (2003). https://doi.org/10.1038/nature01675; Hong, C., Burney, J.A., Pongratz, J. et al. Global and regional drivers of land-use emissions in 1961–2017. Nature 589, 554–561 (2021). https://doi.org/10.1038/s41586-020-03138-y]. The scientist excreted three highest-emitting regions (Latin America, Southeast Asia and sub-Saharan Africa) dominate global emissions growth from 1961 to 2017, driven by rapid and extensive growth of agricultural production and related land-use change. From the point of view of environmental protection, landscape morphology and functioning, economically and socially, land cover is considered to be a very important object of research [2]. Firstly, the stabilizing local surface temperature level and achieve a balance between anthropogenic emissions by sources and removals by sinks of greenhouse gases is very important naw and in the future Globally (Zhou, D.; Xiao, J.; Frolking, S.; Zhang, L.; Zhou, G. Urbanization Contributes Little to Global Warming but Substantially Intensifies Local and Regional Land Surface Warming. Eath’s Future, 2022, Volume 10, Issue 5. DOI: 10.1029/2021EF002401). 

[2] You have used Sentinel data but described details of pre-processing steps. As they are input to your work, you must demonstrate major steps (not too detailed as people working in this area know them well) and classified into different land use categories. What supervised method was used? Why six classes? What criteria was used to decide these six classes? Why only 2018 and 2019 data used? Could include 2016 and 2021? Using a large range could provide better results.

Answer: Thank you for comment. We described the data preparation in the 3.1.1 and 3.1.2 (Materials and Methods) chapters. The basics pre-processing stages are shown in the Figure 2. We have shorted texts and presented what seems most important to other readers.

We used the Random Forest supervised image segmentation method. Information in the text: “Various authors focusing on land cover classification studies using SAR and MSI images, mainly use two main classification algorithms: Support vector machine (SVM) and Random Forest (RF) [39, 43-48]. RF is one of the most popular and effective supervised training algorithms. […] Analysing the work by Kussul, Clerici, Valbuena Calderón and Posada, Haas and Ban, Gerrells, Sun, Denize, and Mercier, which compares the results obtained using the SVM and RF algorithms, shows that the RF algorithm often yields more accurate ground cover classification results; this algorithm was chosen for this study.”

There is an explanation in the text why we used 6 land cover classes “Six classes of land cover (hydrography, forests, vegetation areas, sand dunes, built-up areas, non-vegetation) were selected for the study, because they were recommended by other authors for land cover/land use classification [6, 28 - 37].” Also for our opinion, main six land cover classes are enough to examine all the methods described in the article, but for future works we are planning to apply our described last method to segment land cover and land use as detailed as possible and needed.

It is quite difficult to find 15 % cloudiness images over Lithuania territory. Nevertheless, the main reason is that both the territory and the images for 2018 and 2019 were chosen to test and reject the methods used in the article, which immediately shows that it is not appropriate. In the future, we plan to expand the study in both territorial and temporal terms.

[3] If equations are not developed by you, so refer to works that developed them, you don’t need to put them again to show that they are newly placed in this work.

Answer: Thank you for comment. We left only the most important formulas.

[4] Missing discussion section. Since you are not developing any new method, I would suggest you use existing works on similar topic to strengthen this section.  Though this way you could offer new insight to readers, above works will definitely help you to develop a well discussion part.

Answer: Thank you for comments. We have supplemented the text with one additional conclusion “5. Changes that were falsely identified during the qualitative accuracy check of the identified changes (92.08% of all changes checked) were False Positive results and no False Negative results were observed in the analysis of the images. Although changes are incorrectly identified in some identified cases, visual inspection (especially when potential locations for potential inaccuracies are known) and manual correction would still use less time than not automating all the process.” Moreover, we add the discussion to the lasts chapter 5. New text: In our study the method that provides the most accurate results for land cover segmentation confirms Clerici’s [30] finding that the inclusion of additional indices enriches the image and yields better segmentation results. As many as three indices have been used in this work to highlight the vegetation class (NDVI, S2REP and GNDVI), but questions arise as to whether this is not a surplus. In order to ascertain which vegetation indices are most beneficial for the segmentation result, a broader analysis of the literature and, if necessary, research is provided. Future work will aim to refine the method so that it is optimal in terms of time and quality compared to similar manual work.

The applied image merging techniques also confirm the conclusions of other authors [26-33] that the merging of two different sensors enriches the image information. Also, SAR images may compensate for the low cloud content of MSI images, especially if multiple SAR images are used. However, it is interesting that in all the analysed authors, the percentage accuracy of the fused and separately segmented images varies quite strongly in some cases. This may be influenced by the satellite imagery used, the different ground cover classes used, and the different classification algorithms used. It is also very important to note that in the articles analysed, the research is carried out on different areas that are in different latitudes, which may also be the reason why the results obtained by different researchers are so different.

[5] Check consistency of expressions and grammars throughout the texts. Lulc must be consistently written as “LULC” following an elaboration of the first use then acronym can be used throughout

What are the limitations of this work? Didn’t you face any issue? Putting them will help future workers. How you want to improve this work to address land use change impact of climate?

Answer: We check the expressions, sorry for the mistakes. In the study we do not use LULC acronym because our study subject was only the land cover but not a land use. Some thoughts are added to the discussion chapter (Section 5.).

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The comments and suggestions have already been modified.

Reviewer 2 Report

No