A Geographically Weighted Regression–Compute Unified Device Architecture Approach to Explore the Spatial Agglomeration and Heterogeneity in Arable Land Consumption in Southwest China

Round 1

Reviewer 1 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

The authors adequately addressed the observations.

Author Response

Thank you for your approval.

Reviewer 2 Report (Previous Reviewer 3)

Comments and Suggestions for Authors

It is believed that the discussion of this manuscript is interesting and has the potential to contribute to practical policymaking. Here are two minor suggestions to this manuscript:

1. It would be ideal if the abstract can highlight the research question and give one or two sentences to provide the background. 

2. As Yunnan, Guizhou and Chongqing are not the most common cases in the research community, the manuscript can benefit from (1) providing more detailed information about the study area, (2) explaining why the findings in these areas can be helpful for other areas or countries.

Author Response

1. Summary
Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions/corrections highlighted/in track changes in the re-submitted files.
2. Point-by-point response to Comments and Suggestions for Authors
Comments 1: 1. It would be ideal if the abstract can highlight the research question and give one or two sentences to provide the background.
Response 1: Thank you for pointing this out. We have added some content about the research background and question, from line 10 to 13: “Arable land loss has become a critical issue in China due to rapid urbanization, industrial expansion, and unsustainable agricultural practices. While previous studies have explored the factors contributing to this loss, they often fall short in addressing the challenges of spatial heterogeneity and large-scale dataset analysis.”
Comments 2: As Yunnan, Guizhou and Chongqing are not the most common cases in the research community, the manuscript can benefit from (1) providing more detailed information about the study area, (2) explaining why the findings in these areas can be helpful for other areas or countries.
Response 2: Agree. We have, accordingly, modified section 2 to emphasize the point one. We have added some detail information about why we choose these areas, from line 183 to 190: “Yunnan, Guizhou, and Chongqing are strategically important regions for studying arable land depletion due to their diverse geographical and socio-economic characteristics. These provinces have varied topographies, including mountainous terrains, flat plains, and river valleys, which influence land use patterns. Additionally, they represent a mix of rapidly urbanizing regions (like Chongqing) and more rural, agriculturally reliant areas (like Yunnan and Guizhou). This diversity makes them excellent case studies for understanding the interaction between environmental factors (rainfall, evaporation, slope) and human activities (urbanization, economic development) in driving arable land depletion. Yunnan, Guizhou, and Chongqing are strategically important regions for studying arable land depletion due to their diverse geographical and socio-economic characteristics. These provinces have varied topographies, including mountainous terrains, flat plains, and river valleys, which influence land use patterns. Additionally, they represent a mix of rapidly urbanizing regions (like Chongqing) and more rural, agriculturally reliant areas (like Yunnan and Guizhou). This diversity makes them excellent case studies for understanding the interaction between environmental factors (rainfall, evaporation, slope) and human activities (urbanization, economic development) in driving arable land loss.”   For the point 2, in the summary section, we added the help of this paper to the study of other areas, from line 698 to 713:   “This study not only applies CUDA-enhanced GWR to efficiently analyze large-scale arable land loss but also offers a replicable solution for understanding spatial heterogeneity in regions worldwide. Traditional GWR models often face challenges when handling large datasets, especially in areas with complex interactions between environmental factors and human activities. By utilizing CUDA on high-performance GPU servers, this study enables efficient GWR computation, making it feasible to analyze extensive regions with diverse climatic, topographical, and socio-economic conditions. Regions with high R² clustering indicate strong explanatory power for arable land loss, allowing researchers to leverage this insight to predict future depletion. By examining the influence weights of various independent variables within these clustered regions and correlating them with the stages of economic development and environmental protection measures, researchers can quickly identify spatial distribution patterns of arable land loss and assess its spatial heterogeneity. Furthermore, the method’s visualization and spatial autocorrelation tools offer deeper insights into key drivers such as climate change, land use practices, and policy interventions. This provides a scientific foundation for developing more targeted land management and conservation strategies. “

Reviewer 3 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

The paper shows some basic analysis about spatial changes in arable land in selected area. Authors also claims that novelty of this study is deploying the GWR-CUDA model on a GPU cloud server and try to show advantages of this proposition. Authors try to cover too many topics in this article which provides to extent volume of the text. CUDA computing is just a tool to get the results, and I agree that it may be a kind of novelty but the Authors did not test the efficiency and advantages of this solution in this article, therefore should not cover this topic in the text.

As to a main scope of the article, which is "Spatial Agglomeration and Heterogeneity for Arable Land Consumption" - analysis that are provided are described the way that it is not clear. My major complains are listed below:

1. There is something strange in Figure 1 - according to Table 1 the DEM range varies from 84 to 5452m but legend of Figure 1 shows color scale from -219 (deep blue) to 5558 (red), whereas most of all area of Chongquing province is deep blue colored - is this correct?

2. Line 293 - what is "sub-region"? Is it province? Or is it other part of analyzed area? It should be clearly stated and also maybe showed on Figure 1.

3. Line 295 - it is not clear for me but if I understand well, the grid of sub-region "fishnet" is 300 cells x 300 cells, what are the dimentions of cell? Are all sub-regions of the same size? Are all cells in each sub-regions of the same size? Are the results in each sub-regions comparable? Why exactly 300x300 cells? Or is the "fishnet" grid cell 300m x 300m of size to corespond with raster data pixel size 30x30m? Or is the "fishnet" grid cell 300x300 of pixels? It should be clear for readers.

4. Figure 3 - in second step ("Split the raster...") parts of resulting raster are shown. It may be confusing for readers. Which exactly raster is splitted? Into which parts? Subregions?

5. Figure 3 - in fourth step "Selected fishnet" window covers only part of "fishnet map" from step 3. Morover Authors use "fishnet map" or "fishnet grid" or "fishnet" in the text and in Figures 2 and 3 interchangeably. Please be precise. This part of data preparation process needs more precise explanation. 

6. Figure 2 suggest that whole investigated area are covered with cells that have some values of arable land loss rate. Ofcourse land with no changes in arable land, but even land where there had not been arable land ever before. In turn, Figure 3 (parts 3 and 4) suggests that "fishnet map" and "fishnet" is something else. The latter is grid of cells only with arable land changes. Furthermore Figure 6 have colors on maps that are not included in color scale at the top of the Figure (I see black and grey and white which are not in the scale), what suggest that not whole area was taken into account - only cells with arable land changes. It may be confusing for readers.  Section 3.3 should be definitely rebuilt and Authors have to precise explain data preprocessing. Also how were the independent variables values calculated in each cell for non raster data.

7. Lines from 304 to 327 - I think this should be another section (3.4) of the article, since it describes statistical analysis of spatial results.

8. Line 304 - GWR or GWR-CUDA or Grid-GWR-CUDA? Please be precise. Authors use "GWR-CUDA" in title, keywords and section 3.2, then they use "GWR" sometimes "GWR-CUDA" and also "Grid-GWR-CUDA". Readers of this article will be confused.

9. Section 4.1 definitelly needs maps with "arable land loss" - it could be just two colors (0/1) raster maps. Or "fishnet grid" with arable land loss rates.

10. Section 4.1 definitelly needs all "fishnet grid" maps with independent variables.

11. Due to high volume of this text Authors should cancel Figure 5.

12. Section 4.2.1 - "OLS" please expand the abbreviation before it's first use. This section contains a new method/model (OLS) which was not described neither in "3. Methods" nor in "1. Introduction" or Abstract. Where are the results from? How was it performed? And how could it be compared with "GWR" results that are core of this article? Authors should provide the source of results for OLS calculations and at least some basic informations about the method/model in comparison to GWR. In my opinion this part should be omitted - the Authors try to cover too many topics in this article. I think that this article should focus on spatial analysis using "GWR-CUDA" as a tool (with short explanation about GWR-CUDA method). Some advantages and efficiency gain of the GWR-CUDA over other methods, models and computational systems could be the core of anoter article in a suitable magazine.

13. Figures 6 and 7 - there are labels "DEM" in left-bottom corners of each map over the linear scale - is it correct? Shouldn't there be label "Distance" or just nothing?

14. Figure 7 - due to high volume of this article figure 7 seems to be not neccessary in this text and should be canceled.

15. Table 8 is hardly readable, it should be reorganized.

16. Section 5.1 - typically section "Discussion" should contain some critical review of results. Advantages of CUDA are not results of this work. In my opinion computing with CUDA should be treated as a tool to get the results. Some part of this section could be placed in "3.2. GWR-CUDA" as a supplement to the description of the reason for using the method.

17. Line 584 - "GWR4" - please expand the abbreviation before it's first use or explain what it is.

Best regards.

Author Response

1. Summary

Thank you very much for taking the time to review this manuscript. Please find the detailed responses below.

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: There is something strange in Figure 1 - according to Table 1 the DEM range varies from 84 to 5452m but legend of Figure 1 shows color scale from -219 (deep blue) to 5558 (red), whereas most of all area of Chongquing province is deep blue colored - is this correct?

Response 1: Thank you for your insightful comment. Because of the color scheme, it made the change between the most and the middle values not very obvious, we adjusted it to make the transition between the most and the middle values more obvious.

Comments 2: Line 293 - what is "sub-region"? Is it province? Or is it other part of analyzed area? It should be clearly stated and also maybe showed on Figure 1.

 

Response 2: Thank you for your detailed review, Sub-region is to divide each province into three parts is a part of the province. Adding dividers in Figure 1 is not a good look, so instead we add the appropriate explanation to make it clearer and easier to understand, from  line 318  to 319： 

” (2) Segment and extract the provinces into sub-regions.”

 

Comments 3: Line 295 - it is not clear for me but if I understand well, the grid of sub-region "fishnet" is 300 cells x 300 cells, what are the dimentions of cell? Are all sub-regions of the same size? Are all cells in each sub-regions of the same size? Are the results in each sub-regions comparable? Why exactly 300x300 cells? Or is the "fishnet" grid cell 300m x 300m of size to correspond with raster data pixel size 30x30m? Or is the "fishnet" grid cell 300x300 of pixels? It should be clear for readers.

Response 3:

Thank you for pointing this out, here is the size of each grid: 300m * 300m. All areas have the same grid size. This does not mean there are 300 * 300 grids. The grid size is chosen specifically because the raster data resolution is 30m * 30m. We've added content to explain, from line 319 to 320:

“(3) Create a fishnet with 300m x 300m grid cells to cover the studied sub-regions and generate the corresponding point data and fishnet data.”

 

Comments 4: Figure 3 - in second step ("Split the raster...") parts of resulting raster are shown. It may be confusing for readers. Which exactly raster is splitted? Into which parts? Subregions?

 

Response 4: Thank you for pointing this out. The purpose of partitioning into subareas is to improve operation efficiency and prevent memory overflow. We added the reason for partitioning before studying the process description to explain, from line 309 to 314:

“Additionally, to enhance computational efficiency and prevent memory overflow, we divided the provinces of Guizhou and Yunnan into three sections for the GWR-CUDA analysis. The data for arable land loss in both provinces amounts to around 1 million records. Given the high computational complexity of GWR, even with GPU processing, memory overflow can still occur. Therefore, we opted to partition the data into three parts for more manageable analysis and processing. “

 

Comments 5: Figure 3 - in fourth step "Selected fishnet" window covers only part of "fishnet map" from step 3. Moreover, Authors use "fishnet map" or "fishnet grid" or "fishnet" in the text and in Figures 2 and 3 interchangeably. Please be precise. This part of data preparation process needs more precise explanation. 

 

Response 5: Thank you for your insightful advises, the fourth step in figure 3 is to extract the part that contains data points and omit the part that does not contain data, so that only part of the fishnet map is selected. And we have replaced the “fishnet grid “with “grid cell “, replaced “fishnet map “with “fishnet “.

 

Comments 6: Figure 2 suggest that whole investigated area are covered with cells that have some values of arable land loss rate. Of course land with no changes in arable land, but even land where there had not been arable land ever before. In turn, Figure 3 (parts 3 and 4) suggests that "fishnet map" and "fishnet" is something else. The latter is grid of cells only with arable land changes. Furthermore Figure 6 have colors on maps that are not included in color scale at the top of the Figure (I see black and grey and white which are not in the scale), what suggest that not whole area was taken into account - only cells with arable land changes. It may be confusing for readers.  Section 3.3 should be definitely rebuilt, and Authors have to precise explain data preprocessing. Also how were the independent variables values calculated in each cell for non-raster data.

 

Response 6: Thank you for reminding me, the explanation here is vague. We changed the content so that in fact the data points only include the data of land loss and land retention, and then calculate the land degradation rate, and do not include the data of land that was not previously cultivated. We added an explanation, from line 278 to 286:

“In this context, data points with a value of 1 indicate arable land loss, while data points with a value of 0 indicates that it’s still arable land. By creating fishnet for data points, we can calculate the arable land loss rate of this grid unit by dividing the sum of data points with the value of 1 (arable land loss occurred) by all data points (arable land loss occurred and retained as arable land). Then we replaced the original data points with these loss rate as the target variable of the center point of the fishnet. Subsequently, we will utilize the position-selecting tool to match the coordinates with the center points of the grid cell. Through this process, the attribute table of the fishnet will encompass the coordinates data and the dependent variable dataset. “

In figure 6, Black and white are only the colors of the base map (Dem), Figure 2, Figure 6 and Figure 3 should all have the same meaning, and they all only represent the research of arable land loss area.

 

Comments 7: Lines from 304 to 327 - I think this should be another section (3.4) of the article, since it describes statistical analysis of spatial results.

 

Response 7: Thank you for your advice, I have modified it as a new section.

 

Comments 8: Line 304 - GWR or GWR-CUDA or Grid-GWR-CUDA? Please be precise. Authors use "GWR-CUDA" in title, keywords and section 3.2, then they use "GWR" sometimes "GWR-CUDA" and also "Grid-GWR-CUDA". Readers of this article will be confused.

 

Response 8: Thanks for your suggestion. We have replaced grid-GWR-CUDA with GWR-CUDA. For the use of GWR and GWR-CUDA, both are essentially the result of running GWR, so we only call it the GWR-CUDA model where we introduce CUDA's role in improving the computational efficiency of GWR, and the other parts, such as the results section, just use the GWR model as a default to refer to the results of GWR derived using CUDA.

 

 

Comments 9: Section 4.1 definitelly needs maps with "arable land loss" - it could be just two colors (0/1) raster maps. Or "fishnet grid" with arable land loss rates.

 

Response 9: Thank you for your pointing out this. We have added the fishnet grid with arable land loss rates and its description, from line 376 to389:

“Figure 5 illustrates the overall trend of arable land changes in the study area. A comparison across different time periods reveals that although the general land use of arable land remained relatively stable, since 2010, the loss of arable land around major urban centers has expanded significantly, leading to a gradual reduction in arable land area. For instance, in the southwestern urban core of Chongqing and the central-western region of Kunming in Yunnan Province, both the blank and red areas have increased, indicating a reduction in arable land and an intensifying loss and loss of arable land around urban areas. In contrast, the changes in Guizhou Province are more dispersed. While the total amount of arable land remained relatively stable, the significant expansion of red areas suggests that arable land loss has worsened, and substantial changes have occurred in certain regions. Furthermore, the blue areas, representing an increase in arable land, have shown little spatial variation between the two periods, remaining primarily in densely populated, flat regions. This indicates that the spatial pattern of arable land in these areas has remained relatively stable. “

 

Comments 10: Section 4.1 definitely needs all "fishnet grid" maps with independent variables.

 

Response 10: Thank you for your pointing out this, we do not suggest using the fishnet map with the independent variables, because the zonal statistical tool directly adds the independent variable values to the fishnet map with the rate of arable land loss. We do not need to create fishnet map for the independent variables. On the other hand, making a fishnet mesh of independent variables is very troublesome and laborious, because there are eight independent variables in three regions, plus there are two time periods, a total of 48 charts, which is also impossible to put in 4.1.

 

Comments 11: Due to high volume of this text Authors should cancel Figure 5.

 

Response 11: Thank you for your pointing out this, I have cancelled it and replaced it with fishnet map with arable land loss.

 

Comments 12: Section 4.2.1 - "OLS" please expand the abbreviation before it's first use. This section contains a new method/model (OLS) which was not described neither in "3. Methods" nor in "1. Introduction" or Abstract. Where are the results from? How was it performed? And how could it be compared with "GWR" results that are core of this article? Authors should provide the source of results for OLS calculations and at least some basic informations about the method/model in comparison to GWR. In my opinion this part should be omitted - the Authors try to cover too many topics in this article. I think that this article should focus on spatial analysis using "GWR-CUDA" as a tool (with short explanation about GWR-CUDA method). Some advantages and efficiency gain of the GWR-CUDA over other methods, models and computational systems could be the core of anoter article in a suitable magazine.

 

Response 12: We appreciate your feedback on Section 4.2.1. We agree with your point that comparing the OLS model with the GWR model is indeed not the main focus of this study. Since the results report of the GWR model already includes both OLS and GWR results, we initially compared them to demonstrate the superiority of the GWR model. However, we have now removed this comparison.

 

Comments 13: Figures 6 and 7 - there are labels "DEM" in left-bottom corners of each map over the linear scale - is it correct? Shouldn't there be label "Distance" or just nothing?

 

Response 13: Thank you for your insight comment. DEM represents the base map style, and has nothing to do with the scale, that is, the black and white color in the figure indicates the elevation.

 

Comments 14: Figure 7 - due to high volume of this article figure 7 seems to be not necessary in this text and should be canceled.

 

Response 14: Thank you for your detailed review. Actually, residual spatial analysis is an essential part of this study. To enhance the richness of the paper, we performed spatial analysis on the local results of the GWR, focusing on adjusted R² and residuals. The spatial analysis of residuals serves to validate the reliability of the findings. If the residuals exhibit no spatial autocorrelation, it indicates the randomness and reliability of the results. Therefore, we do not recommend removing Figure 7, as it presents the spatial distribution of residuals. Without this, the spatial analysis of GWR results would become overly simplistic.

 

Comments 15: Table 8 is hardly readable; it should be reorganized.

 

Response 15:  Thank you for your suggestion, we have split Table 8 into two tables by time.

 

Comments 16: Section 5.1 - typically section "Discussion" should contain some critical review of results. Advantages of CUDA are not results of this work. In my opinion computing with CUDA should be treated as a tool to get the results. Some part of this section could be placed in "3.2. GWR-CUDA" as a supplement to the description of the reason for using the method.

 

Response 16:  Thank you for your suggestion, we have revised the "5.1 Advantages of CUDA" to Section "3.2 GWR-CUDA", from line 261 to 270. The description of CUDA's advantages was removed from the discussion.

“This method avoids memory overflow efficiently and reducing the runtime complexity of matrix operations. According to the experimental result, the maximum data volume (812,900) takes about 20 hours and 30 minutes for regression analysis by using single V100 GPU, and the minimum data volume (74,800) takes only 9 minutes to process. To provide a more intuitive comparison, we evaluated the processing time of GWR4, a soft-ware tool for geographically weighted regression (GWR) analysis, against GWR-CUDA. For a data sample of 20,000 observations, GWR4 would have taken approximate 10h to process, while now GWR-CUDA now takes only 28 minutes, achieving a 21.5-fold reduction in time. Besides, four V100 high-performance graphics cards can simultaneously process four data partitions, reducing time costs. “

 

Comments 17: Line 584 - "GWR4" - please expand the abbreviation before it's first use or explain what it is.

 

Response 17: Thank you for pointing this out, we have added the introduction about GWR4, from line 264 to 266:

“To provide a more intuitive comparison, we evaluated the processing time of GWR4, a software tool for geographically weighted regression (GWR) analysis, against GWR-CUDA. “

Round 2

Reviewer 3 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

I see that most of my comments were positively received and included in the changes to text, for which I thank to Authors of this article.

Despite some minor additional comments, which we could still argue about, I believe that at this point the text is clear enough for the Agriculture journal readers and can be published.

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

Comments and Suggestions for Authors

The paper shows some basic analysis about spatial changes in arable land in selected area. This article covers standard methods of spatial analysis that are also shown in many similar articles touching on that kind issues. Authors claims that "This novelty of study, by deploying the GWR-CUDA model on a GPU cloud server to handle large-scale data efficiently" but I am not sure that this is one of the main scope of "Agricilture". Besides in my opinion parralel computing is a kind of tool, such as ArcGIS, Python or Excel sheet to "get the job done". Authors niether provide some special algorithm nor use a "task-based constructed" machine, just use mid-range and outdated consumer graphic card (I think RTX2060 - it is not firmly stated) to enhance computations. I understand the advantages of using "GWR-CUDA" (as the Authors named it), but is it really needed - how often do this kind of analisys are done? And if it is done the way that time is essential, then I would rather use high-end workstation machine.

As to a second scope of the article, which is "Spatial Agglomeration and Heterogeneity for Arable Land Consumption" - I would say that a lot of this kind of analysis are published is "Land" or "Sustainability" and that papers cover it more comprehensive and in depth. Authors have a lot of data but I think they do not "squezze it" as they maight. Analysis that are provided are described the way that it is hard to understand the goal. If I undestand well the concept is: model fitting => R-squared analysis => spatial agglomeration of R2 => discusion on factor affection in high R2 clusters. It is interesting approach to the problem, but for me it is poor described.

In my opinion:

1. In this article a whole sections about CUDA computing could be omitted.

2. Method section should also contain at least some basic informationa about R-squred and Moran-I analysis.

3. In "Grid – Distributed computing" - what are dimentions of the grid cel? "calculate the mean value of the independent variable within each grid" - how was it performed for "GDP" and "pop" variables?

4. A huge part of actual Section 4.1 are not a results - "arable land loss" is at the beginning of algorithm provided in figure 3, it should be described in "introduction" or in "Data acquisition and preprocessing".

5. As the paper describes "geographical" and "spatial" analysis" there is a lack of maps showing changes, variables distribution etc.

6. Why R2 shows mostly low values? Is the design of experiment apropriate? Maybe the Authors shouldn't take the same variables for all area? Maybe the Authors should cross the map of R2 with map of "arable land loss"? Maybe the analysis should be taken only for areas with exact arable land changes not for whole area? Would then it be as computatonal demanding?

7. And finally - "In this study, the GWR model employs 64 CUDA cores for parallel computation, resulting in enhanced efficiency in model fitting" - is it true?  Or it "showcasing a computational efficiency improvement of several thousand to tens of thousands of times compared to the conventional GWR algorithm"?

This article has also some editional issues. The file which I got is not properly formated. There are some active link missing and "Error! Reference source not found" messages. Table 6 is hardly readable. There are some abbreviations that are not firstly fully expalined. 

I believe that this article could possibly be more interesting, but with no doubt it needs major revision and rebuild. I suggest to reconsider the new submission, but after rebuilding this paper.

Best regards.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The article titled "A GWR-CUDA Approach to Explore the Spatial Agglomeration and Heterogeneity for Arable Land Consumption in Southwest China" presents a thorough analysis of a model to identify arable areas in different locations in China using a variety of socioeconomic and environmental variables. It is worth noting that the article meets the standards of an excellently structured and well-presented paper. I congratulate the authors for their valuable contribution. However, I would like to make a few observations to improve the article further:

1. It would be helpful to indicate the specific location of the study area in China. Also, I would recommend highlighting the division of the regions studied using colors and providing a clear legend to help you understand.

2. The use of the term “natural” should be corrected to “environmental” or “environmental data” for greater precision in the description of the variables.

3. The references to the figures that currently have errors (¡Error! Reference source not found) must be corrected.

4. It would be beneficial to add a description of the meaning of the acronyms used in Table 3 for greater clarity.

5. The meaning of the colors in the legend of Figure 6 should be clarified. I think adding a detailed description or an explanatory title would help me understand better.

 

6. He observado que la contribución de cada variable se encuentra en el material complementario. Sería útil incluir un cuadro en el artículo principal que muestre la proporción de la contribución de cada parámetro evaluado, junto con una descripción y un análisis dentro del artículo.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The discussion of this study is very interesting, focusing on spatial agglomeration and heterogeneity for arable land consumption in Southwest China. These discussions are very helpful for the sustainable development of developing countries, especially emerging economies. Here are some specific suggestions:

1. This study focuses on the case of Southwest China, considering Yunnan, Guizhou and Chongqing, but why is Sichuan (one of the most important provinces in Southwest China) not considered?

2. The background introduction in this study uses many local units (such as 10^4 hm^2, ha/person), which are not common in international discussions. In order to improve the readability of the manuscript and facilitate readers from other regional backgrounds, the reviewers suggest that the authors use more common units for these discussions.

3. In addition to GDP, some recent studies have also discussed the role of technology policies and land policies. For example, a recent study "The carbon emission implications of intensive urban land use in emerging regions: insights from Chinese cities" also uses the GWR model and considers the role of capital intensity.

4. There are many inconsistent reference formats and fonts in the manuscript. Please correct them.

5. Figure 3 is somewhat unclear, please provide additional explanation.

6. The manuscript would benefit from discussing how the findings of this study can be generalized to other cases outside of Southwest China.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf