Modeling Spatiotemporal Patterns of Land Use/Land Cover Change in Central Malawi Using a Neural Network Model

Round 1

Reviewer 1 Report

This paper presents a quite straightforward land cover change analysis. Its greatest merit is focusing on a relatively less researched region (Sub-Saharan –Africa) with rapid land cover change processes.

The paper uses well established methods, which are described adequately in the Methodology section. I have only one remark about the methodology: the Land Change Modeller module of Terrset allows the user to declare spatial Constraints and incentives, to inhibit or facilitate certain transitions (protected areas, zones dedicated for development). Based on some features of the case areas (protected areas), I think it would have been justifiable to use this function. However, I found no mention of it in the paper. Since it is a minor shortcoming, I can accept that it is not feasible to modify the research at this phase. However I suggest the use of the Constraints and incentives function in connecting future research.

The results are presented in great detail, and richly illustrated. My biggest concern is that I find the skill measures of the submodels quite low. To my best knowledge, while there is no strict rule what value is acceptable, generally skill measures over 0.6 can be deemed good. (See https://www.sciencedirect.com/science/article/pii/S2405844018317626 and https://www.mdpi.com/2071-1050/12/4/1570/htm for example). This issue should be looked upon, and addressed or explained in some manner.

There is also little word about the role of the included spatial variables. Additional/extended information (either in the results or in the discussion) would be welcomed.

Moreover, Table 3 seems unnecessary. The transitions from to information could be given in text, and since the submodel names are not used later, it is just unnecessary information.

The Discussion does not properly discuss the underlying factors between the differences of the land cover change processes of Dedza and Ntcheu Districts – which mitigates the main point of having more than one case studies. While there some is information in the paper, I think this could be a key finding in highlighting more or less sustainable land use practices/policies.

Also, the discussion could contain more information about the similarities/differences with described/studied land cover change processes from Malawi and from Southwestern Africa. This would give more significance to the findings.

One final small remark: the page numbering should be corrected.

After addressing these relatively minor concerns, I recommend the paper for publication.

Author Response

Thank you for your review comments, please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

I appreciate the aims and research work including very efficient operationalization of analysis using LCM of TerrSet Software. Especially the prediction matrix (Table 10). The paper is very interesting and I read it curious.

However, my main objection is that there is lack of classic elements in almost every Land Use study i.e. transition matrices, which summarize the LU/LC changes. Its form is almost similar to table 10 but contains either the areas in sq .ha (or sq .km) or the shares (percentage) of areas of the registered changes between the land use classes. These matrices should present the transition e.g. 2001-2009, 2009-2019 as well as predictive transitions 2019-2050.

These data are currently available in the paper but dispersed and presented fragmented.

Author Response

Thank you for your review comments, please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript deals with the retrieval and modelling of land use patterns in Central Malawi, a developing city. The study is of significant scientific impacts, but many specific parts of the manuscript can be written in a clearer manner, and some information has to be further provided, before the manuscript is reconsidered for publication:

Major Modifications needed

Paragraphs 2-3 of Introduction (i.e., Paragraphs 1-2 of Page 2): can be shortened, as this is not the main focus of the study.

Paragraph 4 of Introduction (i.e., Paragraph 3 of Page 2): What are these models? This study should more focus on statistical models adopted.

After Paragraph 4 of Introduction: The Introduction should outline and summarize the use of other statistical techniques and algorithms in land use retrievals, especially the description of studies conducted in other developing cities. Some references are as shown:

https://www.mdpi.com/2072-4292/13/16/3337

https://onlinelibrary.wiley.com/doi/abs/10.1111/tgis.12174

https://www.mdpi.com/2072-4292/14/10/2349

https://www.sciencedirect.com/science/article/pii/S0924271621001635

Introduction: References [24] [25], [26], [36] should state the author, instead of writing [24]... Also, references [25] and [26] - How is the predictive model run? Any assumptions were made?

Introduction: The use of MLP-MC sounds amazing based on past studies, but it will be good to add a short summary of statistical results obtained.

Introduction: What are the research gaps of existing studies, especially those which have adopted the CA-Markov Chain model for land use retrieval?

Page 3 (Section 2.1): It seems that Dedza and Ntcheu districts are quite similar in terms of environmental conditions, elevation range etc. What is the significant difference of these 2 districts? What is the main reason that these 2 districts are selected?

Page 4 (the last paragraph of Section 2.1): Socio-economic conditions of the two districts may not be important in this study. Further, the description on poverty conditions should be shortened.

Table 1: How is the resampling of "densely populated areas" attributes conducted? Also, for distance from rivers and distance from major roads, Why a buffer of 50 m or 30 m was selected? Any scientific explanation?

Figure 2: How could the comparison of Summer 2018 field observation datasets with retrieved datasets from satellite imageries in 2001, 2009 and 2019? There may be temporal mismatch.

Section 2.4: The post classification image change analysis should be discussed in more details.

Section 2.5: what is such activation function mapping?

Section 2.5: more statistical / mathematical details of the back-propagation algorithm should be provided in Section 2.5.

Also, how are the training errors being used to adjust the weight for ensuring model accuracy? Some descriptions should be provided.

Section 2.5: Not quite understand how the probability for each pixel to experience changes is calculated. Please kindly provide some probabilistic formulation.

Section 2.6: How about validation in previous 2 time period? 2001 and 2011 respectively?

Section 3.1 and Figure 5: For the pixels that are categorized as "Others", what properties do they possess? For example, how is their vegetative index etc.?

Tables 6 and 7: Are some of the figures really 100? Is it 99.XX, then being rounded off / rounded up? Also, is there a possible explanation of the difference in Producer's accuracy and User's accuracy for Dedza district and Ntcheu district?

Section 3.3 (2nd paragraph): What do you mean by anomalies? Is it possible to be more specific?

Page 19: For the paragraph after Figure 9, still cannot understand how the weight(s) are imposed in the MLP-Markov chain modeling

Table 10: Have you tried to conduct the prediction for a few years later, say, based on datasets of 2001, 2011 and 2019? For example, the land cover changes in 2021 (or 2022), then you will have some validation work. The data of 2050 is rather far away, and is lacking of datasets for validation.

Table 10: How is the assessment conducted with regards to model accuracy?

Section 4 Paragraph 1: The connection between terrain parameters and environment conditions with changes in land use pattern has not been clearly illustrated in the manuscript.

Discussion - the context of future land use modeling in 2050 should not have any repetition when compared with Results (Section 3). Please simplify the first few paragraphs of Discussion

Discussion - The use of 30 m resolution datasets is obviously a problem and a potential technical shortcoming, so how could this be overcome?

Section 5 (Conclusion) - What kind of policy? Need to be more specific

Tables 12, 13, 14 and 17 - Again, is this 100% user's accuracy corrected to some significant figures / decimal places?

Typo / Grammatical Changes

Line 3 of Abstract: "understand"

The last paragraph of Section 2.2: km^2

Table 2: with a bracket in the caption of Table 2.

Table 4: It would be better to state the full form of Ha at the bottom of Table 4.

Figure 6: What do those "black" dots / pixels represent?

Page 15 Line 13: be distinguished

Tables 8 and 9 - should keep consistent number of decimal places in all numerical figures

I think the authors should address all aforementioned problems, before they resubmit the manuscript. Thanks for asking me to review this manuscript, it looks good and interesting.

 

Author Response

Thank you for your review comments, please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The revision and modifications have greatly improved the quality of this work, except the following places.

(1) We understand that the study has adopted regression-based modelling, Cellular Automata and the multilayer perceptron neural network model with Markov chain for land use retrieval, however these ideas come from statistics / enhancement of existing models, therefore it is much better to add a summary of previous studies in land use retrieval, especially those that have made use of statistical techniques and algorithms. This acts as a proper acknowledgment to previous efforts and approaches developed in this discipline. Some references as proposed in our previous review report are as shown:

https://www.mdpi.com/2072-4292/13/16/3337

https://onlinelibrary.wiley.com/doi/abs/10.1111/tgis.12174

https://www.mdpi.com/2072-4292/14/10/2349

https://www.sciencedirect.com/science/article/pii/S0924271621001635

(2) The authors have mentioned that one significant gap is that individual models have limitations when they are applied individually, but are more robust approaches when the models are applied together. Some elaborations and discussions should be included, and this should begin with the references in (1), towards the continuous enhancement of models and statistical framework for land use retrieval.

(3) For the temporal mismatch and comparison, it would be better to include some statistics in the manuscript, showing that even the year does not match, it will be reliable.

Other than that, I think the manuscript is good. Please kindly spend some time to address the aforementioned points and issues. Thank you for your devotion. Enjoy reading it.

Author Response

Please see attachment, thank you

Author Response File: Author Response.docx