Global Maps of Agricultural Expansion Potential at a 300 m Resolution

Round 1

Reviewer 1 Report

see Review_Cengic_land_2023

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

In this paper, the authors create 300m ‘probability’ maps for cropland expansion, at the global-scale. They use long-term ESA CCI land cover datasets to train and validate their models, using ANN approaches. They validated the models with a sub-set of the data for both cross-validation and hindcasting.

General comments: This is a very strong paper and well written. This is an important topic and the authors use best-in-class methodologies and are clear about their approaches and have already made the mapping outputs publicly available. This is one of the best written papers I have read in quite a while. My main critique: I don think 'probability' is an appropriate term for what you are trying to communicate. Essentially, the point is that a certain combination of biophysical and anthropomorphic factors are good predictors of conversion to agriculture. Given the current land-use classification (the three categories you listed for ag), where are the remaining uncultivated areas that fit the historical criteria for conversion to agriculture? The assumption then is that with growing population and the need for food production, these areas may be the ones that become converted, emphasis is on may as there is no empirical use of probability distributions for this assumption. Overall, you have a good story, but the term 'probability' implies some kind of distribution and statistically-driven probability, which is not what the outputs are. For instance, there is not an area of the world with 100% probability of being converted to agriculture, while there are areas that meet 100% of the criteria for historically having been converted to agriculture based on your analysis. These are two distinct things. A better term is needed than 'probability'. Again, I think this paper is very strong and has a lot of potential uses (in fact, I am already planning on using it for an upcoming project – so thanks for the effort!), better to state more clearly what you are deriving. Also, I imagine there is now data from 2016 on. Why was this not used?

Specific comments:

Line 166: Regarding ‘hidden layers’, add a brief description of what this means, one or two sentences. This is a very cryptic term and readers who are not familiar with NN methods may not know what this means.

Lines 177-179: Provide greater detail of the explanation of why this was done. Parameter tuning can also be cryptic.

Line 195: Replace "turned out to be" with "was".

Lines 197-199: Were individual models trained for each of the 6,630 tiles? If so, what are the implications for differing prediction error at the tile-level when aggregated (and presumably smoothed) for the whole prediction mosaic?

Lines 212-214: I imagine there were local differences between individual grid-cells. How did you derive the variable importance at the global-level?

Lines 215-216: With Random Forests, I believe the order of the variables influences the Variable Importance result. I am not familiar with the process for deriving this from ANN, but I imagine it would be similar to RF. I also assume this is why you shuffled the variables, but why only once for each variable? Is there are 'best practice' for doing so?

Lines 224-227: I would like to see how your predictions compare do some of the other global and regional cropland maps. This would add credibility that what you have produced is a least as good as what else is available. For example, the one produced by Tsinghua.

Figure 2: The color chosen for the map may not follow best practices for people which challenges distinguishing between and within colors. There is a great paper by Crameri, et al., ‘The misuse of color in science communication’, 2020, Nature Communications. This paper outlines the best practices in color selection. Give it a look through and see if there is a more appropriate color scheme to use.

Lines 251-257: Directionality is important for many of these. For example, colder or warmer? For pH, more acidic or less acidic? Etc. Please provide these where appropriate.

Lines 301-303: Important caveat, should be included in the data access portal.

Line 319: Regarding the phrase ‘can be reliably used’, I would not go that far. Reliability is totally dependent on things outside of your analysis. For example. National and regional land-use policy, natural disasters, climate change, war, etc. Rather, these data provide users to make accurate assessments of areas that may potentially be converted.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Current global land change models have relatively coarse spatial resolution, which limits environmental impact assessment. To address this issue, the study developed global maps representing the probability of translating to agricultural land at a resolution of 10 arc seconds (~300 m at the equator) and created these maps using an artificial neural network (ANN) model. The probability maps can be used to narrow down the predictions of global land change models to a more refined model of future agricultural expansion, which could be an asset for global environmental assessment.

Recommended for acceptance with modifications.

However, the paper still has shortcomings.

(1) The study period is 2003-2013, and it is suggested that the authors extend the period to about 2020.

(2) The accuracy of the model needs further testing and validation.

Author Response

Please see the attachment.

Author Response File: Author Response.docx