Using Constrained K-Means Clustering for Soil Texture Mapping with Limited Soil Samples
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript presents an application of Constrained K-Means Clustering (CKMC) for soil texture mapping, addressing a gap in digital soil mapping (DSM) where labeled samples are limited.
Key contributions are: i) semi-supervised approach where CKMC leverages limited labeled data alongside abundant unlabeled data, offering a solution for regions with sparse soil sampling; ii) the study compares four distance metrics (Euclidean, Manhattan, Maximum, Canberra) in CKMC against supervised methods (RF, MLP), demonstrating CKMC’s superiority; and iii) combines GF-2 remote sensing imagery and ALOS DEM-derived topographic variables, highlighting MRRTF, LU, TPI, and PlC as key predictors.
Suggestions and recommendations:
- Clarify how CKMC’s constraints (must-link/cannot-link) are derived from soil texture labels to justify its semi-supervised nature.
- Compare findings with semi-supervised DSM studies to contextualize CKMC’s novelty.
- The manuscript acknowledges CKMC’s susceptibility to local optima but does not quantify its impact; a sensitivity analysis (e.g., multiple runs with random initializations) would strengthen robustness claims.
- Discuss how CKMC’s performance might scale to larger areas or more texture classes (beyond four types).
- Define "MRRTF" and "PlC" at first use for non-specialists.
- Soil samples (2021–2025) versus GF-2 imagery (2019). Justify or address eventual land-use changes.
- Excluded climate/parent material due to "homogeneity"; is it possible to quantify this assumption (like spatial autocorrelation tests)?
- Add a correlation matrix of environmental variables to justify excluding collinear predictors (PrC, Asp, …).
- Euclidean outperforms others, but the rationale for initial selection (for example, why not Mahalanobis?) is missing.
- Stratified 80/20 split with 100 repetitions is okay, but consider spatial cross-validation to address spatial autocorrelation.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe article is interesting but needs some improvement.
English language and style are quite fine, only a minor spell check is required.
I think that deciding to discard covariates a priori is rather arbitrary. In my opinion, they should be included when applying VarImp function, and only after the results they can be discarded if their importance is really negligible.
Please give references for all the variables in Table 1 and Table 2.
Line 28: undefined acronyms.
Line 144: undefined acronym.
Line 159: 1 meter.
Line 197: undefined acronym.
There is no need to repeat in the text the same information already reported in Table 3.
There is no need to repeat the results in the Conclusions section, that should present the outcome of the work by interpreting the findings at a higher level of abstraction than the Discussion and by relating these findings to the motivation stated in the Introduction.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsSuggestions in the review pdf file.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsAuthors have improved the quality and have addressed most issues.
