Comparative Analysis of Non-Negative Matrix Factorization in Fire Susceptibility Mapping: A Case Study of Semi-Mediterranean and Semi-Arid Regions

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study proposes an improved Non-negative Matrix Factorization method (SEI-NMF) for forest fire susceptibility mapping in semi-Mediterranean and semi-arid regions and systematically compares it with traditional methods such as PCA, K-means, and IsoData. However, several issues remain:

1）Writing and Presentation Issues:

(1) The introduction is overly long and lacks a clear logical structure. For example, the first paragraph combines the global fire context, climate and anthropogenic factors, and the Kurdo-Zagrosian case study, resulting in a high information density that may overwhelm the reader.

(2) The current literature review only briefly mentions the potential of NMF in fire research and lacks a direct comparison with similar methods (e.g., clearly stating the shortcomings of PCA and K-means). Recent applications of NMF in the remote sensing field should be added to avoid an overemphasis on basic theory.

(3) The innovation points are ambiguously stated. The motivation behind SEI-NMF improvements (sparsity and endmember independence) is not well-established in the introduction, making it difficult for readers to grasp the necessity of the method.

(4) Some of the results lack specific descriptions of experimental findings, such as Figures 5, 6, and 7. The authors describe only the concepts of the figures without detailing the experimental outcomes, which are instead discussed later. This misplacement is inaccurate—discussion sections should interpret the results, not present them. Overall, the writing and logical structure require significant improvement.

2）Equation 12 Parameter Definition: Parameters such as α and γ should have their physical meanings clearly defined upon first appearance to avoid forcing readers to backtrack. A short subsection is recommended to explain the core differences between SEI-NMF and conventional NMF.

3）Formatting Issues: 

(1) Reference formatting is inconsistent, including discrepancies in whether years are bolded, and whether journal titles are abbreviated consistently. These should be corrected according to the Land journal’s formatting guidelines.

(2) Title numbering errors (e.g., “2.3. C. Sparse…” where the “C” is redundant) need to be corrected.

(3) Figure captions and references should be standardized—clarify whether to use “Figure” or “Fig.”

4）Figure-Specific Issues:

(1) Figure 4 (convergence curve) does not clearly differentiate the convergence performance of various NMF variants. It is suggested to explain in the caption why “L1/2-Sparsity” converges faster.

(2) Figure 7’s statistical data would benefit from the addition of error bars or significance tests (e.g., Chi-square test) to validate whether the differences among methods are statistically significant.

5）Discussion of DEM Data Noise: The discussion mentions that “DEM data may introduce noise” but does not analyze how terrain features (e.g., slope, aspect) specifically influence classification. It is recommended to supplement this with an analysis of terrain-spectral data correlations or sensitivity experiments to verify the role of DEM.

6）Data Overlap and Overfitting Risk: The Sentinel-2 data spans from 2020–2023, overlapping partially with the validation data (2021–2023). It should be clarified whether this temporal overlap poses a risk of overfitting.

7）Spatial Overlap Analysis: The specific process of spatial overlap analysis (e.g., buffer size, matching thresholds) should be described in detail. Adding quantitative metrics such as AUC or F1-score would enhance the credibility of the results.

8）ZGI Index Calculation: The ZGI index is only referenced via citation [2], but its specific calculation formula and ecological significance should also be provided explicitly.

9）Discussion of Limitations: The study should clearly outline its limitations, such as SEI-NMF’s computational efficiency on high-dimensional data, sample size constraints, or seasonal biases in semi-arid datasets (e.g., whether the 2020 Sentinel data captures the peak fire season).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

In this study authors propose application of Non-Negative Matrix Factorization (NMF) method, including their proposition a modified Non-Negative Matrix Factorization (NMF) method, for detecting fire-prone areas using satellite-derived data in SM and SA forests. The performance comparison with several unsupervised methods: Principal Component Analysis (PCA), K-means, and IsoData. The presented results demonstrate effectiveness of NMF methods when applied for fire-prone areas identification across large geographic extents typical of SM and SA regions, and also revealed  that when elevation data was included, NMF_L1/2 sparsity method provides superior performance among considered methods. In contrast, the proposed NMF method provided the best results when only Sentinel 2 bands and ZGI were used. 

The paper possess some strengths and potential contribution. Authors are suggested to improve the paper such that:

• the proposed SEI-NMF method offer some minor performance improvements when compared to NMF-L1/2 method when elevation data are excluded (i.e. NMF-L1/2 method has better performance when combined high and average decisions are observed). However, the performance for both of these methods in case when elevation data are included are higher with NMF-L1/2 method being obviously superior - this must be discussed in more detail. Does the proposed method offer some other benefits - such as complexity etc.
• The performances for all method is shown only regarding the area where fire occurred between 2020 and 2023. Are some of these methods prone to overestimate the probability of fire occurrence - this can be viewed if other part of the areas in which we had not observed fire occurrence in the given period. This could be given in similar statistical manner as for the first type of areas.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors of this study have proposed a modified non-negative matrix factorization method to detect fire-prone areas using satellite data in forests of semi-mediterranean and semi-arid regions in western Iran and northern Iraq. Although the paper is very interesting, some wording adjustments need to be made before it can be accepted by the journal.

Introduction
Lines 42-44: The context of the paragraph is about fires and climate change globally, but this particular section talks about the Semi-Mediterranean and Semi-Arid forests of the Kurdish-Zagrosian mountains. It should be relocated to another section or perhaps moved to the second paragraph.

Lines 81-96. Most of this section is quite long and should be removed. It could be included in each of the methods used.

Methods
Lines 307-357. This is the most critical section of the study. Some paragraphs seem more like a discussion than the methodological description of the study, so it has to be almost completely rewritten. In addition, it is mentioned that the validation process was carried out through a spatial overlap analysis, comparing the areas classified as high, medium and low susceptibility to fire with the burned areas recorded after 2020. The authors used the Normalized Burn Ratio (NBR). Although the NBR is obtained from the NIR and SWIR bands that provides a better distinction between burned and unburned areas, as well as an optimal signal to obtain information on the variation of fire severity, it should not be used to validate the study since field data or visual selection of points with satellite images are needed. In addition, a confusion matrix or other statistical measure should be performed.

Results
Lines 359-362. It is not relevant to include this paragraph, better to remove it, since it is expected in this study, the presentation of the maps and the statistical results.

Figures 5 and 6. The colors are confusing. The dark areas are identified as representing the high-fire susceptible areas, while the orange and blue areas represent the average and low-fire susceptible areas. Other color shades should be used. At first glance the dark areas look like shadows.

Figure 7. The presentation of this graph should be improved to make it more understandable.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors addressed my all concerns. 

Author Response

Comments 1: The authors addressed my all concerns. 

Response 1: We sincerely thank the reviewer for their thoughtful evaluation and for accepting our work.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have sufficiently addressed the reviewers' comments, so the manuscript can now be accepted for publication. Congratulations!

Author Response

Comments 1: The authors have sufficiently addressed the reviewers' comments, so the manuscript can now be accepted for publication. Congratulations!

Response 1: We sincerely thank the reviewer for their thoughtful evaluation and for accepting our work