Next Article in Journal
A Numerical Study of the Fire Resistance of Square Steel Tube Columns Embedded in Walls
Previous Article in Journal
Advanced Evaluation of Fire Resistance in Spruce Wood (Picea abies spp.) Treated with Innovative Surface Coatings
 
 
Article
Peer-Review Record

Post-Fire Burned Area Detection Using Machine Learning and Burn Severity Classification with Spectral Indices in İzmir: A SHAP-Driven XAI Approach

by Halil İbrahim Gündüz 1,*, Ahmet Tarık Torun 2 and Cemil Gezgin 1
Reviewer 1: Anonymous
Reviewer 2: Anonymous
Submission received: 27 February 2025 / Revised: 17 March 2025 / Accepted: 20 March 2025 / Published: 21 March 2025
(This article belongs to the Section Fire Science Models, Remote Sensing, and Data)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Post-Fire Burn Severity Detection in İzmir: A Machine Learning and SHAP-Driven XAI Approach with Sentinel-2 Imagery

by

Halil İbrahim Gündüz, Ahmet Tarık Torun and Cemil Gezgin

 

In this manuscript, the authors present the results of a study of using the Sentinel-2-derived spectral indices and four machine-learning algorithms in burned area detection. The study demonstrates that the Random Forest model achieved the highest performance. The results also show that dNBR and dNDVI indices as well as SWIR/NIR bands are the most influential variables when distinguishing between burned and unburned areas. While I believe the results showing the performance of several machine-learning algorithms in burned area detection are of great interest and the manuscript is suitable for the Fire journal, there are issues that require a major revision.

 

Major issues:

The title of presented paper is “Post-Fire Burn Severity Detection in İzmir: A Machine Learning and SHAP-Driven XAI Approach with Sentinel-2 Imagery”, however, machine learning methods were not applied to burn severity detection or classification. When generating train/test datasets the authors used dNDVI and dNBR thresholds only to distinguish between burned and unburned areas (Lines 260-261: “… dNBR > 0.1 and dNDVI > 0.07”). In the Results section, only one figure (Figure 3) is devoted to the burn severity classification, which was obtained without any use of machine learning.

 

It is not clear why the authors used machine learning in their study. If the aim was to compare performance of four machine-learning algorithms in burned area detection, then this should be stated more clearly. If the purpose was to detect burned areas using machine learning, then it is unclear what are the advantages of this approach compared to simple use of spectral indices.

The authors calculated spectral indices to distinguish between burned and unburned areas and used some of the obtained data to train and test the models. Then, using the models and essentially the same data, they again classified the territory into burned and unburned areas. It remains unclear why this approach is better than simply using spectral indices.

 

Minor issues:

Lines 17 – 19: “ML-based classification demonstrated higher accuracy and reliability in detecting burned areas compared to traditional spectral index approaches.” This statement does not follow from the results presented in the article.

 

Lines 23 – 24: “…that spectral characteristics play a crucial role in determining burn scars…”. In my opinion, this statement is obvious. If the burn scar detection is based on the change of spectral characteristics, then the spectral characteristics will play a crucial role in burn scar determining.

 

Lines 142 – 143: “Thus, the aim of this study is to conduct a thorough analysis of the wildfire events…”. I suppose, the authors should discuss the aims of the study in more details, because now it is not clear what exactly they mean by “thorough analysis”.

 

Lines 145 – 164: Please consider moving this text to the Methods section.

 

Figure 1: Please consider removing unnecessary coordinate minutes since they all have values of zero.

 

Lines 201 – 203: “Over the past 15 years … increasing from 225 in 2023 to 298 in 2024”. Since the authors consider 15-year period here, is it possible to provide data on the number of fires since 2010?

 

Line 250: Key & Benson developed their dNBR thresholds for burn severity classification in the boreal forests of North America. Can the authors provide references supporting the applicability of these thresholds for Mediterranean ecosystems?

 

Lines 323 – 338: It remains unclear what dataset was used to tune the hyperparameters. Was it the training set, the test set, or an additional dataset used specifically for hyperparameter tuning?

 

Lines 485 – 486: “The secondary objective of this study is to reveal the spatial distribution and intensity of forest fires…”. The authors, however, do not present any results regarding the fire intensity.

 

Lines 569 – 570: “…that the variables dNBR and dNDVI are the most effective factors in detecting burned areas.” This is probably a result of using these two variables to detect burned areas (Lines 260-261).

 

Lines 642 – 644: “The primary objective of this study is to assess the effectiveness of the dNBR and dNDVI indices in detecting and classifying post-fire burned areas in fire-prone Izmir, while examining the contributions of ML-based approaches to this process.” The objective of the study should be stated in the Introduction section. It also remains unclear how the authors assessed the effectiveness of the dNBR and dNDVI indices in detecting and classifying post-fire burned areas. The authors just used classification thresholds developed for other regions with different ecosystems, and no verification or effectiveness assessment was performed using alternative sources of burn severity data (e.g., field data).

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The paper investigates post-fire burn severity detection in İzmir using Sentinel-2 imagery and machine learning (ML) models, including RF, XGBoost, LightGBM, and Adaboost. It evaluates the effectiveness of traditional spectral indices (dNBR, dNDVI) against ML-based classification for more accurate burned area identification. The study integrates Explainable AI (XAI) with SHAP analysis to improve model interpretability, revealing key spectral features influencing burn severity predictions. Results show that the RF model performs best, providing a robust framework for wildfire monitoring and post-fire ecosystem recovery.

Although the idea is valid, there are several problems that are pointed out below, and the authors need to solve/answer them:  

1. The introduction effectively sets the context and significance of the problem but could benefit from a clearer statement of the contributions and novelty. Add the contributions of the paper at the end of the Introduction section in a bullet point format.

2. In the literature review part,  I highly suggest the author enhance the literature review discussion by adding the limitations of existing approaches more explicitly to better position the novel contributions of your work. Please add the pros and cons of each method.  

3. I highly suggest the author include more recent and cutting-edge research from 2023 to 2025 and especially 2024. Considering the following paper:

  • Gupta, P., Shukla, A. K., & Shukla, D. P. (2024). Sentinel 2 based burn severity mapping and assessing post-fire impacts on forests and buildings in the Mizoram, a north-eastern Himalayan region. Remote Sensing Applications: Society and Environment36, 101279.
  • Boroujeni, S. P. H., Razi, A., Khoshdel, S., Afghah, F., Coen, J. L., O’Neill, L., ... & Vamvoudakis, K. G. (2024). A comprehensive survey of research towards AI-enabled unmanned aerial systems in pre-, active-, and post-wildfire management. Information Fusion, 102369.
  • Tiengo, R., Merino-De-Miguel, S., Uchôa, J., Guiomar, N., & Gil, A. (2025). Burned Areas Mapping Using Sentinel-2 Data and a Rao’s Q Index-Based Change Detection Approach: A Case Study in Three Mediterranean Islands’ Wildfires (2019–2022). Remote Sensing17(5), 830.

4. Justify the choice of Optuna and its impact on model performance.

5. Clarify why specific Sentinel-2 bands were used over others.

6. Compare the processing time and resource usage of ML models.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Post-Fire Burn Severity Detection in İzmir: A Machine Learning and SHAP-Driven XAI Approach with Sentinel-2 Imagery

by

Halil İbrahim Gündüz, Ahmet Tarık Torun and Cemil Gezgin

 

The authors carried out significant revisions to address the concerns raised with the previous version of their manuscript, which this reviewer greatly appreciated. I believe, the manuscript could be recommended for publication in Fire.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors answered all my comments, and I do not have more comments and concerns.

Back to TopTop