Dual-Dataset Deep Learning for Improved Forest Fire Detection: A Novel Hierarchical Domain-Adaptive Learning Approach

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

I have following observations.

1.The introduction is long for readers. The details are important, but the article has to be concise to give an attractive story.

2.To prevent CNN models from overfitting the training set, regular terms are often added. Apart from the use of dropout regularization in the modelling, does the model itself in the manuscript use any kind of other strategy (like cross-validation)?

3.Your parameter settings rely on experience. Is it possible for the parameters to be further optimized? For example grid search or intelligence optimization algorithms.

4.Limited number of performance metrics are used. Additional relevant metrics may be included. Or you can consider the number of parameters and time complexity of the algorithm.

5.Why do you start unfreezing at layer 80? Can you be more specific or add relevant literature?

6.EfficientNetB0 has gone to EfficientNetB7, is the B7 better if accuracy is anything at all? I suggest doing a full comparison. Regarding the section on the structure of the model network, I suggest adding a picture description, in the full text I only see two figures of the dataset, which is inappropriate for a deep learning paper, please do this part. I don't even know if the model uses Batch Normalization.

7.The paper lacks interpretability or negative transfer discussion, which is essential for Math journals. I suggest the authors should at least add and report the section about negative transfer and model interpretability in their limitations. Otherwise, the manuscript looks like just a black box data processing publication. For instance, there are several papers on the subject, just to mention a few:

10.1109/TGRS.2023.3247343

10.3390/rs15133422

10.3390/f14102080 

8.The conclusion could be more specific. And some of the conclusions drawn from the discussion are obvious and do not make much sense.

Author Response

Response to Reviewer 1 Comments

We would like to thank you for the constructive comments and suggestions provided for the manuscript titled "Dual-Dataset Deep Learning for Improved Forest Fire Detection: A Novel Hierarchical Domain-Adaptive Learning Approach". Your insights have greatly helped in enhancing the quality and clarity of our work.

In the revised version of the manuscript, all the modified, added, and revised relevant parts have been highlighted with a light blue background to facilitate your review. This way, you can easily identify the changes and assess how they align with your comments.

I have the following observations.

1.The introduction is long for readers. The details are important, but the article has to be concise to give an attractive story.

Thank you for your feedback on the length of the introduction in our manuscript. Taking your valuable input into consideration, we have undertaken a thorough revision of the introduction to make it more concise while retaining the essential details of our study.

In the revised manuscript, we have significantly shortened the introduction. This was achieved by streamlining the language and focusing more directly on the core aspects of our study. We made sure to maintain a balanced and measured tone throughout, ensuring that the introduction remains engaging and informative. The key changes include:

1. Condensing the Background: We have succinctly summarized the environmental significance of forests and the challenges posed by forest fires, ensuring a clear and quick understanding of the context and relevance of our study.
2. Refining the Description of Previous Work: The discussion of previous advancements in fire detection technology, including the role of deep learning models, has been refined to emphasize their relevance to our research more directly.
3. Highlighting the Novelty of Our Approach: We have more clearly articulated the novel aspects of our study, particularly our hierarchical domain-adaptive learning approach, and how it addresses the challenges in wildfire detection.
4. Reducing Redundancies: We have carefully eliminated any redundant information to enhance the readability and flow of the introduction.
5. Emphasizing Key Contributions: The introduction now more effectively highlights the key contributions and implications of our research, making it more appealing to readers.

Through these revisions, we believe that the introduction now presents a more attractive and succinct story, drawing readers into the study while providing them with all the necessary background information. We hope that these changes address your concerns regarding the length and detail of the introduction.

We are grateful for your insights and suggestions, and we welcome any further feedback you may have.

2. To prevent CNN models from overfitting the training set, regular terms are often added. Apart from the use of dropout regularization in the modelling, does the model itself in the manuscript use any kind of other strategy (like cross-validation)?

Thank you for your insightful query regarding the mitigation of overfitting in our CNN models. Overfitting is indeed a critical concern in model training. To address this, we have conducted a systematic evaluation of regularization techniques on the custom layers added to the base model, as detailed in the added part of the revised '2.4. Training Process' section. Through this evaluation, a dropout rate of 0.2 was identified as optimal, balancing model complexity and performance effectively, and was subsequently adopted for all experiments. The base model itself benefits from the inherent regularization capabilities of the pre-trained EfficientNetB0 architecture, which includes several batch normalization layers. These layers have been shown to aid in model generalization by mitigating internal covariate shift.

Regarding cross-validation, while it is a widely acknowledged method for assessing model generalizability, the computational demands of training with high-resolution image data and the substantial size of our datasets led us to choose a robust single hold-out validation set. This decision is further discussed in the revised '4.1 Study’s Limitations and Future Work' section.

Collectively, these strategies have yielded a model that exhibits consistent performance across training, validation, and test datasets. We believe this multifaceted approach effectively counters overfitting, ensuring that the model remains both robust and generalizable.  

3.Your parameter settings rely on experience. Is it possible for the parameters to be further optimized? For example grid search or intelligence optimization algorithms.

We appreciate the referee’s suggestion to explore more systematic approaches to hyperparameter optimization. The current parameters were indeed chosen based on a combination of empirical evidence from prior research and preliminary experimentation which suggested a good starting point for the model configurations.

While methods such as grid search or more sophisticated optimization algorithms like Bayesian optimization or evolutionary algorithms offer thorough means to explore the parameter space, they also introduce significant computational resources and time, which may not be feasible within the constraints of our current research environment.

In subsequent research, subject to resource availability, we will consider these advanced optimization techniques to refine our model further. This exploration will be crucial for a more detailed understanding of hyperparameter sensitivity and model robustness, a point we have added to the revised '4.1 Study’s Limitations and Future Work' section.

4.Limited number of performance metrics are used. Additional relevant metrics may be included. Or you can consider the number of parameters and time complexity of the algorithm.

Thank you once again for your valuable feedback. We have taken careful consideration of your comments regarding the number of performance metrics and in response to them, we have expanded our suite of performance metrics to include specificity, False Negative Rate (FNR), and Matthews Correlation Coefficient (MCC). These metrics were carefully selected to provide a comprehensive evaluation of the models' performance, addressing different aspects of classification beyond the traditional measures.

As for the previous ones, each metric was computed across five separate runs to ensure robustness, with the average values and standard deviations reported to capture the variability and reliability of the results. Furthermore, to compare the efficacy of the different training approaches, we employed bootstrapping to calculate the 95% confidence intervals for the differences in performance metrics, providing a statistically rigorous assessment.

The inclusion of these additional metrics has enriched the discussion around the operational strengths of the proposed approach and allowed for a more nuanced understanding of its performance. The integration and analysis of these results have been detailed in the revised '2. Materials and Methods' and '3. Results' sections of the manuscript, ensuring that the study reflects a thorough and multifaceted evaluation.

Turning to the reviewer's suggestions regarding architectural complexity, the proposed approach utilizes an innovative training methodology that partitions the base-model architecture into discrete components for the Hierarchical Domain-Adaptive Learning process. Despite this novel training strategy, the final architecture of the resulting model remains consistent with the original EfficientNetB0 in terms of the base-model structure and the regularization techniques applied to the classifier. This consistency ensures that there are no inherent architectural changes that would adversely affect the model's efficiency.

In response, we've conducted a parameters and time complexity analysis, encapsulated in the new section '3.4.1. Time Complexity and Parameter Efficiency Insights' showing comparable efficiency to traditional methods. However, the consistent inference times observed may also be attributed to the GPU support coupled with the size of the test dataset, which consists of 410 images. Future research will extend this analysis to larger datasets, providing deeper insights into the training approach's impact on inference time and learned features. This aspect is of particular importance in scenarios where large-scale model deployment necessitates rapid inference capabilities.

We appreciate the opportunity to enhance our manuscript with this discussion and agree that these points are significant for a comprehensive understanding of the model's practical applications.

5.Why do you start unfreezing at layer 80? Can you be more specific or add relevant literature?

Thank you for your query regarding our approach to fine-tuning the EfficientNetB0-based models. Our decision to divide the model into three sections with almost equal numbers of layers (80, 160, and 238) was driven by the intent to systematically explore the transfer learning capacity of the model across its depth.

We aimed to assess the impact of fine-tuning different proportions of the network to understand how different layers' adaptability contributes to our task of forest fire detection. This was based on the principles of transfer learning where different layers capture features of varying abstraction levels (Yosinski et al., 2014; Zhou et al., 2016; He et al., 2016). Our decision also considered computational efficiency. Balancing computational resources with the potential benefits of fine-tuning additional layers is a crucial aspect in the partition choice, especially given the large-scale and high-resolution nature of our datasets (Pan & Yang, 2010).

1. Starting at Layer 80: The choice of 80 layers represents the first third of the model, where the initial layers capture more basic and general features. By starting to fine-tune from this layer, we aim to adapt the higher and more complex features of the model while retaining the general features learned from the pre-trained dataset.
2. Layer 160 (80*2): This layer marks the two-thirds point in the model. By extending the fine-tuning to this layer, we test the effectiveness of adapting a larger portion of the model, including more specialized features, to our specific task.
3. Layer 238: Fine-tuning the entire model, including all convolutional layers, serves as a baseline to evaluate the extent to which the pre-trained model's capabilities can be leveraged for forest fire detection without significant adaptation of its convolutional architecture.

This structured approach enabled us to empirically determine the depth of fine-tuning that yields the best performance, providing insights into the optimal balance between leveraging pre-trained features and adapting the model to new tasks. The justification of these choices have been detailed in the revised '2.4.1 Hyperparameters Optimization' section.

6.EfficientNetB0 has gone to EfficientNetB7, is the B7 better if accuracy is anything at all? I suggest doing a full comparison. Regarding the section on the structure of the model network, I suggest adding a picture description, in the full text I only see two figures of the dataset, which is inappropriate for a deep learning paper, please do this part. I don't even know if the model uses Batch Normalization.

In response to the reviewer’s commentary on the progression of the EfficientNet architectures, we have included a new section in our manuscript, '3.5. Performance Benchmarking: Hierarchical Domain-Adaptive Learning on EfficientNetB0 vs. EfficientNetB7.' This section provides a comprehensive comparison of the Hierarchical Domain-Adaptive Learning approach when applied to both EfficientNetB0 and EfficientNetB7 architectures. It offers a critical analysis of the performance trade-offs, considering the considered performance metrics and computational efficiency, to discuss the suitability of each architecture for forest fire detection applications. 

In response to your valuable suggestion regarding the inclusion of a visual description to enhance understanding, we have now introduced three illustrative diagrams within ‘Section 2.3’ of the revised manuscript. ‘Figure 4’ presents a detailed view of the architecture customization of EfficientNetB0 for Dual-Task Learning. Following this, ‘Figure 5’ depicts an overview of the final primary model after fine-tuning. Additionally, to further clarify the methodology and the structure of our approach, we have included ‘Figure 3’, a comprehensive method flowchart.

This newly added ‘Figure 3’ methodically illustrates the flow and structure of the dual-task learning approach adopted in our study. It visually represents the initialization of the dual-branch architecture, the distinct processing pathways of the primary and auxiliary branches, the integration of shared layers, and the subsequent steps in the model's training and fine-tuning processes. The flowchart is designed to provide a clear and concise overview of how our model manages dual-task learning and the flow of information through shared and task-specific layers, thereby offering an enhanced comprehension of the complex methodology employed.

We believe that these additions, particularly the method flowchart in ‘Figure 3’, will greatly aid in elucidating the intricacies of our model’s structure and its operational flow to the readers, aligning with the goals of our research and addressing the concerns raised in your commentary.

7.The paper lacks interpretability or negative transfer discussion, which is essential for Math journals. I suggest the authors should at least add and report the section about negative transfer and model interpretability in their limitations. Otherwise, the manuscript looks like just a black box data processing publication. For instance, there are several papers on the subject, just to mention a few:

10.1109/TGRS.2023.3247343

10.3390/rs15133422

10.3390/f14102080 

Thank you for your valuable feedback on the importance of discussing negative transfer and model interpretability in our study. We appreciate your emphasis on these aspects as they are indeed crucial, particularly in a field as application-oriented as forest fire detection.

Negative Transfer Discussion: We recognize the importance of addressing the potential for negative transfer in the dual-task learning approach. Our method, involving shared layers processing both forest and non-forest fire scenarios, aims to enhance forest fire detection. However, we acknowledge the possibility of negative transfer where features relevant to non-forest scenarios could adversely affect our primary task of forest fire detection. To mitigate this, we carefully selected the depth of the shared layers and fine-tuned the model specifically on the primary dataset. However, despite these measures, we recognize that the risk of negative transfer cannot be entirely eliminated. 

We agree that elaborating on this aspect in our manuscript will provide a more comprehensive understanding of the approach and its considerations. For this reason, we included a discussion in the revised ‘Discussion' section, backed by relevant literature including the references you have kindly provided [10.1109/TGRS.2023.3247343].

Interpretability Discussion: Thank you for highlighting the importance of interpretability in our study. We agree that this aspect is crucial, especially for a Math journal. To address this, we have added a new paragraph in the ‘3.4. Comparative Analysis of the Baseline with the Study’s Proposed Approach’ section, backed by relevant literature including the references you have kindly provided [10.3390/f14102080], where we discuss our preliminary efforts using Gradient-weighted Class Activation Mapping (Grad-CAM) to explore interpretable techniques. While this initial analysis has been insightful, we recognize the need for a more systematic feature analysis. We propose to delve deeper into this aspect in future work, aiming to enhance the transparency and comprehensibility of our model's decision-making process in forest fire detection.

8.The conclusion could be more specific. And some of the conclusions drawn from the discussion are obvious and do not make much sense.

Thank you for your constructive feedback on the conclusion of our manuscript. We have taken your comments into serious consideration and revised the conclusion to be more specific and directly relevant to the core findings of our study.

In the revised conclusion, we have succinctly emphasized the significance and potential impact of the Hierarchical Domain-Adaptive Learning Approach, specifically in the context of forest fire detection. The revisions include:

1. Conciseness and Focus: We have streamlined the conclusion to directly highlight the study's key findings and their implications in a concise manner. This ensures the conclusion is not only informative but also engaging and directly relevant.
2. Specificity of Contributions: The revised conclusion clearly articulates the specific advancements our approach brings to the field, including its statistical validation and potential for real-world applications in environmental monitoring.
3. Highlighting Real-World Applicability: We have underscored the practical implications of our research, particularly its ability to handle diverse datasets and adapt to specific fire scenarios. This emphasis aligns with the transformative impact our approach can have on future innovations in forest fire detection.
4. Direct Addressing of Referee’s Comments: We have specifically focused on addressing your concerns about the original conclusion being too general and obvious. The revised conclusion now presents a more measured and focused summary of our research contributions.

We believe these revisions effectively address your concerns and enhance the overall impact and clarity of the conclusion. We are committed to ensuring that our manuscript communicates its findings and their significance in the most effective way possible.

Thank you once again for your valuable feedback.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript is skillfully composed. It is worth mentioning that it elaborates on the Novel Hierarchical Domain-Adaptive Learning Approach, and relevant key concepts of the proposed method have been carefully explained . Furthermore, the manuscript also provides a detailed description of the baseline approaches. However, there are several issues that need to be addressed before the paper can be considered for publication. If these concerns are adequately resolved, I believe the paper's valuable contribution will become even more significant. 

Major Comment: In Section 2.3, "Architecture Customization of EfficientNetB0 for Dual-Task Learning," it would be more reasonable to include a model architecture diagram and a method flowchart to provide readers with a clearer understanding of the approach.

Minor Comments: In Table 13, bolding the superior results would enhance visual clarity for a more intuitive comparison.

Author Response

Response to Reviewer 2 Comments

We would like to thank you for the encouraging and constructive comments and suggestions provided for the manuscript titled "Dual-Dataset Deep Learning for Improved Forest Fire Detection: A Novel Hierarchical Domain-Adaptive Learning Approach". Your insights have greatly helped in enhancing the quality and clarity of our work.

In the revised version of the manuscript, all the modified, added, and revised relevant parts have been highlighted with a light blue background to facilitate your review. This way, you can easily identify the changes and assess how they align with your comments.

The manuscript is skillfully composed. It is worth mentioning that it elaborates on the Novel Hierarchical Domain-Adaptive Learning Approach, and relevant key concepts of the proposed method have been carefully explained . Furthermore, the manuscript also provides a detailed description of the baseline approaches. However, there are several issues that need to be addressed before the paper can be considered for publication. If these concerns are adequately resolved, I believe the paper's valuable contribution will become even more significant. 

Major Comment: In Section 2.3, "Architecture Customization of EfficientNetB0 for Dual-Task Learning," it would be more reasonable to include a model architecture diagram and a method flowchart to provide readers with a clearer understanding of the approach.

Thank you for your valuable feedback. In response to your suggestion regarding the inclusion of a visual description to enhance understanding, we have now introduced three illustrative diagrams within ‘Section 2.3’ of the revised manuscript. ‘Figure 4’ presents a detailed view of the architecture customization of EfficientNetB0 for Dual-Task Learning. Following this, ‘Figure 5’ depicts an overview of the final primary model after fine-tuning. Additionally, to further clarify the methodology and the structure of our approach, we have included ‘Figure 3’, a comprehensive method flowchart.

This newly added ‘Figure 3’ methodically illustrates the flow and structure of the dual-task learning approach adopted in our study. It visually represents the initialization of the dual-branch architecture, the distinct processing pathways of the primary and auxiliary branches, the integration of shared layers, and the subsequent steps in the model's training and fine-tuning processes. The flowchart is designed to provide a clear and concise overview of how our model manages dual-task learning and the flow of information through shared and task-specific layers, thereby offering an enhanced comprehension of the complex methodology employed.

We believe that these additions, particularly the method flowchart in ‘Figure 3’, will greatly aid in elucidating the intricacies of our model’s structure and its operational flow to the readers, aligning with the goals of our research and addressing the concerns raised in your commentary.

Minor Comments: In Table 13, bolding the superior results would enhance visual clarity for a more intuitive comparison.

Thank you for your valuable suggestion. We have updated ‘Table 14’ by bolding the superior results to improve visual clarity and facilitate a more intuitive comparison. We've applied the same formatting to all newly added comparative tables across the manuscript for consistency and ease of understanding. We believe these changes significantly enhance the manuscript's readability.

Reviewer 3 Report

Comments and Suggestions for Authors

Authors used hierarchical domain-adaptive learning approach for improved forest fire detection from dual-dataset using deep learning method. But the paper fails to address the following questions and needs modification for the better understanding of the readers

1. The manuscript need proper care for making free from typo and grammatical mistake.
2. It will be better if the literature work will be discussed in form of table with their limitations.
3. A proper block diagram for the proposed method is needed with explanation for better understanding.

4.The title suggest about the forest fire detection, but the figure 1 represent the non forest fire situation. How this model is robust for finding the fire in different situation (fire in the building, offices, fire in industries, or fire during a celebration of X-mas).

5. Authors used the hyperparameters (learning rate, optimizer, and epoch). How this values are optimally decided to obtain better accuracy.
6. In this work, transfer learning concept is used by considering EfficientNetB0 model.  What makes the author to use EfficientNetB0 though there are many pretrained networks are available. The working of the pretrained model and their comparisons are discussed in below papers. 

 https://doi.org/10.1016/j.bea.2022.100069

https://doi.org/10.1002/ima.22761

7. It will be better to understand the effectiveness of the work if a comparison table will be presented with respect to the related works.

8.In this manuscript, only five matrices have been computed for the unbalanced dataset. whereas, MCC is a better and can be accessed from the below mentioned paper.

https://doi.org/10.1016/j.bea.2023.100089 

9. The database need to discussed with detail.

Comments on the Quality of English Language

Moderate editing of English language required. The manuscript needs to be checked properly.

Author Response

Response to Reviewer 3 Comments

We would like to thank you for the constructive comments and suggestions provided for the manuscript titled "Dual-Dataset Deep Learning for Improved Forest Fire Detection: A Novel Hierarchical Domain-Adaptive Learning Approach". Your insights have greatly helped in enhancing the quality and clarity of our work.

In the revised version of the manuscript, all the modified, added, and revised relevant parts have been highlighted with a light blue background to facilitate your review. This way, you can easily identify the changes and assess how they align with your comments.

Authors used hierarchical domain-adaptive learning approach for improved forest fire detection from dual-dataset using deep learning method. But the paper fails to address the following questions and needs modification for the better understanding of the readers

1. The manuscript need proper care for making free from typo and grammatical mistake.

Thank you for highlighting the importance of ensuring the manuscript is free from typographical and grammatical errors. In response to your feedback, we have undertaken a thorough review of the manuscript to correct any such errors.We have carefully scrutinized each section for typos, grammatical mistakes, and clarity issues.

2. It will be better if the literature work will be discussed in form of table with their limitations.

Thank you for your valuable suggestion to enhance the literature review section of our manuscript. In response to your comment, we have introduced ‘Table 1’ in the revised manuscript's introduction. This table provides a detailed analysis of reference [15], outlining its approach, the results obtained from the study, and its limitations.

We specifically focused on reference [15] in ‘Table 1’ because it is the only study, to our knowledge, that has utilized the newly published wildfire dataset. This makes it particularly relevant and significant to our research. While we acknowledge the potential benefits of presenting other cited literature in a tabular format, we have opted to limit this detailed tabular discussion to reference [15] for a few reasons.

Firstly, providing a tabular summary for all the literature cited would be extensive and could detract from the main focus of our manuscript. Our aim is to maintain a clear and concise presentation of relevant literature while ensuring that the primary contributions of our own research are highlighted effectively. Additionally, the scope of our manuscript is specifically tailored to address certain aspects of wildfire detection, and we believe that an in-depth table comparison of all previous literature might dilute this focus.

We hope that the inclusion of ‘Table 1’ will provide a valuable and succinct overview of the most directly relevant prior work, and we believe this approach strikes a balance between a comprehensive literature review and the specific aims of our study.

Thank you again for your feedback.

3. A proper block diagram for the proposed method is needed with explanation for better understanding.

In response to your valuable suggestion regarding the inclusion of a visual description to enhance understanding, we have now introduced three illustrative diagrams within ‘Section 2.3’ of the revised manuscript. ‘Figure 4’ presents a detailed view of the architecture customization of EfficientNetB0 for Dual-Task Learning. Following this, ‘Figure 5’ depicts an overview of the final primary model after fine-tuning. Additionally, to further clarify the methodology and the structure of our approach, we have included ‘Figure 3’, a comprehensive method flowchart.

This newly added ‘Figure 3’ methodically illustrates the flow and structure of the dual-task learning approach adopted in our study. It visually represents the initialization of the dual-branch architecture, the distinct processing pathways of the primary and auxiliary branches, the integration of shared layers, and the subsequent steps in the model's training and fine-tuning processes. The flowchart is designed to provide a clear and concise overview of how our model manages dual-task learning and the flow of information through shared and task-specific layers, thereby offering an enhanced comprehension of the complex methodology employed.

We believe that these additions, particularly the method flowchart in ‘Figure 3’, will greatly aid in elucidating the intricacies of our model’s structure and its operational flow to the readers, aligning with the goals of our research and addressing the concerns raised in your commentary.

4.The title suggest about the forest fire detection, but the figure 1 represent the non forest fire situation. How this model is robust for finding the fire in different situation (fire in the building, offices, fire in industries, or fire during a celebration of X-mas).

Thank you for your insightful comments and for the opportunity to clarify the focus and methodology of our study. As you correctly noted, the title of our manuscript emphasizes forest fire detection. This specific focus is consistently developed throughout Sections 2, 3, and 4 of the manuscript, where we aim to enhance fire detection capabilities specifically in forest contexts.

Regarding Figure 1 and other images in our dataset that do not depict forest fires, I would like to clarify their role in our study. Our approach utilizes a dual-dataset strategy: the primary dataset comprises images of forest contexts, and the auxiliary dataset includes non-forest images. The inclusion of the auxiliary dataset is not intended to broaden the scope of our study to other types of fires, such as those in buildings, offices, or industrial settings. Rather, its purpose is solely to augment the robustness of our model in accurately identifying forest fires. By exposing the model to a diverse range of non-forest fire scenarios, we aim to reduce the likelihood of false positives and enhance the model's precision in forest fire contexts.

This methodology allows us to rigorously test and validate our model's effectiveness in detecting forest fires, which is the primary focus of our research. 

We hope this clarification aligns with the intended scope of our study and addresses your concerns regarding the use of non-forest images in our dataset. We believe that this focused approach will contribute significantly to advancements in forest fire detection technologies.

5. Authors used the hyperparameters (learning rate, optimizer, and epoch). How this values are optimally decided to obtain better accuracy.

Thank you for your inquiry regarding the selection of hyperparameters in our study. I would like to direct your attention to Section 2.4.1, "Hyperparameters Optimization," where we detail our systematic approach to determining the optimal hyperparameters for our models.

In this section, we describe a validation strategy using a held-out dataset to assess performance on unseen data. We experimented with various learning rates (10^-2, 10^-3, 10^-4, 10^-5), ultimately selecting the most effective one based on its performance on the test dataset.

We set the batch size at 32 to optimize computational efficiency and gradient stability. The training epochs were varied in 5-epoch increments to determine the optimal number for each configuration. 

In this study, all experiments employed the Adaptive Moment Estimation (ADAM) optimizer as the default choice. Due to limitations in resources, we did not explore alternative optimizers, focusing instead on maximizing the efficiency and effectiveness within the constraints of our available resources.

Each of these steps in our hyperparameter optimization process was designed to ensure that our model achieves the highest possible accuracy in detecting forest fires. The chosen hyperparameters for each approach, including learning rates, number of epochs, batch sizes, and the number of frozen layers, are summarized in the corresponding tables in the results section (Section 3).

We hope this explanation provides clarity on our methodology for selecting hyperparameters for our approach. 

6. In this work, transfer learning concept is used by considering EfficientNetB0 model.  What makes the author to use EfficientNetB0 though there are many pretrained networks are available. The working of the pretrained model and their comparisons are discussed in below papers. 

 https://doi.org/10.1016/j.bea.2022.100069

https://doi.org/10.1002/ima.22761 

Thank you for your valuable feedback on the selection of the base model for our study. We appreciate your interest in the rationale behind choosing EfficientNetB0 and have accordingly revised ‘Section 2.2, Base Model Selection: EfficientNetB0’. In this updated section, we delve deeper into the specific attributes of EfficientNetB0 that make it an optimal choice for our application in forest fire detection, particularly its streamlined architecture without complex interconnections between blocks and its balance of computational efficiency and accuracy.

It's important to note that the primary aim of our study is to showcase the benefits of our novel training approach, rather than to compare different architectural models. Therefore, the selection of EfficientNetB0 is strategic, focusing on its suitability for real-world deployment and its compatibility with our training methodology. This model's lightweight and scalable design, devoid of intricate block interconnections, facilitates the implementation of our training approach, emphasizing efficiency and practicality in operational settings. Furthermore, we have included a reference you have proposed to the comparative studies that illustrate EfficientNetB0's fair performance in terms of accuracy. This contextualizes our choice within the broader landscape of available architectures and underscores its relevance to the objectives of our research.

7. It will be better to understand the effectiveness of the work if a comparison table will be presented with respect to the related works.

We are grateful for your insightful commentary. As suggested, we have incorporated 'Table 22' to the discussion section of the revised version of the manuscript, where we compare our results with those of the only other study, reference [15], which has utilized the newly published wildfire dataset so far. This table, alongside the preceding explanatory paragraph, contrasts the results achieved in the present study using the EfficientNetB0 model with the MobileNetV3 results from study [15]. This comparison highlights the benefits and improvements our method offers over the existing model, with particular emphasis on the unique application of the wildfire test dataset. 

We believe this addition substantially strengthens our manuscript by providing a clear benchmark against the current literature.

8.In this manuscript, only five matrices have been computed for the unbalanced dataset. whereas, MCC is a better and can be accessed from the below mentioned paper.

https://doi.org/10.1016/j.bea.2023.100089 

Thank you for your valuable feedback. We have taken careful consideration of your comments regarding the number of performance metrics and in response to them, we have expanded our suite of performance metrics to include specificity, False Negative Rate (FNR), and Matthews Correlation Coefficient (MCC). These metrics were carefully selected to provide a comprehensive evaluation of the models' performance, addressing different aspects of classification beyond the traditional measures.

As for the previous ones, each metric was computed across five separate runs to ensure robustness, with the average values and standard deviations reported to capture the variability and reliability of the results. Furthermore, to compare the efficacy of the different training approaches, we employed bootstrapping to calculate the 95% confidence intervals for the differences in performance metrics, providing a statistically rigorous assessment.

The inclusion of these additional metrics has enriched the discussion around the operational strengths of the proposed approach and allowed for a more nuanced understanding of its performance. The integration and analysis of these results have been detailed in the revised '2. Materials and Methods' and '3. Results' sections of the manuscript, ensuring that the study reflects a thorough and multifaceted evaluation.

9. The database need to discussed with detail.

In response to your commentary on the need for detailed discussion of the database, we have expanded our previous descriptions within sections '2.1.1. Auxiliary Dataset' and '2.1.2. Primary Dataset'. These sections now include comprehensive statistics on image resolutions, sources of image procurement emphasizing diversity, geographical and environmental ranges of the images, and a discussion on the consideration of confounding elements specific to wildfire scenarios. We believe this additional information significantly enriches the context and robustness of our dataset descriptions.

Comments on the Quality of English Language: Moderate editing of English language required. The manuscript needs to be checked properly.

 Thank you for highlighting the importance of ensuring the manuscript is free from typographical and grammatical errors. In response to your feedback, we have undertaken a thorough review of the manuscript to correct any such errors.We have carefully scrutinized each section for typos, grammatical mistakes, and clarity issues.

Reviewer 4 Report

Comments and Suggestions for Authors

1.The work is presented well. 

2. The results carried out by both datasets can be compared with the existing work of the literatures discussed earlier. The benefits or improvement over those existing methods can be discussed.
3. Detecting fire in the night can be included in the work

Author Response

Response to Reviewer 4 Comments

We would like to thank you for the encouraging and constructive comments and suggestions provided for the manuscript titled "Dual-Dataset Deep Learning for Improved Forest Fire Detection: A Novel Hierarchical Domain-Adaptive Learning Approach". Your insights have greatly helped in enhancing the quality and clarity of our work.

In the revised version of the manuscript, all the modified, added, and revised relevant parts have been highlighted with a light blue background to facilitate your review. This way, you can easily identify the changes and assess how they align with your comments.

1.The work is presented well. 

2. The results carried out by both datasets can be compared with the existing work of the literatures discussed earlier. The benefits or improvement over those existing methods can be discussed.

We are grateful for your insightful commentary. As suggested, we have incorporated 'Table 22' to the discussion section of the revised version of the manuscript, where we compare our results with those of the only other study, reference [15], which has utilized the newly published wildfire dataset so far. This table, alongside the preceding explanatory paragraph, contrasts the results achieved in the present study using the EfficientNetB0 model with the MobileNetV3 results from study [15]. This comparison highlights the benefits and improvements our method offers over the existing model, with particular emphasis on the unique application of the wildfire test dataset. 

We believe this addition substantially strengthens our manuscript by providing a clear benchmark against the current literature.

3. Detecting fire in the night can be included in the work

We appreciate your suggestion to include nocturnal fire detection in our study. We acknowledge the importance of detecting fires at night and have analyzed the performance of the proposed model on the 17 night images contained within the wildfire test dataset. Our error analysis reveals that the proposed approach outperforms the training from scratch model, reducing false predictions from 3 to just 1 on these night images. This underscores the potential of the Hierarchical Domain-Adaptive Learning model under nocturnal conditions. Nevertheless, we recognize the necessity for more extensive research focused on nighttime fire detection and consider this an important direction for future work. 

In the manuscript, commentary on night image fire detection has been incorporated into Section '3.4. Comparative Analysis of the Baseline with the Study’s Proposed Approach,' highlighting the model's improved accuracy in such conditions. Additionally, we acknowledge the need for further investigation into night-time fire detection as a part of future research directions outlined in Section '4.1 Limitations and Future Work.'