Object-Oriented Convolutional Neural Network for Forest Stand Classification Based on Multi-Source Data Collaboration

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article presents an Object-oriented Convolutional Neural Network (OCNN) for forest stand classification using multisource data, integrating red edge bands and canopy height information to enhance classification accuracy. The study incorporates Sentinel-2 and RapidEye data, along with LiDAR-derived canopy height, to improve spectral information and feature extraction. The OCNN method is compared with traditional object-oriented classification methods, showing superior performance with an accuracy of 85.68%. The approach addresses the underutilization of red edge spectral information and the limitations of manual feature selection in traditional methods, demonstrating the potential of integrating object-oriented techniques with deep learning for forest classification.

 

Strengths:

·         Innovative integration of red edge bands and LiDAR-derived canopy height information for improved classification accuracy.

·         Comprehensive comparison between traditional methods and OCNN, showcasing the latter's superior performance.

·         Effective utilization of multisource remote sensing data, enhancing spectral information for classification.

Limitation:

·         The study's generalizability might be limited to similar forest environments and data quality.

·         The complexity of the OCNN model could pose challenges in terms of computational resources and implementation.

Missing Aspects:

·         A broader validation across different forest types and conditions could strengthen the findings.

 

·         Further exploration into the scalability of the OCNN model and its application to other remote sensing tasks.

 

Here is the Review and Analysis of the Article Section by Section

1.       Introduction

 

Review:

The introduction effectively sets the stage for the research by highlighting the importance of accurate forest stand classification and its relevance to forest management and ecosystem sustainability. It acknowledges the advancements in remote sensing technologies and the need for improved classification methods. The introduction also identifies the gaps in current research, particularly the underutilization of red edge spectral information and the limitations of traditional object-oriented classification methods, thus justifying the need for the proposed Object-oriented Convolutional Neural Network (OCNN) approach.

 

Strengths:

Clearly outlines the significance of forest stand classification and the potential of remote sensing in this domain. Identifies specific gaps in the current research landscape, setting a solid foundation for the novelty of the study.

 

Weaknesses:

The introduction could benefit from a more detailed discussion on the challenges and limitations of existing classification methods, providing a stronger argument for the proposed approach.

 

Missing Aspects:

A brief review of previous attempts to integrate deep learning with remote sensing for forest classification could enrich the context. The introduction could elaborate on how the proposed method intends to overcome the identified gaps.

 

2.       Study Area and Data

Review:

This section provides a detailed description of the study area, including its geographical location, climatic conditions, and forest composition. It also comprehensively outlines the remote sensing data used, including multispectral images from RapidEye and Sentinel-2A and airborne LiDAR data, along with their preprocessing steps. This thorough detailing ensures clarity and reproducibility of the research.

 

Strengths:

Detailed and clear description of the study area and the remote sensing data used, enhancing the clarity and reproducibility of the research. Comprehensive data preprocessing steps are outlined, ensuring data quality and reliability.

 

Weaknesses:

The section could benefit from a discussion on why these specific data sources were chosen over others, highlighting their advantages in the context of this study.

Missing Aspects:

A discussion on the potential limitations or challenges associated with the chosen data sources or the study area's specific characteristics could provide a more balanced view.

 

3.       Methods

Review:

The methods section is meticulously detailed, covering image fusion, segmentation, and the structure of the CNN used in the study. It explains the rationale behind the chosen techniques and their expected contribution to improving classification accuracy. The experimental design is clearly laid out, including the selection of training samples and CNN sample set construction.

 

Strengths:

Comprehensive detailing of the methods used, from image fusion to CNN architecture, ensuring transparency and reproducibility. Logical flow and clear explanation of the rationale behind each methodological choice.

 

Weaknesses:

While the section is well-detailed, it could be overwhelming for readers unfamiliar with some of the technical aspects. A simplified summary or conceptual diagram could aid in understanding.

 

Missing Aspects:

The potential limitations or challenges of the chosen methods and how they were addressed or could impact the results are not discussed.

 

 4.       Results

Review:

The results section presents a clear and detailed analysis of the findings from the image fusion, segmentation, and classification processes. It provides a comparative analysis of the OCNN method against traditional classification methods, supported by quantitative data and visual illustrations.

 

Strengths:

Detailed presentation of results with supportive quantitative data and visual illustrations, facilitating a clear understanding of the findings. Effective comparison between the proposed OCNN method and traditional methods, highlighting the improvements brought by OCNN.

 

Weaknesses:

The section is heavily loaded with technical details, which might be challenging for readers to follow. Summarizing key findings in a less technical manner could enhance readability.

 

Missing Aspects:

A deeper analysis of why certain methods performed better than others, beyond the quantitative results, could provide valuable insights into the strengths and limitations of the OCNN approach.

 

5.       Discussion and Conclusion

Review:

The discussion synthesizes the results, offering insights into the reasons behind OCNN's high classification accuracy and comparing it with traditional methods. It acknowledges the advantages of integrating CNN with object-oriented methods and highlights the significance of data augmentation. The conclusion summarizes the key findings and contributions of the study, emphasizing the effectiveness of the proposed method in forest stand classification.

 

Strengths:

Provides a comprehensive synthesis of the results, offering plausible explanations for the findings. Effectively ties the study's contributions to broader research and practical implications in forest classification.

 

Weaknesses:

 The discussion could benefit from a more critical examination of the study's limitations and the potential implications for future research.

 

Missing Aspects:

 

Suggestions for future research directions, particularly in addressing the identified limitations or extending the application of the OCNN method to other areas, could make the conclusion more impactful.

To enhance the scientific value and quality of the article, several major scientific issues need to be addressed:

 

·         Validation and Generalization: The study demonstrates the effectiveness of the OCNN approach in the specific study area but lacks broader validation across diverse forest types and geographical regions. To improve the work scientifically, it would be beneficial to test the method's applicability and robustness across different ecosystems and environmental conditions, ensuring the model's generalization capability.

·         Comparison with State-of-the-Art Methods: While the paper compares OCNN with traditional object-oriented methods, it does not extensively compare it with other state-of-the-art deep learning approaches used in similar contexts. Including such comparisons could significantly strengthen the scientific basis of the paper, highlighting OCNN's advantages and potential limitations.

 

·         Feature Extraction and Selection Process: The discussion section highlights the advantage of CNNs in automatically extracting and learning features from the data. However, a more detailed exploration of which features are most salient for the classification task and how these features contribute to the classification accuracy would provide deeper insights into the model's workings and potential areas for improvement.

·         Impact of Misclassifications: The discussion also points out that certain categories, like 'Other' and pine trees, have higher numbers of discriminative sample points than actual sample points, indicating misclassifications. A more thorough investigation into the causes of these misclassifications and potential strategies to mitigate them would enhance the scientific rigor of the paper. This could involve exploring the model's sensitivity to the input data quality, the representation of minority classes in the training data, or the spatial resolution of the input images.

·         Methodological Limitations: The paper should address the limitations of the OCNN method more critically. This includes discussing the challenges related to the model's complexity, computational demands, and any assumptions made during the development of the model. Understanding these limitations is crucial for future applications and improvements of the method.

·         Future Research Directions: Finally, the paper could benefit from a clearer outline of future research directions. This includes potential methodological improvements, applications to other remote sensing tasks, or integration with additional data sources. Highlighting these areas would not only provide a path forward for subsequent research but also position the current work within the larger context of remote sensing and forest management research.

Comments on Figures & Tables.

·         Figure 1: Overview of Gaofeng forest farm. While this figure provides a good geographic context, it could be improved by adding specific annotations or markers to highlight areas of interest mentioned in the text, such as different forest stand types or locations where significant data were collected.

·         Table 1: Band characteristics of Sentinel-2A and RapidEye. This table is essential for understanding the spectral resources utilized in the study. Including a brief explanation in the caption about how these specific bands contribute to forest stand classification, particularly the red edge and NIR bands, could make it more informative.

·         Figure 2 (Line 489): Sample-plot layout. It would be beneficial to provide more details in the caption regarding the scale and how these plots were distributed across the study area to ensure representativeness of different forest types.

·         Figure 3: Overall technical route. This figure is crucial for understanding the workflow of the study. Ensuring that each step is clearly labeled and perhaps providing a brief description of each step in the caption could aid readers in following the methodology more easily.

·         Figure 4: The flowchart of the MV fusion method. The figure is key for understanding the image fusion process. Enhancing the figure with clearer, more detailed annotations for each step and explaining the significance of the MV method in the context of this research could add value.

·         Figure 5: The flowchart of the ODSD segmentation method. Similar to Figure 4, this figure would benefit from more detailed annotations and a caption that explains why the ODSD method was chosen over other segmentation methods.

·         Figure 6: CNN basic network structure. A more detailed caption explaining the function of each layer in the context of forest stand classification and how this architecture is suited for handling multisource data could make this figure more informative.

·         Figure 7: Comparison of MV and GS fusion results. Providing a discussion in the caption about the implications of the differences observed in the fusion results on the final classification accuracy would be helpful. It might also be beneficial to discuss any limitations observed in the fusion process.

·         Table 4: Comparison of Sentinel-2 RE2 band fusion quality evaluation metrics. Expanding the caption to explain how each metric contributes to evaluating the fusion quality and why these particular metrics were chosen could provide readers with a better understanding of the evaluation process.

 

Addressing these issues would not only strengthen the current work's scientific foundation but also provide valuable insights for future research in remote sensing and forest stand classification

Comments on the Quality of English Language

The paper is generally well-written with clear and coherent language. However, there are occasional grammatical errors, awkward phrasings, and inconsistencies in terminology that could be refined to improve readability and professionalism. For instance, the use of technical terms and the consistency in the formatting of names, acronyms, and technical jargon could be standardized throughout the document.

Some specific suggestions for improvement include:

·         Ensuring consistent use of either American or British English spelling conventions throughout the document.

·         Reviewing the use of articles ("a", "an", "the") where they are missing or used unnecessarily.

·         Checking for subject-verb agreement in complex sentences to ensure grammatical correctness.

·         Simplifying complex sentences to enhance clarity and understanding for readers who may not be native English speakers or who may be unfamiliar with the technical jargon of the field.

·         Ensuring consistent punctuation, particularly in the use of commas and semicolons in lists and complex sentences.

 

·         Clarifying ambiguous pronouns to ensure it is clear what or whom they are referring to.

Author Response

Dear Reviewer：

Thank you very much for your careful review and constructive suggestions with regard to our manuscript "Object-oriented Convolutional Neural Network for Forest Stand Classification Based on Multisource Data Collaboration"(ID：forests-2873022). Those comments are all valuable and very helpful for revising and improving our paper, as well as the importance guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in blue in the paper. The main corrections in the paper and the responds to the reviewer's comments are as following:

 

Missing Aspects:

1. A broader validation across different forest types and conditions could strengthen the findings.

Thank you very much for your suggestion. The previous manuscript did not provide a detailed introduction to the research area. Considering your suggestion, we have strengthened the explanation in this section to explain this issue.In fact, this study has been extensively validated, due to the research area covers a rich variety of forest types, and is one of the most complex forest stands situations in China. The research focus of this research is to explore the use of red edge band to solve the problem of further improving the classification accuracy in the case of complex stands. The universality of this method and the application in the next step are the key problems to be solved in the future. Your opinion has great guiding significance to my future research direction. As shown in line129-135:

“It includes artificial forests, mainly composed of Eucalyptus and secondary forests, mainly mixed with coniferous and broad-leaved trees. The research area covers a rich variety of forest types, mainly Illicium verum, Pinus massoniana, Pinus elliottii, Eucalyptus grandis, Cunninghamia lanceolata and broad-leaved tree species. It is a typical subtropical climate forest. Moreover, in the Chinese region, its forest types, tree species, and stands are one of the most complex, which are very representative.”

2. Further exploration into the scalability of the OCNN model and its application to other remote sensing tasks.

Thank you very much for your suggestion. The scalability of the OCNN model is exactly what we need to focus on in the future. This issue has already been statement in the discussion. As shown in line552-561:

“While the creation of a multi-source remote sensing dataset for forest stands has validated the efficacy of the ResNet_18 model, our study has some limitations. Firstly, Some forest type samples have relatively few samples, which will affect the feature learning process for that category. Secondly, the model’s generalizability under varying geographical, climatic, or ecological conditions still requires further verification. Thirdly, the black box nature of CNN makes it impossible to know the key features for forest classification. Future research should focus on increasing the size and diversity of the training data set to improve the accuracy and robustness of the machine learning models. Besides, there is a need to strengthen the explainability of automatic quantitative visual representation, and explore the feature learning process of deep neural networks through automation.”

Here is the Review and Analysis of the Article Section by Section

1. Introduction

Weaknesses:

The introduction could benefit from a more detailed discussion on the challenges and limitations of existing classification methods, providing a stronger argument for the proposed approach.

Thank you very much for your suggestion. A detailed explanation of the limitations of existing methods does provide strong argument for the proposed method in this article. We have added this section of explanation as per your suggestion. As shown in line "48-52"：

“While many studies have used one of the approaches described above, there is a growing literature highlighting the benefits of LiDAR data, and ignoring targeted research on spectral data, particularly there is no specific research on the impact of red edge band fusion on forest classification accuracy.”

Missing Aspects:

A brief review of previous attempts to integrate deep learning with remote sensing for forest classification could enrich the context. The introduction could elaborate on how the proposed method intends to overcome the identified gaps.

 Your opinion is very constructive. A review of the application of deep learning in forest classification does can enrich the context provide strong arguments for the proposed method in this article. This point, which was overlooked in previous manuscripts, has been added to the article. As shown in line106-110:

“ In the realm of CNN for tree species classification, the majority of research on tree species classification centers on identifying single tree species or individual species within forest stands. There is a relative scarcity of studies addressing the identification of complex forest structures and mixed forests with multiple tree species[43-45].”

 

2. Study Area and Data

Weaknesses:

The section could benefit from a discussion on why these specific data sources were chosen over others, highlighting their advantages in the context of this study.

 Your opinion is very constructive. We have strengthened the explanation of the reasons for selecting data sources in the manuscript, including their advantages and applicability in this study. As shown in line60-68:

“To date, only several satellites have been launched (RapidEye, WordView-2, Sentinel-2, etc.) that are equipped sensors with red-edge bands[19]. Among them, RapidEye is the world's first multispectral commercial satellite that provides a 710nm red edge band which is conducive to monitoring vegetation and suitable for agricultural, forestry, and environmental monitoring. RapidEye can provide 5-meter spatial resolution image data, which is sufficient for the classification of forest species in this study. Sentinel-2 is the only data with multiple bands within the red edge range, which is very effective for monitoring vegetation health information and easy to obtain[20]. ”

Missing Aspects:

A discussion on the potential limitations or challenges associated with the chosen data sources or the study area's specific characteristics could provide a more balanced view.

Your opinion is very constructive. The challenge associated with the study area lies in the complexity of forest stand types, which were not described in detail before and have been supplemented in the manuscript. As shown in line129-135:

“It includes artificial forests, mainly composed of Eucalyptus and secondary forests, mainly mixed with coniferous and broad-leaved trees. The research area covers a rich variety of forest types, mainly Illicium verum, Pinus massoniana, Pinus elliottii, Eucalyptus grandis, Cunninghamia lanceolata and broad-leaved tree species. It is a typical subtropical climate forest. Moreover, in the Chinese region, its forest types, tree species, and stands are one of the most complex, which are very representative.”

3. Methods

Weaknesses:

While the section is well-detailed, it could be overwhelming for readers unfamiliar with some of the technical aspects. A simplified summary or conceptual diagram could aid in understanding.

Your opinion is very constructive. Technical details have been added to each conceptual diagram and explanatory text has been added to the caption to help readers understand. Please refer to the specific answer for chart modification in the following text.

Missing Aspects:

The potential limitations or challenges of the chosen methods and how they were addressed or could impact the results are not discussed.

Your opinion is very constructive. The previous manuscript overlooked the description of the challenges and solutions faced by the selected method. These descriptions has been supplemented in the manuscript. As shown in line268-274, line332-339, and line 360-365:

“Multi-source remote sensing data fusion presents a certain level of complexity and difficulty. Data generated by different sensors differ in format, resolution, data quality, and spatial reference systems. Multi-source data require preprocessing, registration, and fusion. The demands on algorithms are high, making data fusion extremely challenging. Although the MV fusion method used in this study has solved these problems to some extent, it still involves uncertainty. By strengthening the control and evaluation of data quality to enhance data reliability.”

“Owing to the complexity and diversity of high-resolution images, image segmentation is challenging due to the need for multi-scale expression and multi-source feature representation. The ODSD segmentation algorithm, employed in this study, has been refined to overcome the challenges posed by existing seed point selection methods—namely, the excessive number of seed points, suboptimal selection efficiency, and insufficient selection in target finer details. However, the outcomes still exhibit inherent jagged edges and objects that are either under- or over-segmented. By continuously adjusting parameters, one can opt for the most optimal segmentation result.”

“The challenge faced by the CNN employed in this study is to enhance the model's generalization capabilities. Not all models can be applied to a completely new dataset. This can be achieved through more diverse training data, regularization techniques, and trying multiple networks to find the best model suitable for forest classification. The limitation of CNN is that it is often seen as a "black box" and it is difficult to explain its internal working principle.”

4. Results

 Weaknesses:

The section is heavily loaded with technical details, which might be challenging for readers to follow. Summarizing key findings in a less technical manner could enhance readability.

Your opinion is very constructive. The language expression has been adjusted in the text to make the result description more acceptable to readers.

Missing Aspects:

A deeper analysis of why certain methods performed better than others, beyond the quantitative results, could provide valuable insights into the strengths and limitations of the OCNN approach.

Your opinion is very constructive. The reasons for ResNet_18's better performance have been explained in the “Discussion”. As shown in line268-274:

“There are two reasons why OCNN has high classification accuracy. One reason is the combination of CNN and object-oriented methods utilizes the peripheral information of forest objects and bounding rectangles, and extracts deep abstract information through a large number of convolutional layers, while traditional classification only uses a few manually selected statistical information. Another reason is CNN increases a large number of learning samples through data augmentation, while traditional classification only selects a few hundred samples. This result demonstrates the effectiveness of OCNN in forest classification and can to some extent improve the accuracy of forest classification. ResNet_18 achieves high classification accuracy among the three classic CNN networks due to its inherent advantages: its unique residual unit addresses the issue of gradient vanishing as the network depth increases, significantly simplifying the model optimization process, cause the magnitude of the training error to be inversely proportional to the number of network layers, thereby achieving satisfactory testing performance.”

 

5. Discussion and Conclusion

Weaknesses:

 The discussion could benefit from a more critical examination of the study's limitations and the potential implications for future research.

Your opinion is very constructive. More critical examination of the study's limitations has been incorporated into the discussion. As shown in line 552-561:

“While the creation of a multi-source remote sensing dataset for forest stands has validated the efficacy of the ResNet_18 model, our study has some limitations. Firstly, Some forest type samples have relatively few samples, which will affect the feature learning process for that category. Secondly, the model’s generalizability under varying geographical, climatic, or ecological conditions still requires further verification. Thirdly, its learning process can be seen as a "black box". Future research should focus on increasing the size and diversity of the training data set to improve the accuracy and robustness of the machine learning models. Besides, there is a need to strengthen the explainability of automatic quantitative visual representation, and explore the feature learning process of deep neural networks through automation. ”

Missing Aspects:

 Suggestions for future research directions, particularly in addressing the identified limitations or extending the application of the OCNN method to other areas, could make the conclusion more impactful.

Your opinion is very constructive. Suggestions for future research directions has been incorporated into the discussion. As shown in line 557-561:

“ Future research should focus on increasing the size and diversity of the training data set to improve the accuracy and robustness of the machine learning models. Besides, there is a need to strengthen the explainability of automatic quantitative visual representation, and explore the feature learning process of deep neural networks through automation. ”

To enhance the scientific value and quality of the article, several major scientific issues need to be addressed:

1. Validation and Generalization: The study demonstrates the effectiveness of the OCNN approach in the specific study area but lacks broader validation across diverse forest types and geographical regions. To improve the work scientifically, it would be beneficial to test the method's applicability and robustness across different ecosystems and environmental conditions, ensuring the model's generalization capability.

Your opinion is very constructive. The previous manuscript did not provide a detailed introduction to the research area. Considering your suggestion, we have strengthened the explanation in this section to explain this issue.In fact, this study has been extensively validated, due to the research area covers a rich variety of forest types, and is one of the most complex forest stands situations in China. The research focus of this research is to explore the use of red edge band to solve the problem of further improving the classification accuracy in the case of complex stands. The universality of this method and the application in the next step are the key problems to be solved in the future. Your opinion has great guiding significance to my future research direction. As shown in line 129-135, line 557-561:

“It includes artificial forests, mainly composed of Eucalyptus and secondary forests, mainly mixed with coniferous and broad-leaved trees. The research area covers a rich variety of forest types, mainly Illicium verum, Pinus massoniana, Pinus elliottii, Eucalyptus grandis, Cunninghamia lanceolata and broad-leaved tree species. It is a typical subtropical climate forest. Moreover, in the Chinese region, its forest types, tree species, and stands are one of the most complex, which are very representative.”

“ Future research should focus on increasing the size and diversity of the training data set to improve the accuracy and robustness of the machine learning models. Besides, there is a need to strengthen the explainability of automatic quantitative visual representation, and explore the feature learning process of deep neural networks through automation. ”

2. Comparison with State-of-the-Art Methods: While the paper compares OCNN with traditional object-oriented methods, it does not extensively compare it with other state-of-the-art deep learning approaches used in similar contexts. Including such comparisons could significantly strengthen the scientific basis of the paper, highlighting OCNN's advantages and potential limitations.

Your opinion is very constructive. This study did not use the State-of-the-Art deep learning approaches, considering that the selected model should be more classic, portable, and robust, and currently has good application effects in other directions. Your suggestions are very helpful for our future research on classification methods.

3. Feature Extraction and Selection Process: The discussion section highlights the advantage of CNNs in automatically extracting and learning features from the data. However, a more detailed exploration of which features are most salient for the classification task and how these features contribute to the classification accuracy would provide deeper insights into the model's workings and potential areas for improvement.

Your opinion is very constructive. The advantage of the CNN used in this study is its powerful feature learning ability. However, its learning process can be seen as a "black box". This study did not focus on model improvement, but your suggestions for exploring classification features to improve the model are very helpful to us and an important exploration area. In the future, efforts will be made to enhance the interpretability of automated quantitative visual representations and explore the feature learning process of deep neural networks through automation. It has been explained in the outlook for future research. As shown in line 559-561:

“Besides, there is a need to strengthen the explainability of automatic quantitative visual representation, and explore the feature learning process of deep neural networks through automation. ”

4. Impact of Misclassifications: The discussion also points out that certain categories, like 'Other' and pine trees, have higher numbers of discriminative sample points than actual sample points, indicating misclassifications. A more thorough investigation into the causes of these misclassifications and potential strategies to mitigate them would enhance the scientific rigor of the paper. This could involve exploring the model's sensitivity to the input data quality, the representation of minority classes in the training data, or the spatial resolution of the input images.

Your opinion is very important to us. As you said, the quality of training samples in CNN is very important, which directly affects classification accuracy. The article explores the possible reasons for classification errors, but it is still not comprehensive. Further investigation of the causes of errors is the key to improving accuracy. An error analysis has been added to the manuscript. As shown in line 549-551:

“Although we have tried to select pure forests of various categories when selecting training samples, it is inevitable that a small number of non pure forests will be selected, which may also affect classification accuracy.”

5. Methodological Limitations: The paper should address the limitations of the OCNN method more critically. This includes discussing the challenges related to the model's complexity, computational demands, and any assumptions made during the development of the model. Understanding these limitations is crucial for future applications and improvements of the method.

Your opinion is very important to us. Understanding the limitations of OCNN is crucial for further improving methods and future applications. The limitations of the OCNN method have been supplemented in the discussion. As shown in line 553-557:

“Firstly, Some forest type samples have relatively few samples, which will affect the feature learning process for that category. Secondly, the model’s generalizability under varying geographical, climatic, or ecological conditions still requires further verification. Thirdly, its learning process can be seen as a "black box". ”

6. Future Research Directions: Finally, the paper could benefit from a clearer outline of future research directions. This includes potential methodological improvements, applications to other remote sensing tasks, or integration with additional data sources. Highlighting these areas would not only provide a path forward for subsequent research but also position the current work within the larger context of remote sensing and forest management research.

Your opinion is very constructive. More critical examination of the study's limitations has been incorporated into the discussion. As shown in line 557-561:

“ Future research should focus on increasing the size and diversity of the training data set to improve the accuracy and robustness of the machine learning models. Besides, there is a need to strengthen the explainability of automatic quantitative visual representation, and explore the feature learning process of deep neural networks through automation. ”

Comments on Figures & Tables.

Figure 1: Overview of Gaofeng forest farm. While this figure provides a good geographic context, it could be improved by adding specific annotations or markers to highlight areas of interest mentioned in the text, such as different forest stand types or locations where significant data were collected.

Thank you very much for your suggestion. Although the previous image had a good geographic context, it lacked specific areas of interest. We have marked them in the image according to your suggestion.

 

Table 1: Band characteristics of Sentinel-2A and RapidEye. This table is essential for understanding the spectral resources utilized in the study. Including a brief explanation in the caption about how these specific bands contribute to forest stand classification, particularly the red edge and NIR bands, could make it more informative.

Thank you very much for your suggestion. The highlighted red edge bands have been introduced in the title of Table 1. As shown in line 175-179:

“ Table 1. Band characteristics of Sentinel-2A and RapidEye: given that different forest types exhibit diverse reflection properties in the red edge band, this band plays a pivotal role in differentiating tree species. Specifically, bands 5, 6, 7, and 8A from Sentinel-2A, along with band 4 from RapidEye, are identified as the red edge bands. These bands are of primary interest to our research. ”

• Figure 2 (Line 489): Sample-plot layout. It would be beneficial to provide more details in the caption regarding the scale and how these plots were distributed across the study area to ensure representativeness of different forest types.

Thank you very much for your suggestion. The details of the plot settings have been added to the title of Figure 2. As shown in line 204-206:

“Figure 2. Sample-plot layout: each plot measures 30m x 30m, with an internal division into nine 10m x 10m quadrats. These plots are evenly distributed within the study area, and the number of plots for each type is reasonably arranged based on the scale of each tree species.”

Figure 3: Overall technical route. This figure is crucial for understanding the workflow of the study. Ensuring that each step is clearly labeled and perhaps providing a brief description of each step in the caption could aid readers in following the methodology more easily.

Thank you very much for your suggestion. We have adjusted the flowchart of Figure 3 and provided a brief description for each step in the caption. As shown in line 224-230:

 

“Figure 3. Overall technical route. The overall process is divided into five sections, ①Data aquisition: obtain RapidEye, Sentinel-2A, and LiDAR data required for research; ②Data pre-processing: preprocess three types of images separately; ③Information fusion: utilizes RapidEye to fuse the RE2 and RE4 bands of Sentinel-2A, enhancing the resolution of these two bands to produce the following outputs: RapidEye+S2A_RE2, RapidEye+S2A_RE2+S2A_RE4 and RapidEye+S2A_RE2+S2A_RE4+CHM; ④OCNN classification：consists of two steps---image segmentation and image classification; ⑤ Accuracy evaluation and analysis.”

Figure 4: The flowchart of the MV fusion method. The figure is key for understanding the image fusion process. Enhancing the figure with clearer, more detailed annotations for each step and explaining the significance of the MV method in the context of this research could add value.

Thank you very much for your suggestion. The MV fusion flowchart in Figure 4 has been adjusted, and a brief description has been provided for each step in the caption. The reason for choosing MV fusion has been added in the text. As shown in line 243-246 and line 279-282:

“ However, in the absence of panchromatic images, conventional methods cannot be directly applied for fusion. Unlike general fusion, this study is a fusion between different data sources, and commonly used fusion methods are not applicable.”

 

“Figure 4. The flowchart of the MV fusion method. The process is primarily comprised of three steps. ①The high-resolution band, RapidEye, is processed to extract its high-frequency information. ②The low-resolution bands, namely RE2 and RE4, are processed to establish the relationship with the low-frequency information of RapidEye.③Image fusion produces the final integrated result.”

• Figure 5: The flowchart of the ODSD segmentation method. Similar to Figure 4, this figure would benefit from more detailed annotations and a caption that explains why the ODSD method was chosen over other segmentation methods.

Thank you very much for your suggestion. The ODSD segmentation flowchart in Figure 5 has been adjusted and details have been added. As shown in the following figure:

 

• Figure 6: CNN basic network structure. A more detailed caption explaining the function of each layer in the context of forest stand classification and how this architecture is suited for handling multisource data could make this figure more informative.

Thank you very much for your suggestion. A brief description has been provided for CNN basic network structure in the caption. As shown in line 367-375:

“Figure 6. CNN basic network structure. ①Input layer. Input our prepared sample set here. Inputs can be multidimensional. ②Convolutional layer. The primary objective of convolutional operations is to extract distinct features from the input. The initial convolutional layer may capture only basic features like edges, lines, and corners.Subsequent layers within the network progressively extract more intricate features, building upon those found in the preceding layers. ③Pooling layer. Pooling layers are typically inserted between convolutional layers, progressively downsizing the spatial dimensions of the data. This reduces the number of parameters and computational load, thereby mitigating overfitting to some extent. ④Fully connected layer. The fully connected layer integrates feature extraction with the classification and regression stages, transforms multidimensional features into one-dimensional vectors, and applies linear transformations and activation functions to produce the final output. ⑤Output layer. Output classification results.”

• Figure 7: Comparison of MV and GS fusion results. Providing a discussion in the caption about the implications of the differences observed in the fusion results on the final classification accuracy would be helpful. It might also be beneficial to discuss any limitations observed in the fusion process.

Thank you very much for your suggestion. The impact of the two fusion results on subsequent classification has been explained in the caption of Figure 7. As shown in line 449-452:

“After MV fusion, the texture and spatial structure information of the Sentinel-2A forest are clearer, and the stand features are well preserved, ensuring that the extracted object features are relatively complete in subsequent classification. The GS fusion excels in non-forest areas such as roads, water bodies, and buildings, which are not the focus of subsequent classification and offer limited assistance for stand classification.”

 

• Table 4: Comparison of Sentinel-2 RE2 band fusion quality evaluation metrics. Expanding the caption to explain how each metric contributes to evaluating the fusion quality and why these particular metrics were chosen could provide readers with a better understanding of the evaluation process.

Thank you very much for your suggestion. The significance of each evaluation indicator has been explained in the caption of Table 4, indicating the reasons for selecting these indicators. As shown in line 453-458:

“M: The mean represents the brightness of an image, with optimal image quality maintaining a suitable range of mean values, avoiding excessively high or low values. STD: Measures the dispersion of information within an image; a larger standard deviation indicates a greater amount of information. EN: An indicator of the richness of image information, with a higher value indicating greater information content. AG: Reflects the fineness of local details in an image, measuring how distinct the details appear. SF: Measures the overall activity level of an image in the spatial domain, reflecting the rate of grayscale value changes. CEN: the dissimilarity between corresponding positions in the fused and source images.”

Addressing these issues would not only strengthen the current work's scientific foundation but also provide valuable insights for future research in remote sensing and forest stand classification.

 

Comments on the Quality of English Language

The paper is generally well-written with clear and coherent language. However, there are occasional grammatical errors, awkward phrasings, and inconsistencies in terminology that could be refined to improve readability and professionalism. For instance, the use of technical terms and the consistency in the formatting of names, acronyms, and technical jargon could be standardized throughout the document.

Some specific suggestions for improvement include:

• Ensuring consistent use of either American or British English spelling conventions throughout the document.
• Reviewing the use of articles ("a", "an", "the") where they are missing or used unnecessarily.
• Checking for subject-verb agreement in complex sentences to ensure grammatical correctness.
• Simplifying complex sentences to enhance clarity and understanding for readers who may not be native English speakers or who may be unfamiliar with the technical jargon of the field.
• Ensuring consistent punctuation, particularly in the use of commas and semicolons in lists and complex sentences.
• Clarifying ambiguous pronouns to ensure it is clear what or whom they are referring to.

We appreciate your attention to the language quality in our manuscript. We have carefully revised the language through out the manuscript.

 

 

we appreciate reviewers' warm work earnestly, and hope the correction will meet with approval.

 

Yours sincerely,

Xiaoqing Zhao

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

General comments:

A lot of work has been done on data collection (especially in the field) and data classification. However, the study is limited to one year, 2020, and apparently one date. The main problem with the manuscript is that object classification models built from remote sensing data do not perform well on data from other years or locations. In this connection, the question arises why the authors did not test the ready-made model on data obtained from the same satellites in other years. This did not require labour-intensive field studies.

 

Specific comments:

Lines 20-21. Keywords should not repeat terms from the title of the manuscript, e.g. “Object-oriented; convolutional neural network”.

Lines 8-10. The proposal is unclear.

Lines 27, 34, 53, 66, 68, … References needed.

Lines 55-56. “… domestic and foreign optical remote sensing …” This phrase in an international journal is not clear.

Line 121. “images were captured in 2020”. The dates of data received from RapidEye, Sentinel-2A and Lidar should be specified.

Line 143. What was the time frame in which the field research was conducted?

Line 143. Section 2.2.2. “Data collected in the field.” Specify the accuracy of the GPS. Why is the sample plot size chosen with a side length of 30 m? How many RapidEye and Sentinel-2A pixels were per one sample plot, one quadrat and one sample (e.g. Eucalyptus)?

Lines 308-312. The proposal is unclear.

Line 347. Section 4.2. “Image Segmentation Results”. What is the object for further classification and how are its boundaries established?

Author Response

Dear Reviewer：

Thank you very much for your careful review and constructive suggestions with regard to our manuscript "Object-oriented Convolutional Neural Network for Forest Stand Classification Based on Multisource Data Collaboration"(ID：forests-2873022). Those comments are all valuable and very helpful for revising and improving our paper, as well as the importance guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in blue in the paper. The main corrections in the paper and the responds to the reviewer's comments are as following:

 

General comments:

A lot of work has been done on data collection (especially in the field) and data classification. However, the study is limited to one year, 2020, and apparently one date. The main problem with the manuscript is that object classification models built from remote sensing data do not perform well on data from other years or locations. In this connection, the question arises why the authors did not test the ready-made model on data obtained from the same satellites in other years. This did not require labour-intensive field studies.

Your opinion is very constructive. The previous manuscript did not provide a detailed introduction to the research area. Considering your suggestion, we have strengthened the explanation in this section to explain this issue.In fact, this study has been extensively validated, due to the research area covers a rich variety of forest types, and is one of the most complex forest stands situations in China. The research focus of this research is to explore the use of red edge band to solve the problem of further improving the classification accuracy in the case of complex stands. The universality of this method and the application in the next step are the key problems to be solved in the future. Your opinion has great guiding significance to my future research direction.

 

Lines 20-21. Keywords should not repeat terms from the title of the manuscript, e.g. “Object-oriented; convolutional neural network”.

Your opinion is very constructive. The keywords have been modified to “forest remote sensing; RapidEye; Sentinel-2; red-edge band; canopy height ”

Lines 8-10. The proposal is unclear.

Your opinion is very constructive. The research objectives in the abstract have been expressed more clearly. As shown in line 8-16:

“Accurate classification of forest stand is crucial for the protection and management needs. However, forest stand classification remains a great challenge because of the high spectral and texture similarity of different tree species. Although existing studies have used multiple remote sensing data for forest identification, the effects of different spatial resolutions and combining multi-source remote sensing data for automatic complex forest stands identification using deep learning methods still require further exploration. Therefore, this study proposed an Object-oriented Convolutional Neural Network(OCNN) classification method , leveraging data from Sentinel-2, RapidEye, and LiDAR to explore the classification accuracy of using OCNN to identify complex forest stands.”

Lines 27, 34, 53, 66, 68, … References needed.

Thank you for your suggestion. We have carefully reviewed recent literature and included references to relevant studies in the revised manuscript.

Lines 55-56. “… domestic and foreign optical remote sensing …” This phrase in an international journal is not clear.

Thank you for your suggestion。This sentence has been modified, as shown in line 8-16:

“To date, only several satellites have been launched (RapidEye, WordView-2, Sentinel-2, etc.) that are equipped sensors with red-edge bands[19].”

Line 121. “images were captured in 2020”. The dates of data received from RapidEye, Sentinel-2A and Lidar should be specified.

Thank you for your suggestion. The data acquisition time has been supplemented in the text. As shown in line 143-143 and line 157-158:

“The RapidEye image used in the study was obtained on 15 July 2020 and Sentinel-2A image used in the study was obtained on 18 July 2020, the qualities are good with almost no cloud cover. The study area was extracted from the selected satellite imagery for classification. ”

“The flight mission was carried out on 20 July 2020, with clear and cloudless weather and wind speeds less than 3.0 m/s. ”

Line 143. What was the time frame in which the field research was conducted?

Thank you for your suggestion. The time for field investigation has been added in the text. As shown in line 182:

“A ten-day field survey was carried out at the end of July 2020,”

Line 143. Section 2.2.2. “Data collected in the field.” Specify the accuracy of the GPS. Why is the sample plot size chosen with a side length of 30 m? How many RapidEye and Sentinel-2A pixels were per one sample plot, one quadrat and one sample (e.g. Eucalyptus)?

Your opinion is very constructive. A detailed description of GPS accuracy and sample plots has been added to the manuscript. The reason for choosing a side length of 30 meters is because Sentinel-2 pixels are 10 meters, which ensures that each square corresponds exactly to one pixel. According to the relationship between the species area curve, especially in subtropical regions, the number of tree species in a sample plot of 30 meters by 30 meters tends to approach saturation. And the forestry survey standard is also 30 meters, so it is set to 30 meters * 30 meters, and in this case, it can include 9 sample plots, which can increase the sample size. The specific description has been added in the manuscript, As shown in line 182-194:

“the collected of field point data is carried out through sample-plot survey, using GPS navigation, which positioning accuracy is around 10 meters, compass, and measuring rope to determine the center of the sample-plot. Drawing upon the trends depicted in the species area curve, particularly in subtropical regions where a 30m x 30m sample plot typically approaches saturation, and considering the forestry survey standard, this study has opted for a plot edge length of 30m x 30m. The sample-plot is a square plot with a side length of 30m, and finally divided into 9 quadrats with an area of 10-meter-by-10-meter shown in Figure 2. Each plot contains 6 x 6 RapidEye pixels and 3 x 3 Sentinel-2 pixels. If the forest stands within 30 meters around the center belong to the same type, then the center serves as the plot's core. Otherwise, move the center to the appropriate position. If the plot cannot be reduced to a single tree type no matter how it is shifted, it may contain two types but not both as forest and non-forest land.”

 

Lines 308-312. The proposal is unclear.

Your opinion is very constructive. The objectives in the article have been clearly defined.

 

Line 347. Section 4.2. “Image Segmentation Results”. What is the object for further classification and how are its boundaries established?

Your opinion is very constructive. The objects for further classification include four types of data sources: RapidEye, RapidEye+RE2, RapidEye+RE2+RE4, RapidEye+RE2+RE4+CHM, which are classified into six types. However, the focus of this study is not on how to obtain more accurate segmentation results. Therefore, to control variables, the boundary determination method for each classification object is based on RapidEye's image segmentation results.

 

we appreciate reviewers' warm work earnestly, and hope the correction will meet with approval.

 

Yours sincerely,

Xiaoqing Zhao

 

 

 

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Abstract

 

Line 11 to 14

The study objective is lacking. Is this line about the objectives? If yes, the authors should explicitly state that. The knowledge gap is already precisely presented. A relevant study objective will complement that. 

 

Line 17

Maybe the authors should state what traditional object-oriented classification methods were compared against. 

 

Line 18 to 19

The variable that was classified should be mentioned. 

 

Introduction 

 

Line 24 to 37

There are multiple arguments in the first paragraph of the Introduction. Yet, citations are needed to support these claims. 

 

Line 28 to 29

Other than tree species, the estimated biomass is also important. In silviculture, wood production receives the utmost concern. 

 

Line 69 to 70

More information about why other satellites are not used can be provided in order to justify the choice of the RapidEye and Sentinel-2. 

 

Line 89

The authors should review the improvement brought by CNN over older machine learning approaches. It does not have to be about satellite image recognition. 

 

Study Area and Data

 

Line 114 to 115

It is suggested that the scientific name of each tree species can be provided.

 

Also, are there any existing tree surveys that document the species composition?

 

Line 139 to 140

This difference is directly influenced by the biological structure of trees. For instance, trunk flare of trees may interfere with the elevation measurement. Have any check been done to remove the related errors?

 

Methods

 

Line 161

Does it mean that the forest stand and species were classified using canopy height? In mixed-species stands, can this classification strategy work? 

 

Results

 

Line 380

In Table 6, from the accuracy values, it is clear that the classification accuracy increased with the use of more and more images. I believe that it is a normal phenomenon becuase a model with more predictors tend to outperform the one with less.

 

Line 390 to 398

If the classification accuracy is not so different, can a simpler classification method with less required inputs be used?

 

Discussion 

 

Line 407 to 409

Have any comparisons been conducted in terms of the computation time of the classification methods?

 

Comments on the Quality of English Language

Some polishing of the language will benefit the presentation of the findings.

Author Response

Dear Reviewer：

Thank you very much for your careful review and constructive suggestions with regard to our manuscript "Object-oriented Convolutional Neural Network for Forest Stand Classification Based on Multisource Data Collaboration"(ID：forests-2873022). Those comments are all valuable and very helpful for revising and improving our paper, as well as the importance guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in blue in the paper. The main corrections in the paper and the responds to the reviewer's comments are as following:

 

Abstract

Line 11 to 14

The study objective is lacking. Is this line about the objectives? If yes, the authors should explicitly state that. The knowledge gap is already precisely presented. A relevant study objective will complement that. 

Your opinion is very constructive. The research objectives of this article have been clarified and have been revised. As shown in lines 8-16：

“Accurate classification of forest stand is crucial for the protection and management needs. However, forest stand classification remains a great challenge because of the high spectral and texture similarity of different tree species. Although existing studies have used multiple remote sensing data for forest identification, the effects of different spatial resolutions and combining multi-source remote sensing data for automatic complex forest stands identification using deep learning methods still require further exploration. Therefore, this study proposed an Object-oriented Convolutional Neural Network(OCNN) classification method , leveraging data from Sentinel-2, RapidEye, and LiDAR to explore the classification accuracy of using OCNN to identify complex forest stands. ”

Line 17

Maybe the authors should state what traditional object-oriented classification methods were compared against. 

Thank you very much for your suggestion. The traditional methods for comparing with CNN have been listed in the manuscript, as shown in line 20:

The text is as follows:“including SVM, DTC, MLC and KNN.”

Line 18 to 19

The variable that was classified should be mentioned. 

Thank you very much for your suggestion. The variable "forest stand" has been added to the text

Introduction 

Line 24 to 37

There are multiple arguments in the first paragraph of the Introduction. Yet, citations are needed to support these claims. 

Thank you very much for your suggestion. We have carefully reviewed recent literature and included references to relevant studies in the revised manuscript to support the argument.As shown in line 26-31:

“The precise identification of forest stands distribution and tree species types is an important factor in forest stands investigation, playing a crucial role in forest resource planning and management, forest biomass, carbon storage assessment, habitat, ecosystem and socio-economic sustainable development[1,2]. Mapping vegetation at the species level can help monitor their growth characteristics and spatial distributions and design specific modeling for different tree species existing in an area[3]. ”

 

Line 28 to 29

Other than tree species, the estimated biomass is also important. In silviculture, wood production receives the utmost concern. 

I completely agree with your opinion. This sentence has been revised in the manuscript. As shown in line 26-29:

“The precise identification of forest stands distribution and tree species types is an important factor in forest stands investigation, playing a crucial role in forest resource planning and management, forest biomass, carbon storage assessment, habitat, ecosystem and socio-economic sustainable development[1,2].”

 

Line 69 to 70

More information about why other satellites are not used can be provided in order to justify the choice of the RapidEye and Sentinel-2. 

Thank you very much for your suggestion。The manuscript provides additional explanations on the reasons for data selection. As shown in line 60-73：

“To date, only several satellites have been launched (RapidEye, WordView-2, Sentinel-2, etc.) that are equipped sensors with red-edge bands[19]. Among them, RapidEye is the world's first multispectral commercial satellite that provides a 710nm red edge band which is conducive to monitoring vegetation and suitable for agricultural, forestry, and environmental monitoring. RapidEye can provide 5-meter spatial resolution image data, which is sufficient for the classification of forest species in this study. Sentinel-2 is the only data with multiple bands within the red edge range, which is very effective for monitoring vegetation health information and easy to obtain[20]. At present, there are some studies using Sentinel-2 and RapidEye data for forest classification respectively[21-26], but few studies have combined RapidEye with Sentinel-2 to maximize the use of the red edge band. And currently, the multispectral data used for forest classification in collaboration with LiDAR data is mostly SPOT5, IKONOS, Landsat, and other data lacking the red edge band, which do not meet the requirements for conducting research on the red edge band in this study. ”

 

The authors should review the improvement brought by CNN over older machine learning approaches. It does not have to be about satellite image recognition. 

Thank you very much for your suggestion. The manuscript clarifies the technical advantages and wide range of applications of CNN compared to traditional classification methods, and has achieved good results. As shown in line 98-104:

“Compared with traditional methods, CNN has better ability to process image and sequence data, has good learning ability for high-dimensional features, and has better generalization ability. It is widely used to solve various multi-level complex problems, such as road traffic monitoring[26], image classification, face recognition, semantic segmentation, etc[24-26], and has been successfully applied in remote sensing image scene classification, object detection, image retrieval, and other fields [27-30], these classification effects have been greatly improved. ”

 

Study Area and Data

Line 114 to 115

It is suggested that the scientific name of each tree species can be provided.

Your opinion is very constructive. We have consulted relevant information to modify the name of each tree species. As shown in line 132-133:

“Illicium verum, Pinus massoniana, Pinus elliottii, Eucalyptus grandis, Cunninghamia lanceolata”

 

Also, are there any existing tree surveys that document the species composition?

 Yes, there are records of relevant field investigations, and the table below shows some of the data in the files.

Sample plot number	Dominant tree species	Total square plots
		Number of sample trees/hectare	Average diameter/cm	Average high /m
1	Mytilaria laosensis Lecomte	900	19.79	19.34
2	Acacia crassicarpa Benth	756	17.84	14.46
3	Eucalyptus urophylla	1433	10.60	15.01
4	Cunninghamia lanceolata	1389	16.66	14.93
5	Tilia tuan Szyszyl	2178	13.42	13.76
6	Hybrid pine	989	17.68	12.62
7	Eucalyptus grandis	2278	11.13	17.03
8	Cunninghamia lanceolata	1278	16.84	15.27
9	Cunninghamia lanceolata	1344	16.95	14.19
10	Michelia odora	1267	15.82	17.21
11	Cunninghamia lanceolata	1100	18.57	16.17
12	Cunninghamia lanceolata	1378	13.93	12.01
13	Eucalyptus grandis	1456	10.38	13.45
14	Cunninghamia lanceolata	1356	15.42	14.82
15	Cunninghamia lanceolata	2133	13.31	11.99
16	Illicium verum	1589	10.81	8.77
17	Michelia macclurei	1244	17.19	15.34
18	Mytilaria laosensis Lecomte	1500	16.72	19.41
19	Cunninghamia lanceolata	978	14.59	10.17
20	Cunninghamia lanceolata	3733	11.15	11.09
21	pinus elliottii	1567	13.84	8.79
22	Castanopsis hystrix	1711	12.39	12.68
23	Eucalyptus grandis	1711	12.49	16.98
24	Eucalyptus urophylla	1644	9.70	13.52
25	Magnolia sumatrana	900	20.88	17.84
26	Cunninghamia lanceolata	1467	16.43	14.29
27	Cunninghamia lanceolata	1000	17.74	15.94
28	Illicium verum	1056	13.22	8.89
29	Cunninghamia lanceolata	2144	11.80	10.28
30	Eucalyptus grandis	2722	10.33	15.27
31	Cunninghamia lanceolata	1678	13.99	13.56
32	Illicium verum	822	9.80	6.87
33	Illicium verum	1600	8.49	6.52
34	Illicium verum	678	19.22	10.17
35	Broad-leaved tree	978	23.08	9.70
36	Eucalyptus grandis	1789	10.20	13.70
37	Cunninghamia lanceolata	1222	14.09	13.09
38	Cunninghamia lanceolata	2500	11.15	10.63
39	Pinus massoniana	578	27.82	16.25
40	Eucalyptus grandis	1633	11.66	16.35
41	Eucalyptus grandis	1144	8.68	10.35
42	Cunninghamia lanceolata	1700	16.03	15.14
43	Cunninghamia lanceolata	1356	13.67	13.78
44	Eucalyptus grandis	1789	11.93	17.96
45	Illicium verum	2478	8.88	8.03
46	Illicium verum	1467	10.24	7.25
47	Cunninghamia lanceolata	1122	16.60	13.32
48	Eucalyptus grandis	2456	8.48	12.16
49	Broad-leaved tree	1367	17.67	13.03
50	Eucalyptus grandis	1444	11.35	16.20
51	Eucalyptus grandis	2444	12.47	23.56
52	Eucalyptus grandis	500	10.16	10.79
53	Eucalyptus grandis	2056	13.70	21.35
54	Cunninghamia lanceolata	1333	13.93	13.65
55	Cunninghamia lanceolata	1389	14.04	13.64
56	Broad-leaved tree	1256	16.67	13.37
57	Illicium verum	1178	15.63	9.36
58	Illicium verum	1844	8.60	7.57
59	Eucalyptus grandis	1900	11.71	16.87
60	Pinus massoniana	1578	13.10	11.77
61	Pinus massoniana	989	19.18	13.11
62	Eucalyptus grandis	2844	9.51	14.49
63	Pinus massoniana	1167	16.49	13.19
64	Illicium verum	1700	10.52	7.79
65	Broad-leaved tree	1700	12.92	9.01
66	Cunninghamia lanceolata	878	15.55	10.13
67	Broad-leaved tree	1289	15.79	11.26
68	Cunninghamia lanceolata	1900	13.67	12.46
69	Pinus massoniana	1444	16.20	13.09
70	Eucalyptus grandis	1200	12.81	14.60
71	Eucalyptus urophylla	1911	10.32	14.83
72	Pinus massoniana	622	21.28	14.56
73	Pinus massoniana	633	22.33	14.66
74	Pinus massoniana	578	21.31	14.82
75	Eucalyptus urophylla	1867	10.01	15.02
76	Pinus massoniana	1700	18.26	15.35
77	Illicium verum	767	17.17	10.91
78	Illicium verum	1122	15.49	10.21
79	Castanopsis hystrix	567	23.05	18.12
80	Eucalyptus grandis	2111	11.43	17.19
81	Cunninghamia lanceolata	811	19.88	15.95
82	Cunninghamia lanceolata	1478	16.38	14.78
83	Eucalyptus urophylla	2644	7.84	11.86
84	Pinus massoniana	1422	12.72	6.96
85	Illicium verum	1556	9.37	8.38
86	Pinus massoniana	544	25.81	16.73
87	Cunninghamia lanceolata	1256	16.84	13.99
88	Illicium verum	1411	17.55	13.28
89	Eucalyptus grandis	1811	13.76	20.98
90	Eucalyptus grandis	1400	14.84	22.31
91	Eucalyptus grandis	1456	12.44	19.91
92	Pinus massoniana	1111	17.18	15.09
93	Broad-leaved tree	1444	14.99	9.30
94	Pinus massoniana	1356	15.98	15.05
95	Broad-leaved tree	1033	13.38	11.02
96	Pinus massoniana	856	20.76	16.91
97	Broad-leaved tree	1200	16.99	14.59
98	Pinus massoniana	922	18.34	16.34
99	Pinus massoniana	933	18.44	16.42
100	Pinus massoniana	1111	17.18	15.09
101	Pinus massoniana	1300	20.97	16.60
102	Pinus massoniana	1878	18.75	16.41
103	Pinus massoniana	556	22.61	17.59
105	Pinus massoniana	756	21.88	18.22
107	Pinus massoniana	1511	14.54	8.12
108	pinus elliottii	1678	13.96	8.86
109	Pinus massoniana	1322	12.90	6.20
110	Pinus massoniana	1544	14.08	8.42
111	Pinus massoniana	1667	13.58	7.90
112	Pinus massoniana	1511	16.33	8.96

 

Line 139 to 140

This difference is directly influenced by the biological structure of trees. For instance, trunk flare of trees may interfere with the elevation measurement. Have any check been done to remove the related errors?

Thank you very much for your suggestion. Yes, when processing LiDAR data, we have taken into account the noise issue you mentioned. Tree trunk flares are also a form of noise, and relevant explanations have been added in manuscript. As shown in line 161-170:

“During the acquisition process of point cloud data, noise is commonly introduced for various reasons, such as sensor errors and environmental factors. This noise is typically manifested as isolated points, drifting points, and redundant points, which can compromise the precision of modeling and information extraction. Therefore, it is essential to remove this noise. Common methods for noise removal include median filtering, mean filtering, Gaussian filtering, and statistical filtering, etc. In this study, statistical filtering is employed to remove noise. Specifically, for each point in the point cloud, the average value and standard deviation of points within a certain radius are calculated. This method can effectively remove noises, but it requires adjusting appropriate radius and multiplier parameters.”

 

Methods

Line 161

Does it mean that the forest stand and species were classified using canopy height? In mixed-species stands, can this classification strategy work?

Thank you very much for your opinion. It doesn’t mean that the forest stand and species were classified using canopy height. Not only does it use forest height, but also spectral and texture information data, with different species having different spectral and texture information. Due to the fact that the forest types at the regional scale are mixed forests, mainly secondary forests, the same species exhibit a spatial pattern of aggregation and distribution. Therefore, we were able to achieve canopy and species classification between 10m by 10m small plots, which can help us classify different species between 10m by 10m small plots within 30m by 30m. We also used object-oriented methods, and convolutional neural networks learned the average value of a forest stand, which can then achieve mixed forest classification.

Results

 

Line 380

In Table 6, from the accuracy values, it is clear that the classification accuracy increased with the use of more and more images. I believe that it is a normal phenomenon becuase a model with more predictors tend to outperform the one with less.

I completely agree with your opinion. The addition of more information is likely to improve classification accuracy, but the improvement in accuracy also proves that the red edge band can improve classification accuracy. Your opinion provides us with a way to find which feature is the key to classification in the classification process. This part will be further explored in future research.As shown in line 554-558:

“ Future research should focus on increasing the size and diversity of the training data set to improve the accuracy and robustness of the machine learning models. Besides, there is a need to strengthen the explainability of automatic quantitative visual representation, and explore the feature learning process of deep neural networks through automation. ”

 

Line 390 to 398

If the classification accuracy is not so different, can a simpler classification method with less required inputs be used?

 Your feedback is very constructive, and a classification method with high computational efficiency and good classification results is something we have been exploring. The OCNN method used this time has improved the accuracy of forest classification. Future research will focus on optimizing the model, identifying key features for forest classification, reducing input, and achieving good classification results.

 

 

Discussion 

Line 407 to 409

Have any comparisons been conducted in terms of the computation time of the classification methods?

Thank you very much for your opinion. Yes, we have compared the classification efficiency. In traditional classification methods, the MLC method is slightly more efficient, but the classification performance is not the best. The running speed of the model in the OCNN experiment is related to the set number of iterations, and it is necessary to find the optimal number of iterations. In this experiment, 50 iterations can achieve the highest classification accuracy, and the time is slightly higher than the SVM classification method.

 

Comments on the Quality of English Language

Some polishing of the language will benefit the presentation of the findings.

We appreciate your attention to the language quality in our manuscript. We have carefully revised the language through out the manuscript.

we appreciate reviewers' warm work earnestly, and hope the correction will meet with approval.

 

 

Yours sincerely,

Xiaoqing Zhao

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

1. The authors have diligently addressed most of the specific raised issues in my review. They have revised sections where clarity was lacking and improved the overall flow of the manuscript.

   The revised manuscript is notably clearer and more comprehensible for readers. Complex concepts have been explained succinctly, enhancing the scientific content.

   Additional Improvements:

   While the authors have made substantial progress, I recommend a few additional enhancements:

   ·        Clarify the methodology section further by providing more details on data collection and analysis.

   ·        Consider expanding the discussion section to explore implications and future research directions.

   In summary, the authors have made commendable efforts to address my comments. However, I encourage them to incorporate the suggested improvements to elevate the manuscript’s quality. 

Author Response

Dear Reviewer：

Thank you for approving our previous revisions and providing us with suggestions for further improving the quality of our manuscript "Object-oriented Convolutional Neural Network for Forest Stand Classification Based on Multisource Data Collaboration"(ID：forests-2873022). We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in blue in the paper. The main corrections in the paper and the responds to the reviewer's comments are as following:

 

1. Clarify the methodology section further by providing more details on data collection and analysis.

Thank you very much for your suggestion. We have added details about data collection and analysis. As shown in line 141-147 and line 151-163:

 

“The purpose of this study is to explore the classification potential of red edge band in complex forest stand species and to further improve classification accuracy by utilizing the canopy height information from LiDAR data. Select RapidEye, Sentinel-2A images with red edge bands, and LiDAR data obtained through unmanned aerial vehicle. By setting up sample plots in the field and conducting field investigations on forest stand types in the study area, a basis is provided for the subsequent production of training samples and verification of classification accuracy.”

 

“RapidEye images are obtained through purchase, and the purchased RapidEye product level is 3A. It has undergone radiation correction, sensor correction, and geometric correction, with a spatial resolution of up to 5 meters. One scene image can cover the research area without the need for image mosaic. After obtaining data, register with Sentinel-2A and extract the study area through clip. On the official website of the European Space Agency (https://scihub.copernicus.eu/dhus/#/home) download Sentinel-2A images that meet the conditions. The downloaded product level is L1C, which is an atmospheric apparent reflectance product that has undergone orthophoto correction and geometric precision correction, and has not undergone atmospheric correction. Atmospheric correction is currently being carried out in the Sen2cor plugin released by ESA, which specializes in producing L2A level data. Subsequently, the image will be resampled to 10 meters and registered with RapidEye to extract the study area through clip.”

 

2. Consider expanding the discussion section to explore implications and future research directions.

Thank you very much for your suggestion. We have already clarified the significance of this study in the discussion section and expanded on the shortcomings to lead to future research directions. As shown in line 561-567 and line 570-577:

 

“Correct forest stand species identification is essential for forest inventories, as it is the basis for forest biomass calculations, carbon storage assessment, habitat, ecosystem and socio-economic sustainable development. Our method will complement other approaches to identify stand species based on remote sensing, will help to increase overall accuracy, and thus, will support the more efficient creation of forest inventories. This assumption is based on the future possibility of achieving high-precision automatic or semi-automated classification processes. Our research aims to explore this process.”

 

“Firstly, Some forest types have relatively few training samples and are prone to getting stuck in local optima, which will affect the feature learning process of the category and thus affect classification accuracy. Secondly, Although the study area of this experiment is already a complex forest type, the universality of the model under different geographical, climatic or ecological conditions still needs further verification. Thirdly, its learning process can be seen as a "black box". Although it has strong adaptability to data, this leads to weak interpretability in the operation process, making it difficult to grasp the key parameters in the feature learning process. ”

 

we appreciate reviewers' warm work earnestly again, and hope the correction will meet with approval.

 

Yours sincerely,

Xiaoqing Zhao

 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors!

Thank you very much for answering all my questions, I have no more questions regarding the manuscript.

Author Response

Dear Reviewer：

Thank you for approving our previous revisions for our manuscript "Object-oriented Convolutional Neural Network for Forest Stand Classification Based on Multisource Data Collaboration"(ID：forests-2873022).

we appreciate reviewers' warm work earnestly again.

Yours sincerely,

Xiaoqing Zhao

 

 

Author Response File: Author Response.pdf