Examining the Roles of Spectral, Spatial, and Topographic Features in Improving Land-Cover and Forest Classifications in a Subtropical Region

Round 1

Reviewer 1 Report

In Introduction authors work almost only with local literature (authors from China). Land cover images clasiffication theme is processed worldwidely, by many authors from many regions and universities (Europe, USA...). This part must be improved.

In Methodology it was appropriate to list all the software used.

Comparison table in row 416 contains many data and they are confusing. Could be devided.

Author Response

Reviewer#1

In Introduction authors work almost only with local literature (authors from China). Land cover images clasiffication theme is processed worldwidely, by many authors from many regions and universities (Europe, USA...). This part must be improved.

Reply: Thanks for your comment. Yes, many authors from different countries have conducted studies related to land cover classification. We added 28 new references.

In Methodology it was appropriate to list all the software used.
Reply: Thanks, We mainly used ENVI and ArcGIS to conduct this research, and indicate the software in proper places.

Comparison table in row 416 contains many data and they are confusing. Could be devided.

Reply: Thanks for your suggestion. Yes, this table is really too long and contains too many data. Therefore, this table is separated into two, one is to compare the classification results by incorporation of Sentinel-2 data, and another is to compare the classification results by incorporation of Landsat data.

Reviewer 2 Report

The article examines the roles of different features that affect the classification accuracy in forest and land-cover targeting.

The article sounds proper and I recommend accept for publication after minor revisions listed below:

• Fig. 2 attempts to explaining the methodology used in the fusion data and further classification procedures. Although the methodology is explained and enumerated in the text (subsection 2.1), Fig. 2 is confused. Should be necessary to relate the numerated methodology with the diagram shown in this figure.
• From Table 1 it can be seen that the acquisition date of the dataset is not coincident (varies from February 2017 to March 2018). May this gap-time affect (positively or negative) in the classification? It should be noted that the spectral and spatial (principally) patterns can change during this date. Please, provide some explanation to the readers indicating if this gap-time produce or not an effect on the results.
• line 238: replace the sentence "...spatial resolution). ST10..." to ": "...spatial resolution), ST10..." (comma instead of a dot).

Author Response

Reviewer #2

The article examines the roles of different features that affect the classification accuracy in forest and land-cover targeting.

The article sounds proper and I recommend accept for publication after minor revisions listed below:

Reply: Thank you very much for your positive comments. We carefully revised this manuscript according to your comments.

• Fig. 2 attempts to explaining the methodology used in the fusion data and further classification procedures. Although the methodology is explained and enumerated in the text (subsection 2.1), Fig. 2 is confused. Should be necessary to relate the numerated methodology with the diagram shown in this figure.

Reply: good suggestion. This figure is re-organized so that the major steps such as data preparation, data fusion, selection of variables, classification and evaluation of results are clearly related to the specific subsections in the Method section.

• From Table 1 it can be seen that the acquisition date of the dataset is not coincident (varies from February 2017 to March 2018). May this gap-time affect (positively or negative) in the classification? It should be noted that the spectral and spatial (principally) patterns can change during this date. Please, provide some explanation to the readers indicating if this gap-time produce or not an effect on the results.

Reply: Good comments. Yes, this difference between image acquisition dates may affect the fused results between Landsat and ZY-3 data, thus could further influence classification result. When conducting the classification process, we carefully examined the selection of samples for training and validation respectively, so that the difference of image acquisition dates will not affect classification and accuracy assessment considering the possible land cover change during the different time period. We added one paragraph in the Discussion section (subsection 4.4) to provide some explanations on impacts of different acquisition dates of satellite data on potential classification results.

• line 238: replace the sentence "...spatial resolution). ST10..." to ": "...spatial resolution), ST10..." (comma instead of a dot).

Reply: corrected. Thanks.

Reviewer 3 Report

This study conducted the land-cover classification in the subtropical forest using the multi-source data (ZY-3, Snetinel-2, and Landsat 8 OLI)and studied the role of spectral, spatial, and topographic features in enhancing the land-cover classification. The topic of this research is very interesting and the paper is almost well-written. However, there are still some issues that need to be addressed.

Here are some comments for improving the quality of this paper:

(1)-Lines 18-20: As far as I know, previously, various studies investigated the effects of spectral and spatial features of remotely sensed data and topographic characteristics on classification results (please correct me if I am wrong). Please modify this sentence in the way that it reflects the specific contributions of your research (from the perspective of the employed data etc.).

(2)-Lines 20-23: As far as I understood the main focus of your paper is on assessing the impact of spectral, spatial, and topographic features (not the impacts of different data sources...) on the quality of the land use classification.  In this context, and to implement your proposed framework, and conduct the classification, you used RF classifier (I think it is also important to mention to the adopted classifier in this study) and the multi-source data( ZiYuan-3 (ZY-3), Sentinel-2, Landsat 8, and topography). If my understanding is correct, please modify this part in the way that it reflects these issues.

(3)-Line 99: segmentation and texture---> do you mean object-based image analysis approach and using the textural information?

(4)-Lines 125-126: "Optical sensor data have spectral, spatial, and temporal features, but the effectiveness of using these features and their roles in improving forest classification have not been fully examined". I suggest you to be a bit more specific. Please see my first comment.

(5)-Line 143-144: Guangxi Zhuang 143 Autonomous Region---> Please add the country name at the end: Guangxi Zhuang 143 Autonomous Region, China

(6)-Line 156-173: This should be explained under a new subtitle such as 2.2. The proposed methodology, the proposed framework etc. 

(7)-Line 174: 2.2. Data preparations---> I think it should be changed into 2.2.1 Data preparations

(8)- Line 178: 2.2.1. Collection of field survey data and design of a land cover classification system---> I think it should be changed into 2.2.1.1. Data preparations

(9)- Line 181: "into GIS software" --->please clearly mention to the name of the employed software

(10)-Line 189: "---> What do you mean by the term "special class"? Please clarify this a bit more.

(11)-Line 238-239: "with 10 m spatial resolution". I think you resampled the bands with the spatial resolution other than 10m into 10 m. Please mention this in your paper (I could not find any description on this).

(12)-Line 240-242: "Here the abbreviations ZY, ST, and LS represent ZiYuan-3, Sentinel-2, 240 and Landsat 8 OLI; MS and PAN represent multispectral bands and panchromatic band; PC1 represents 241 the first component from the principal component analysis of the ZY-3 multispectral bands." ----> I suggest you move these lines (that includes the acronyms) to somewhere at the beginning of the section.

(13)-Why did you select only the PC1 ? Please clarify this a bit more. Maybe you can provide more details on "How much of the variance is explained by your first principal component?"

(14)- I find your discussion section very interesting.

Recommendation:

As you discussed about the  importance of using multiple data sources (derived from the technical sensors) to improve land-cover and forest classification in section 4.4 (lines 602- ), I recommend you to also discuss briefly about the application (and possibility and benefits) of fusion of human sensor data (Volunteered Geographic Information--VGI) with technical sensor data (optical remotely sensed data, LiDAR data etc.) in extracting the useful contextual information (that might be useful in the classification process) for improving the quality of classification, improving the quality of classification by increasing the number of training samples that can be extracted from open data sources (such as OSM) and enhancing the automation of the classification process (the training and validation samples can be extracted from the VGI). You can use and refer to the following references (as well as other related studies in this area) for further information :

-Vahidi, H.; Klinkenberg, B.; Johnson, B.A.; Moskal, L.M.; Yan, W. Mapping the Individual Trees in Urban Orchards by Incorporating Volunteered Geographic Information and Very High Resolution Optical Remotely Sensed Data: A Template Matching-Based Approach. Remote Sens. 2018, 10, 1134.

-Johnson, B.A.; Iizuka, K. Integrating OpenStreetMap crowdsourced data and Landsat time-series imagery for rapid land use/land cover (LULC) mapping: Case study of the Laguna de Bay area of the Philippines. Appl. Geogr. 2016, 67, 140–149.

Author Response

Reviewer #3

This study conducted the land-cover classification in the subtropical forest using the multi-source data (ZY-3, Snetinel-2, and Landsat 8 OLI)and studied the role of spectral, spatial, and topographic features in enhancing the land-cover classification. The topic of this research is very interesting and the paper is almost well-written. However, there are still some issues that need to be addressed.

Reply: Thank you very much for your positive comments. We have carefully revised this paper according to your comments.

Here are some comments for improving the quality of this paper:

(1)-Lines 18-20: As far as I know, previously, various studies investigated the effects of spectral and spatial features of remotely sensed data and topographic characteristics on classification results (please correct me if I am wrong). Please modify this sentence in the way that it reflects the specific contributions of your research (from the perspective of the employed data etc.).

Reply: Yes, I completely agreed with you that various studies investigated the effects of spectral and spatial features of RS data and topographic characteristics on classification results, but mainly based on individual sensor data. A comprehensively comparative analysis of these features from different sensor data with various spatial resolutions have not been conducted. Therefore, this research explored the roles of spectral and spatial features from ZY-3, Sentinel-2 and Landsat with spatial resolution from 2 m, 6 m, 10m, 15 m and 30 m, and topographic factors in influencing land cover classification in a mountainous region. Through this comparative analysis, we can better understand how to select suitable spectral and spatial features corresponding to specific remotely sensed data to obtain the best classification results. Based on your comments, we re-wrote this sentence.

(2)-Lines 20-23: As far as I understood the main focus of your paper is on assessing the impact of spectral, spatial, and topographic features (not the impacts of different data sources...) on the quality of the land use classification.  In this context, and to implement your proposed framework, and conduct the classification, you used RF classifier (I think it is also important to mention to the adopted classifier in this study) and the multi-source data( ZiYuan-3 (ZY-3), Sentinel-2, Landsat 8, and topography). If my understanding is correct, please modify this part in the way that it reflects these issues.

Reply: yes, you are right. The above comments (1) and (2) are related, so we re-wrote these sentences according to your comments. 

(3)-Line 99: segmentation and texture---> do you mean object-based image analysis approach and using the textural information?

Reply: Thanks for your comments. We did not clearly describe how to use segmentation and texture in this sentence. We try to say that segmentation and textures are two common approaches to use spatial features. Use of segmentation is often involved in the object-oriented classification approach, which segmentation is one of the critical steps, while texture is calculated from a specific band based on selected window size such as 3x3 using certain approaches such as variance and contrast. This sentence is re-written and more texts were added to explain these terms.

(4)-Lines 125-126: "Optical sensor data have spectral, spatial, and temporal features, but the effectiveness of using these features and their roles in improving forest classification have not been fully examined". I suggest you to be a bit more specific. Please see my first comment.

Reply: Good comments, we added more texts to describe this issue by adding different sensor data with various spatial resolutions. In this way, the research problem is clearly described.

(5)-Line 143-144: Guangxi Zhuang 143 Autonomous Region---> Please add the country name at the end: Guangxi Zhuang 143 Autonomous Region, China

Reply: added. Thanks

(6)-Line 156-173: This should be explained under a new subtitle such as 2.2. The proposed methodology, the proposed framework etc. 

Reply: good suggestion. A subtitle “2.2 The proposed framework” was added

(7)-Line 174: 2.2. Data preparations---> I think it should be changed into 2.2.1 Data preparations

Reply: Thanks, this subsection was changed to “2.3 Data preparations”

(8)- Line 178: 2.2.1. Collection of field survey data and design of a land cover classification system---> I think it should be changed into 2.2.1.1. Data preparations

Reply: it was changed as 2.3.1

(9)- Line 181: "into GIS software" --->please clearly mention to the name of the employed software

Reply: “ArcGIS” was added.

(10)-Line 189: "---> What do you mean by the term "special class"? Please clarify this a bit more.

Reply: I just try to highlight the importance of eucalyptus in this study. In the revised version, this sentence was removed.

(11)-Line 238-239: "with 10 m spatial resolution". I think you resampled the bands with the spatial resolution other than 10m into 10 m. Please mention this in your paper (I could not find any description on this).

Reply: Thanks for your suggestion. The spectral bands with 20 m spatial resolution was resampled to 10 m when conducting atmospheric correction using Sen2Res, as described in subsection 2.3.2 “Collection and preprocessing of different remotely sensed data.” More texts were added to indicate the spatial resolution issue.

(12)-Line 240-242: "Here the abbreviations ZY, ST, and LS represent ZiYuan-3, Sentinel-2, 240 and Landsat 8 OLI; MS and PAN represent multispectral bands and panchromatic band; PC1 represents 241 the first component from the principal component analysis of the ZY-3 multispectral bands." ----> I suggest you move these lines (that includes the acronyms) to somewhere at the beginning of the section.

Reply: good suggestion. This sentence was moved to the beginning of (1) in this subsection.

(13)-Why did you select only the PC1 ? Please clarify this a bit more. Maybe you can provide more details on "How much of the variance is explained by your first principal component?"

Reply: The following sentence was added to explain the use of PC1 in this research. The PC1 from ZY-3 multispectral image was used here is that it concentrates most information from the multispectral bands (over 80% of total variance explained in this research) and only one band with high spatial resolution is required in the HPF data fusion.

(14)- I find your discussion section very interesting.

Recommendation:

As you discussed about the  importance of using multiple data sources (derived from the technical sensors) to improve land-cover and forest classification in section 4.4 (lines 602- ), I recommend you to also discuss briefly about the application (and possibility and benefits) of fusion of human sensor data (Volunteered Geographic Information--VGI) with technical sensor data (optical remotely sensed data, LiDAR data etc.) in extracting the useful contextual information (that might be useful in the classification process) for improving the quality of classification, improving the quality of classification by increasing the number of training samples that can be extracted from open data sources (such as OSM) and enhancing the automation of the classification process (the training and validation samples can be extracted from the VGI). You can use and refer to the following references (as well as other related studies in this area) for further information :

-Vahidi, H.; Klinkenberg, B.; Johnson, B.A.; Moskal, L.M.; Yan, W. Mapping the Individual Trees in Urban Orchards by Incorporating Volunteered Geographic Information and Very High Resolution Optical Remotely Sensed Data: A Template Matching-Based Approach. Remote Sens. 2018, 10, 1134.

-Johnson, B.A.; Iizuka, K. Integrating OpenStreetMap crowdsourced data and Landsat time-series imagery for rapid land use/land cover (LULC) mapping: Case study of the Laguna de Bay area of the Philippines. Appl. Geogr. 2016, 67, 140–149.

Reply: Thank you very much for your nice suggestions. I completely agreed with you that use of crowdsourced data are valuable for improving land cover classification, and this will be more important in the future as more and more open data sources are available that can be used as training samples and validation samples, as well as the more and more data sources are used as input variables for land cover classification using advanced algorithms such as deep learning. I also thank you for recommending these important references, they were cited in the Discussion section. Based on your suggestion, I added one paragraph in subsection 4.4 to discuss use of open data sources in land cover classification.

Round 2

Reviewer 3 Report

OK