Mapping and Monitoring Forest Plantations in São Paulo State, Southeast Brazil, Using Fraction Images Derived from Multiannual Landsat Sensor Images

Round 1

Reviewer 1 Report

I have outlined my comments below. Considering them will help authors to provide improved readership.

-        - What is the novelty on this work. what contribution does this paper add to the literature?

-        - You have used Landsat TM and OLI data but there are wavelength variations between these sensors. What implications did you used when using these differing wavelengths to classify the forest plantations

-        - Table 1 is unnecessary as the information is open to anyone.

-       -  To enhance international significance, you need to exemplify works from different areas in the introduction section

-       -  I suggest the authors to add a flowchart figure to present clearly their work.

-        -  Discussion part is poor and not enough on this manuscript.

Author Response

Reviewer 1 - Comments and Suggestions for Authors

I have outlined my comments below. Considering them will help authors to provide improved readership.

Thank you very much for your comments and suggestions that really improved the manuscript. We followed all them, e.g., we deleted Table 1, included the flowchart figure, added more references, and answered your questions.

-- What is the novelty on this work. what contribution does this paper add to the literature?

Thanks for your question. The novelty of this work is the idea to explore the plantation cycles similar to the agricultural phenological cycles. While agriculture assessment uses the annual time series, tree plantation needs the longer time series (here we used ten years that were enough for detecting Eucalypt plantation growth cycle but not for Pine plantation).

-        - You have used Landsat TM and OLI data but there are wavelength variations between these sensors. What implications did you used when using these differing wavelengths to classify the forest plantations

Thanks for your comments. For our study, TM and OLI presented similar information, considering that we used the same spectral bands and the spectral indices were also derived from corresponding spectral bands of both sensors (e.g. NDVI = (NIR – Red) / (NIR + Red)).

-        - Table 1 is unnecessary as the information is open to anyone.

Thanks for your comment about this information. We deleted Table 1 in the revised manuscript.

Table 1. Characteristics of Landsat TM and OLI sensors

TM Bands	TM Wavelength (µm)	TM Res. (m)	OLI Bands	OLI Wavelength (µm)	OLI Res. (m)
			Band 1 - Coastal aerosol	0.43-0.45	30
Band 1 - Blue	0.45-0.52	30	Band 2 - Blue	0.45-0.51	30
Band 2 - Green	0.52-0.60	30	Band 3 - Green	0.53-0.59	30
Band 3 - Red	0.63-0.69	30	Band 4 - Red	0.64-0.67	30
Band 4 - Near Infrared (NIR)	0.76-0.90	30	Band 5 - Near Infrared (NIR)	0.85-0.88	30
Band 5 - SWIR 1	1.55-1.75	30	Band 6 - SWIR1	1.57-1.65	30
Band 7 - SWIR 2	2.08-2.35	30	Band 7 - SWIR 2	2.11-2.29	30
			Band 8 - Panchromatic	0.50-0.68	15
			Band 9 - Cirrus	1.36-1.38	30
Band 6 - Thermal	10.40-12.50	120

 

-       -  To enhance international significance, you need to exemplify works from different areas in the introduction section

Thanks for the excellent suggestion. We added more references in the revised manuscript.

-       -  I suggest the authors to add a flowchart figure to present clearly their work.

Thanks for your suggestion. The flowchart figure was added in the revised manuscript.

Figure 4. Flowchart of the methodological approach.

-        -  Discussion part is poor and not enough on this manuscript.

Thanks for your comments. We extended the Discussion part in the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

I reviewed the paper by Shimabukuro et al. I found the contribution of the paper very limited as only well-known methodologies have been implemented. Furthermore, it is not clear what the objective and contribution of the paper are, especially in comparison to previous studies. More importantly, the obtained accuracies are low, making it doubtful to accept the applicability of the paper. Moreover, other critical issues existed in the paper that prevented me from recommending it for publication. I have provided some specific concerns below:

1- Although a two-class problem was considered, the obtained accuracies are too low. This makes it difficult to accept the applicability of the paper. Furthermore, the authors did not attempt to provide further information regarding the low accuracy of maps. It is necessary to discuss the possible reasons for obtaining low accuracies and the solutions to overcome this issue.

2- As the obtained accuracies are low, especially too low for forest classes, the associated uncertainties would propagate into subsequent conclusions, reducing the robustness of the conclusions and derived information.

3- It is not clear why the Pine was not considered as an endmember, especially for the pilot site where it existed.

4- Nearly 10 years of satellite imagery was considered to produce one land cover map with two classes. For instance, satellite imagery from 1985 to 1995 was considered to produce the land cover map of 1995. Why? An important comment is why only one year or two years of satellite imagery close to 1995 was not considered. What issue has been solved by considering 10 years of data? The high spectral variability for each pixel will be high when considering a large time period for generating one land cover map. What uncertainties have been introduced by considering 10 years of satellite imagery? This issue becomes more serious when only reference samples for the year 1995 were incorporated for evaluation. Maybe considering 10 years of satellite imagery is the possible reason for obtaining low classification accuracy.

5- Why only 1000 samples have been used for accuracy assessment? Why the whole MapBiomas land cover data was not considered for evaluation and agreement calculation? This could make the evaluation more robust. Furthermore, the specification of MapBiomas land cover map is missing. What is its spatial resolution? What is its accuracy? What are its land cover classes? How it has been produced?

 

 

 

Author Response

Reviewer 2 - Comments and Suggestions for Authors

I reviewed the paper by Shimabukuro et al. I found the contribution of the paper very limited as only well-known methodologies have been implemented. Furthermore, it is not clear what the objective and contribution of the paper are, especially in comparison to previous studies. More importantly, the obtained accuracies are low, making it doubtful to accept the applicability of the paper. Moreover, other critical issues existed in the paper that prevented me from recommending it for publication.

Thanks for your comments and suggestions that really improved the manuscript.

I would say that there are a few works about classification of forest plantations using remote sensing, especially considering the spectral responses capable of detecting the growth cycle of forest plantation. The reason is that is not easy to map the extended area of forest plantations due to their specific characteristics. The forest cover can not be mapped using only one or a few years of images because the forest plantation rotation time. It is necessary to make use of the time series as has been done by a few papers available in the literature. However, these papers use vegetation indices as the primary information to get the rotation time and are used to classify these areas. The novelty of our work is that besides the original spectral bands and well-known vegetation indices, we are using fraction images derived from LSMM. The vegetation fraction highlights the vegetation cover, the soil fraction highlights the forest non-forest areas, and the shade fraction highlights the forest structure and is used in this work as information that helped to discriminate Eucalypt from Pine plantations. Regionally (the whole São Paulo state) we compared our results with the forest plantations areas extracted from the MapBiomas LULC global map. We also evaluated our results in the pilot areas using the field information available from the Sylvamo Company. The agreement is very good, but we only showed it for the phenological analysis of the Pine and Eucalypt. We made the evaluation process clearer in the revised manuscript.

I have provided some specific concerns below:

1-Although a two-class problem was considered, the obtained accuracies are too low. This makes it difficult to accept the applicability of the paper. Furthermore, the authors did not attempt to provide further information regarding the low accuracy of maps. It is necessary to discuss the possible reasons for obtaining low accuracies and the solutions to overcome this issue.  As the obtained accuracies are low, especially too low for forest classes, the associated uncertainties would propagate into subsequent conclusions, reducing the robustness of the conclusions and derived information.

Thanks for the comments. In order to show the potential of the proposed method we included the visual comparison and the area estimated between our results and the results extracted from MapBiomas. We also increased the number of sample plots as suggested. We increased the number of samples to check the accuracy of the classification as presented in the revised manuscript. We created 20,000 stratified random points within each class based on the pixel classification. The stratified random points strategy is based on the proportional distribution to each relative area of the classes. Then, we generated a confusion matrix using MapBiomas LULC classes as the reference to estimate the accuracy of the classification results of the proposed method.

 

3- It is not clear why the Pine was not considered as an endmember, especially for the pilot site where it existed.

Thanks for the comment. The Pine was not considered as an endmember, because Eucalypt has the higher spectral curve than Pine that is required by the LSMM concepts and could be used for the whole time series analyzed. We included this statement in the revised manuscript.

4- Nearly 10 years of satellite imagery was considered to produce one land cover map with two classes. For instance, satellite imagery from 1985 to 1995 was considered to produce the land cover map of 1995. Why? An important comment is why only one year or two years of satellite imagery close to 1995 was not considered. What issue has been solved by considering 10 years of data? The high spectral variability for each pixel will be high when considering a large time period for generating one land cover map. What uncertainties have been introduced by considering 10 years of satellite imagery? This issue becomes more serious when only reference samples for the year 1995 were incorporated for evaluation. Maybe considering 10 years of satellite imagery is the possible reason for obtaining low classification accuracy.

Thanks for the comments. According to le Maire et al. a classification using only one annual time series may classify the Eucalypt plantations as “evergreen broadleaf forest” or “bare ground/low vegetation cover”, “deciduous forest” because of clear-cuts. Thus, it is the reason to analyze multiannual time series to map areas occupied by Eucalypt plantations as done in this work. We used 10 years of satellite imagery since the methodology is based on the rotation cycle of forest plantation areas, especially Eucalypt plantation that has a short rotation time (6 to 8 years on average). The resulting map shows the forest plantation areas for this time periods (1985- 1995 and 2013 -2021, i.e., the mapped areas were occupied by forest plantations during the whole periods. We did not use one or two years because we would not be able to map the areas with forest plantations as showed in the following figures. We can observe that the methodology allowed to detect the different phenological stages of these forest plantations especially Eucalypt plantation during the study time periods. In addition, we can see that the maximum vegetation, shade and soil fraction images highlighted the whole area occupied by forest plantations then helping to classify these areas. We included these statements in discussion section of the revised manuscript.

 

Figure 17. Landsat TM dataset (1985 1995).

 

Figure 18. Maximum (vegetation, shade, soil) fraction images derived from TM time series (1985-1995).

5- Why only 1000 samples have been used for accuracy assessment? Why the whole MapBiomas land cover data was not considered for evaluation and agreement calculation? This could make the evaluation more robust. Furthermore, the specification of MapBiomas land cover map is missing. What is its spatial resolution? What is its accuracy? What are its land  cover classes? How it has been produced?

Thanks for the questions. We increased the number of samples to check the accuracy of the classification as presented in the revised manuscript. We created 20,000 stratified random points within each class based on the pixel classification. The stratified random points strategy is based on the proportional distribution of each relative area of the classes. Then, we generated a confusion matrix using MapBiomas LULC classes as the reference to estimate the accuracy of the classification results of the proposed method.

MapBiomas is a multi-disciplinary network to reconstruct annual land use and land cover information from 1985 to present for Brazil, based on random forest  algorithm applied to Landsat archive using Google Earth Engine. Mapbiomas mapped five major classes: forest, non-forest natural formation, farming, non-vegetated areas, and water. These classes were broken into two sub-classification levels leading to the most comprehensive and detailed mapping for the country at a 30 m pixel resolution. The average overall accuracy of the land use and land cover time-series, based on a stratified random sample of 75,000 pixel locations, was 89% ranging from 73 to 95% in the biomes (Souza Jr. et al. 2020). For comparison with our results, we reclassified MapBiomas classes into only two classes: forest plantation and all the other LULC classes together (non-forest plantation).

 

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

All the comments have been addressed  by the authors.

Reviewer 2 Report

I appreciate the authors' efforts in improving the manuscript. The authors addressed the comments, and the paper could be accepted in its current form.