Combining Radar and Optical Imagery to Map Oil Palm Plantations in Sumatra, Indonesia, Using the Google Earth Engine

Round 1

Reviewer 1 Report

This paper proposed combining radar and optical imagery to map oil palm plantations. The experiments are conducted using Google Earth Engine. However, the novelty of the paper is not clearly stated. 1. The authors claim that one of the conclusion is that combining optical and radar dataset improved the accracy of oil palm land-cover classification. I have to say, this is not a novel phenomenon. Please highlight the originality and novelty. 2. What's the rational behind the choice of the features for classification. Since some of the features are correlated, I thnk some feature extraction process like PCA may help to improve the discriminality and improve the classifiation accuracy. Please elaborate the possibility.

Author Response

This paper proposed combining radar and optical imagery to map oil palm plantations. The experiments are conducted using Google Earth Engine. However, the novelty of the paper is not clearly stated.

#1. The authors claim that one of the conclusion is that combining optical and radar dataset improved the accuracy of oil palm land-cover classification. I have to say, this is not a novel phenomenon. Please highlight the originality and novelty.

Response: We thank the reviewer for this comment. We agree with the reviewer that there have been other studies which use optical and radar datasets to map land cover, including oil-palm cover, and we have cited these studies. However, we believe that our research is novel in terms of completing the processing (especially for the speckle filtering and conversion to normalized radar cross-section steps for our SAR mosaic), classification and mapping of oil palm land cover using SAR and Landsat imagery entirely on Google Earth Engine. We have prepared the code and made it publicly available so that it can be freely utilized and built upon by the research community. We edited our Discussion to better express the novelty of our paper:

L439-440: “Moreover, we developed a code to preprocess radar and optical datasets as well as classify land-cover entirely in Google Earth Engine”.

In addition, we contribute to the ongoing discussion about the performance of automated vs visual classification for oil palm land cover and we explained in the following sentence:

L500-504: “However, our study also highlights limitations of automated mapping techniques for oil palm plantations and seconds the conclusion by Miettinen et al. (2018) about the need to combine use of visual and automated oil palm mapping approaches for monitoring oil palm expansion.”

#2. What's the rational behind the choice of the features for classification. Since some of the features are correlated, I think some feature extraction process like PCA may help to improve the discriminality and improve the classifiation accuracy. Please elaborate the possibility.

Response: We think that the features of choice in our classification process are fairly common for optical and radar datasets. These have been used in a previous study on mapping land cover in Myanmar by De Alban et al. (2018) Rem Sens. Also, our classifier of choice, random forests, is known to be robust to highly correlated predictors (Matsuki et al. 2016; https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4710485/). Hence, we do not think a PCA is strictly needed here since our focus is on analysing which map produces the highest accuracies especially for oil palm land cover and we are not focused on understanding which features are more important in our models (i.e., our focus is on classification and not examining variable/feature importance). Nevertheless, we thank the reviewer for this comment and will take this suggestion up for future studies.

Reviewer 2 Report

See the attached file.

Comments for author File: Comments.pdf

Author Response

#1. In the Abstract, random forest algorithm should be mentioned.

Response: Thank you for this comment. We mentioned random forest in the Abstract:

L19-21: “We evaluated the use of a semi-automated approach with random forest as a classifier and combined optical and radar datasets to classify oil palm land-cover in 2015 in Sumatra, Indonesia, using Google Earth Engine.”

 

#2. Lines 19-20: “Here, we evaluated the use of an automated approach of combining optical and radar datasets to classify oil palm land-cover in 2015 …” The manuscript just used a random forest algorithm for classification. This step might be automated. However, the image pre-processing, prepare of training and Testing ROIs, and majority filter for classification and so on were conducted step by step. These steps were far from “automated”!

Response: We thank the reviewer for this comment. We have changed automated to semi-automated under the Abstract:

L21-23: “We compared our map with two existing remotely sensed oil palm land-cover products that utilized visual and semi-automated approaches for the same year.”

 

#3. Lines 21-22: “We compared our map with two existing remotely sensed oil palm land-cover products that utilized visual and automated approaches for the same year”. For the automated approach for comparison, there is no information about “automated” in the manuscript!

I have read this article and found that it just used an unsupervised ISODATA clustering method and some post-processing steps. And the article pointed that it used a semi-automated classification method!

Besides, using the map that based on an unsupervised method for comparison did not demonstrate the advantage of the used method in the manuscript! In my opinion, the supervised methods might always achieve better results than the unsupervised ones!

Response: Thank you for pointing this out, we changed the term “automated” to “semi-automated” in the abstract and in necessary parts of the manuscript.

We agree with the reviewer that supervised methods will always achieve better results than unsupervised ones. However, we compared the semi-automated map from Miettinen et al. (2016) with our maps in order to know whether methods and spatial resolution influenced the results. We contrasted our semi-automated method to a visual method used by Austin et al. (2017).

We included the following to highlight how our results support the use of supervised classification:

L472-473: “Our results support previous studies that found higher accuracies using supervised methods [63-65]”

63. Mohammady, M.; Moradi, H.R.; Zeinivand, H.; Temme, A.J.A.M. A comparison of supervised, unsupervised and synthetic land use classification methods in the north of Iran. International Journal of Environmental Science and Technology 2015, 12, 1515–1526.
64. Krishna Bahadur, K.C. Improving landsat and irs image classification: Evaluation of unsupervised and supervised classification through band ratios and dem in a mountainous landscape in Nepal. Remote Sensing 2009, 1, 1257–1272.
65. Ismail, M.H. Evaluating supervised and unsupervised techniques for land cover mapping using remote sensing data. Geografia - Malaysian Journal of Society and Space 2009, 5.

 

#4. Lines 29-30: “However, our Combined map overestimates the extent of planted oil palm and is limited in interpreting land-use characteristics of oil palm plantations as compared to a combination of visual and automated approaches”

In my opinion, the manuscript did not obtain this conclusion.

Response: We agree with the reviewer that this sentence in the Abstract was incorrect because Miettinen et al’s land-cover maps did not employ an automatic procedure. There is also a lack of clarity in the second half of the sentence. We have replaced this sentence to the following to improve clarity:

L27-31: “The amount of oil palm land-cover in our Combined map is closer to official government statistics than the two existing land-cover products that used visual interpretation techniques. Our analysis of the extents of disagreement in oil palm land-cover indicated our map has comparable accuracy to one of them and higher accuracy than the other.”

 

#5. According to the analysis in Introduction, the combined data will better than radar-only and optical-only. Then their comparisons could not provide any more results and conclusions. Apparently, combined map will outperform the other two. Maybe some other machine learning algorithms should be used for comparison.

Response: We thank the reviewer for this comment. We followed up with this comment by performing the classification using a CART classifier in Google Earth Engine. We obtained the highest accuracies once again under the Combined image compared to Landsat and Radar only images. We have included this information under our Supplementary Methods: “Supplementary Table 7”

 

#6. In the Section of study site, the location and imagery of the study areas should be provided.

Response: We thank the reviewer for this comment. For SAR imagery we used a seamless yearly composite created by Google Earth Engine based on the Japanese Space Agency (JAXA) PALSAR-2 data. We modified our manuscript:

L195-197: “Our 2015 SAR mosaic was obtained from the seamless yearly mosaic available in Google Earth Engine database, and clipped it around Sumatra, our area of interest.”

For Landsat, we listed the Landsat 8 images we used in the Supplementary Methods (Supplementary Table 6) and we specified in the manuscript the number and references of images we used in our image composite.

L169-171: “To remove cloud cover in Landsat images, Landsat image composites were based on time series where the best available observation from 40 Landsat 8 images of the path 127 and rows 59 and 60 from the year 2015 were selected through pixel-based image compositing (Figure 2).”

 

#7. In the manuscript, six land-cover classes were defined. Then, did the maps for comparison have the same land-cover classes?

Response: The reviewer is correct – we defined six land-cover classes. However, because the focus of our paper is on fusing Landsat and SAR to classify oil-palm, we compared our oil-palm classification with the oil-palm land-cover in two existing land-cover mapping products (Miettinen2016 and Austin2017). To do this, we sampled extents of disagreement in oil-palm classification between our Combined map and each of the other maps (Miettinen2016 and Austin2017). As explained in L326, extents of disagreement are pixels classified as oil-palm in one of the two maps in each comparison but not the other map. We then randomly selected 100 points of these points, and recorded the “true” land-cover based on high resolution Google Earth Pro imagery.

 

#8. In the Section of accuracy assessments, the manuscript mentioned that statistical tests were used. However, there were no information for the results of statistical tests!

Response: We deleted “statistical tests”, the manuscript now reads:

L293-294: “We evaluated our classification results using standard accuracy assessment metrics such as the error matrix, overall accuracy, and user’s, and producer’s accuracies.”

 

#9. The section of conclusions did not contain the main contents of the manuscript and should be revised.

Response: We added the following line in the conclusion to highlight the result we obtained for combining radar and optical data:

L493-495: “Specifically, our study highlights the advantages of combining radar and optical data to improve the accuracy of oil palm land-cover classification.”

 

Reviewer 3 Report

See attached file

Comments for author File: Comments.pdf

Author Response

#1. For this paper I have to remark the fact that authors must cite relevant works concerning the polarimetric SAR classification like PolSAR, PolInSAR and MCA-PolInSAR, this to give an answer to this question given by the author: (i) Does the use of radar and optical imagery for oil palm classification outperform radar-only and optical-only classification? Give evidence of general problems of miss-classification when using only SAR images for oil-palm classification.

Response: We added two more references on past studies which used polarimetric SAR classification to map oil palm:

L82-83: “Past studies have successfully used radar data to detect new fronts of deforestation [25], as well as tree crop plantations in particular oil palm [17,18,26,27]”

[26] Dong, X.; Quegan, S.; Yumiko, U.; Hu, C.; Zeng, T. Feasibility Study of C- and L-band SAR Time Series Data in Tracking Indonesian Plantation and Natural Forest Cover Changes. IEEE Journal of selected topics in applied earth observations and remote sensing 2015, 8, 3692–3699.

[27] Ramdani, F. Recent expansion of oil palm plantation in the most eastern part of Indonesia: feature extraction with polarimetric SAR. International Journal of Remote Sensing 2019, 40, 7371–7388.

We also included the following line on the general problem of misclassification while using SAR images:

L448-449: “In general, classification using SAR data overlook new oil palm plantation because they have similar backscatter as herbaceous vegetation and bare land [17].”

 

#2. Figure 1 is the overall work-flow for mapping oil palm plantations in Riau, Jambi and South Sumatra. I please ask authors to give more information about this computational scheme. It is not clear.

Response: Thank you for your comment. The overall workflow summarizes the process we went through to create and validate our map. Computations were carried on in Google Earth Engine using the built-in functions of the software. We modified the legend for Figure 1 to provide a brief explanation of our process.

L163-166: “Figure 1. Overall workflow for mapping oil palm plantations in Riau, Jambi and South Sumatra which includes pre-processing of our PALSAR-2 and Landsat 8 images, image classification using a random forest algorithm, developing accuracy assessments for our oil palm land cover map and comparing our oil palm land cover map with two other oil palm map sources.”

 

#3. Figure 2: insert the reference system (Lat, Lon or range, azimuth)?

Response: Thank you for pointing this out. We have modified the Figure 2 to include the latitude and longitude (L190). We have also added the location in the legend of Figure 2: “Kampar, Riau province, 101°25’00” E 0.50°00’00” N” (L192).

 

#4. Give exact information about the record numbers of SAR and optical images you used for performing all the research work.

Response: We thank the reviewer for this comment. For SAR imagery we used a seamless yearly composite created by Google Earth Engine based on the Japanese Space Agency (JAXA) PALSAR-2 data. We modified our manuscript:

L195-197: “Our 2015 SAR mosaic was obtained from the seamless yearly mosaic available in Google Earth Engine database, and clipped it around Sumatra, our area of interest.”

For Landsat, we listed the Landsat 8 images we used in the Supplementary Methods (Supplementary Table 6) and we specified in the manuscript the number and references of images we used in our image composite.

L169-171: “To remove cloud cover in Landsat images, Landsat image composites were based on time series where the best available observation from 40 Landsat 8 images of the path 127 and rows 59 and 60 from the year 2015 were selected through pixel-based image compositing (Figure 2).” 

 

#5. In the experimental results it is not clear what you have found. Figure 3 (a) classification result is different from Figure 3 (b) and Figure 3 (c) explain better why.

Response: The three inset maps you can see in Figure 3 are the result of three different semi-automated classifications - one combining both radar and optical images (A), one using only radar image (B) and one using only optical image (C). The results are different because each imagery product detects oil palm and other land-cover with different performance. The performance of each classification is quantified in Table 3 with Users, Producers and Overall accuracies. We added the following to better explain our results:

L354-356: “Overall, combining SAR and Landsat imageries outperforms single imagery (SAR-only and Landsat-only) classification in mapping oil-palm cover due to the superior producer‘s and user’s accuracy of the former technique (Table 3).”

 

#6. Give a graphical representation about the ground truth and give information about the false alarm, miss, correct rejection and hit probabilities

Response: We thank the reviewer for this comment, however, we believe our error matrices under Figure 4 and the Producer and User accuracies under Table 3 adequately represent the ‘false alarm, miss, correct rejection and hit probabilities’ requested by the reviewer.

Round 2

Reviewer 1 Report

1. In table 3, the classification accuracy for SAR data is as low as 27% and 24%. However, if very discriminative features are extracted or selected, much better results could be obtained. There are many literatures focusing on PolSAR data classification and achieves high accuracy. Therefore, I am doubting the fairness of the research design. Please elaborate it more. 

2. I still think the conclusion that "a combination of optical and radar data outperforms the use of optical-only or  radar-only datasets" is not the main novelty. Some descriptions should revised in the abstract and conclusion to highlight the innonvation work. 

Reviewer 2 Report

See the attatched file.

Comments for author File: Comments.docx

Reviewer 3 Report

Dear authors,

in the present form, this paper can be accepted.

Good work!