A Parallel-Cascaded Ensemble of Machine Learning Models for Crop Type Classification in Google Earth Engine Using Multi-Temporal Sentinel-1/2 and Landsat-8/9 Remote Sensing Data

Round 1

Reviewer 1 Report

I congratulate you on the originality and design of the work. My only concern about the manuscript is the location of the tables and figures. I think you should reconsider the journal's rules on this subject. Also, the tables are not compatible with each other. Table 4 is different from the others. Images should be re-examined.

Author Response

Dear Reviewer 1,

Thank you for dedicating your time and consideration to reviewing our manuscript. We express our gratitude for your valuable feedback and thoughtful evaluation. We have taken your comments into careful consideration and made the necessary revisions. We have re-examined all the figures and tables, ensuring that they now adhere to a consistent format. Additionally, we have double-checked their placement within the manuscript to ensure accuracy. Once again, we sincerely appreciate your efforts in reviewing our work.

Sincerely yours,

Authors

Reviewer 2 Report

Please see the attached file.

Comments for author File: Comments.pdf

Author Response

Reviewer 2:

In this manuscript, a novel Parallel-Cascaded ensemble structure (Pa-PCA-Ca) with seven target classes for crop type classification was proposed in Google Earth Engine (GEE). It’s an interesting research and has important outcomes however needs to be improved for better understanding. I have some comments following:

Authors' Response:

We sincerely appreciate your time and effort in reviewing our manuscript. We would like to express our gratitude for your valuable feedback and thoughtful evaluation. We have carefully considered all your comments and made the necessary revisions. Please find our point-to-point response to your comments:

1. What do you mean by “another ML model” in the abstract? You have not addressed any many learning models such as RF or SVM and it’s not clear what another means in this context.

Authors' Response:

Thank you for your comment. As mentioned in the paper, our proposed method involves two levels of machine learning models for classification: base models and a Meta-model. The outputs of the base models, after applying PCA, are then fed into the Meta-model. Therefore, when we refer to "another ML model," we are specifically referring to the Meta-model. We have provided additional clarification in the manuscript based on your comment. It is worth highlighting that the selection of both the base models and the Meta-model involved a 5-fold cross-validation on the training data, considering RF, SVM, CART, and GBT as options.

2. Why did you use VIs for the classification process? If this is a need for accurate crop type classification, please clearly mention it in the manuscript.

Authors' Response:

We appreciate your comment. As indicated in the initial version of the manuscript (now highlighted in lines 285-300), we have provided a clear explanation as to why we considered using spectral indices. The reason for their inclusion is that spectral indices have proven to be effective in previous studies for crop type classification, as they are designed to highlight specific land objects. Therefore, their utilization can improve classification accuracy, as demonstrated in our article.

3. Regarding the SAR/PolSAR data for crop type mapping, please check the following papers:
Tamiminia, H., Homayouni, S., McNairn, H., & Safari, A. (2017). A particle swarm optimized kernel-based clustering method for crop mapping from multi-temporal polarimetric L-band SAR observations. International Journal of Applied Earth Observation and Geoinformation, 58, 201-212.
Ustuner, M.; Balik Sanli, F. Polarimetric Target Decompositions and Light Gradient Boosting Machine for Crop Classification: A Comparative Evaluation. ISPRS Int. J. Geo-Inf. 2019, 8, 97.

Authors' Response:

Thank you for your comment. The papers you mentioned have made valuable contributions to PolSAR classification of crops. However, as you rightly pointed out, these papers utilized PolSAR data and different decomposition techniques for crop type classification. In our study, we used S1 data, which are not polarimetric (only VV and VH bands), and we did not employ any decomposition techniques. The two mentioned papers were cited in the manuscript to enrich the literature. Thank you.

4. The authors mentioned that “there is a limited number of studies that have classified crop types using a multi-source combination of multi-temporal S1/2 and L8/9 data”. This is because, apart from the spatial resolution, both multispectral data might cover almost same spectral region for particular (red, green, blue, NIR, SWIR) bands therefore it is understood that people could choose one of these sensors rather than choosing both. Please specify or clearly indicate what differs S2 from L8 except RedEdge Bands and higher spatial resolution? Otherwise, abundant information might overfit the ML model? Please
see the following paper:
Gerald Forkuor, Kangbeni Dimobe, Idriss Serme & Jerome Ebagnerin Tondoh (2018) Landsat-8 vs. Sentinel-2: examining the added value of sentinel-2’s red-edge bands to land-use and land-cover mapping in Burkina Faso, GIScience & Remote Sensing, 55:3, 331-354, DOI: 10.1080/15481603.2017.1370169

Authors' Response:

We appreciate your comment. You are correct in noting that S2 and L8/9 primarily acquire data in the same spectral bands. However, since they capture data at different times, the simultaneous use of both datasets can increase the density of the time series. This denser time series allows for a better consideration of the phenological characteristics of crops, resulting in improved accuracy. Additionally, in our proposed method, each S2 and L8/9 dataset is used in a separate branch. Therefore, there is no redundant information within each branch. We have provided further explanations in the introduction, which are highlighted in lines 60-65. Your suggested paper is very useful and interesting, and we cited it in the manuscript.

5. It seems that Spectral Indices (SIs) enhanced the classification accuracy. That’s great but it would be better to check and demonstrate the “feature ranking” among SIs. It’s not clear which spectral indices contributes more than another. Please check the following papers:
Koley, S., & Chockalingam, J. (2022). Sentinel 1 and Sentinel 2 for cropland mapping with special emphasis on the usability of textural and vegetation indices. Advances in Space Research, 69(4), 1768-1785.

Authors' Response:

Thank you for your comment and suggesting this interesting article which we carefully reviewed it and added it to our references. In our manuscript, we propose a novel Pa-PCA-Ca model that incorporates five different feature collections in each branch. Among these five branches, two branches are related to spectral indices (one for S2 and one for L8/9). The output probability maps from each branch are then transformed using PCA, and the top components of each class are fed into the Meta-model. The final classification is a result of the cumulative effect of all five branches and the PCA transformation. Therefore, the spectral indices do not have a direct impact on the classification, and the classification is performed based on the cumulative information from all branches, rather than solely relying on the spectral indices. As a result, the feature importance and ranking of spectral indices may not provide any additional useful information since there are other data sources considered as well. Moreover, the spectral indices we utilized are among the most widely used indices in the remote sensing community, and their importance has been mentioned in previous papers. Furthermore, the effect of incorporating spectral indices was discussed in the results section (Figure 6), highlighting their significance in crop classification.

Once again, we sincerely thank you for your valuable feedback, which has greatly contributed to the improvement of our manuscript.

Reviewer 3 Report

Dear authors:

All my comments are attached.

Comments for author File: Comments.pdf

Author Response

Reviewer 3.

The study focuses on crop classification based on time-series high-resolution
remote sensing images. The authors propose a technical framework similar to ensemble learning, the proposed ensemble framework achieved better performance.

Authors' Response:

Dear Reviewer,

We sincerely appreciate the time and effort you have dedicated to reviewing our manuscript. On behalf of all the authors, we would like to express our gratitude for your valuable feedback and thoughtful evaluation. We have carefully considered your comments and have made the necessary revisions. Please find our point-by-point response to your comments below:

1. List the number of features in Feature Collection 1-5 in Table 3.

Authors' Response:

Thank you for your comment. We have made the changes to the manuscript, and these revisions are now highlighted in Table 3.

2. Hyper-parameter tuning is a necessary task for machine learning models. To ensure the completeness of research results, the final confirmed hyper-parameter values for grid search should be listed in Table 4, rather than just providing candidate values.

Authors' Response:

We appreciate your positive comment. We have added detailed information in the Results section (lines 430-435 and 450) that describes the optimum values of the hyperparameters for the Random Forest (RF) model. This model was found to be the best in all branches of the Pa structure and also the best Meta-model in the CA structure.

3. In section 4.3, as the number of features increases, the classification performance shows a gradually improving trend, which seems too perfect. Previous studies have shown that as the number of features increases, the model performance gradually deteriorates, especially in cases with a large number of features. The authors need to provide sufficient evidence to demonstrate that the "Hughes" phenomenon does not occur without using PCA.

Authors' Response:

Thank you for your positive comment. You are absolutely correct. However, it is important to note that our method involves the development of a Pa-PCA-Ca approach. In the Pa section, five branches of RS data sources are used separately, and they are not stacked together to create a high-dimensional dataset. Furthermore, in the Ca stage, Principal Component Analysis (PCA) is implemented to generate uncorrelated features. As a result, our method is robust against the Hughes phenomenon. The impact of Hughes can be observed in Figure 11, where the case of "Stacked features without PCA" shows that when all the features are stacked together and fed to a machine learning model, the correlated features affect the performance and result in lower accuracy. We have added more relevant information to the manuscript (discussion, lines 615-620).

4. The discussion section can consider compression, only discussing content related to this study.

Authors' Response:

Thank you for your comment. We have thoroughly reviewed the discussion section and have made sure to only include information that is directly related to the study. The Discussion part was divided into two different sections for being more comfortable to study by readers. We believe that in the current version, all the information is relevant and necessary. Thank you once again.

Once again, we sincerely appreciate your efforts in reviewing our work.