Cropland Mapping in Tropical Smallholder Systems with Seasonally Stratified Sentinel-1 and Sentinel-2 Spectral and Textural Features

Round 1

Reviewer 1 Report

The manuscript "Cropland mapping in tropical smallholder systems with seasonally stratified Sentinel-1 and Sentinel-2 spectral and textural features" is submitted to the esteemed Remote Sensing journal. The authors use several datasets (i.e., from high to coarser optical remotely sensed data systems; and SAR) in Ghana's agricultural areas. 

Some challenges were explored and included giving insights for crop calendars and seasonality discussion. However, it would be interesting to justify better how and why datasets were used. In fact, please provide the reasons for better replicability and suggest further studies employing some specific dataset separability. 

Another aspect that would require some attention is the discussion section which contains very few references from the literature. Finally, there is a need to shorten some sentences and improve the quality of the figures for clarity. The manuscript could be recommended considering these issues above. 

Other comments:

L65: "almost" unaffected;

L110:  together with rigorous quality checks? (please detail better);

Figure 2: increase the font size and add subsections numbers for better visibility;

Figure 3: increase the size for better visibility;

L220: version of the software (check from previous manuscripts how to refere a software);

Figure 9: how would different datasets perform isolated? (SAR alone, Sentinel alone, etc; at least suggest for further studies);

Author Response

Dear Reviewer, 

Thank you for your detailed comments and valuable feedback. Please see my addressed edits for each of the points mentioned:

I have justified each sensor dataset used in the intro (123-130) and made modifications to the discussion section by adding a paragraph and making sentences simple. Further point-wise edits are below:

L65: "almost" unaffected; - Added the word in the manuscript

L110:  together with rigorous quality checks? (please detail better); - Added the sentence - "In this research, the process of ensuring quality is referred to as cross-checking consensus PS labels, which is performed by experts of the field, which is further discussed in section 3.5"

Figure 2: increase the font size and add subsections numbers for better visibility; - Change the figure

Figure 3: increase the size for better visibility; - Done

L220: version of the software (check from previous manuscripts how to refere a software); - added as - ENVI version 5.3 (Exelis Visual Information Solutions, Boulder, Colorado)

Figure 9: how would different datasets perform isolated? (SAR alone, Sentinel alone, etc; at least suggest for further studies); - Added sentences 481-486

Reviewer 2 Report

The authors make an efficient use of S1/S2 imagery for cropland mapping for smallholder farming in a seasonally stratified manner using polarimetric, spectral and textural features, with a focus on study of chosen test region in Ghana.

Overall remark:

Good work and detailed ground covered to fully explore and investigate the the topic of study with a novel classifier scheme mainly in terms of use of features from different parts of EM the spectrum.

A minor suggestion to include one or two lines about:

-what benefit radar vegetation indices could add to the future studies?

-why the random forest classifier was chosen as the preferred ML technique, over other more sophisticated deep learning techniques?

-how geography and different climatic conditions (say non-tropical) affect the generalization of this study to other areas?

Author Response

Dear reviewer, 

Thank you so much for your valuable comments; please see the detailed edits I made for each point below:

-what benefit radar vegetation indices could add to the future studies?

I have added a few lines (486-489) with references in the file justifying that dual-pol RVI also utilizes the degree of polarisation in addition to backscatter intensity, increasing sensitivities of S1 to crop growth and structure.  

-why the random forest classifier was chosen as the preferred ML technique, over other more sophisticated deep learning techniques?

A few lines are added (371-374) in the model training section, saying that due to easy interpretability, feature importance, computational efficiency, and Partial dependence plots, RF was chosen over other sophisticated methods. 

-how geography and different climatic conditions (say non-tropical) affect the generalization of this study to other areas?

I have mentioned that in the temperate region, which typically exhibits more pronounced seasonal variations with distinct growing and dormant periods, generalization of pixel-wise seasonal stratification may not require, and also the temperate climates, where larger agricultural holdings, organized field patterns, and climate-specific cultivation practices are prevalent the importance o texture and topography needs further verification. (lines 516-520; 538 - 543)

Reviewer 3 Report

This study utilizes MODIS, Sentinel-1 and 2 to classify arable field areas in the Eastern Region of Ghana. The paper was well organized and well written. The method and results are clear. What I am confused is the study try to improve the classification accuracy using MODIS data to stratify finer Sentinel-2 reflectance and Sentinel-1 backscatter image features on a per-pixel basis. However, I did not find the results support the objective. The variable importance did not show the MODIS played any role in the random forest model. Did I misunderstand？Section 4.1 seems to have little connection with other sections of the results. More information or explanation should be added. I think the paper is good if it can explain why they need MODIS data and how they improved the classification accuracy by MODIS stratify process.

Author Response

Dear reviewer, 

Thank you for your valuable comments and feedback.

MODIS was used as a dataset with a proven track record (Ali 2013) to capture annual seasonality utilizing the long-term average phenology (largely climate-driven), which tends to change more gradually over space than annual seasonality. The power of this technique was to capitalize on the advantages of high/low spatial/temporal information. The MODIS disaggregated stats which is a separate process (section 3.3) than seasonal stratification, which claim to be very important in the literature, did not show up in RF feature importance because finer S1 +S2 features were included in the paper (discussion lines 519-543). Even though landscape stratification was not the primary goal, it was observed that it indeed captured the different seasonality of the region, which in future studies needs to be incorporated to capture better wet and dry differences. I did make some changes in the intro (124-130) and in the discussion (538 - 545) regarding the importance of stratification in different climates, which may clarify more about the stratification process.

 

Do let me know if I need to add or remove anything.