Dense Time Series of Harmonized Landsat Sentinel-2 and Ensemble Machine Learning to Map Coffee Production Stages

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors' work on coffee mapping using HLS data combined with machine learning algorithms is highly meaningful, and the manuscript is well-written with rigorous logic. However, I have the following concerns:
(1) Vegetation cover mapping via machine learning is a common approach; the authors should clarify the novelty of this study.
(2) The study conducts vegetation mapping at four levels, but only Level 4 serves as the final objective. Are the preceding three levels redundant?
(3) The sampling methodology requires more detailed descriptions to facilitate reproducibility and reference for related studies.

Additionally, specific comments are provided below:

1. Lines 155–156: How were the "remote inspection" and "ground survey" conducted?

2. Line 226: What do these numerical values represent?

3. Line 255: How were GLCM features extracted? Specify the window size and domain size. Moreover, GLCM textures are typically extracted from high-resolution imagery (e.g., UAV/ground observations). Explain the validity of using 30m-resolution data for texture analysis.

4. Line 222: Which algorithm was used for downscaling? Clarify the difference between HLS.L30 and HLS.S30.

5. Section 3.1: This section describes remote sensing data and should be relocated to the Methods section.

Author Response

Please refer to the following documents.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This work proposes a method for mapping coffee crop stages utilizing rich satellite images and ML.  The combination of Harmonized Landsat-Sentinel (HLS) data and ensemble learning algorithms (RF and XGBoost) demonstrates methodological strength, particularly with high sensitivity in coffee detection (>95%).  However, the study's classification performance for the stumping stage (62%) reveals a significant deficiency that requires more investigation, possibly due to a lack of ground truth data or spectral similarities with previous stages.  While the hierarchical classification approach is admirable, the manuscript could use a more in-depth discussion of model generalizability to various areas and cropping systems.  Overall, this work provides useful tools for climate-resilient crop monitoring; nevertheless, stage separation, figures’ quality and validation procedures could be improved, and some suggestions and questions can further improve the quality of the manuscript.

Specific comments:

Please clearly state the objectives at the end of the introduction section, as a, b, and c.

Sentinel-2A and 2B, which one was used?

The ground truth samples are very limited, it makes the study vague.

Fig. 6, 7: Improve the quality and color combinations.

The manuscript makes frequent use of the pronoun “we,” which, while acceptable in scientific writing, appears somewhat repetitive in its current form. I recommend reducing its usage where possible by rephrasing sentences in a more objective or passive tone to improve the flow and maintain a more formal scientific style. This adjustment will enhance the readability and professionalism of the manuscript.

What is the relevance of Fig, 8 in this study?

The results are enough and valid, but for write-up, please give in scientific manuscript look instead of thesis.

Fig. 13. Modify for clear visibility

All figures need improved color, clarity, and annotation improvements.

In discussion,

• Be concise and analytical, avoid restating results. Instead, interpret their significance in light of existing literature.
• Support key findings with references to highlight how your work aligns with or diverges from prior studies.

Concise the conclusion section.

Author Response

Please refer to the following documents.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors utilized machine learning with the HLS dataset to map different coffee production stages, which holds certain guiding significance for coffee production management and aligns with the journal's scope. However, the following issues require revision before further review can proceed:

(1) The article formatting is poor. There is excessive white space on the right side, causing the main text lines to be unnaturally narrow. This significantly hinders the review process.

(2) The article's organizational structure is problematic. Readers are generally less interested in the classification results at Level 1 and Level 2; presenting these results is unnecessary. Readers are primarily concerned with the classification evaluation results at Level 3 and Level 4. Including Level 1 and Level 2 results distracts readers.

(3) Text in many figures appears blurry, for example in Figures 1 and 7. The authors need to carefully check the clarity of text in every figure.

(4) In Figure 4, the observation samples for the various coffee stages at Level 4 are poorly displayed. Furthermore, the number of samples per stage appears low. Is this dataset truly sufficient for machine learning?

(5) The training data and test data are not clearly defined. There is no detailed explanation of which data constituted the training set and which the test set, nor how the data was partitioned.

(6) The accuracy for the ST (Senescent and Terminal) stage is relatively low. This needs to be discussed in detail within the discussion section, including potential reasons and directions for future improvement.

(7) The article mentions using both RF (Random Forest) and XGBoost for Level 4 classification, but Figure 12 does not provide classification maps for these two different methods.

(8) XGBoost Parameter Tuning: The grid search range was narrow (e.g., max_depth tested only 3 values), potentially limiting performance optimization.

(9) The potential application of deep learning methods (such as U-Net) for crop stage classification was not discussed.

Author Response

Please refer to the following documents.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

1. Introduction

The current citations related to coffee research are insufficient, with only two references cited between lines 63–66. It is recommended to expand the literature review by incorporating additional scholarly sources focusing on coffee yield assessment, nutritional status evaluation, pest and disease detection, and water stress analysis. Furthermore, when discussing heterogeneity in lines 71–78, it is advised to include relevant literature to substantiate the claims made.

The rationale for selecting the study area has not been adequately justified. To enhance the contextual background, the uniqueness of climatic conditions and topographical features of the region should be included to emphasize its representativeness.

2. Materials and Methods

In lines 226–228, the reference to a simple time linear interpolation method requires further clarification. Please specify the exact algorithm or procedure used for interpolation and provide details on how the accuracy of this method was validated.

3. Feature Space Combinations

The section title should be labeled as “2.4” to maintain consistency with the overall document formatting structure. Additionally, please clarify the specific parameters used for the GLCM textural metrics.

4. Feature Importance Analysis

There appears to be an inconsistency in the numbering of this section. A thorough review of the entire document’s section numbering sequence is recommended to ensure accuracy. In lines 543–547, while GNDVI demonstrates high importance within the Random Forest (RF) model, it does not significantly enhance accuracy at Level 4. Please elaborate on the potential reasons for this discrepancy.

5. Advantages of HLS Data for class Separability

Although the study highlights the impact of climate change on coffee production, the discussion section lacks an analysis of how specific climatic conditions in 2023—particularly the spatial and temporal distribution of precipitation—may have influenced classification accuracy across different growth stages. Since the model was developed using data from 2023, its applicability to other years also remains unaddressed and should be discussed.

6. Charts

The resolution of the figures is relatively low, and the annotations lack clarity. It is recommended to improve both the visual quality and labeling of all charts to ensure they meet publication standards.

Comments for author File: Comments.pdf

Author Response

Please refer to the following documents.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have well revised the manuscript.

Author Response

We thank the reviewer for the time dedicated to the work and for the valid suggestions.