Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript addresses an important gap in mapping diversified farming systems in tropical regions, focusing on the Ribeira Valley, Brazil. The integration of Sentinel-2 dense time series, gap-filling, and hierarchical classification provides novel and valuable insights. The study is well-structured and methodologically sound, but several points could be improved:

Introduction

• What is the primary methodological challenge: persistent cloud cover or crop heterogeneity? The introduction currently mixes both without a clear hierarchy.

Methods

• How do the authors ensure that the fixed 15-day interpolation interval does not introduce bias during periods of rapid phenological change?
• Why was the Daubechies (db4) wavelet chosen for gap-filling? Were other smoothing approaches (e.g., Savitzky–Golay, spline) considered?

• Is the number of field and remote samples (83 banana, 74 peach palm, 100 non-perennial crops) sufficient for training a Random Forest model with more than 70 temporal variables?

Results

• Why did NDWI alone outperform combined indices at Level 2? Could the authors provide a technical explanation for this result?

Discussion

• How transferable is this hierarchical classification approach to other regions with diversified smallholder agriculture?

• The Sete Barras (Rio Preto) example is illustrative but takes up much space. Could this be streamlined and tied more directly to the methodological contribution?

Comments for author File: Comments.pdf

Author Response

What is the primary methodological challenge: persistent cloud cover or crop heterogeneity? The introduction currently mixes both without a clear hierarchy.

Response: We sincerely thank you for the attention dedicated to our work and for your precise observation. We recognize that both the excessive cloud cover and the heterogeneity and complexity of the study area are equally relevant challenges. This justifies our choice of Sentinel- 2 imagery, as it provides the most suitable spatial resolution to address the heterogeneity and non-structured patterns typical of the region. In addition, its frequent revisit time increases the likelihood of acquiring usable imagery throughout the year. Accordingly, we have reorganized the paragraph in lines 71-78 of the Introduction to make this argumentation more straightforward.

“In recent years, Sentinel-2 imagery has proven effective for monitoring heterogeneous and fragmented tropical landscapes [22]. Its high spatial resolution (10–20 m), short revisit interval (up to five days), and rich spectral range make it especially suitable for detecting land cover dynamics even under persistent cloud cover [23-24]. However, the limitations of conventional mapping methods remain evident. For instance, in the Ribeira Valley, one of Brazil’s leading banana-producing regions, national initiatives such as the MapBiomas Project failed to correctly classify perennial crop areas, labeling them mostly as “Mosaic of Uses” [25-26].

Methods

• How do the authors ensure that the fixed 15-day interpolation interval does not introduce bias during periods of rapid phenological change?

Response: Dear Reviewer, thank you for this valid observation. We agree that resampling Sentinel-2 data to 15-day intervals directly affects the ability to capture rapid phenological changes. To address this issue, this choice was therefore explicitly justified in the Methodology section (lines 231-241), along with a discussion of the associated trade-off (lines 499-501), as shown below.

“The choice of a 15-day resampling interval was motivated by both environmental and methodological considerations. First, due to the climatic characteristics of the study area, marked by persistent cloud cover, it is not feasible to reliably generate cloud-free composites at 5-day intervals. Second, increasing temporal resolution also expands the dimensionality of the feature space; in order to avoid overfitting and maintain a balanced signal-to-noise ratio given the available sample size, a 15-day interval was adopted. Third, the region does not exhibit extensive multi-cropping systems; instead, land use and cover are largely stable throughout the year, dominated by forests, perennial crops, and pastures, all favored by high water availability. Finally, it was previously demonstrated that even with bimonthly intervals, Sentinel-2 imagery can yield agricultural classifications with error rates below 25% in tropical environments [37].

“Adopting a 15-day temporal resolution reduced feature dimensionality and stabilized the model, but it also limited detection of rapid phenological changes [67-68]”.

• Why was the Daubechies (db4) wavelet chosen for gap-filling? Were other smoothing approaches (e.g., Savitzky–Golay, spline) considered?

Response: Important note, thank you. To address your concerns, we briefly justify our choice of method between lines 252-256 in section 2.4 of Materials and Methods.

“DWT was chosen over filters such as Savitzky–Golay because it more effectively suppresses high-frequency noise while better preserving the original phenological curve, particularly when employing db4 (i.e., four coefficients), which provides an intermediate level of smoothing [39]”.

• Is the number of field and remote samples (83 banana, 74 peach palm, 100 non-perennial crops) sufficient for training a Random Forest model with more than 70 temporal variables?

Response: We acknowledge that the Level 3 sample dataset is relatively limited. However, the production plots were visited in the field with the assistance of local producers, and any remote supplementation was applied only where land-use classification could be confirmed with certainty. Despite the limited sample size, the models achieved auspicious performance, as demonstrated by the five-fold cross-validation. Additionally, we ensured a minimum of 20 validation samples for each class (banana and hearts of palm), which exceeds the sample sizes reported in many comparable studies. We have explicitly acknowledged this limitation and its implications in the Discussion section (Please lines 527-529).

Results

• Why did NDWI alone outperform combined indices at Level 2? Could the authors provide a technical explanation for this result?

Response: Thank you for this critical observation. To enhance understanding of the relevance of NDWI, particularly at Level 2, we have developed a discussion paragraph that includes a separability analysis. Please see lines 417-424.

“At Level 2 (perennials vs. non-perennials), NDWI alone delivered 89% accuracy, with commission and omission errors between 5% and 16%. ROC analysis confirmed its superior class separability (mean AUC = 0.837), clearly outperforming NDVI (AUC = 0.751) and BSI (AUC = 0.169). This reflects the ecological setting: in humid, rainforest-dominated landscapes, perennial crops maintain stable canopy water content detectable by NDWI, whereas NDVI tends to saturate under high biomass, and BSI contributes little under dense cover. Future research may explore NDWI in combination with indices less prone to saturation, such as EVI [47]”.

Discussion

• How transferable is this hierarchical classification approach to other regions with diversified smallholder agriculture?

Response: To address this relevant issue, we have added a paragraph in the discussion, between lines 523-527.

“Transferring hierarchical schemes to other regions depends on prior knowledge of local conditions. Broad categories (e.g., vegetated vs. non-vegetated) are generally transferable across different contexts. However, finer distinctions require contextual understanding of plot size, crop rotations, intercropping, and management practices [49-50]”.

• The Sete Barras (Rio Preto) example is illustrative but takes up much space. Could this be streamlined and tied more directly to the methodological contribution?

Response: Dear reviewer, during the restructuring of the discussion and methodology sections, suggested by the other reviewers, the example of Sete Barras was removed from the text.

Reviewer 2 Report

Comments and Suggestions for Authors

Overall comments:

Soares et al. employed a random forest based hierarchical classification to classify and map anthropogenic agricultural areas and native vegetation (Level 1), permanent crop lands and non-permanent crop lands from the agricultural areas (Level 2), and banana and peach palm farms from the permanent crop lands (Level 3), in the Brazilian Atlantic Forest region using multi-temporal data from the Sentinel-2 satellite images. Their results showed a good performance in the hierarchical classification for the three levels. Overall, the manuscript is well-written and structured. I have some comments below for the author’s consideration.

 

Specific comments:

1. Line45, “this same spatial diversity” is unclear to me. Is it something referred to in previous paragraphs?
2. Line108, Figure 1. Could you add the legend of the color for the DEM? It looks like a RGB composition imagery rather than a DEM showing the terrains to me.
3. Line111, “This is a region where nature still speaks loudly”, I don’t think this is a proper writing that should be in a scientific paper.
4. Just wondering if Figure 1 and Figure 2 can be combined together.
5. Line253, could you explain this sentence: “Each point was linked to 25 values per index, corresponding to 15-day intervals”. What are the 25 values?
6. Line336, Figure 7, I recommend reversing the y-axis so that the most important variables appear at the top. In addition, looking at Figure 7, I am unsure how the authors determined the best VI combinations in Table 1. For example, VIs should normally be associated with specific dates, yet Table 1 presents the combinations regardless of dates. Clarification on this point would be helpful.
7. Does error propagation at each level of the hierarchical classification affect overall classification performance? The authors may consider providing justification or discussion on this point.

 

Author Response

Line 45, “this same spatial diversity” is unclear to me. Is it something referred to in previous paragraphs?

Response: In fact, this expression became generic and had no direct connection to the previous paragraph. Therefore, we reworded it to maintain textual cohesion.

“Nevertheless, mapping tropical agroforestry systems with orbital remote sensing remains technically challenging due to their spatial heterogeneity [18]. The coexistence of multiple species within small properties, overlapping canopies surrounded by native forests, and similar phenological cycles often result in spectral confusion, particularly in medium-resolution sensors such as Sentinel-2 [19-21]. (Lines 66-70)

• Line108, Figure 1. Could you add the legend of the color for the DEM? It looks like a RGB composition imagery rather than a DEM showing the terrains to me.

Response: Thank you very much for noticing this mistake. The caption of Figure 1B was inadvertently incorrect. Panel B does not show a Digital Elevation Model (DEM); instead, it presents an RGB composite of the Jacupiranga municipality created in Google Earth Engine to facilitate visualization of terrain and land cover. The caption has been corrected accordingly:

Figure 1. (A) Location of the study area, which corresponds to the municipality of Jacupiranga, São Paulo, Brazil. (B) RGB composite of the study area created using Copernicus data in the Google Earth Engine to facilitate visualization of terrain and land cover.

• Line111, “This is a region where nature still speaks loudly”, I don’t think this is a proper writing that should be in a scientific paper.

Response: Thank you for this important observation. The sentence in question has been reworded to meet academic standards. Please see the Introduction section.

• Just wondering if Figure 1 and Figure 2 can be combined together.
Response: We agree that the figures complement each other and therefore we unite them as a

single element (Figure 1).

• Line 253, could you explain this sentence: “Each point was linked to 25 values per index, corresponding to 15-day intervals”. What are the 25 values?

Response: Dear reviewer, we have rewritten this section to clarify its content (please see lines 298-302).

“Raster values were extracted for each sample point using geographic coordinates, projected to the raster CRS. For every vegetation index, each point was associated with 25 consecutive time-series values obtained at 15-day intervals, representing its complete temporal profile. These temporal vectors were then organized into .csv files for each classification level”.

• Line 336, Figure 7, I recommend reversing the y-axis so that the most important variables appear at the top. In addition, looking at Figure 7, I am unsure how the authors determined the best VI combinations in Table 1. For example, VIs should normally be associated with specific dates, yet Table 1 presents the combinations regardless of dates. Clarification on this point would be helpful.

Response: The authors thank the reviewer for the helpful suggestion. Figure 7 has been updated by reversing the y-axis so that the most important variables appear at the top, thereby improving readability. As for Table 1, it reports the overall performance of Random Forest models trained with entire annual time-series stacks of each VI combination (NDVI, NDWI, BSI, and their possible combinations) for 2024, after temporal interpolation and smoothing. These combinations were evaluated as complete temporal signatures, not as single-date inputs, which is why specific acquisition dates are not listed in the table. To complement this aggregated evaluation, Figure 7 shows the variable importance of the best-performing model at each classification level (NDVI+NDWI+BSI for Levels 1 and 3; NDWI only for Level 2); where the individual VIs are labeled with their corresponding dates. This figure illustrates which specific dates within the full time-series contributed most to the classifier’s performance, clarifying the temporal dimension absent from Table 1. To clarify this issue for readers and avoid confusion, we have changed the mention of the analyzed combinations in the materials and methods section, as well as the caption of Figure 7.

In section 2.5: “Seven vegetation index (VI) combinations were tested at each classification level using their complete annual time-series stacks (NDVI, NDWI, and BSI individually; NDVI + NDWI, NDVI + BSI, BSI + NDWI, and NDVI + NDWI + BSI). Each stack contained all 2024 acquisition dates after interpolation and smoothing, providing a full temporal profile for model training” (lines 304-308).

Figure 7. Top 20 most important variables (identified by acquisition dates) based on Mean Decrease Accuracy (MDA) for the best-performing model at each classification level: NDVI + NDWI + BSI full time series for Levels 1 and 3, and NDWI full time series for Level 2.

In section

• Does error propagation at each level of the hierarchical classification affect overall classification performance? The authors may consider providing justification or discussion on this point.

Response: Thank you for the suggestion. Indeed, hierarchical classification systems are susceptible to top-down error propagation, which goes undetected by evaluation metrics. To address this limitation of the method, we've added a paragraph to the discussion section (lines 517-522).

“Hierarchical classification systems, despite their benefits, are vulnerable to error propagation [75-76]. Misclassifications at higher levels prevent accurate assignment at lower levels. Grouping spectrally similar classes and

limiting hierarchy depth mitigated this effect. In this study, we deliberately avoided using deep or overly complex hierarchies, opting instead to group spectrally similar classes. Nevertheless, some degree of error propagation is an expected outcome of the approach”.

Reviewer 3 Report

Comments and Suggestions for Authors

Title: Mapping Diversified Farming Systems in the Brazilian Atlantic Forest Using Multitemporal Remote Sensing.

Overall, the article addresses the important and timely issue of mapping agricultural systems in a complex tropical environment using Sentinel-2 satellite data and a Random Forest classifier. The authors present a valuable and proven methodological approach to distinguishing perennial crops in agricultural mosaics. The work is largely consistent with the title and abstract, and the results obtained are relevant to the development of digital agriculture, spatial planning and sustainable development.

The main objective of the work is achieved reliably, but section 4.2, The importance of diversification for agricultural resilience, goes beyond the empirical material presented. The methodology (lines 133–139) clearly shows that banana monoculture dominates in the study area, which does not justify broad considerations of crop biodiversity in the context of agricultural resilience. This type of argument, unsupported by data from the studied area, may be misleading, and I suggest shortening it or moving it to the introductory/discussion section as background literature.

In the research methodology, it is recommended to expand the description of the procedure for interpolating missing pixels. It is worth indicating:  The scale of the phenomenon (percentage of missing data),  Criteria for selecting adjacent pixels or images,  The impact of this operation on the final quality of the classification.

Standardisation of the method of recording geographical coordinates: the methodology (line 97) uses the degrees-minutes-seconds format, while the figures (Figs. 1, 2, 6, 8) use degrees-minutes. For clarity and consistency, a uniform format should be used.

The results are generally presented correctly and are legible. A few minor editorial errors do not significantly affect the perception of the data.

The discussion needs to be rewritten to better align with the aim and title of the paper. The conclusions directly resulting from the spatial classification of banana and whip palm cultivation should be highlighted, rather than focusing on general considerations about agricultural diversification.

The conclusions should be expanded and more strongly related to the presented results. Emphasising the importance of the methods used (multi-temporal analysis, Random Forest) for future research and agricultural practice in tropical regions will strengthen the message of the work.

Overall, the article makes an important contribution to research on the use of remote sensing in tropical agriculture. After implementing the suggested changes – in particular, clarifying the methodology, standardising the notation and revising the chapter on diversification – the work will be a valuable publication that may be of interest to both GIS and remote sensing specialists and researchers involved in sustainable agricultural development.

Author Response

The main objective of the work is achieved reliably, but section 4.2, The importance of diversification for agricultural resilience, goes beyond the empirical material presented. The methodology (lines 133–139) clearly shows that banana monoculture dominates in the study area, which does not justify broad considerations of crop biodiversity in the context of agricultural resilience. This type of argument, unsupported by data from the studied area, may be misleading, and I suggest shortening it or moving it to the introductory/discussion section as background literature.

Response: Dear reviewer, we agree with your assessment that the way the biodiversity issue was discussed led the reader to a different understanding than intended. Our objective was not to highlight crop diversity, but rather the biodiversity context inherent to the region, which preserves its native forest, as well as the heterogeneity typical of Brazilian family farming systems. Therefore, we removed the indicated section from the discussion. Part of its content was incorporated into the introduction and methodology.

• In the research methodology, it is recommended to expand the description of the procedure for interpolating missing pixels. It is worth indicating: The scale of the phenomenon (percentage of missing data), Criteria for selecting adjacent pixels or images, The impact of this operation on the final quality of the classification.

Response: Thank you for the observation. To clarify the time-series interpolation procedure, we have revised the corresponding section in the methodology, incorporating the requested information. Please see lines 210-130.

“Due to cloud contamination and acquisition gaps in the Sentinel-2 time series, several missing observations were identified within the annual index stacks. On average, 67.8% of observations were missing across the study area in 2024. These gaps were addressed exclusively in the temporal domain. For each pixel, a one-dimensional linear interpolation was applied to its time series using the scipy interp1d function, based on the Julian day integers of the original acquisition dates. This procedure connects the

nearest valid observations before and after each missing point, ensuring temporal consistency without resorting to spatial neighbors or external images.

To guarantee full annual coverage, each time series was resampled to a regular 15-day interval from January 1st to December 31st, thereby producing 25 observations per index and reducing data dimensionality. Pixels with fewer than two valid input observations were flagged and excluded to prevent unreliable extrapolation. To further improve the quality of the reconstructed series, a discrete wavelet transform (db4) with soft thresholding was applied to the interpolated profiles, attenuating high-frequency noise while preserving seasonal trends. The final output consisted of cloud-free and temporally smoothed stacks of NDVI, NDWI, and BSI for 2024, stored as raster series at 15-day intervals. This preprocessing step reduces the influence of missing data and noise on the subsequent classification, although it may slightly attenuate abrupt short-term changes. All metadata and Python notebooks with the code for these procedures are available to the reader.

• Standardisation of the method of recording geographical coordinates: the methodology (line 97) uses the degrees-minutes-seconds format, while the figures (Figs. 1, 2, 6, 8) use degrees-minutes. For clarity and consistency, a uniform format should be used.

Response: Coordinate standards have been standardized.

• The results are generally presented correctly and are legible. A few minor editorial errors do not significantly affect the perception of the data. The discussion needs to be rewritten to better align with the aim and title of the paper.

Response: We agree with the reviewer's assessment. The discussion was restructured, expanding the comparison with the literature to better align with the results obtained (Please see Section 4).

• The conclusions directly resulting from the spatial classification of banana and whip palm cultivation should be highlighted, rather than focusing on general considerations about agricultural diversification. The conclusions should be expanded and more strongly related to the presented results. Emphasising the importance of the methods used (multi-temporal analysis, Random Forest) for future research and agricultural practice in tropical regions will strengthen the message of the work.

Response: Dear reviewer, as with the discussion section, the conclusions have been restructured to better reflect the findings of the work, avoiding generic considerations (Please see lines 531- 552).

Reviewer 4 Report

Comments and Suggestions for Authors

The manuscript maps two perennial crops in Atlantic Forest region, Brazil using Sentinel-2 time-series vegetation indices and a hierarchical Random Forest to separate perennial vs. non-perennial and then banana vs. peach palm. Results are clearly presented. The study addresses an important domain, but the paper's contribution and novelty need sharper articulation, and several methodological details are required.

Major comments

As written, this reads primarily as a case study. If that's the intent, expand the significance of Jacupiranga/Ribeira Valley (socio-environmental importance, management needs) so the case itself justifies publication. If you claim a generalizable workflow novelty, demonstrate it: compare against simpler baselines, include ablations (e.g., with/without hierarchy; indices vs. full bands; etc.).

Given "frequent clouds", justify the exclusive use of Sentinel-2. Did you test Sentinel-1 SAR or simple texture/structure metrics (e.g., entropy) to separate perennials from native vegetation? That would strengthen the claim of robustness.

Accuracy assessment, please specify the number of validation samples per class, and other relevant information in the results section, such as sampling scheme, any spatial blocking.

Minor comments

Line 196: adjust font and formatting.

I would keep the font consistent in all maps and tables. Certain map labels can be improved, for example, the graticule labels do not need so many zeros. 49°30,000'W can simply be 49°30'W. 

Line 121-127. Why do all degree symbols have an extra short line?

Author Response

Major comments

 

• As written, this reads primarily as a case study. If that's the intent, expand the significance of Jacupiranga/Ribeira Valley (socio-environmental importance, management needs) so the case itself justifies publication. If you claim a generalizable workflow novelty, demonstrate it: compare against simpler baselines, include ablations (e.g., with/without hierarchy; indices vs. full bands; etc.).

Response: Dear reviewer, thank you for this valuable observation. While our workflow has produced novel results, our primary focus is on the Ribeira Valley itself and its unique socio- environmental context, challenges, and management needs. We therefore aim to strengthen the characterization and discussion of these particularities, especially within the introduction and study area sections, so that the case’s relevance and significance are clear even if readers view it primarily as a case study.

• Given "frequent clouds", justify the exclusive use of Sentinel-2. Did you test Sentinel-1 SAR or simple texture/structure metrics (e.g., entropy) to separate perennials from native vegetation? That would strengthen the claim of robustness.

Response: Thank you for your feedback. One of the core objectives of our study was to assess the potential of Sentinel-2's temporal resolution in regions affected by persistent cloud cover. We believe that optical data like Sentinel-2's offer important advantages: they are lighter to store and process, more scalable for application over large areas or long time series, and typically easier to interpret for many users than, for example, some radar or specialized sensor data. Nevertheless, we recognized the role of radar data in these environments by adding a paragraph weighting the results with optical data (see lines 502-510).

“While optical data remain advantageous for scalability, storage, and interpretability, adding microwave observations could further improve crop discrimination under persistent cloud cover. Combining Sentinel- 1 SAR with Sentinel-2 imagery offers particular promise for capturing subtle phenological variations [69- 71]. Sentinel-2’s spatial resolution is appropriate for regional assessments. However, it can hinder the identification of crops in small plots or intercropped systems, which are typical of family farms and agroforestry in the Ribeira Valley. Classification uncertainties may also arise from spectral overlap between certain crops and native vegetation at specific growth stages [71]”.

• Accuracy assessment, please specify the number of validation samples per class, and other relevant information in the results section, such as sampling scheme, any spatial blocking.

Response: As suggested, the information was further detailed in the methodology and discussion sections, as shown below:

Table 2. Confusion matrices and commission and omission errors (CE and OE, respectively) from classification levels 1, 2, and 3. Matrix cell values indicate the number of validation samples.

“Sampling followed a convenience scheme, as specific production farms were purposefully selected in collaboration with local producers to ensure access and accurate field information. Peach palm (Bactris gasipaes) samples were restricted to the visited farms and were not extrapolated beyond these areas, given the challenges of accurately interpreting this type of cultivation through photointerpretation. This approach ensured data reliability while acknowledging the limited spatial representativeness of the sample set” (Lines 286-292).

“Finally, the absence of spatial blocking and the relatively small dataset are other limitations of the study, which may constrain broader generalizations. These issues should be addressed in future studies” (Lines 527-529)

Minor comments

• Line 196: adjust font and formatting. Response: font and formatting were adjusted.

• I would keep the font consistent in all maps and tables. Certain map labels can be improved, for example, the graticule labels do not need so many zeros. 49°30,000'W can simply be 49°30'W.

Response: The authors appreciate the reviewer’s helpful suggestion regarding the map and table fonts and the graticule labels. In the revised manuscript, the font style and size have been standardized across all maps and tables to improve visual consistency. Additionally, the graticule labels have been simplified.

• Line 121-127. Why do all degree symbols have an extra short line? Response: The extra short lines were removed.