Advanced Machine Learning Methods as a Planning Strategy in the Capellanía Wetland

Jury 1

Observation

Correction

Page

Lines

1

Limitations of model assumptions: Relying on linear relationships (such as the linear regression assumed in NDVI predictions) may overlook nonlinear features of wetland dynamics (such as sudden climate events or policy interventions).

In the section 2.6.1 Although this method assumes an approximately linear relationship over the analyzed period (2013–2024), potentially underestimating non-linear behaviors associated with climatic events or anthropogenic interventions, its application is appropriate given the limited number of observations per polygon and the need to obtain continuous values to feed the main model. This approach has been employed in previous studies on wetland and landscape dynamics [35]. Integration with subsequent non-linear models (such as Random Forest) mitigates this limitation and has been validated in studies of landscape and wetland dynamics [36]

8

261-265

2

Data dependency: The model requires at least two observations to support it, which limits its application in data-poor areas, such as wetlands where long-term monitoring is lacking.

Section 2.2 includes the field survey and land-cover identification.

5

178

3

Doubts about regional universality: The study focuses on the Capellanía wetland in Bogotá, and the conclusion has not been validated for generalization to other wetland types or geographic environments, such as coastal or alpine wetlands.

The conclusion includes some limitations of the research and makes some recommendations for further research.

20

637-666

4

Chaotic distribution of chapters: The discussion (4. Discussion) and conclusion (5. Conclusions) sections overlap, and both describe NDVI changes and compare them with other cases (such as Lima and Concepción), which weakens the independence and generalizability of the conclusion. Key information about the methodology, such as the MOLUSCE platform and the random forest model, are not explained in separate chapters, but are scattered throughout the discussion and conclusion sections, making it difficult to understand the results. rencia.

The sections have been reorganized to improve clarity:

Appropriate clarifications are made in the Discussion and Conclusions.

 The following sections are explained in the methodology.

2.6.2 Estimation Using Random Forest

2.6.4 Urban Expansio Simulation with molusce

542

8

9

542 and 638

268

305

5

Late analysis of model limitations: Limitations of the MOLUSCE model, such as its reliance on fixed rules and ignoring policy factors, were only mentioned at the end of the discussion and were not explained in advance in the methodology or experimental design sections, which may affect readers' judgments about the credibility of the results.

Some limitations of the Molusce model are explained in section 2.6.4

9

306

6

Insufficient connection between central themes and conclusions: Research objectives (such as "The potential impact of ALO road construction on wetlands") were only specifically discussed during the debate, while the introduction only briefly mentioned "The threat of urban expansion to wetlands," resulting in a lack of coherence in logical progression.

The introduction was reinforced with a final paragraph explicitly linking the general threat of urbanization to the specific case of the Avenida Longitudinal de Occidente (ALO) and announcing the comparative scenarios (with/without ALO) to address the main objective of the study. Likewise, the conclusion was expanded with a final paragraph that synthesizes the progressive recovery of vegetation (2013–2032), highlights that the ALO would act as a catalyst for urbanization by reducing non-urbanized areas, and links these findings directly to the scenarios analyzed, reinforcing their implications for environmental and road planning in Bogotá.

16

513

7

Redundancy and ambiguity in expression: Key terms are used repeatedly (such as "NDVI classes 3 and 4" are described multiple times in the same sentence structure), without simplified or alternative expressions. Some conclusions are expressed vaguely, such as "the results support the incorporation of spatial modeling into decision-making," without specifying how it works, which reduces its feasibility.

The wording of the Results, Discussion, and Conclusions sections was revised to eliminate unnecessary repetition and improve flow. Synonyms and clearer language were used to refer to NDVI classes, avoiding the repeated use of "classes 3 and 4" and replacing it with terms such as "moderate and dense vegetation."

10, 17 20

366,542-637

8

Formatting Issues and Linguistic Details: Paragraph numbers (such as 359-386, 387-428) are interspersed within the main text, suggesting formatting errors that affect reading flow. Some sentences have complex structures (such as nested sentence structures that combine... and use...), making comprehension difficult.

The document's formatting was reviewed and corrected, eliminating paragraph numbers that appeared in the main text and could confuse readers. In addition, long sentences were simplified, breaking them into shorter, clearer phrases to improve readability. Punctuation and sentence structure were adjusted to facilitate understanding, especially in the Discussion and Conclusions, avoiding overly long or nested sentences.

The entire document

Jury  2

Observation

Correction

Page

Lines

1

The work has little field validation; the study is based primarily on remote sensing data. The lack of field verification (ground truth) can introduce uncertainty into the classification. I suggest the authors seek in situ data sources to validate their findings, even if only for short periods.

Section 2.2 Field survey and land-cover identification is incorporated, validating land cover with in-situ observations. In April 2024, a field visit was made to the Capellanía wetland, where plant species and landscape elements were recorded and used to generate and update a KMZ file (dense vegetation, grasslands, herbaceous areas, and bodies of water). This information was compared with NDVI spectral values to adjust and support the classification.

5

178

2

Limited spatial scale: The study focuses solely on the Capellanía Wetland. Comparisons with other wetlands in Bogotá could broaden the generalizability and strengthen recommendations for the entire city. Would this be possible? If so, please do so. If not, justify its application to this location only. Why this location and not others/all?

Sections 1 and 2.1 justify the selection of the Capellanía Wetland in the Introduction and Study Area. It was explained that Capellanía is one of the most impacted urban wetlands in Bogotá, with a loss of nearly 88% of its area due to road infrastructure and urban expansion, and with additional risk due to the planned Western Longitudinal Avenue (ALO). These conditions make it representative and a priority for evaluating vegetation dynamics under anthropogenic pressure.

3 y 2

98 and 143

3

Lack of climate change scenarios: The simulations only consider urban growth and infrastructure, but don't include climate variables (rainfall, drought, and temperatures over time), which are crucial for wetland dynamics. I believe there is a weather station in Bogotá linked to a public sector that provides data. I suggest including graphics showing the local climate.

Section 2.5 and Table 3 explain why climatic variables are not integrated.

7 y 8

236 and 248

4

Limited Discussion: The discussion is too brief, addresses few articles in the literature, and does not highlight the study's findings. Please improve. Lack of climate change scenarios: The simulations only consider urban growth and infrastructure, but do not include climate variables (rainfall, drought, and temperatures over time), which are crucial for wetland dynamics. I believe there is a weather station in Bogotá linked to a public sector that provides data. I suggest including graphics showing the local climatology.

The Discussion section has been expanded to strengthen the interpretation of results and their connection to recent literature. Additional relevant studies from different contexts (Pakistan, Sindh, Lima, Concepción, Barranquilla, Cauca) have been incorporated, highlighting the usefulness of Random Forest and MOLUSCE in urban environments and comparing observed trends with other Latin American wetlands.

17

542

Jury 3

Observation

Correction

Page

Lines

1

In line 168 of Table 2, NDVI values between -1 and -0.1 are classified as bare soil. However, in urban wetlands, negative NDVI values are often associated with water surfaces or surfaces with very low reflectance. This classification can easily lead to systematic misclassification, for example, in situations where open water appears as bare soil. How do the authors view this?

We appreciate the observation. Indeed, negative NDVI values may correspond to water bodies rather than bare soil. To avoid this confusion, we incorporated a water mask based on water-sensitive indices (MNDWI, NDWI, and NDMI), classifying water as an independent category. This ensures a precise separation between aquatic surfaces and non-vegetated soils throughout the analysis.

7

204

2

In the methodology description (line 180), the NDWI/NDMI/MNDWI indices are listed as predictors, while in another (line 193), the model was trained with the variables "NDMI and MNDMI." Please clarify this inconsistency..

NDVI was used as the dependent variable, while NDWI, NDMI, MNDWI, and the seasonal variables (sine and cosine of the month) were used as predictors. In the spatial validation section, the reference to “NDMI and MNDMI” was a typographical error and has been removed; the variables are now consistently listed throughout the methodological section.

8

256

3

The pseudo-R² value of 1.00000 reported for MOLUSCE (line 208) should be interpreted: Is this an indicator of calibration exclusively within the sample, or is it actually the result of out-of-sample validation? Please clarify this and complement the analysis with a confusion matrix.

It was clarified that the pseudo-R² = 1.00 corresponds to the internal calibration in MOLUSCE and not to an out-of-sample validation. Since the tool does not generate confusion matrices, transition matrices were incorporated as an equivalent resource (Tables 5a and 6a), following the methodology proposed by Muhammad et al. (2022, Land, 11(3), 419. https://doi.org/10.3390/land11030419

9

293

4

In line 249 (Table 3), there is an extreme correlation between NDVI and NDWI, r = −0.99. This is likely due to the nature of the sampling frame or to strong collinearity between the predictors. To clarify this, the authors recommend supplementing the analysis with a predictor elimination test (ablation), or presenting feature importance or SHAP results to show the true contribution of each variable to model performance.

The extreme correlation between NDVI and NDWI is due to their spectral nature and the use of common bands. To rule out biases, we applied a SHAP analysis, which revealed the differentiated importance of the predictors, keeping NDVI solely as the dependent variable and ensuring the consistency of the model.

12

403

5

Respecto al valor MSE = 0,397 en la línea 314 ¿no se considera que es particularmente grande en la escala NDVI?

After the reclassification and methodological adjustment, the model metrics improved substantially, yielding R² = 0.991, RMSE = 0.0214, and MAE = 0.0127 on the NDVI scale. These values indicate a very low error and confirm the high predictive capacity of the Random Forest in the multitemporal analysis.

1 y 18

14-568

6

  Regarding the MSE value = 0.397 on line 314, is it not considered to be particularly large on the NDVI scale?

After the reclassification and methodological adjustment, the model metrics improved substantially, yielding R² = 0.991, RMSE = 0.0214, and MAE = 0.0127 on the NDVI scale. These values indicate a very low error and confirm the high predictive capacity of the Random Forest in the multitemporal analysis.

18

568-569

7

How were the training and test sets separated in space and time to prevent data leakage? For example, how was it ensured that neighboring pixels from the same date were not simultaneously included in the training and test sets?

The dataset was split into 80% for training and 20% for spatial validation using GroupShuffleSplit, ensuring independence between polygons (fid) and preventing the simultaneous inclusion of neighboring pixels in both sets.

8

267

8

ALO is inconsistent: Western Longitudinal Avenue (abbreviation) vs. Western Longitudinal Avenue (text)

The requested adjustment was made to ensure consistency in the use of the term. In the first mention, only the abbreviation ALO is used, and in subsequent mentions, the full name followed by the abbreviation is employed: Avenida Longitudinal de Occidente (ALO).

1

21

9

Abbreviations: include MAE and MSE.

Included in the abbreviations.

18

528

10

Line 122: Study

Line 160: Remote Sensing and Spectral Correlation

Line 209: Spatial

Line 368: Metric Analysis

21

674

11

References are not cited in the text from position 26 onwards.

All references in the article have been corrected.

12

I suggest supplementing the reference list with relevant literature related to the methodology section (e.g., the use of remote sensing indices, Random Forest validation strategies, the use of MOLUSCE, or other cellular automata-based urban growth models). This would strengthen the scientific basis of the manuscript and create a closer connection between the chosen methods and previous research.

Recent references were added to strengthen the methodological section: for the use of spectral indices in urban wetlands [31, 33], for the application of Random Forest in remote sensing and spatial validation strategies [17, 38, 48], and for the use of MOLUSCE and other cellular automata models in urban growth simulation [40–43]. These additions reinforce the justification of the chosen methods and their consistency with the previous literature.

21,22 and 23

687-801

13

In line 168 of Table 2, NDVI values between -1 and -0.1 are classified as bare soil. However, in urban wetlands, negative NDVI values are often associated with water surfaces or surfaces with very low reflectance. This classification can easily lead to systematic misclassification, for example, in situations where open water appears as bare soil. How do the authors view this?

Adjustments are made see table 2

7

233

14

In the methodology description (line 180), the NDWI/NDMI/MNDWI indices are listed as predictors, while in another (line 193), the model was trained with the variables "NDMI and MNDMI." Please clarify this inconsistency.

Adjustments are made and information is presented in section 2.7.1

8

255

15

  The pseudo-R² value of 1.00000 reported for MOLUSCE (line 208) should be interpreted: Is this an indicator of calibration exclusively within the sample, or is it actually the result of out-of-sample validation? Please clarify this and complement the analysis with a confusion matrix.

Adjustments are made to paragraphs 2.7.4

9

318

16

In line 249 (Table 3), there is an extreme correlation between NDVI and NDWI, r = −0.99. This is likely due to the nature of the sampling frame or to strong collinearity between the predictors. To clarify this, the authors recommend supplementing the analysis with a predictor elimination test (ablation), or presenting feature importance or SHAP results to show the true contribution of each variable to model performance.

Adjustment is made

12

403

17

Regarding the MSE value = 0.397 on line 314, is it not considered to be particularly large on the NDVI scale?

Adjustment is made

1 y 18

14-568

18

  Line 372 shows an MAE of 0.018, while elsewhere it shows an MAE of 0.0018. Please clarify which result is valid.

Adjustment is made Adjustment is made

18

568-569

19

How were the training and test sets separated in space and time to prevent data leakage? For example, how was it ensured that neighboring pixels from the same date were not simultaneously included in the training and test sets?

In the model, a random partition of pixels was not performed; instead, a grouped split by polygon was applied using GroupShuffleSplit. In this way, all pixels from the same polygon were assigned exclusively to either the training or the test set, thus preventing spatial leakage, i.e., avoiding that neighboring pixels were mixed across both sets. Although a strict temporal blocking was not implemented —meaning that the same acquisition date could appear in different polygons across training and test— this strategy was consistent with the main objective of the study: to evaluate the model’s generalization capacity toward unseen polygons.

8

280

20

ALO is inconsistent: Western Longitudinal Avenue (abbreviation) vs. Western Longitudinal Avenue (text)

Adjustment is made

1

21

21

Abbreviations: include MAE and MSE.

Abbreviations are incorporated

21

685

22

Line 122: Study

Line 160: Remote Sensing and Spectral Correlation

Line 209: Spatial

Line 368: Metric Analysis

Adjustments are made to the bibliographical

21,22 and 23

686-802

23

References are not cited in the text from position 26 onwards.

24

I suggest supplementing the reference list with relevant literature related to the methodology section (e.g., the use of remote sensing indices, Random Forest validation strategies, the use of MOLUSCE, or other cellular automata-based urban growth models). This would strengthen the scientific basis of the manuscript and create a closer connection between the chosen methods and previous research.

Adjustments are made to the bibliographical

21,22 and 23

686-802