Review Reports - Mineral Prospectivity Prediction in the Mayoumu Area, Tibet, Based on Multi-Source Exploration Information and Ensemble Learning Models

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors：Specific review comments are provided in the attached document.

Comments for author File: Comments.pdf

Author Response

Dear Reviewer,

Re: Manuscript ID: remotesensing-4126724

Thank you for your comments concerning our manuscript entitled “Mineral Prospectivity Prediction in the Mayoumu Area, Tibet, Based on Multi-Source Exploration Information and Ensemble Learning Models”. Those comments are valuable and very helpful. We have read through comments carefully and have made corrections. Based on the instructions provided in your letter, we uploaded the file of the revised manuscript. Revisions in the text are shown using red highlight for additions, and strikethrough font for deletions. The responses to the reviewer's comments are marked in red and presented following.

Below, we respond to each comment point by point.

Q1. Insufficient details on remote sensing data preprocessing. The masking thresholds for Sentinel-2 data are not provided, and the masking order and logical relationships for different interference factors (vegetation, water bodies, snow, shadows) are not explained. In the preprocessing of GF-5 hyperspectral data (Lines 189-191), only “spectral smoothing and stripe removal” are mentioned, with no description of parameter settings(e.g., window size). A similar issue appears in the ILR transformation of geochemical data in Section 4.1.3 (Lines 207): the element combination method is not specified, nor is the method for determining the anomaly threshold in RPCA analysis (Lines 465-469).

Response: Thank you for the valuable suggestion. We agree that the original manuscript lacked sufficient operational details and parameter transparency. To improve reproducibility, we revised the preprocessing description in three ways: (1) Sentinel-2 masking: We now clearly describe the masking sequence and compositing logic (union rule and overlap handling) using SCL-based cloud/cloud-shadow removal, followed by NDWI/NDVI water/vegetation exclusion, and density slicing for snow/shadow, with additional constraint from mapped Quaternary cover. (2) GF-5 parameters: We added explicit settings for spectral smoothing and destriping: Savitzky–Golay filter (window length = 11 bands; polynomial order = 4) and a global column-wise destriping (full column as statistical window; gain/offset estimated from column mean and standard deviation; column-by-column linear normalization), with before examples shown in Fig. 2c–2d. (3) Geochemistry: We clarified the ILR element-composition scheme and specified the RPCA map display strategy using Jenks natural breaks (7 classes) for consistent visualization and comparison.

Q2. Lack of transparency in learning model construction and optimization. The hyperparameter grid search ranges for the ensemble learning models (GBDT, XGBoost, LightGBM) developed in Chapter 3 are not clearly defined. For example, the specific step size set in the experiment is not specified based on the provided formulas, and specific parameter intervals (e.g., learning rate, tree depth, number of leaf nodes) are lacking. The division ratio of the training set and test set is not stated, which may lead to overfitting. What is the stratification strategy for the 5-fold cross-validation in Chapter 4? (Whether the spatial distribution characteristics of mineralized samples are considered), which may cause bias in the validation results.

Response: Thank you for this important comment. We agree that the original manuscript did not describe the training and evaluation workflow with sufficient clarity. In the revised manuscript, we corrected the “grid search” statement: no grid search was performed; instead, we adopted a complexity-constrained manual-tuning strategy based on recommended defaults and overfitting control under limited samples, and we now report the final key parameter settings and rationale. We also clarified that model evaluation used stratified 5-fold cross-validation (StratifiedKFold, n_splits=5, shuffle=True, random_state=42), with an approximate 80%/20% train–validation split per fold, stratified by class label, and continuous predictors were standardized using StandardScaler. Regarding potential optimistic bias from spatial autocorrelation, we added discussion noting that sample construction used a 750 m buffer to reduce spatial mixing between positives and negatives, and non-mineralized samples were randomly generated outside the buffer (also considering negative-anomaly areas) to improve representativeness. We explicitly acknowledge that the current evaluation is not a strict spatial blocked CV, list it as a limitation, and propose spatial blocking/sensitivity analyses for future work. Revised text (Lines 609-629).

Q3. Ambiguous logic in multi-source data fusion. The resampling method used in the rasterization (500m×500m) of multi-source data on the ArcGIS platform is not explained (e.g., 30m GF-5 hyperspectral, 10m Sentinel-2 multispectral, 1:200,000 regional geochemical data). The basis for selecting 500m is not stated, which may affect prediction accuracy. The descriptions of the 10 predictors are inconsistent; and the logic for handling dimensional differences among different variable types (remote sensing anomalies, geochemical indicators, fault distance, etc.) is unclear (Lines 496-499). In addition, the scientific basis for adopting a 750m buffer radius when selecting mineralized/non-mineralized samples is not provided (e.g., whether based on the influence range of known deposits, width of alteration zones, etc.) (Line 514).

Response: Thank you for this critical comment. We agree that the original manuscript did not sufficiently describe the spatial-support selection, the grid-based assignment workflow, and the rationale for sample construction, which may affect clarity and reproducibility. In the revised manuscript, we have substantially expanded the “data fusion and sample construction” description, with the following key updates: (1) Rationale for the 500 m grid: We now explain that the 500 m × 500 m grid represents a trade-off between spatial representation and data stability. It captures regional-scale spatial variations of anomalies while avoiding extensive no-data and unstable assignment/interpolation that may arise from overly fine grids, and it is also consistent with the subsequent sampling buffer scale. (2) Unified multi-source workflow and value assignment: We added a clear ArcGIS-based workflow: all predictors were converted to raster layers, unified in projection and spatial support, and then assigned to grid units by extracting raster values at representative locations (grid centroids). This procedure produced 11,211 independent prediction units and a quantitative dataset with 11,211 samples and 10 predictors. (3) Handling dimensional differences: To improve comparability across predictors and reduce potential scale effects on model convergence and feature-contribution interpretation, we state that continuous predictors were standardized prior to machine-learning modeling. (4) Justification of the 750 m buffer and sample independence: We clarified that the 750 m buffer was used to reduce spatial mixing between positive and negative samples while avoiding an overly large buffer that would excessively shrink the negative-sample candidate space, and it is compatible with the 500 m grid (~1–2 grid cells). Grid units within the buffer were treated as positive samples. Non-mineralized samples were randomly generated outside the buffers, with additional constraints using geochemical negative-anomaly areas. To further mitigate potential spatial autocorrelation caused by clustering of negative samples, a minimum inter-point distance constraint (> 1000 m) was applied. Ultimately, a balanced dataset of 101 mineralized and 101 non-mineralized units was constructed. Revised text (Lines 568–607).

Q4. Insufficient scientific basis for extraction of alteration anomaly information and structural
information. The method for extracting alteration anomalies based on PCA and the “iCrosta”
thresholding scheme lacks clear mineralogical-spectral justification and may be overly arbitrary. The structural information extraction in Section 4.1.4 is also insufficient: the geometric and kinematic characteristics of regional structures are not adequately represented. Given that Tibet is one of the most important active tectonic regions globally, structural systems from major thrust faults to oblique secondary faults may control mineralization processes and fluid migration. The authors are advised to supplement structural descriptions, provide the spatial distribution of the extracted faults/structures, and clarify how these structural parameters are incorporated into the prediction model.

Response: Thank you for this important comment. We agree that the original manuscript did not adequately justify the alteration mapping from a mineralogical–spectral perspective, nor did it sufficiently describe the regional structural geometry/kinematics and its parameterization for modeling. Accordingly, we have substantially strengthened the revised manuscript: in Section 4.1.1, we added an explicit “mineralogy–spectroscopy–PC component” linkage supported by USGS spectral reflectance curves (new Figure 4), clarified the Sentinel-2 PCA band sets and the alteration meaning of the inverted component (−PC4), and refined the anomaly extraction using an iCrosta-based standard-deviation grading scheme (μ+3σ, μ+2.5σ, μ+2σ) followed by median filtering to improve reproducibility and spatial continuity (Figure 5). Revised text (Lines 427–465). In Section 4.1.4, we expanded the tectonic description to emphasize hierarchical fault systems and key geometric discontinuities (e.g., intersections, bends, step-overs) that enhance permeability and fluid focusing in southern Tibet, and we parameterized fault vectors into a “fault distance” raster predictor using Euclidean distance analysis (Figure 7f) for model input. In addition, we updated the geological background map (Figure 2) by overlaying the major faults/structural traces to better visualize the structural framework and its spatial correspondence with alteration and prospectivity products. Revised text (Lines 542–566).

Q5. Insufficient verification of prediction results. The delineated high-potential areas lack detailed geological interpretation and validation. The manuscript should further discuss metallogenic conditions of newly identified targets, and provide quantitative verification (e.g., overlap rate, buffer hit rate, etc.) rather than only qualitative descriptions.

Response: Thank you for this constructive comment. We agree that validation should be strengthened by linking the predicted high-potential zones to metallogenic conditions and, where possible, quantitative measures. In the revised manuscript, we have expanded the Discussion by (i) providing a more explicit geological interpretation of the newly delineated targets through comparison with the geological/structural map (lithologic units, major faults and their intersections/bends), and (ii) adding field verification results based on route-based investigations and representative outcrop checks for two target areas (Investigation Areas I and II; locations shown in Figure 12). The field observations document mineralization– alteration assemblages (e.g., Pb–Sb mineralization, silicification/iron staining, epidote alteration, and secondary Cu minerals) and structural controls (fracture corridors, shear-related brecciation and quartz veins/lenses), which provide independent geological evidence supporting the reliability of the prediction. These additions are presented in the revised Discussion. Revised text (Lines 750–867)

Q6. The section title “3.1.5 Sample Dataset Preparation” should be “4.1.5” based on the current manuscript structure.

Response: Thank you for pointing this out. We have corrected the section numbering to be consistent with the revised manuscript structure.

Q7. The formula formatting is non-standard: Equations (3), (4), etc. have missing symbols and layout errors. Please correct the equation expressions and ensure consistency in notation.

Response: We appreciate this comment. We have carefully checked and corrected the formatting of the equations, including missing symbols and layout issues, and standardized the notation to ensure consistency and readability throughout the manuscript.

Q8. In the Introduction, the sentence in Line 60 “Numerous studies…” cites only three references[7-9], which is insufficient. Please add more references to support the statement.

Response: Thank you for the suggestion. We agree that the original citation support was limited. We have added additional relevant references to strengthen the literature basis of this statement in the Introduction.

Q9. In Figure 2a, the study area extent is not prominent relative to the main controlling structural unit. Please enlarge and highlight the study area appropriately.

Response: Thank you. We revised Figure 2a to improve visual clarity by enlarging and highlighting the study-area extent relative to the main controlling structural unit.

Q10. There is a lack of transitional sentences between Chapter 2 and Chapter 3. Please add a
bridging paragraph to connect data description and modeling methods.

Response: We agree and appreciate this helpful suggestion. We have added a short bridging paragraph to connect the data description (Chapter 2) with the modeling methodology (Chapter 3), improving the logical flow of the manuscript.

Q11. The terms “mineral prospectivity prediction” and “mineralization probability” are used
alternately; please standardize terminology throughout the manuscript.

Response: Thank you for this comment. We have standardized the terminology throughout the manuscript by using “mineral prospectivity prediction” consistently and revising all inconsistent expressions accordingly.

Q12. ILR is first defined in the Abstract; there is no need to repeat “isometric log-ratio (ILR)
transformation” again later (Line 649). Please revise for conciseness.

Response: We agree. To improve conciseness, we removed the repeated full definition and retained only “ILR transformation” where appropriate, while keeping the first definition in the Abstract.

We have carefully reviewed the full text and corrected errors wherever necessary. We sincerely thank you for giving us the opportunity to resubmit the revised manuscript and for your valuable time and effort in reviewing our work. We hope that the revisions meet your expectations.Thank you and best regards.

Sincerely.

Kai Qiao

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Comments on the text:

Line 133: I would like to suggest to add in this part of the text to add that GBDT,XGBoost and LightGBM are Boosting strategies.

Lines 193 to 196: At this point, the text becomes a little bit confusing as you are referring also to Landsat-8 dataset. So, it was used also a Landsat-8 dataset, except from S2 and GF-5 in the pre-processing and it is not included in the final processing? Please have again a look in this part of the text and do the required improvements.

Line 202: I would like to ask you why did you decide to work S2 with spatial resolution 10m and GF-5 with 30m, instead of bringing S2 to 30m spatial resolution and you have heterogeneous datasets in terms of spatial resolutions for your comparison?

Line 204: I would like to ask you why are used winter images for the mosaic from November 2024 and you didn't keep up with summer acquisitions? The sun elevation differs in summer and in winter and this can create disparate approach of a dataset.

Line 248: I would like to suggest to add that GBDT is a Boosting strategy as you describe it in the Introduction mainly for the people that are more focused in Remote Sensing and now they are getting familiar with Machine Learning new approaches.

Line 555-556: I would like to suggest also for future work to examine thermal datasets of ASTER for example in order to see the temperature variations on the surface of the hydrothermal alteration field (if there are any very significant).

Line 665-668: This part I would suggest to be described in a more prominent way within the whole paper with more comparisons with the geological map in order the reader to be convinced that there is the connection between the different products or also even to do statistical analysis on how many pixels of the mineralization of different types are included in the different lithologies of the hydrothermal alteration field.

Comments on the Figures:

Figure 1: I would like to propose to include somewhere the spatial resolution of your datasets as for S2 you have 10m spatial resolution and for GF-5 you have 30m.

Figure 2: I would suggest to pass the faults on the maps with the hydrothermal alteration results based on the tectonic map in order to be able to see the tectonic control on the surface of the hydrothermal alteration field.

Figure 3: Very interesting the results after the bad line correction and stripe removal. Which model/technique is used for the stripe removal?

Figure 4: I would like to suggest in general to keep a homogeneous description of your legends in the alteration maps in order to be easy for the reader to compare them. Between the legends of S2 results and GF-5 results there are differences, so this makes it difficult to compare and see possible improvements. Furthermore, I would like to suggest to describe the terms of first-order, second-order and third order ferric iron alteration and ferric hydroxyl alteration respectively in the text of 4.1.1.

Figure 5: I would like to suggest and highlight again that the legend of Figure 4 and Figure 5 could be presented with the same legend terminology in order to be more easily comparable. I am taking in consideration that the one is a multispectral dataset and the other is hyperspectral, which means of course that with the second one you can go more in details on the geochemical composition of the hydrothermal alteration but it is suggested to find a way to connect the 2 results and describe it more easily and comprehensive also in the text.

Figure 10: I would like to suggest once again (as the tectonic activity is controlling the mineralisation) to add the tectonic faults of the area in the map. Furthermore, for Figure 10 I would like to clarify what is the spatial resolution of the MPD map, as the two satellite datasets that are used have different spatial resolution. Did any resampling took place before the final product in order to have the same spatial resoltuion and compare homogeneous things? Moreover, I would like to suggest again to keep a common terminology on the legend of the map in order to be able to be comparable with Figures 4 and 5.

Author Response

Dear Reviewer,

Re: Manuscript ID: remotesensing-4126724

Below, we respond to each of the suggestions and questions raised by the reviewer in turn.

Q1. Suggestions from reviewer: Line 133: I would like to suggest to add in this part of the text to add that GBDT,XGBoost and LightGBM are Boosting strategies.

Response: Thank you for this helpful suggestion. We agree that clarifying the learning paradigm improves readability for remote-sensing-oriented readers. We have revised the text at Line 133 to explicitly state that GBDT, XGBoost, and LightGBM are Boosting-based ensemble learning strategies, and we adjusted the surrounding sentence accordingly for consistency and clarity. Revised text (Lines 130–133).

Q2. Suggestions from reviewer: Lines 193 to 196: At this point, the text becomes a little bit confusing as you are referring also to Landsat-8 dataset. So, it was used also a Landsat-8 dataset, except from S2 and GF-5 in the pre-processing and it is not included in the final processing? Please have again a look in this part of the text and do the required improvements.

Response: Thank you for pointing this out. We agree that the original wording in Lines 193–196 could mislead readers into thinking that Landsat-8 data were included in the final workflow. In our earlier work, Landsat-8 was indeed tested/processed as a multispectral dataset; however, for the present study we ultimately selected Sentinel-2 as the representative multispectral source due to its higher spatial resolution and better suitability for alteration mapping in the study area. Therefore, Landsat-8 was not used as an input to the final processing and modeling in this manuscript. The mention of “Landsat-8” in Lines 193–196 was an incorrect/legacy description and has now been removed (or corrected) to clearly state that the multispectral preprocessing and subsequent alteration extraction were conducted using Sentinel-2, together with GF-5 hyperspectral data. This revision eliminates ambiguity and aligns the text with the actual datasets used.

Q3. Suggestions from reviewer: I would like to ask you why did you decide to work S2 with spatial resolution 10m and GF-5 with 30m, instead of bringing S2 to 30m spatial resolution and you have heterogeneous datasets in terms of spatial resolutions for your comparison?

Response: Thank you for this insightful and technically important comment. We fully understand the reviewer’s concern regarding the use of datasets with different spatial resolutions.In this study, Sentinel-2 (10 m) and GF-5 (30 m) data were used for different but complementary purposes. Sentinel-2 multispectral imagery was employed for alteration anomaly extraction using PCA-based iCrosta analysis. In this context, the higher spatial resolution (10 m) is advantageous for preserving fine-scale surface alteration features and delineating narrow mineralized zones. GF-5 hyperspectral data (30 m), in contrast, were mainly used for mineral identification and MTMF-based alteration mineral mapping, where spectral resolution is more critical than spatial resolution.It should be clarified that we did not conduct a direct pixel-by-pixel comparison between Sentinel-2 and GF-5 outputs at their original resolutions. Instead, all derived evidential layers were subsequently integrated within a unified spatial framework during the mineral prospectivity modeling stage. Prior to model construction, the datasets were spatially aligned and resampled to a consistent grid resolution to ensure comparability and to avoid scale-induced bias.Resampling Sentinel-2 data to 30 m at the preprocessing stage would have reduced valuable spatial detail without improving the spectral characteristics of the GF-5 dataset. Therefore, we retained the native resolutions during feature extraction to preserve maximum information content and harmonized them only during the modeling stage.

Q4.Suggestions from reviewer: I would like to ask you why are used winter images for the mosaic from November 2024 and you didn't keep up with summer acquisitions? The sun elevation differs in summer and in winter and this can create disparate approach of a dataset.

Response: Thank you for this important comment regarding seasonal differences and solar elevation effects. Within the past five years, the GF-5 AHSI scenes that simultaneously satisfied the requirements of (1) complete coverage of the study area, (2) low cloud contamination, and (3) suitability for hyperspectral mineral analysis were limited to four acquisitions: 7 June 2023, 7 June 2024, 7 November 2024, and 14 November 2024. No additional summer-season scenes meeting these criteria were available in the archive. Therefore, both June and November data were incorporated to ensure spatial completeness and data quality. We acknowledge that differences in solar elevation between summer and winter may influence surface reflectance. However, several factors reduce the potential impact in this study. First, the study area is located in a high-altitude, sparsely vegetated region where seasonal spectral variability is relatively limited. Second, all scenes underwent radiometric calibration and atmospheric correction to ensure spectral consistency. Terrain correction was also applied to mitigate topographic illumination effects related to varying solar angles. In addition, relative radiometric normalization was conducted during mosaicking to minimize inter-scene discrepancies. Importantly, the GF-5 hyperspectral data were primarily used for mineral identification and MTMF-based alteration mineral extraction. These analyses rely mainly on diagnostic absorption features rather than absolute reflectance magnitude. Consequently, moderate seasonal variations in illumination conditions have limited influence on mineral detection results.

Q5.Suggestions from reviewer: Line 248: I would like to suggest to add that GBDT is a Boosting strategy as you describe it in the Introduction mainly for the people that are more focused in Remote Sensing and now they are getting familiar with Machine Learning new approaches.

Response: Thank you for this helpful suggestion. We agree that explicitly clarifying the learning paradigm improves readability for remote-sensing-oriented readers. We have revised the description of GBDT to state that it is a Boosting-based ensemble learning method, and we briefly explained that Boosting builds a strong learner by sequentially adding weak learners to correct previous errors. The revised text can be found in Lines 288–293 of the updated manuscript.

Q6.Suggestions from reviewer: Line 555-556: I would like to suggest also for future work to examine thermal datasets of ASTER for example in order to see the temperature variations on the surface of the hydrothermal alteration field (if there are any very significant).

Response: Thank you for this helpful suggestion. We agree that incorporating thermal infrared (TIR) observations (e.g., ASTER thermal products) could provide additional constraints on surface thermal patterns potentially related to hydrothermal alteration, thereby strengthening multi-source prospectivity assessment. Because the current manuscript focuses on multispectral/hyperspectral alteration mapping, geochemical indicators, and structural proxies within an ensemble-learning framework, a dedicated ASTER-TIR analysis is beyond the scope of this revision. Nevertheless, we will consider integrating ASTER-based thermal indicators (e.g., land surface temperature and thermal anomaly characterization) in future work to complement the mineral prospectivity mapping (MPM) and further improve the robustness of the evaluation.

Q7.Suggestions from reviewer: Line 665-668: This part I would suggest to be described in a more prominent way within the whole paper with more comparisons with the geological map in order the reader to be convinced that there is the connection between the different products or also even to do statistical analysis on how many pixels of the mineralization of different types are included in the different lithologies of the hydrothermal alteration field.

Response: Thank you for this constructive suggestion. We expanded this section in the revised manuscript (Lines 782–799) to more explicitly compare the mineralization potential results with the geological map (lithologic units and major faults/structures) and to strengthen the geological linkage among different products. We also added field verification results in the Discussion (route-based validation and representative outcrops) to independently support the consistency among high-potential zones, mineralization–alteration assemblages, and structural controls. We acknowledge that a rigorous statistical cross-tabulation would require additional scale harmonization and uncertainty handling; therefore, we consider this quantitative analysis as a follow-up task after further addressing scale effects and threshold sensitivity.

Q8.Suggestions from reviewer: Figure 1: I would like to propose to include somewhere the spatial resolution of your datasets as for S2 you have 10m spatial resolution and for GF-5 you have 30m.

Response: Thank you for this helpful suggestion. We have revised Figure 1 and its caption to explicitly report the spatial resolutions of the input datasets, indicating that Sentinel-2 imagery has a spatial resolution of 10 m and GF-5 AHSI hyperspectral data have a spatial resolution of 30 m.

Q9.Suggestions from reviewer: Figure 2: I would suggest to pass the faults on the maps with the hydrothermal alteration results based on the tectonic map in order to be able to see the tectonic control on the surface of the hydrothermal alteration field.

Response: Thank you for this helpful suggestion. We agree that overlaying structural faults on the hydrothermal-alteration map can more clearly demonstrate the structural control on the surface distribution of alteration zones. Therefore, in the revised manuscript we have updated Figure 2 by superimposing the major faults (including the principal fault zones and key subsidiary faults) based on the regional structural map/fault interpretation results, and we have also refined the legend and annotations. This revision makes the spatial correspondence between alteration anomalies and fault corridors more explicit, thereby improving the interpretability and robustness of the figure.

Q10.Suggestions from reviewer: Figure 3: Very interesting the results after the bad line correction and stripe removal. Which model/technique is used for the stripe removal?

Response: Thank you for this helpful comment. We have revised the manuscript to explicitly describe the destriping technique used for the GF-5 hyperspectral imagery. Specifically, stripe noise was removed using a global column-wise destriping correction. The method treats each image column as the statistical window, estimates gain and offset from the column-wise mean and standard deviation, and performs column-by-column linear radiometric normalization to suppress striping artifacts while preserving image structures.

Q11-Q12.Suggestions from reviewer: Figure legends for the alteration maps should be more homogeneous to facilitate comparison between Sentinel-2 and GF-5 results. In addition, please explain in Section 4.1.1 the meaning and basis of the first-/second-/third-order ferric iron alteration and ferric hydroxyl (hydroxyl) alteration.

Response: Thank you for this constructive comment. Please note that the figures have been renumbered in the revised manuscript: the former Figures 4 and 5 are now Figures 5 (Sentinel-2 alteration anomaly grading) and 6 (GF-5 MTMF alteration-mineral mapping). In Section 4.1.1, we have clarified the meaning and basis of the “first-/second-/third-order” levels by explicitly describing the iCrosta grading procedure applied to the PCA outputs: anomaly intensity is classified using thresholds of μ + 3σ (Level I), μ + 2.5σ (Level II), and μ + 2σ (Level III), and the results are further refined using median filtering to enhance spatial continuity. To improve comparability between Sentinel-2 and GF-5 products, we standardized the figure presentation and, more importantly, added an explicit cross-sensor correspondence in the text/captions: Sentinel-2 ferric-iron/iron-staining anomalies broadly indicate iron-oxide/hydroxide alteration and therefore relate to GF-5 iron-oxide/hydroxide minerals (e.g., limonite/hematite), whereas Sentinel-2 hydroxyl anomalies correspond to hydroxyl-bearing alteration minerals mapped from GF-5 (e.g., muscovite/epidote). Because the two products represent different outputs (statistical anomaly grading versus mineral-type mapping), we did not enforce fully identical legend terminology to avoid conceptual ambiguity; instead, we ensured a traceable and consistent comparison through harmonized layout and explicit linkage between the two results.

Q13.Suggestions from reviewer: I would like to suggest once again (as the tectonic activity is controlling the mineralisation) to add the tectonic faults of the area in the map. Furthermore, for Figure 10 I would like to clarify what is the spatial resolution of the MPD map, as the two satellite datasets that are used have different spatial resolution. Did any resampling took place before the final product in order to have the same spatial resoltuion and compare homogeneous things? Moreover, I would like to suggest again to keep a common terminology on the legend of the map in order to be able to be comparable with Figures 4 and 5.

Response: Thank you for the suggestion. Please note that the former Figure 10 has been renumbered as Figure 12 (Mineralization Potential Map, MPM) in the revised manuscript. Regarding the request to overlay faults, we agree that structural context is essential; however, adding dense fault traces onto the MPM reduces map legibility and visual clarity at the regional scale. Therefore, we present the fault network on Figure 2 (geological/structural map) and explicitly discuss the spatial correspondence between the high-potential zones in Figure 12 and the major fault corridors in the text, so that readers can perform a clear cross-comparison without overloading the MPM. We also expanded Section 4.1.5 (Sample Dataset Preparation) to clarify the multi-source fusion workflow: Sentinel-2 (10 m) and GF-5 (30 m) were used for complementary feature extraction and were not compared pixel-by-pixel at their native resolutions; before modeling, all evidential layers were harmonized and assigned to a 500 m × 500 m fishnet grid (11,211 cells). Accordingly, the final MPM (Figure 12) has a spatial resolution of 500 m. Finally, because Figure 12 reports mineralization potential classes (probability/potential levels), which are different in nature from alteration-anomaly grading or mineral-mapping products, we did not enforce identical legend terminology; instead, we clarified the meaning and role of each product in the revised text. If the Editor/Reviewer prefers, we can provide an alternative version of Figure 12 with faults overlain.

We have also carefully checked the manuscript for clarity and consistency and revised accordingly. We sincerely appreciate your time and consideration, and we hope the revised manuscript meets your expectations.

Thank you and best regards.

Kai Qiao

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript presents a Mineral Prospectivity Mapping (MPM) workflow for orogenic gold in the Mayoumu area, Tibet, integrating GF-5 hyperspectral, Sentinel-2 multispectral, and 1:200,000 scale geochemical data. The use of Ensemble Learning (GBDT, XGBoost, LightGBM) combined with SHAP for model interpretability is a sound technical choice. While the results show high predictive accuracy (AUC > 0.9), several critical issues regarding data scaling, sample independence, and geological interpretation need to be addressed before the paper is suitable for publication. But there are still some issues:

The paper conducted mineral prospectivity prediction using multi-source data and combined with GBDT, XGBoost and LightGBM. The results were analyzed from multiple aspects. The introduction of data processing, methods principles, and experimental results is relatively comprehensive. While the results show high predictive accuracy (AUC > 0.9), several critical issues regarding data scaling, sample independence, and geological interpretation. But there are still some issues:

Description

(1)In Lines 89-91, I believe that deep learning, neural networks, and convolutional neural networks are not in a parallel relationship. It is recommended that the author revise the expression.

Introduction:

The current representative methods in Lines 89-91 is insufficient. It is suggested that more new methods (such as graph neural networks, Transformer, etc.) from the past two years be incorporated into the Introduction, and more references related to these methods should be appropriately introduced. Such as:

(1)Graph neural network

[1] Lou Y, Liu Y. Mineral Prospectivity Mapping Based on a Novel Self-Ensembling Graph Convolutional Network[J]. Mathematical Geosciences, 2025: 1-28.

[2] Xu Y, Zuo R. An interpretable graph attention network for mineral prospectivity mapping[J]. Mathematical Geosciences, 2024, 56(2): 169-190.

(2)Transformer:

[1] Ning Y, Wang Y, Lu J, et al. Mineral prospectivity mapping for multi-source geoscience data: A novel unsupervised deep learning method[J]. Ore Geology Reviews, 2025: 106866.

[2] Gao L, Gopalakrishnan G, Nasri A, et al. Transformer–GCN Fusion Framework for Mineral Prospectivity Mapping: A Geospatial Deep Learning Approach[J]. Minerals, 2025, 15(7): 711.

That is, the introduction to the new methods proposed in 2024 or 2025 as mentioned above.

Samples

(1)In Lines 517-518, a detailed explanation of how the 101 mineralized units were obtained should be provided.

(2)Training Sample Independence：The paper mentions "101 mineralized units" and "101 non-mineralized units."

In Lines 570-578, the role of the confusion matrix is introduced. Please include the results of the confusion matrix and conduct an analysis on them.
Figures

(1)It is suggested that latitude and longitude information be added to Figures 4, 5, 7 and 10.

(2)The labels in Figure 2 and Figure 5 are too small; increase the font size for readability.

Lines 633-641 indicate that the focus of this research is on the fusion and utilization of multi-source data. It is suggested that the authors elaborate on the data processing procedure in combination with the details of the three methods in Section 3, and highlight the data fusion operation to reflect the inherent multi-factor coupling in the orogenic mineral systems.

It is recommended that a comparative analysis of the advantages and disadvantages of these three methods be incorporated into Section 4.2.2 based on the experimental results.

In Lines 656-657, it is mentioned that ensemble learning can effectively capture the coupled effects of alteration footprints, geochemical enrichment, and structural focusing. It is suggested that the author should elaborate on this in Section 3 by dataset and model details.

Discussion on Practical Application

(1)The "Target Areas" identified in Figure 13 look promising, but the discussion is somewhat academic.

(2)Are any of the new high-potential zones located in areas previously dismissed by traditional prospecting?

(3)A brief comparison with existing geological maps or recent field reports would significantly strengthen the claim that this ensemble learning approach adds value over standard exploration methods.

Additionally, explain why Sb and Hematite were identified by SHAP as the top predictors from a metallogenic perspective for this specific deposit type in Tibet.

It is suggested that the author summarize the paper and add a conclusion section.

11.Terminology

Ensure "Ensemble Learning" is used consistently rather than switching between "Integrated Learning" and "Machine Learning" haphazardly.

12.References

Several citations (e.g., Ref 50) are listed with 2025 dates. Ensure these are "in press" or update the metadata if they have been officially released.

Comments on the Quality of English Language

It is better to improve before publication.

Author Response

Dear reviewer

Re: Manuscript ID: remotesensing-4126724

In the following, I will respond to each of the suggestions and questions raised by the reviewers in turn.

Q1. Suggestions from reviewer: (1) In Lines 89-91, I believe that deep learning, neural networks, and convolutional neural networks are not in a parallel relationship. It is recommended that the author revise the expression.

Response: Thank you for the helpful comment. We agree that deep learning, neural networks, and convolutional neural networks are not parallel concepts, as convolutional neural networks are a class of neural networks and belong to deep learning architectures. We have revised the wording in Lines 89–91 accordingly by restructuring the sentence to reflect the correct hierarchical relationship and avoid listing them as equivalent categories.

Q2. Suggestions from reviewer: The current representative methods in Lines 89-91 is insufficient. It is suggested that more new methods (such as graph neural networks, Transformer, etc.) from the past two years be incorporated into the Introduction, and more references related to these methods should be appropriately introduced. Such as: (1)Graph neural network [1] Lou Y, Liu Y. Mineral Prospectivity Mapping Based on a Novel Self-Ensembling Graph Convolutional Network[J]. Mathematical Geosciences, 2025: 1-28. [2] Xu Y, Zuo R. An interpretable graph attention network for mineral prospectivity mapping[J]. Mathematical Geosciences, 2024, 56(2): 169-190. (2)Transformer: [1] Ning Y, Wang Y, Lu J, et al. Mineral prospectivity mapping for multi-source geoscience data: A novel unsupervised deep learning method[J]. Ore Geology Reviews, 2025: 106866. [2] Gao L, Gopalakrishnan G, Nasri A, et al. Transformer–GCN Fusion Framework for Mineral Prospectivity Mapping: A Geospatial Deep Learning Approach[J]. Minerals, 2025, 15(7): 711. That is, the introduction to the new methods proposed in 2024 or 2025 as mentioned above.

Response: Thank you for this constructive suggestion. We agree that the original Introduction (Lines 87–88) did not sufficiently cover recent advances in mineral prospectivity mapping, particularly the rapid development of graph neural networks and Transformer-based geospatial deep-learning frameworks in the past two years. Following your recommendation, we have revised the Introduction to (i) expand the overview of state-of-the-art approaches, and (ii) incorporate representative recent studies published in 2024–2025, including self-ensembling graph convolutional networks, interpretable graph attention networks, and Transformer/Transformer–GCN fusion frameworks for multi-source geoscience data.

Q3. Suggestions from reviewer: (1) In Lines 517-518, a detailed explanation of how the 101 mineralized units were obtained should be provided. (2)Training Sample Independence：The paper mentions "101 mineralized units" and "101 non-mineralized units."

Response: Thank you for these important comments. We agree that the original manuscript did not sufficiently explain how the 101 mineralized units were generated, nor did it clearly address sample-selection independence and potential spatial autocorrelation. In the revised manuscript, we have clarified the sample-construction procedure as follows. (1) How the 101 mineralized units were obtained: We first generated a regular fishnet grid (500 m × 500 m) over the study area. Then, a 750 m buffer was delineated around each of the 14 known Au–Sb–Pb deposits/occurrences. Grid cells whose centroids fall within the buffer zones were labeled as mineralized (positive) units. Under this rule, the buffered areas intersected 101 grid cells in total; therefore, 101 mineralized units were obtained for model training. (2) Sample selection, spatial independence, and spatial autocorrelation: Non-mineralized (negative) samples were not chosen arbitrarily. Instead, candidate negative regions were defined as (i) areas outside the 750 m buffers and (ii) geochemical negative-anomaly zones, which together represent locations with relatively low mineralization potential. Negative samples were then randomly generated from these candidate regions, and to reduce spatial clustering of negatives, we enforced a minimum inter-point spacing of >1000 m between any two selected negative samples. Revised text (Lines 585–607).

Q4. Suggestions from reviewer: In Lines 570-578, the role of the confusion matrix is introduced. Please include the results of the confusion matrix and conduct an analysis on them.

Response: Thank you for this helpful suggestion. We have now added the confusion-matrix results and a concise interpretation based on the 5-fold cross-validation. Specifically, we included a new figure (Figure 10: Confusion-Matrix Results for GBDT, XGBoost, and LightGBM Models) and reported the aggregated counts for each model: GBDT (TN = 82, FP = 19, FN = 3, TP = 98), XGBoost (TN = 81, FP = 20, FN = 4, TP = 97), and LightGBM (TN = 82, FP = 19, FN = 3, TP = 98). The results show consistently low false negatives across all three models, indicating strong capability in identifying mineralized units and thus reducing the risk of missing prospective targets. Meanwhile, the false positives remain moderate, suggesting that the additional field-verification workload is manageable under practical exploration constraints. Revised text (Lines 679–689).

Q5. Suggestions from reviewer: Figures (1)It is suggested that latitude and longitude information be added to Figures 4, 5, 7 and 10. (2)The labels in Figure 2 and Figure 5 are too small; increase the font size for readability.

Response: Thank you for these helpful comments. We have revised the figures to improve map readability and cartographic completeness. Specifically, we added latitude/longitude graticules (or coordinate ticks) to Figures 4, 5, 7, and 10. In addition, we increased the font size of labels in Figures 2 and 5 (including legend text and annotations where applicable) to ensure clear readability in the final printed format.

Q6. Suggestions from reviewer: Lines 633-641 indicate that the focus of this research is on the fusion and utilization of multi-source data. It is suggested that the authors elaborate on the data processing procedure in combination with the details of the three methods in Section 3, and highlight the data fusion operation to reflect the inherent multi-factor coupling in the orogenic mineral systems.

Response: Thank you for this important suggestion. We agree that the original manuscript did not sufficiently emphasize the operational steps of multi-source data fusion. In the revised manuscript, we have expanded the workflow description in Section 4.1.5 (Sample Dataset Preparation) to clearly present how multi-scale predictors from different sources were harmonized and fused prior to model training. Specifically, we now describe: (i) regular fishnet gridding at 500 m × 500 m to define consistent prediction units (11,211 cells); (ii) conversion of all predictor layers to raster format and harmonization of spatial resolution and map projection; (iii) assignment of the ten predictors to each grid cell by extracting values at the cell centroid, thereby unifying the spatial support scale while preserving the geological meaning and spatial heterogeneity of each variable; and (iv) standardization of continuous predictors prior to machine-learning modeling to ensure comparability and stable convergence across the three Boosting models (GBDT, XGBoost, and LightGBM). These revisions make the fusion procedure explicit and traceable, and highlight the multi-factor coupling logic required for prospectivity modeling in an orogenic metallogenic setting. (Lines 751–773).

Q7. Suggestions from reviewer: It is recommended that a comparative analysis of the advantages and disadvantages of these three methods be incorporated into Section 4.2.2 based on the experimental results.

Response: Thank you for this helpful suggestion. We have revised Section 4.2.2 to include a concise, result-driven comparison of GBDT, XGBoost, and LightGBM. Specifically, we discuss the trade-offs among the three models using the experimental outcomes, including the confusion-matrix statistics (FN/FP) and key performance metrics (Precision, F1-score, and AUC). We emphasize that all three models yield low FN counts (GBDT: 3; XGBoost: 4; LightGBM: 3), while LightGBM provides the most balanced Precision–F1 performance and the highest AUC (0.94), supporting its selection for subsequent mapping.

Q8. Suggestions from reviewer: In Lines 656-657, it is mentioned that ensemble learning can effectively capture the coupled effects of alteration footprints, geochemical enrichment, and structural focusing. It is suggested that the author should elaborate on this in Section 3 by dataset and model details.

Response: Thank you for the reviewer’s constructive suggestion. We agree that it is necessary to more fully clarify the correspondence among the “coupling of multi-source ore-controlling factors, the model learning mechanism, and the geological interpretation.” To this end, we have supplemented and expanded the relevant text in the revised manuscript, further explaining how multi-source factors are learned by the models and linked to the underlying geological processes. These revisions have been incorporated in the corresponding paragraph of the revised manuscript. Revised text (Lines 774–805).

Q9. Suggestions from reviewer: Discussion on Practical Application (1)The "Target Areas" identified in Figure 13 look promising, but the discussion is somewhat academic. (2)Are any of the new high-potential zones located in areas previously dismissed by traditional prospecting? (3)A brief comparison with existing geological maps or recent field reports would significantly strengthen the claim that this ensemble learning approach adds value over standard exploration methods.

Response: We sincerely thank the reviewer for the constructive comments on strengthening the practical discussion of the target areas (Figure 13). We agree that the original Discussion was overly academic and did not sufficiently demonstrate the incremental exploration value of the identified targets relative to conventional approaches. Accordingly, we expanded the Discussion to provide a more application-oriented interpretation of the LightGBM-based mineralization potential map, and we added new field verification results based on route-based investigations and representative outcrop checks (see the revised Discussion; locations are shown in Figure 12). Revised text (Lines 822–867).

Q10. Additionally, explain why Sb and Hematite were identified by SHAP as the top predictors from a metallogenic perspective for this specific deposit type in Tibet.

Response: We appreciate the reviewer’s request for a metallogenic interpretation of the SHAP results. The prominence of Sb and hematite is consistent with the Au–Sb (±Pb) hydrothermal system in southern Tibet rather than being a purely statistical artifact. Sb is a key ore-related/pathfinder element (commonly hosted by stibnite) and typically forms high-contrast regional anomalies, making it highly discriminative for mineralized units. Hematite represents ferric-iron alteration/oxidized halos that commonly develop along fault-controlled hydrothermal pathways and can be robustly detected by multispectral/hyperspectral data via diagnostic VNIR features. Under the unified 500 m grid, the co-occurrence of Sb enrichment and iron-oxide alteration (often proximal to faults) reflects the coupled “geochemistry–alteration–structure” process; boosting tree models capture such nonlinear interactions, and SHAP accordingly assigns high contributions. We have added this explanation in the revised Discussion.

Q11. Suggestions from reviewer: It is suggested that the author summarize the paper and add a conclusion section.

Response: Thank you for this helpful suggestion. We agree that a dedicated conclusion improves the clarity and completeness of the manuscript. In the revised version, we have added a Conclusion section. The manuscript has been revised accordingly (see the newly added Conclusion section). Revised text (Lines 868–908).

Q12. Suggestions from reviewer: Terminology. Ensure “Ensemble Learning” is used consistently rather than switching between “Integrated Learning” and “Machine Learning” haphazardly.

Response: Thank you for the suggestion. We have carefully checked the manuscript and standardized the terminology throughout. Specifically, we consistently use “Ensemble Learning” when referring to the modeling framework (GBDT, XGBoost, and LightGBM), and we use “Machine Learning” only as the broader umbrella term when discussing the general methodological category. Inappropriate or inconsistent uses of “Integrated Learning” have been corrected accordingly.

Q13. Suggestions from reviewer: Several citations (e.g., Ref 50) are listed with 2025 dates. Ensure these are "in press" or update the metadata if they have been officially released.

Response: Thank you for the reminder. We have re-checked all references listed with 2025 publication dates. In particular, Ref. 50 has been confirmed as not yet formally released, and we have therefore revised its status to “in press” in the reference list.

We have also carefully re-checked the full manuscript for clarity, consistency, and formatting, and revised accordingly. We sincerely appreciate your time and constructive feedback, which have helped improve the quality of the manuscript. We hope that the revised version will meet your expectations.

Thank you and best regards.

Sincerely.

Kai Qiao

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Dear authors, my main comments regarding your article proposel are the following:

-In the Highlights section, remove the default text.

-The Introduction is very long, consider separating the metallogenic/tectonic part to the Geological Profile section while maintaining the prospectivity methods in the Introduction.

-Figure 3 needs scale to understand the dimension of the noise.

-I do not find the explanation for the different orders of alteration used in Figure 4, nor the alteration projected in Figure 5. Please detail in the text and describe better the approach in the text.

Best regards

Author Response

Dear Reviewer,

Re: Manuscript ID: remotesensing-4126724

Below, we respond to each of the suggestions and questions raised by the reviewer in turn.

Q1. Suggestions from reviewer: In the Highlights section, remove the default text.

Response: Thank you for the suggestion. We have removed the default/placeholder text from the Highlights section.

Q2. Suggestions from reviewer: The Introduction is very long, consider separating the metallogenic/tectonic part to the Geological Profile section while maintaining the prospectivity methods in the Introduction.

Response: Thank you for this helpful suggestion. We agree that the original Introduction was overly detailed. In the revised manuscript, we have removed the metallogenic and tectonic background description from the Introduction and retained only the content directly related to mineral prospectivity prediction and methodological motivation. The removed geological background information has been streamlined and relocated to the Geological Profile section to improve the manuscript structure and readability.

Q3. Suggestions from reviewer: Figure 3 needs scale to understand the dimension of the noise.

Response: Thank you for the suggestion. We have added a scale bar to Figure 3 in the revised manuscript to clearly indicate the spatial dimension of the noise and improve interpretability.

Q4. Suggestions from reviewer: I do not find the explanation for the different orders of alteration used in Figure 4, nor the alteration projected in Figure 5. Please detail in the text and describe better the approach in the text.

Response: We thank the reviewer for this important comment. Please note that the relevant figures have been renumbered in the revised manuscript: the original Figure 4 and Figure 5 have been updated to Figure 5 (Sentinel-2 alteration anomaly grading results) and Figure 6 (GF-5 MTMF alteration-mineral mapping results), respectively. In response to the concerns regarding the explanation of alteration orders/levels and the mapping/projection procedure, we have expanded and clarified the text as follows: (1) Basis for alteration orders/levels (revised Figure 5): We added a detailed iCrosta grading workflow for the PCA outputs. Histogram statistics were computed for the principal-component pixel values related to hydroxyl- and ferric iron–associated anomalies to derive the mean (μ) and standard deviation (σ). A dynamic thresholding range was defined using μ + nσ, and anomaly intensity was classified into three levels: Level I (μ + 3σ), Level II (μ + 2.5σ), and Level III (μ + 2σ). This explicitly links the “orders/levels” to their statistical meaning and anomaly strength. A median filter was then applied to suppress isolated speckle pixels and enhance spatial continuity, yielding the final anomaly maps (Figure 5). (2) GF-5 alteration-mineral mapping (revised Figure 6): We clarified the MTMF mapping and screening logic. Following endmember identification, MTMF was applied to the GF-5 hyperspectral data, and pixels with high MTMF scores and low infeasibility were selected as reliable regions of interest (ROIs). Based on these ROIs, spatial distribution maps of the major alteration minerals (limonite, hematite, muscovite, and epidote) were produced for the study area (Figure 6).

We have also carefully checked the manuscript for clarity and consistency and revised accordingly. We would like to thank you for allowing us to resubmit a revised copy of the manuscript, and we sincerely appreciate your time and consideration. We hope that the revised manuscript meets your expectations.

Thank you and best regards.

Kai Qiao

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

This revised manuscript has been substantially revised and improved in accordance with my comments. It is recommended that this version could be accepted for publication in the journal of Remote Sensing.

Author Response

Dear Reviewer,
Thank you very much for your time and effort in reviewing our revised manuscript. We sincerely appreciate your positive evaluation and recommendation for acceptance. Your constructive comments have greatly helped us improve the quality and clarity of the manuscript.
With kind regards,
Kai Qiao
On behalf of all authors

Reviewer 2 Report

Comments and Suggestions for Authors

Not further comments from my side. Excellent work after the revision.

Author Response

Dear Reviewer,
Thank you very much for your positive feedback. We sincerely appreciate your recognition of our revision and are grateful that you have no further comments. Your support and constructive input have been invaluable in improving the manuscript.
Kai Qiao
On behalf of all authors

Reviewer 3 Report

Comments and Suggestions for Authors

Abstract:
1) In line 38, the phrase "predict mineral prospectivity prediction" is redundant. Please revise it to "predict mineral prospectivity" or "conduct mineral prospectivity prediction."
2)"Ensemble learning" is a key term appearing in the title, highlights, and keywords, but it is not mentioned in the abstract. It is recommended to briefly introduce this core concept in the abstract to improve consistency and readability.

Keywords:
1) "multi-sourcen" is incorrect.
2) Abbreviation formatting of keywords are different.

Introduction:
1)Some citations are consistently placed at the end of sentences such as [1-6], [7-12] and [13-18], they'd better to place at corresponding position.

Methodology:
1)The manuscript proposes an integrated ensemble-learning prospectivity framework. However, Figure 1 illustrates the three models working independently and in parallel. It is unclear how they interact or are integrated. The figure should be revised to better reflect the interaction mechanism (e.g., stacking, voting, or blending) within the ensemble.
2)Some words(Structure data) are not so clear, not same abbreviation formatting(Band selection, ILR transform) in the first line of Figure 1. The last step "Output" includes "Mineral prospectivity map", there is colorful "Mineral prospectivity ma" picture above it, are they same same or not?
3) The specific role of each model within the proposed framework is not clearly defined. Please clarify whether all models work together or each model serves a distinct role.
4) Is it enough without comparing with other mineral prospectivity mapping methods such as random forest, support vectore machine, or CNN, GNN?

Figure Problems:
1) The north arrow, the map coordinates are necessary in Figure 3.
2) The geosptial extent of Figure 3 differs from figure 2(b).
3) Abbreviation formatting of figures' titles must be same and correct. Figure1-8, Figure 9-12.
4) There is an inconsistency in subfigure numbering. The subfigure labels in Figures 9 and 10 (e.g., a, b, c) do not follow the same pattern as those in Figure 8. Additionally, what is currently labeled as Figure 12 should be Figure 13.
5) The current "Figure 12" (Field verification photographs) should be renumbered as Figure 13. Please also ensure that any abbreviations used in this caption are consistent with other figure captions.

Text Issues：
1)There is a wrong paragraph break by the end of Line 534.
2)The formatting of all section headings should be completely consistent throughout the manuscript.
3)Secondary headings are noticeably different, for example, section 3.2 and section 4.1.
4)Tertiary titles' font and capitalization are noticeably different, for example, the subsection of 3.2.1 and the subsection of 4.1.1.
5)Many abbreviations are different in the manuscript，such as line 83-87, NDWI of Line 198, NDVI of line 199, SWIR of line 227, line 237 and line 248. Line 83 omitted SVM.
6) It is repeated "predict to do prediction" in line 38.

Formula Issues:
1) The numbering of all formulas should be placed at the far right of the formula line.

References:
1) Typical and cutting-edge papers about MPM need to be added such as random forest, support vector machine.

Comments on the Quality of English Language

It still necessary to improve before publication.

Author Response

Response to Reviewer Comments

Dear Reviewer

Re: Manuscript ID: remotesensing-4126724

In the following, I will respond to each of the suggestions and questions raised by the reviewers in turn.

Q1. Suggestions from reviewer: Abstract: 1) In line 38, the phrase "predict mineral prospectivity prediction" is redundant. Please revise it to "predict mineral prospectivity" or "conduct mineral prospectivity prediction." 2)"Ensemble learning" is a key term appearing in the title, highlights, and keywords, but it is not mentioned in the abstract. It is recommended to briefly introduce this core concept in the abstract to improve consistency and readability.

Response: Thank you for the helpful comments. We have revised the Abstract accordingly. (1) We removed the redundant phrase “predict mineral prospectivity prediction” and revised it to “predict mineral prospectivity” to improve conciseness and grammar. (2) To ensure consistency with the title, highlights, and keywords, we explicitly introduced ensemble learning in the Abstract by stating that the proposed integrated multi-source workflow is developed within an ensemble-learning framework and that three Boosting-based ensemble models (GBDT, XGBoost, and LightGBM) were trained for mineral prospectivity prediction. These changes have been incorporated in the revised Abstract.

Q2. Suggestions from reviewer: Keywords: 1) "multi-sourcen" is incorrect. 2) Abbreviation formatting of keywords are different.

Response: Thank you for the helpful comments. We have revised the Keywords accordingly. Specifically, we corrected the typo “Multi-sourcen” to “Multi-source” and standardized the capitalization/formatting across all keywords for consistency. The revised Keywords are: “Hyperspectral; Multi-source; Ensemble learning; Mineral prospectivity prediction; Tibet.”

Q3. Suggestions from reviewer: Introduction:1)Some citations are consistently placed at the end of sentences such as [1-6], [7-12] and [13-18], they'd better to place at corresponding position.

Response: Thank you for the helpful suggestion. We have revised the Introduction by relocating citations from sentence ends to the specific clauses/statements they support, so that each reference is placed at its most relevant position. This adjustment improves readability and strengthens traceability of the supporting literature.

Q4. Suggestions from reviewer: Methodology:1)The manuscript proposes an integrated ensemble-learning prospectivity framework. However, Figure 1 illustrates the three models working independently and in parallel. It is unclear how they interact or are integrated. The figure should be revised to better reflect the interaction mechanism (e.g., stacking, voting, or blending) within the ensemble. 2) Some words(Structure data) are not so clear, not same abbreviation formatting(Band selection, ILR transform) in the first line of Figure 1. The last step "Output" includes "Mineral prospectivity map", there is colorful "Mineral prospectivity ma" picture above it, are they same same or not? 3) The specific role of each model within the proposed framework is not clearly defined. Please clarify whether all models work together or each model serves a distinct role. 4) Is it enough without comparing with other mineral prospectivity mapping methods such as random forest, support vector machine, or CNN, GNN?

Response: Thank you for these constructive comments. We agree that the original Figure 1 could be misinterpreted as an “ensemble integration” (e.g., stacking/voting/blending) among models. In this study, the term ensemble learning refers to each individual boosting-based ensemble model (GBDT, XGBoost, and LightGBM) rather than an additional meta-ensemble that combines their outputs. Accordingly, we revised Figure 1 and the related text to make the workflow explicit as independent model training → benchmarking/comparison → best-model selection. Specifically: (1) We clarified the data-fusion step as “Multi-source feature fusion (feature stacking & normalization)”, emphasizing that stacking is performed at the feature level (concatenation of multi-source predictors into a unified feature matrix) rather than model-level stacking. (2) We refined the wording and abbreviation formatting in the first row (e.g., “Structural data”, “ILR transform”), and we standardized acronyms across the figure. (3) We removed ambiguity in the final output by explicitly labeling the map as the final mineral prospectivity map produced by the selected best model (LightGBM), so the “Output” and the map panel refer to the same product. Regarding the suggestion to compare with additional MPM methods (e.g., RF, SVM, CNN/GNN), we acknowledge that broader benchmarking can be informative. However, our primary objective here is a controlled comparison among three representative boosting algorithms that are widely used and highly competitive for small-to-moderate sample sizes, non-linear feature interactions, and tabular multi-source predictors. Including additional model families (particularly deep-learning models such as CNN/GNN/Transformer) would require substantial extra work (e.g., architecture design, hyperparameter tuning, training-data augmentation, spatial/blocked CV, and fairness controls) to avoid biased comparisons, which is beyond the scope of the current revision.

Q5. Suggestions from reviewer: Figure Problems: 1) The north arrow, the map coordinates are necessary in Figure 3. 2) The geosptial extent of Figure 3 differs from figure 2(b). 3) Abbreviation formatting of figures' titles must be same and correct. Figure1-8, Figure 9-12. 4) There is an inconsistency in subfigure numbering. The subfigure labels in Figures 9 and 10 (e.g., a, b, c) do not follow the same pattern as those in Figure 8. Additionally, what is currently labeled as Figure 12 should be Figure 13. 5) The current "Figure 12" (Field verification photographs) should be renumbered as Figure 13. Please also ensure that any abbreviations used in this caption are consistent with other figure captions.

Response: Thank you for these valuable suggestions. We have revised the figures accordingly. First, we added a north arrow and coordinate information (graticules/coordinate ticks) to Figure 3, and clarified that Figure 3 covers a local subset of the area shown in Figure 2(b). Second, we standardized the abbreviation and capitalization formatting across all figure titles/captions and adopted consistent sequential numbering (e.g., Figure 1–8, Figure 9–13). In addition, we adjusted the subfigure labeling in Figures 9 and 10 to follow the same (a–f) scheme as Figure 8. Finally, we corrected the figure numbering: the former Figure 12 (field verification photographs) has been renumbered as Figure 13, and all corresponding in-text citations and caption abbreviations have been updated to ensure consistency throughout the manuscript.

Q6. Suggestions from reviewer: Text Issues：1)There is a wrong paragraph break by the end of Line 534. 2)The formatting of all section headings should be completely consistent throughout the manuscript. 3)Secondary headings are noticeably different, for example, section 3.2 and section 4.1. 4)Tertiary titles' font and capitalization are noticeably different, for example, the subsection of 3.2.1 and the subsection of 4.1.1. 5)Many abbreviations are different in the manuscript，such as line 83-87, NDWI of Line 198, NDVI of line 199, SWIR of line 227, line 237 and line 248. Line 83 omitted SVM. 6) It is repeated "predict to do prediction" in line 38.

Response: Thank you for these careful editorial comments. We have revised the manuscript accordingly: (1) corrected the erroneous paragraph break around Line 534; (2) standardized the formatting of all section and subsection headings to ensure consistent numbering style, font, and capitalization across the entire manuscript (including secondary and tertiary headings); (3) harmonized abbreviation usage and formatting throughout the text (e.g., NDWI, NDVI, SWIR), and added the missing “SVM” where appropriate (around Lines 83–87); and (4) removed the redundant phrasing in Line 38. These changes improve overall readability and consistency.

Q7. Suggestions from reviewer: Formula Issues: 1) The numbering of all formulas should be placed at the far right of the formula line.

Response: Thank you for the helpful suggestion. We have reformatted all equations so that the equation numbers are consistently aligned to the far right margin of each equation line throughout the manuscript, following standard journal typesetting conventions.

Q8. Suggestions from reviewer: References: 1) Typical and cutting-edge papers about MPM need to be added such as random forest, support vector machine.

Response: Thank you for this valuable suggestion. We agree that adding typical and cutting-edge studies on mineral prospectivity mapping (MPM), such as those using Random Forest (RF) and Support Vector Machine (SVM), helps strengthen the literature background and improve the contextual comparison of methods. Accordingly, we have updated the reference list by adding relevant MPM studies (Refs. 20–23 in the revised manuscript).

Thank you and best regards.

Sincerely.

Kai Qiao

Author Response File: Author Response.pdf