A Web-Based National-Scale Coastal Tidal Flat Extraction System Using Multi-Algorithm Integration on AI Earth Platform

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study presents a nationwide tidal flat monitoring system based on remote sensing, demonstrating significant value in engineering implementation, operational application, and localized services. The methodology is rigorous, and the validation is comprehensive. However, the theoretical innovation in algorithms is somewhat limited, and certain technical details require further clarification. I recommend acceptance after minor revisions.

Specific Suggestions for Improvement:

1. Clarify whether the sample distribution in misclassified areas (e.g., regions affected by suspended sediments or human modification) biases the overall accuracy assessment. A stratified accuracy analysis (by land cover type) would strengthen the conclusions.
2. Provide additional validation across different seasons/years, particularly during high-sediment-load periods, to assess the system’s temporal robustness.
3. Evaluate system performance under extreme conditions (e.g., post-typhoon imagery) to verify its operational resilience.
4. Enhance the discussion by linking tidal dynamics (e.g., tidal cycles) to threshold segmentation variability, improving interdisciplinary relevance.
5. Lines 46–59:While summarizing existing advances, explicitly highlight key limitations and unresolved challenges to better contextualize the study’s contributions.
6. Conclusion Streamlining:The conclusion is overly detailed. Condense it to emphasize only the most critical findings and implications.

The study is well-executed but would benefit from addressing the above points to fully solidify its impact.

Author Response

Comments 1: Clarify whether the sample distribution in misclassified areas (e.g., regions affected by suspended sediments or human modification) biases the overall accuracy assessment. A stratified accuracy analysis (by land cover type) would strengthen the conclusions.

Response 1: We sincerely appreciate the reviewer’s valuable suggestion regarding the potential bias in accuracy assessment due to sample distribution in misclassified areas. In our study, we addressed one major source of misclassification—turbid estuaries—by substituting the shortwave infrared band with the near-infrared band, which significantly reduced errors in these regions. Nevertheless, a small number of misclassifications still occur. For example, aquaculture facilities and moored fishing vessels or other floating objects on the sea surface may present spectral characteristics similar to land in water indices. However, as our analysis was based on year-round data, such floating objects are not present throughout the year, and therefore their corresponding non-water frequency values tend to fall between 0 and 1, inevitably introducing minor errors.

Moreover, our classification is binary—distinguishing only between tidal flats and non-tidal flats—so a stratified analysis by land cover type is not essential in this specific context. While we acknowledge the reviewer’s concern, we believe the current approach sufficiently captures the classification performance without additional stratification.

 

Comments 2: Provide additional validation across different seasons/years, particularly during high-sediment-load periods, to assess the system’s temporal robustness.

Response 2: We sincerely thank the reviewer for this insightful suggestion regarding temporal robustness. As the proposed system is designed to extract tidal flats on an annual basis, the classification integrates multi-temporal imagery from the entire year rather than focusing on a single season. This approach inherently accounts for seasonal variations, including periods of high sediment load, and minimizes the influence of short-term anomalies.

In the revised manuscript, we have supplemented the results with the accuracy assessment for 2024 in addition to the original 2020 evaluation. This provides a direct comparison between two distinct years and further validates the system’s stability and robustness over time. (see Section 3.1.1, Lines 459 - 469).

In addition, we have incorporated an interannual comparative analysis of tidal flat changes between 2020 and 2024, with detailed case studies in Hangzhou Bay and Yueqing Bay. The results show that Hangzhou Bay experienced a net expansion, with 139.90 km² of stable tidal flats, 101.12 km² of newly formed tidal flats, and 48.78 km² of lost tidal flats, driven by sediment deposition, hydrodynamic changes, and anthropogenic activities. In contrast, Yueqing Bay exhibited relative stability, with 114.43 km² of stable tidal flats, 5.63 km² gained, and 18.25 km² lost, reflecting a slight net contraction likely influenced by its semi-enclosed hydrodynamic setting and lower human disturbance. These findings highlight regional differences in tidal flat morphological evolution and provide spatially explicit insights to support targeted coastal management (see Section 3.2.3, Lines 539 - 563).

 

Comments 3: Evaluate system performance under extreme conditions (e.g., post-typhoon imagery) to verify its operational resilience.

Response 3: We sincerely thank the reviewer for this valuable suggestion and agree that assessing system performance under extreme conditions would provide additional insights. According to data from the China Meteorological Administration, nine typhoons passed through or affected China in 2020, impacting 23 Sentinel-2 scenes with cloud cover exceeding our threshold, and thirteen typhoons occurred in 2024, affecting 55 scenes. In total, our analysis used 6,901 valid scenes in 2020 and 6,572 in 2024, meaning that typhoon-affected imagery accounted for only a small proportion of the annual dataset.

Although typhoons can cause temporary gaps in usable imagery due to high cloud cover, our year-long time series still provided sufficient cloud-filtered scenes (<60% cloud cover) to capture both high-tide and low-tide conditions. Therefore, these short-term losses did not materially affect the extraction of annual tidal flat extents. We acknowledge that focused analyses using post-typhoon imagery could further test the system’s operational robustness, and this limitation has been noted in the discussion section. (Section 4, Lines 596–598).

 

Comments 4: Enhance the discussion by linking tidal dynamics (e.g., tidal cycles) to threshold segmentation variability, improving interdisciplinary relevance.

Response 4: In the revised discussion, we have re-examined the causes of inter-method differences to avoid conflating tidal effects with other confounding factors. Our updated analysis shows that the lower performance of the Random Forest (RF) method in Hangzhou Bay and Yueqing Bay is primarily due to classification bias in the GWL_FCS30 training dataset, whose official overall accuracy is reported as 86.44%. Because GWL_FCS30 is based on Landsat imagery with a 16-day (or 5-day with multiple satellites) revisit cycle, its temporal coverage is less frequent than Sentinel-2, limiting the availability of images capturing both high- and low-tide conditions within the same period. This can lead to omission errors for tidal flats present only during certain tidal stages, which are then propagated into the RF classification. We have discussed how these limitations influence the classification results in highly dynamic estuarine and aquaculture environments and outlined potential improvements, including cleaning label noise, supplementing with manually validated samples, and integrating multi-source data to better capture tidal variability. (see Section 3.2.2, Lines 524 - 537; Section 4, Lines 598–620).

 

Comments 5: Lines 46–59: While summarizing existing advances, explicitly highlight key limitations and unresolved challenges to better contextualize the study’s contributions.

Response 5: We sincerely appreciate the reviewer’s insightful suggestion. Following this advice, we have revised the introduction to more clearly articulate the key limitations and unresolved challenges, thereby providing stronger context for the study’s contributions. In the revised text, we explicitly state that while traditional wetland monitoring approaches relying on field surveys and aerial photography are costly and spatially constrained, multispectral satellite-based methods—although offering consistent, large-scale coverage—can still face spectral confusion in complex coastal environments with dynamic water-land boundaries. We have also added that many existing open-source tools, such as Google Earth Engine-based mapping platforms and regional Chinese approaches relying on single-algorithm workflows, either provide limited spatial coverage or lack standardized, scalable pipelines suitable for nationwide, long-term operations, with localization and compliance constraints further complicating deployment in mainland China. (see Section 1, Lines 110 - 121)

 

Comments 6: Conclusion Streamlining: The conclusion is overly detailed. Condense it to emphasize only the most critical findings and implications.

Response 6: We appreciate the reviewer’s suggestion regarding the length of the conclusion. In response, we have condensed the section to focus only on the most critical findings and implications, removing redundant methodological details and secondary descriptions. The revised conclusion now emphasizes the system’s key contributions, performance outcomes, comparative advantages, and practical significance for coastal wetland monitoring. (see Section 5, Lines 622 - 638).

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript titled “A Web-Based National-Scale Coastal Tidal Flat Extraction System Using Multi-Algorithm Integration on AI Earth Platform” develops a web-based, interactive tidal flat extraction system on Alibaba's AI Earth platform, combining remote sensing indices with machine learning for large-scale, user-friendly mapping and real-time visualization, reducing technical barriers for coastal management in China. This study explores the application of AI Earth in long-term remote sensing detection, providing new ideas and methods for large-area surface monitoring in the future. However, there are some minor issues that need to be addressed, as follows:

1. Introduction: Please review the relevant literature, specifically research on wetland classification and explain why you chose the machine learning (Random Forest) combined approach.

2. 2.1.2 Sentinel-2 Data: In the data section, please describe the time period for the acquired Sentinel-2 data.

3. 2.3.1: What method was used to select samples for tidal flats with vegetation cover?

4. 2.3.2: "The Random Forest classification approach utilizes the tidal flat class (ID 187) from the GWL_FCS30 wetland product as reference training data." Please clarify whether this training data and your classification data are from the same time period. If not, how did you address the issue of temporal mismatch between training samples and data?

5. 2.5.2 and 2.5.3: I suggest simplifying these two sections and including them in the results section instead. They could be summarized briefly.

6. 3. Results: The titles in "1.1. Validation of National Tidal Flat Extraction Results" and "1.2. Verification of Key Tidal Flat Extraction Results in Zhejiang Province's Key Bay Area" are incorrect. Should they be 3.1 and 3.2?

7. In the 3. Results section, the validation methods should be described in Section 2, while this section should only focus on the validation results. In the results, only accuracy validation is described. Were specific classification results also presented?

8. There are too many figures, some of which are unnecessary. Please optimize the figures based on this feedback.

Author Response

Comments 1: Introduction: Please review the relevant literature, specifically research on wetland classification and explain why you chose the machine learning (Random Forest) combined approach.

Response 1: Thank you for the suggestion. We have revised the Introduction to add relevant literature on wetland classification, particularly studies using machine learning methods, and to explain why we chose the combined Random Forest and thresholding approach. (see Section 1, Lines 64 - 73).

 

Comments 2: 2.1.2 Sentinel-2 Data: In the data section, please describe the time period for the acquired Sentinel-2 data.

Response 2: We thank the reviewer for this valuable suggestion. In the revised manuscript, we have updated Section 2.1.2 to explicitly describe the time period of the Sentinel-2 data used in this study. Specifically, we clarify that while the dataset spans multiple years, the present analysis focuses on the years 2020 and 2024 as representative temporal snapshots for comparative assessment. We have also added the total number of usable Sentinel-2 Level-2A scenes for each year (6,901 in 2020 and 6,572 in 2024, with cloud cover <60%), and included a monthly distribution figure (Section 2.1.2, Lines 189 – 199, Figure 1) to illustrate the temporal availability of imagery.

 

Comments 3: 2.3.1: What method was used to select samples for tidal flats with vegetation cover?

Response 3: We appreciate the reviewer’s suggestion and have clarified the scope of our mapping. In this study, “tidal flats” specifically refer to unvegetated intertidal flats—bare intertidal surfaces. Vegetated intertidal areas (e.g., mangroves, saltmarshes, seagrass meadows) are excluded from the target class and treated as non-target categories (see Section 2.3, Lines 272–273). Consistent with this definition, the thresholding approach was applied to delineate unvegetated surfaces, and the Random Forest model was trained using “Tidal flat” (ID 187) samples from the GWL_FCS30 product after visual screening to retain only unvegetated pixels (see Section 2.3.2, Lines 338–342). This clarification has been incorporated into the revised manuscript.

 

Comments 4: 2.3.2: "The Random Forest classification approach utilizes the tidal flat class (ID 187) from the GWL_FCS30 wetland product as reference training data." Please clarify whether this training data and your classification data are from the same time period. If not, how did you address the issue of temporal mismatch between training samples and data?

Response 4: We appreciate the reviewer’s comment and agree that temporal consistency between training and classification data is important. In our study, the GWL_FCS30 wetland product used as reference training data corresponds to the same time period as the classification imagery. Therefore, there is no temporal mismatch between the training samples and the data used for classification. The manuscript has been revised accordingly (see Section 2.3.2, Lines 339–342).

 

Comments 5: 2.5.2 and 2.5.3: I suggest simplifying these two sections and including them in the results section instead. They could be summarized briefly.

Response 5: We appreciate the reviewer’s suggestion. We have simplified Sections 2.5.2 and 2.5.3 by summarizing the key points and moved the content to the Results section as a new subsection titled “Visualization and Statistical Analysis of Results”. The descriptions have been condensed to focus on the main visualization features and statistical outputs, and the related figures have been retained to illustrate the results (see Section 3.3, Lines 564–580).

 

Comments 6: 3. Results: The titles in "1.1. Validation of National Tidal Flat Extraction Results" and "1.2. Verification of Key Tidal Flat Extraction Results in Zhejiang Province's Key Bay Area" are incorrect. Should they be 3.1 and 3.2?

Response 6: We appreciate the reviewer’s careful observation. We have corrected the numbering in the Results section. The titles “1.1. Validation of National Tidal Flat Extraction Results” and “1.2. Verification of Key Tidal Flat Extraction Results in Zhejiang Province’s Key Bay Area” have been updated to “3.1” and “3.2” respectively in the revised manuscript. (see Section 3.1, Line 458; Section 3.2, Line 488).

 

Comments 7: In the 3. Results section, the validation methods should be described in Section 2, while this section should only focus on the validation results. In the results, only accuracy validation is described. Were specific classification results also presented?

Response 7: We appreciate the reviewer’s constructive suggestion. In the revised manuscript, the description of validation methods has been moved entirely to Section 2.6 (“Validation Framework and Methods”), while Section 3 now focuses solely on presenting the validation results. (see Section 2.6, Lines 422 - 452).

 

Comments 8: There are too many figures, some of which are unnecessary. Please optimize the figures based on this feedback.

Response 8: We appreciate the reviewer’s suggestion regarding figure optimization. In the revised manuscript, we carefully reviewed all figures and removed non-essential ones, including the original system functional module diagram, the multi-temporal comparison visualization from the visualization interface section, and the validation sample distribution figure. As the sample distribution can be sufficiently described in the text, a separate figure was deemed unnecessary. For clarity and conciseness, we now retain only one representative screenshot per functional page.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper ​​designs and builds​​ an automatic coastal tidal flat extraction system based on the ​​AI Earth platform​​. This system has ​​important new ideas​​ and ​​practical value​​. The research ​​combines two methods​​ (thresholding method and Random Forest) and uses ​​WebGIS technology​​. The paper was generally well written and showed  ​​good experimental tests​​. But some details ​​need improvement​​.
1. The paper ​​only checks​​ data from the year 2020. This ​​cannot show​​ how tidal flats change over time. 
2. The paper ​​did not directly compare​​ the system's performance with ​​Google Earth Engine (GEE)​​ when checking accuracy. It is suggested to ​​add tests​​ that compare both platforms, e.g., testing the ​​same area and time period​​, then compare two metrics: Kappa coefficient​​ (accuracy score) and Time cost​​ (how fast they process data).
3. Can applications based on AI Earth only rely on the user interface, or can they be implemented programmatically, like Google Earth Engine (GEE)?
​​4. Missing reference. ​​The ​​Peak Start Threshold (PST) method​​ is mentioned, but ​​no reference​​ is given. ​​GWL_FCS30 wetland product​​ needs a citation.
5. ​​Typo correction​​ in Line 502: The word "​​thresholding​​" is repeated.

Author Response

Comments 1: The paper only checks data from the year 2020. This cannot show how tidal flats change over time.

Response 1: We thank the reviewer for this helpful suggestion and agree that incorporating a temporal perspective is important. In response, we added an interannual change analysis at the local scale for Hangzhou Bay and Yueqing Bay (see Section 3.2.3, Lines 539 - 563). The results show that Hangzhou Bay experienced a net expansion of tidal flats (stable: 139.90 km², gained: 101.12 km², lost: 48.78 km²), whereas Yueqing Bay remained relatively stable with a slight contraction (stable: 114.43 km², gained: 5.63 km², lost: 18.25 km²). This new comparison directly addresses the reviewer’s concern by providing spatially explicit evidence of tidal flat change between 2020 and 2024, enhancing the temporal relevance of our findings.

 

Comments 2: The paper did not directly compare the system's performance with Google Earth Engine (GEE) when checking accuracy. It is suggested to add tests that compare both platforms, e.g., testing the same area and time period, then compare two metrics: Kappa coefficient (accuracy score) and Time cost (how fast they process data).

Response 2: We appreciate the reviewer’s suggestion to compare the proposed system’s performance with Google Earth Engine (GEE). The main objective of this study is the development and application of the tidal flat extraction system on the AI Earth (AIE) platform, rather than a systematic performance benchmarking across platforms. Nevertheless, we conducted a small-scale comparative test using the same extraction method on both AIE and GEE for two representative areas—Hangzhou Bay (165 Sentinel-2 scenes from 2020, cloud cover <60%) and Yueqing Bay (241 scenes from 2020, cloud cover <60%)—under a Windows 11 24H2 environment with an Intel i7-13700F CPU and NVIDIA T1000 GPU.

The results showed similar Kappa coefficients between the two platforms, with AIE achieving slightly higher accuracy and GEE showing faster processing times. Since this comparison was not the focus of our work, and the analysis was intended only for internal reference, we decided not to include it in the revised manuscript.

 

Comments 3: Can applications based on AI Earth only rely on the user interface, or can they be implemented programmatically, like Google Earth Engine (GEE)?

Response 3: We thank the reviewer for the question. The AI Earth platform itself provides both a visual toolbox for interactive operations and a code-based interface similar to Google Earth Engine (GEE). However, the web-based tidal flat extraction system developed in this study is designed to support only the visual, user interface–based operations to make it more accessible to non-programming users.

 

Comments 4: Missing reference. The Peak Start Threshold (PST) method is mentioned, but no reference is given. GWL_FCS30 wetland product needs a citation.

Response 4: We thank the reviewer for this comment. We have already provided a brief explanation of the Peak Start Threshold (PST) principle in Section 2.3.1 (see Lines 319 - 329), while the full methodological details will be published in a separate dedicated paper.

For the GWL_FCS30 wetland product, we have added the appropriate reference in the revised manuscript. (see Section 2.3.2, Line 339).

 

Comments 5: Typo correction in Line 502: The word "thresholding" is repeated.

Response 5: We appreciate the reviewer’s careful observation. The repeated word “thresholding” in Line 502 has been removed in the revised manuscript. (see Section 3.2.2, Line 511).

Reviewer 4 Report

Comments and Suggestions for Authors

Comments and Suggestions for Authors The authors provide an interactive web-based Coastal Tidal Flat Extraction system. They provide wide-ranging data types for the user all in open-source platform. This is a wonderful idea and looks to be very well executed. I applaud the authors and their efforts. This will be very useful for all future studies on coastal geomorphology. I trust your data portal will be used by many. I also think this paper is very well-written. I have a handful of line edits below. I suggest minor comments.

Comments 1: Line 58:  Add abbreviations, while the Wetland Area and Dynamics for Methane Modeling (WAD2M) global wetland dataset developed by Zhang's team

Comments 2: Line 63-64:  I would suggest, “By harnessing distributed computing capabilities to process vast collections satellite datasets including Landsat and MODIS imagery”

Comments 3: Line 275: The thresholding process employs two complementary approaches to convert the continuous index values into binary water/non-water classifications. (What kind of threshold you calculated here?? Can you explain in details)

Comments for author File: Comments.docx

Author Response

Comments 1: Line 58: Add abbreviations, while the Wetland Area and Dynamics for Methane Modeling (WAD2M) global wetland dataset developed by Zhang's team

 Response 1: We thank the reviewer for the suggestion. In the original manuscript, this sentence appeared at Line 58. In the revised manuscript, due to text adjustments and restructuring, it now appears at Lines 62–63. We have added the abbreviation “WAD2M” after the first mention of the Wetland Area and Dynamics for Methane Modeling global wetland dataset accordingly.

 

Comments 2: Line 63-64:  I would suggest, “By harnessing distributed computing capabilities to process vast collections satellite datasets including Landsat and MODIS imagery”

Response 2: We appreciate the reviewer’s suggestion. The suggested revision for Lines 63–64 in the reviewer comment now corresponds to Lines 78–79 in the revised manuscript, due to changes in paragraph structure and content adjustments. We have incorporated the reviewer’s suggestion accordingly.

 

Comments 3: The thresholding process employs two complementary approaches to convert the continuous index values into binary water/non-water classifications. (What kind of threshold you calculated here?? Can you explain in details)

Response 3: We thank the reviewer for raising this point and for the opportunity to clarify our method. We acknowledge that our original wording “two complementary approaches” could lead to misunderstanding. In our system, the Otsu and Peak Start Threshold (PST) algorithms are two independent and alternative thresholding methods, and the user selects one of them for each classification task; they are not combined in a single workflow.

Otsu Method:

This is a widely used global thresholding technique that analyses the histogram of the water index image and selects the threshold that maximizes the inter-class variance between water and non-water classes. It is particularly effective when the histogram exhibits a bimodal distribution, which is common in many coastal environments. However, in complex coastal areas with turbid waters or mixed pixels, the bimodal assumption may not hold, leading to reduced performance.

Peak Start Threshold (PST) Method:

The PST method, developed in this study, aims to improve classification in heterogeneous coastal zones. It is based on NDVI histogram analysis from Sentinel-2 imagery and UAV hyperspectral data. In a typical coastal NDVI histogram, three peaks are often observed: (i) low values for water bodies, (ii) medium values for bare tidal flats, and (iii) high values for vegetated land. PST selects the starting point of the main land-related peak (ii) closest to zero NDVI as the water/land threshold. This is because the “low-value valley” between peaks (i) and (ii) corresponds to the shoreline and transition zones, as confirmed by UAV validation. Using this “peak start” threshold enables robust separation even where water and land classes overlap spectrally.

In both methods, the determined threshold converts continuous water index values into binary water/non-water classifications. This classification is performed on each image, and multi-image results are aggregated over the specified time window to generate the “Non-Water Body Occurrence Probability Map”, which quantifies tidal flat persistence.

We have revised Section 2.3 (See Line 274) in the manuscript to explicitly state that Otsu and PST are alternative algorithms.

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have fully addressed all my concerns.