Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning

Meng, Jie; Xu, Duanyang; Tao, Zexing; Ge, Quansheng

doi:https://doi.org/10.5194/essd-2025-264

Preprints

https://doi.org/10.5194/essd-2025-264

Preprints

24 Jun 2025

| 24 Jun 2025

Status: a revised version of this preprint is currently under review for the journal ESSD.

Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning

Jie Meng, Duanyang Xu, Zexing Tao, and Quansheng Ge

Abstract. Sandy beaches provide essential ecological and economic services, but their functions are increasingly threatened by human activities. Analyzing the spatial distribution of China's sandy beaches and the impacts of human activities offers valuable insights for coastal resource management and ecological protection. However, remote sensing technologies face challenges such as limited data sources and tidal influences, which affect recognition accuracy. Therefore, integrating multi-source remote sensing data and reducing the impact of tidal fluctuations to improve recognition accuracy remains a key challenge. This study proposes an innovative approach utilizing multi-source data and an ensemble learning model to identify sandy beaches in China (2016–2023). By integrating Sentinel-1/2 satellite data, terrain data, and nighttime light data, along with spectral, terrain, texture, and polarization features, sandy beaches were identified across multiple years, and the results were consolidated into a single-year dataset to analyze spatial patterns and risks from human infrastructure squeeze. (1) High-precision classification identified 2984 sandy beaches in China, covering a total area of 260.70 km². Guangdong had the largest number, area, and perimeter, while Shanghai had the widest sandy beaches. (2) In Fujian, Guangdong, and Taiwan, the identified sandy beaches covered 149.68 km², with perimeters of 5155.91 km and widths of 49.50 m, 32.83 m, and 50.70 m, respectively. These results were significantly better than those from reference datasets. (3) From 1990 to 2023, the area at risk from human infrastructure squeeze increased from 109.95 km² to 245.58 km², a rise of 135.63 km², with the most significant increase occurring between 1990 and 2000. Guangdong and Fujian showed growth rates of 1.05 km²/year and 0.73 km²/year, respectively. This study provides an up-to-date dataset on China's sandy beaches. It assesses their spatial patterns and human impact risks, contributing to research and policy for the sustainable development of coastal zones (https://doi.org/10.5281/zenodo.15307240, Meng et al., 2025).

Received: 07 May 2025 – Discussion started: 24 Jun 2025

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 2358 KB)

Supplement (1401 KB)

Download & links

Jie Meng, Duanyang Xu, Zexing Tao, and Quansheng Ge

Status: final response (author comments only)

RC1:
'Comment on essd-2025-264', Anonymous Referee #1, 21 Aug 2025

I have carefully read the manuscript entitled « Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning » by Jie Meng et al. and tested associated datasets available in Zenodo. It presents original information on sandy beach locations throughout China, taking the form of a shape file produced using ensemble learning and multi-source data, from which various beach spatial characteristics (e.g., number, width, area) and risk analysis of coastal squeeze can be assessed and analyzed. It is not the first time sandy beaches were mapped in China using satellite imagery - with already existing datasets mentioned and assessed in comparison by the authors - yet it is the first time ensemble learning is used, which according to the authors improved detection accuracy. Besides, when previous studies used Sentinel-2 data only, Sentinel 1 and Google Earth imagery were also included in this study. I have found the manuscript well written and illustrated, although additional careful proofreading by the authors could have avoided several mistakes such as identical section and figure titles, repetitions and typos. Figure captions in particular must be improved. I attach an annotated version of the manuscript, listing a number of technical corrections and specific comments, that the authors should take into account while preparing their revision.
My main comment relates to the evaluation of the data produced and concerns the truth data used for the evaluation, as well as the quality metrics selected. It is not clear how you produced the truth data (testing set) and why this differs from your dataset 1 which is also produced through visual interpretation (since the same method was used, results should be same, however Table 6 indicates this is not the case). I think this requires further explanation in the manuscript. You mention 14,694 features were labelled for testing. However, this spans 5 land cover classes with various repartitions from year to year. This means that in general, less than 400 sandy beaches were used for evaluating your dataset for each year, which is a lot less than 14,694!
Concerning quality metrics, why were they chosen, and are they really complementary to each other? Tables 5 and 6 suggest similar interpretations can be made for all metrics as results do not differ significantly. Additional information should be provided to guide readers in how to interpret these quality metrics and the results obtained.
Likewise, the authors evaluate other (independent) datasets (some produced by the authors themselves, eg. Dataset 1) which they name reference datasets. I think « reference » is misused in this context as it generally implies that the data were used for validation (truth data), which is not the case here.
Additional information could be provided and discussed in regard to the liability and capacities of the machine learning method used. What is the minimum beach size that can be detected? Looking at the dataset, it seems small pocket beaches can remain undetected. In contrast, elongated features (some artificial) may be erroneously detected as sandy beaches. What is the impact of tide range (potentially variable across China) and having images obtained at different tide levels on data consistency and how can this be improved? It is said using annual averages (composite images) helps mitigate this issues, but it is not clear how, particularly as Figure 2 shows an heterogeneous spatial distribution of satellite images and certainly, with regular revisits, this means images are obtained at different tide levels throughout the country. From this could arise systematic biases in the spatial characteristics deduced from the dataset (eg., beach width and area). Was this, and how, mitigated for this study?
The temporal change in coastal squeeze is assessed over the period 1990-2024, yet beach distribution data outside 2016-2024 are not available. Thus, which data did you use for sandy beach spatial coverage? Currently, this is not explained in the text.
In conclusion, despite I see potential in the manuscript, I recommend moderate to major revisions be undertaken before it can be fully considered for eventual publication in ESSD. I hope you will find my comments useful for preparing your revised manuscript.

Citation: https://doi.org/10.5194/essd-2025-264-RC1
- AC1:
  'Reply on RC1', Jie Meng, 30 Oct 2025
  Dear Reviewers:
  Thank you very much for your thoughtful and detailed review. Your suggestions have provided us with important and constructive insights, which have significantly improved the manuscript. We have carefully considered all of your comments and have made substantial revisions to the manuscript based on your feedback. We have done our best to enhance the manuscript and hope that the revised version will meet your approval. A point-by-point response to the outstanding comments is attached to this manuscript. The major revisions are summarized as follows:
  Response to Comments by Reviewer 1:
  
  I have carefully read the manuscript entitled « Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning » by Jie Meng et al. and tested associated datasets available in Zenodo. It presents original information on sandy beach locations throughout China, taking the form of a shape file produced using ensemble learning and multi-source data, from which various beach spatial characteristics (e.g., number, width, area) and risk analysis of coastal squeeze can be assessed and analyzed. It is not the first time sandy beaches were mapped in China using satellite imagery - with already existing datasets mentioned and assessed in comparison by the authors - yet it is the first time ensemble learning is used, which according to the authors improved detection accuracy. Besides, when previous studies used Sentinel-2 data only, Sentinel 1 and Google Earth imagery were also included in this study. I have found the manuscript well written and illustrated, although additional careful proofreading by the authors could have avoided several mistakes such as identical section and figure titles, repetitions and typos. Figure captions in particular must be improved. I attach an annotated version of the manuscript, listing a number of technical corrections and specific comments, that the authors should take into account while preparing their revision.
  
  Response:
  We would like to express our gratitude to the reviewer for their valuable comments. Based on the suggested revisions in the attached document, we have made the following adjustments to the manuscript format:
  1.L53: We have changed "technologies" to "techniques" in the manuscript.
  2.L57: Small grayscale differences refer to situations where the variation in brightness (grayscale value) between different regions or pixels in the image is minimal. For instance, certain areas may display nearly identical grayscale values, which can lead to poor performance of segmentation algorithms in these regions. This typically happens in low-contrast images, regardless of whether the image is grayscale or color.
  3.L62: We have changed "rough" to "exploratory" in the manuscript.
  4.L70: We have changed "tidal image interference" to "the impact of tidal variations on sandy beach extraction from remote sensing images" in the manuscript.
  5.L72: We have changed "is" to "are" in the manuscript.
  6.L76: We have changed "contributing over" to "contributing to over" in the manuscript.
  7.L80: We have modified the sentence from: "However, despite the rapid economic development in these regions, there is a lack of nationwide dynamic monitoring tools for sandy beaches, and the risks posed by human infrastructure squeeze are not well understood." to: "However, despite the rapid economic development in these regions, there is a lack of nationwide dynamic monitoring tools for sandy beaches, and the risks posed by human infrastructure squeeze—particularly due to urban development and coastal expansion—are not well understood."
  8.L86: We have modified the sentence from: "Multi-temporal data from 2016 to 2023 are used to build an annual representative beach dataset, reducing tidal fluctuation impacts and supporting precise mapping and monitoring;" to: "Multi-year sandy beach data extracted using ensemble learning from 2016 to 2024 are merged to construct an annual representative sandy beach dataset, reducing tidal fluctuation impacts and supporting precise mapping and monitoring."
  9.L95: Sandy beaches are typically located along the coastline. To ensure the study area includes as many relevant beaches as possible, while aligning with the recognized coastal zone boundaries defined by experts, we established a buffer zone. Specifically, the study area incorporates a buffer zone extending 10 km inland and 20 km offshore. The primary goal of this design is to maintain the integrity of the sandy beach area, preventing the buffer zone from being too narrow, which could lead to certain beach areas being truncated or excluded from the study region. Such a design would result in an underestimation of China’s sandy beach area, not accurately reflecting the true extent of the coastline.
  10.L101: We have changed "Figure 1: Location of China’s coastal zone and distribution of partial test points for result verification" to "Figure 1: Location of China’s coastal zone and distribution of partial typical land feature points for results verification."
  11.L119: We have changed "Figure 2: Spatial distribution of Sentinel-1 and Sentinel-2 images used in this study: (a) Image Count of Sentinel-1, (b) Image Count of Sentinel-2" to "Figure 2: Spatial distribution of Sentinel-1 and Sentinel-2 images used in this study from 2016 to 2024: (a) Image Count of Sentinel-1, (b) Image Count of Sentinel-2."
  12.L125: Regarding the issue you raised, this is indeed the case. The phrase "this is the range of pixel grayscale values retained" refers to the valid range of grayscale values we retained during the processing. This approach ensures the validity and consistency of the image data while also minimizing potential noise interference.
  13.L142: We have changed "Li et al., 2022" to "Miao et al., 2022" in the manuscript.
  14.L145: We have changed "Study area and materials" to "Methodology" in the manuscript.
  15.L150: Regarding the repetition you mentioned, I included the sentence "To accurately monitor the current status of sandy beaches in China, this study integrates multi-source data from 2016 to 2024" to make the technical process appear more complete and help readers better understand my methodology. This section aims to clearly outline the background and data sources of the study, making the subsequent research methods and steps easier to comprehend.
  16.Regarding Figure 3: First, the field surveys mentioned in L164 were used to determine a subset of training and testing points for beach extraction. Additionally, the "Public data" in the figure was updated to include beach data from the OSM database for comparison. Finally, we have changed the caption of Figure 3 from "Spatial distribution of Sentinel-1 and Sentinel-2 images used in this study: (a) Image Count of Sentinel-1, (b) Image Count of Sentinel-2" to "Figure 3: The technical framework of the study."
  17.L164: The field surveys refer to a simple process where we go to the field to check if the area is a sandy beach. If it is, we record the GPS coordinates (latitude and longitude). It is a straightforward operation, so it was not described in detail here.
  18.L165: This section just establishes a sample library of test points for subsequent validation. These are point data used to verify the minimum pixel at the location of the point in the classification results (which is 10m). The specific points differ each year, and the number of points for each year is shown in Figure 4.
  19.L170: Since I have previously conducted similar work in Fujian Province, and based on the features referenced in the literature, I selected these features as the input variables for the model. The paper "Meng, J., Xu, D., Tao, Z., and Ge, Q.: Sandy Beach Extraction Method Based on Multi-Source Data and Feature Optimization: A Case in Fujian Province, China, Remote Sens., 17(16), 2754, doi:10.3390/rs17162754, 2025" discusses in detail how each feature affects the model. Additionally, we included the importance analysis results in L245, which show the contribution of each input feature to the results each year.
  20.L179: Due to the large number of input features, providing a detailed description would take up a lot of space. As many studies have already used these features, we have provided the calculation methods for each input feature along with the corresponding references for readers to consult.
  21.L190: The parameters here are set to default values. We have replaced them with better base learner models for beach extraction. The reason for listing the parameters is to help readers effectively replicate and use this method.
  22.L193: We have changed the manuscript from "Pixel-based classification algorithms inevitably produce salt-and-pepper noise, and some small patches—such as buildings and bare land—are difficult to distinguish from environmentally influenced sandy beaches (Mattson et al., 2024)" to "Pixel-based classification algorithms inevitably produce salt-and-pepper noise because sandy beaches share similar spectral characteristics with certain areas, such as buildings and bare land, making them difficult to distinguish (Mattson et al., 2024)."
  23.L207: Due to the large number of equations and input parameters, providing a detailed description would take up a lot of space. Since many studies have already used these evaluation metrics for accuracy assessment, we have abbreviated the formulas and cited the corresponding articles. Readers can refer to these articles for more detailed formulas for accuracy evaluation.
  24.L215: Sandy beaches are generally stable features, with minimal changes over time. However, tidal influences can cause some beaches to be misclassified as water bodies due to water level fluctuations, resulting in variations in the extracted beaches from year to year. Therefore, we used multi-year beach extraction results and combined them to create a merged multi-year beach dataset. This approach significantly reduces the impact of tidal fluctuations on beach extraction, providing a more accurate reflection of the real situation regarding the distribution, data, and area of sandy beaches nationwide. Based on this, we used impermeable surfaces from the land-use data of 1990-2024 to create a 100-meter buffer zone for clipping, producing the human infrastructure squeeze risk areas for 1990-2024.
  25.L217: The detailed explanation of Sen's slope would require considerable space. Since this method is widely used and has been employed in many studies, we have cited the relevant references for readers to consult.
  26.L219: We have changed "risk areas across regions" to "risk zones across different regions" in the manuscript.
  27.L222: We have changed the manuscript from "The performance of the ensemble learning algorithm was validated by calculating PA, UA, F1-score, OA, and Kappa coefficient, and we obtained the validation results (Table 5)" to "The ensemble learning algorithm's performance was evaluated using key metrics such as Accuracy, Precision, Recall, sandy beach F1-score, and AUC, with the results presented in Table 5."
  28.L225: In the initial version, I used a relatively simple set of metrics, which led to minimal differences in the results. Based on your suggestions, I have updated the metrics to include Accuracy, Precision, Recall, sandy beach F1-score, and AUC. These updated metrics now provide more distinct results, helping to better assess the model's performance.
  29.L228: The "merged data" refers to the combined beach extraction results from 2016 to 2024. Tidal influences led to some beaches being misclassified as water bodies, causing the beach area to be underestimated. Since sandy beaches are relatively stable features, merging multi-year data helps reflect the current state of the beaches, significantly reducing the impact of tidal fluctuations on beach extraction.
  30.L236: The inconsistency between Figure 5 and its corresponding conclusion was due to an error in drawing the figure. I have corrected this now. Additionally, due to the large longitudinal extent of Guangdong, multiple bars need to be stacked together for comparison.
  31.L236: The data presented here are based on the merged beach data from 2016 to 2024. The width represents the average width of each province, while the perimeter and area reflect the total perimeter and area of each province.
  32.L246: We have changed the manuscript from "Figure 6: Spatial distribution of sandy beaches in China: (a) Spatial distribution of sandy beach numbers, (b) Spatial distribution of sandy beach length, (c) Spatial distribution of sandy beach width, (d) Spatial distribution of sandy beach area" to "Figure 6: Spatial distribution of sandy beaches in China: (a) sandy beach numbers, (b) sandy beach length, (c) sandy beach width, and (d) sandy beach area."
  33.L263: We have replaced "reference" with "published" and made corresponding changes throughout the manuscript. Additionally, we have made the necessary adjustments to Figure 8 as well.
  34.L270: Previously, the manuscript compared various datasets. We have now changed it to a comparison of models. It was not reasonable to equate datasets using the accuracy evaluation method I established, so we removed the accuracy comparison between datasets, as it was meaningless. Instead, we have focused on comparing popular models to highlight the advantages of our model and demonstrate the effectiveness of our sandy beach extraction. However, comparisons of beach distribution and sandy beach attributes across different datasets have still been retained.
  35.L277: We have changed "a clear advantage" to "provides a certain supplementary advantage."
  36.L293: Dataset 3 shows significantly higher values because a large number of other land cover types were misclassified as sandy beaches. Even with other scales, the results remain hard to read. We will include a table in the appendix listing the number of beaches in each province for every dataset.
  37.L295: We have changed the manuscript from "in the risk area of human infrastructure squeeze in the study area" to "in the risk areas of human infrastructure squeeze in the study area."
  38.L304: The unit for the y-axis has been moved closer to the y-axis title.
  39.L311: We have changed the manuscript from "high coastal development intensity" to "have high coastal development density."
  40.L321: We have changed the manuscript from "Figure 10: Spatial changes of human infrastructure squeeze risk area: (a) Current risk area of human infrastructure squeeze, (b) Spatial changes of human infrastructure squeeze risk area from 1990 to 2023" to "Figure 11: Spatial changes of human infrastructure squeeze risk area: (a) current risk area, (b) changes in the risk area from 1990 to 2024."
  41.L340: Regarding the texture features of sandy beaches, at the image scale, certain characteristics behave as expected, but this pattern may not always hold true in different contexts or broader analyses. However, at the 10m scale in this study, this pattern is consistent.
  42.Figure 12: We have modified Figure 12 by retaining only one value.
  43.L368: We have changed the manuscript from "strong applicability" to "robust generalization."
  44.L381: We have changed the manuscript from "Overall, the distribution of sandy beaches in China reflects the combined effects of sediment supply, coastal type, and hydrodynamic conditions, resulting in more sandy beaches in the north and south, and fewer on the central coast" to "Overall, the distribution of sandy beaches in China is shaped by sediment supply, coastal type, and hydrodynamic conditions, with more beaches in the north and south, and fewer along the central coast."
  45.We have revised the manuscript to change "Figure 13: Spatial changes of various factors from 1990 to 2023: (a) Spatial change of per capita GDP, (b) Spatial change of resident population, (c) Spatial change of built-up area, (d) Spatial change of road area" to "Figure 14: Spatial changes of various factors from 1990 to 2024: (a) per capita GDP change, (b) resident population change, (c) built-up area change, (d) road area change."
  
  I have carefully read the manuscript entitled « Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning » by Jie Meng et al. and tested associated datasets available in Zenodo. It presents original information on sandy beach locations throughout China, taking the form of a shape file produced using ensemble learning and multi-source data, from which various beach spatial characteristics (e.g., number, width, area) and risk analysis of coastal squeeze can be assessed and analyzed. It is not the first time sandy beaches were mapped in China using satellite imagery - with already existing datasets mentioned and assessed in comparison by the authors - yet it is the first time ensemble learning is used, which according to the authors improved detection accuracy. Besides, when previous studies used Sentinel-2 data only, Sentinel 1 and Google Earth imagery were also included in this study. I have found the manuscript well written and illustrated, although additional careful proofreading by the authors could have avoided several mistakes such as identical section and figure titles, repetitions and typos. Figure captions in particular must be improved. I attach an annotated version of the manuscript, listing a number of technical corrections and specific comments, that the authors should take into account while preparing their revision.
  
  Response:
  Thank you for your valuable feedback; it has been extremely helpful. I will now respond to each of your comments:
  1.Firstly, regarding the test set, we constructed a sample set through field surveys and visual interpretation using Sentinel-2 and Google images. Once the sample set was obtained, we randomly divided it into a 7:3 ratio for training and testing. Of course, it is essential to ensure that the proportion of sandy and non-sandy beach samples in both the training and test sets, as well as the sample distribution across provinces, remain consistent.
  2.Since our product is first generated using ensemble learning followed by post-processing and visual interpretation corrections, the workload is minimal, and it better compensates for the sandy beach areas, avoiding the issue of underestimating the sandy beach area due to missed identification during visual interpretation. In contrast, Dataset 1 was labeled without any prior knowledge, resulting in many missing sandy beach areas. You are correct in pointing out that relying solely on Dataset 1 seems unreasonable, so we replaced it with a more reliable source. The current dataset is derived from the OpenStreetMap (OSM) database, and we have made updates in Figure 3, Table 1, and Line 138 of the manuscript. The revised text now reads: “In this study, we used three datasets to evaluate our identified sandy beach dataset (Table 1): (1) The China sandy beach dataset, data directly obtained from the OpenStreetMap (OSM) database; (2) The 2022 China 10m sandy beach dataset identified by Ni et al. (Ni et al, 2024) using a support vector machine method based on Sentinel-2 imagery; (3) The 2020 China coastal land use dataset at 10m resolution, identified by Miao et al. (Miao et al, 2022) using an object-oriented classification method based on Sentinel-2 imagery.” The updates can be seen in the revised figures and tables. Additionally, we have modified the comparison between subsequent datasets (L262, L275). The updated text reads: “To further assess the accuracy and reliability of the dataset, this study compared three published datasets in selected areas of Fujian, Shandong, and Guangdong (Fig. 8). Dataset 1, directly obtained from the OpenStreetMap (OSM) database, is highly subjective and tends to misclassify non-beach areas as sandy beaches, while also missing some actual sandy beach areas. Dataset 2, constructed using a support vector machine on Sentinel-2 imagery, shows high consistency with our dataset but still misses some sandy beach areas. Dataset 3, created using an object-oriented approach, demonstrates high accuracy for other land cover types but faces significant misclassification issues with bare land and urban areas. The results show that our dataset provides higher accuracy in sandy beach classification, significantly reducing misclassification.” “According to the comparison results, our dataset shows significant advantages over published datasets 1, 2, and 3 in several key metrics, particularly in terms of sandy beach area, perimeter, width, and number (Fig. 9). Overall, our dataset provides a supplementary advantage in sandy beach area coverage, with larger areas in all regions: Fujian (54.57 km²), Guangdong (78.88 km²), and Taiwan (46.60 km²), significantly surpassing published datasets 1 (30.35, 20.59, and 19.07 km²) and 2 (29.17, 58.35, and 25.51 km²). Regarding perimeter, our dataset closely matches actual sandy beach boundaries: Fujian (1435.89 km), Guangdong (2849.39 km), and Taiwan (1324.98 km), compared to published datasets 1 (581.95, 826.40, and 509.22 km) and 2 (756.92, 1856.62, and 906.63 km). In terms of width, our dataset also outperforms published datasets 1 (52.21, 20.94, and 48.42 m) and 2 (45.18, 32.58, and 27.56 m), with values of 54.91 m in Fujian, 38.92 m in Guangdong, and 57.17 m in Taiwan. The number of identified sandy beaches in our dataset is higher than in published datasets 1 and 2, further highlighting the reduced misclassification and noise in our results (Fig. 9). Moreover, published dataset 3 has significantly higher area, perimeter, width, and number values than the other datasets and actual values, leading to many non-sandy beaches being incorrectly identified as sandy beaches.” These changes are reflected in Figures 8 and 9.
  3.Additionally, we realized that it was unreasonable to compare different products using the test set we constructed. Therefore, this part has been removed from the manuscript and replaced with a comparison between the current mainstream models and our ensemble learning model. In this updated version, we replaced the base learner for the ensemble learning model with tree-based models, which are less sensitive to parameters and better suited for handling large datasets and wider areas. The revised text is as follows (L183): " Stacking is a powerful ensemble learning method that uses predictions from multiple base learners as inputs to a meta-learner for final prediction. It combines the strengths of different models to overcome individual limitations, enhancing accuracy, stability, and generalization (Chen et al., 2024). In this study, Random Forest (RF), Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM) were selected as base learners. Their output classification probabilities were used as input features for the meta-learner, which adopted Logistic Regression (LR) to integrate probabilities, calculate distances to target classes, and produce final results (Table 3). For performance comparison, Support Vector Machine (SVM), K-Nearest Neighbor (KNN), RF, Classification and Regression Tree (CART), and Convolutional Neural Network (CNN) models were also employed as benchmark models." Furthermore, the parameters for the base learners were also updated, and the revised parameters are reflected in Table 3. Table 6 has been updated to show a comparison between different models, where our ensemble learning model yields the best results (2016-2024 average). The accuracy, precision, recall, F1-score, and AUC are 0.9335, 0.9160, 0.9014, 0.9084, and 0.9802, respectively, outperforming the combined results of SVM, KNN, RF, CART, and CNN models. In addition, we have added the corresponding discussion in L257. The updated text reads: "We evaluated the performance of our model and several other machine learning algorithms (SVM, KNN, RF, CART, and CNN) using accuracy, precision, recall, F1-score, and AUC on the validation set (Table 6). Our model outperformed all others, achieving the highest accuracy (0.9335), precision (0.9160), and AUC (0.9802). It showed significantly better classification performance, especially in reducing misclassification. In comparison, the SVM model had good precision but lower recall and accuracy, while KNN performed the worst across all metrics." These results are reflected in Table 6.
  4.Previously, there were indeed only 14,694 points used for testing between 2016 and 2023, covering 5 land cover categories. Each year, the dataset was redefined with less than 400 sandy beach points per year, which was clearly insufficient. Now, we have updated the study to cover a 9-year period from 2016 to 2024, as reflected throughout the manuscript. The time span mentioned in L15 and other sections referring to the years has been updated from 2016-2023 to 2016-2024. Additionally, the number of Sentinel-1 and Sentinel-2 images used has been revised accordingly, with the updated numbers being 23,859 and 80,759, respectively, as reflected in L108, L111, and also shown in Figure 2. Furthermore, we have updated the land use data to cover the period from 1990 to 2024 (L212), and we revised the method for establishing buffer zones. The updated approach creates a buffer zone around sandy beach data and then performs the clipping, which results in more accurate sandy beach extraction than clipping based on land use data. The revised text now reads: "Therefore, this study established a 100-meter buffer zone around the Chinese land use dataset and conducted an overlay analysis using the obtained sandy beach data. The analysis generated annual human infrastructure squeeze risk areas to evaluate the squeeze effects of infrastructure on sandy beaches from 1990 to 2024."
  Regarding the issue of sample size, we recognized that the previous sample set was insufficient. Therefore, we changed the problem to a binary classification task, merging the other four land cover classes into one, with an emphasis on labeling sandy beach data. A total of 341,057 points were labeled from 2016 to 2024, with 249,098 points for the training set and 49,1959 points for the testing set. The breakdown for sandy beach and non-sandy beach samples for both the training and testing sets from 2016 to 2024 is as follows:
  2016-2024 training set (sandy beach): 6776, 7308, 7841, 8231, 8722, 9549, 10632, 11716, 14130
  
  2016-2024 testing set (non-sandy beach): 19290, 19431, 18822, 18568, 18510, 17981, 18138, 18136, 16470
  
  2016-2024 testing set (sandy beach): 2901, 3128, 3358, 3525, 3735, 4088, 4550, 5017, 6054
  
  2016-2024 testing set (non-sandy beach): 4175, 4200, 4797, 4756, 4738, 5352, 5399, 5398, 7056
  
  This has been reflected in Figure 4. Due to the changes in both the years and the sample points, some data in the results section of the article will be updated accordingly. For example, in L236, the sandy beach area, perimeter, width, and number have changed. The revised text now reads: "From a provincial perspective, Guangdong has the most sandy beaches, with 1,096 sandy beaches, and also ranks first in both total length (2,849.39 km) and total area (78.88 km²). In terms of sandy beach width, Hebei has the widest sandy beaches, with an average width of 68.77 m. Other regions, such as Fujian (sandy beach area 54.57 km², total length 1,435.89 km, width 54.91 m) and Hainan (sandy beach area 51.65 km², total length 1,977.96 km, width 41.02 m), also show significant sandy beach resources."
  
  Concerning quality metrics, why were they chosen, and are they really complementary to each other? Tables 5 and 6 suggest similar interpretations can be made for all metrics as results do not differ significantly. Additional information should be provided to guide readers in how to interpret these quality metrics and the results obtained.
  
  Response:
  Thank you for the reviewer’s valuable feedback. Indeed, the five metrics (PA, UA, F1-score, OA, and Kappa coefficient) previously selected in Tables 5 and 6 did not show significant differences in results, and thus were not effective in distinguishing the performance of different models. However, these metrics are commonly used in multi-class classification tasks. To better evaluate the performance of our binary classification model, I have converted the problem into a binary classification task. As a result, we replaced these metrics with more suitable evaluation metrics for binary classification: Accuracy, Precision, Recall, Sandy Beach F1-score, and AUC (Area Under the Curve). These metrics are more effective in reflecting the model’s performance in binary classification tasks. In the revised Tables 5 and 6, the differences between the newly selected five metrics can be observed (see L227 and L270). Due to space limitations, we have only referenced the relevant literature in the manuscript without providing a detailed introduction to these evaluation metrics, as they are mainstream and commonly used in the field.
  1.Accuracy is the most basic evaluation metric, measuring the proportion of correct predictions among all predictions. It is calculated by dividing the number of correct predictions by the total number of samples. Although accuracy is straightforward and intuitive, it may not fully reflect model performance, especially in cases of imbalanced data. In tasks like sandy beach extraction, where the ratio between positive (sandy beach) and negative (non-sandy beach) samples may vary greatly, accuracy may not be an adequate measure of performance.
  2.Precision measures the proportion of actual positive samples among those predicted as positive by the model. It reflects the confidence of the model when predicting positive class instances. Precision is crucial in sandy beach extraction tasks, as it helps assess how well the model avoids misclassifying non-beach areas as beaches. A low precision indicates that the model is making many false positives, which could lead to errors in beach management and protection efforts.
  3.Recall is the proportion of correctly predicted positive samples among all actual positive samples. A higher recall indicates that the model is better at identifying positive samples (i.e., sandy beach regions). However, this could lead to an increase in false positives, as non-beach regions might be misclassified as beaches. Therefore, recall needs to be evaluated together with precision to assess the model's performance comprehensively.
  4.F1-score is the harmonic mean of precision and recall, providing a balance between the two. F1-score is especially useful for tasks with imbalanced class distributions. In sandy beach extraction applications, F1-score helps us balance precision and recall, ensuring the model minimizes misclassification while identifying as many sandy beach areas as possible.
  5.AUC (Area Under the Curve) is a critical metric for evaluating a binary classification model’s discriminative ability. It is derived from the ROC (Receiver Operating Characteristic) curve and measures the model’s ability to distinguish between positive and negative samples. A higher AUC value indicates that the model is better at distinguishing between the classes. AUC is particularly robust in imbalanced datasets and provides an overall performance evaluation, especially when dealing with multiple threshold values.
  With these updated evaluation metrics, we are now able to assess the performance of the binary classification model more accurately, especially in the context of sandy beach extraction. These metrics not only assist in more reliable identification of sandy beach regions but also help reduce misclassifications, providing more trustworthy results for the model.
  
  Likewise, the authors evaluate other (independent) datasets (some produced by the authors themselves, eg. Dataset 1) which they name reference datasets. I think « reference » is misused in this context as it generally implies that the data were used for validation (truth data), which is not the case here.
  
  Response:
  Thank you for your valuable feedback. The dataset 1 is no longer the one I created; instead, we have replaced it with a sandy beach dataset from the OpenStreetMap (OSM) database. Additionally, we believe that the term "reference dataset" is not appropriate, so we have changed it to "published dataset." It is important to note that these three datasets are only used for comparison with our own dataset, not for accuracy validation.
  
  Additional information could be provided and discussed in regard to the liability and capacities of the machine learning method used. What is the minimum beach size that can be detected? Looking at the dataset, it seems small pocket beaches can remain undetected. In contrast, elongated features (some artificial) may be erroneously detected as sandy beaches. What is the impact of tide range (potentially variable across China) and having images obtained at different tide levels on data consistency and how can this be improved? It is said using annual averages (composite images) helps mitigate this issues, but it is not clear how, particularly as Figure 2 shows an heterogeneous spatial distribution of satellite images and certainly, with regular revisits, this means images are obtained at different tide levels throughout the country. From this could arise systematic biases in the spatial characteristics deduced from the dataset (eg., beach width and area). Was this, and how, mitigated for this study?
  
  Response:
  Thank you for your valuable feedback. Indeed, the impact of tides on beach extraction is a complex issue, especially when dealing with beach data on a national scale. The differences in tidal times across various regions can cause dynamic changes in beach areas, which in turn affect the accuracy of the extraction. To address this, we have synthesized the results of beach extractions over multiple years to minimize the influence of tidal fluctuations and construct a more representative dataset reflecting the actual beach distribution. Firstly, regarding the minimum beach size, we have set the threshold at 1,000 square meters. This threshold helps avoid retaining small areas below this size during post-processing, as these small areas are often misidentified as beaches due to noise or fluctuations. This setting ensures the reliability of the beach extraction and improves the overall quality of the dataset. However, the effect of tides cannot be ignored, particularly in areas where there is a clear transition between beaches and water bodies. Each year, due to tidal changes, certain beach areas may be submerged by water, leading to misclassification as water bodies in the imagery. Although beaches themselves are relatively stable as natural features and do not exhibit significant expansion or contraction, tidal influence causes yearly variations in the beach extraction results, especially between high and low tides. Additionally, due to varying tidal times in different regions, satellite images are obtained at different times, further complicating data consistency. For example, some regions may show beaches covered by water in satellite images obtained at one time, while others may show exposed beaches during the same period.
  To better address this issue, we decided to merge the beach extraction results over multiple years (from 2016 to 2024) to create a more representative beach dataset. Although annual composite images may be affected by tidal fluctuations, combining these results helps capture the widest extent of beach coverage each year. This method not only reduces misclassification due to tidal changes but also provides a more stable dataset that better reflects beach distribution, particularly one that is closer to the low tide beach distribution. Since low tide exposes the largest area of the beach, combining multiple years of data can effectively capture the beach area during low tide, avoiding omissions and misclassification that may occur when relying on data from only one year. Furthermore, using multiple years of data helps mitigate the systematic bias caused by tidal fluctuations in any single year. For example, if high tide in a given year causes part of the beach to be submerged and misclassified as water, using only data from that year could result in an underestimation of the beach area. However, by combining data from multiple years, other years’ images might show these areas as beaches, thus helping to reconstruct the actual beach extent. This approach allows us to create a more accurate beach distribution map and reduces errors in beach extraction. It is important to note that while merging data from multiple years can minimize the impact of tidal fluctuations on beach extraction, this method is not perfect. Due to differences in tidal times between regions, composite images may still show some local discrepancies. In regions where the transition between high and low tide occurs, beach boundaries may still be inaccurately represented due to the timing of the satellite images. Therefore, we adopted the approach of synthesizing the maximum beach coverage from each year’s extraction results to generate a representative beach distribution map. This method helps minimize errors caused by tidal fluctuations and provides the best possible estimate of beach distribution. Through this approach, we aim to obtain a more stable and reliable beach distribution map, especially during low tide, ensuring that the data better reflects the actual beach situation. This not only helps in accurately estimating beach area, perimeter, and width, but also provides reliable foundational data for beach management and protection. For example, many beach conservation efforts require monitoring and evaluation during low tide, so obtaining accurate low tide beach data is crucial for implementing effective protection measures. In conclusion, while tidal fluctuations do have an impact on beach extraction, combining multi-year beach extraction results allows us to produce a more stable and accurate dataset. This approach effectively alleviates errors caused by tidal changes and provides a more accurate basis for long-term monitoring and protection of beaches.
  
  The temporal change in coastal squeeze is assessed over the period 1990-2024, yet beach distribution data outside 2016-2024 are not available. Thus, which data did you use for sandy beach spatial coverage? Currently, this is not explained in the text.
  
  Response:
  We sincerely appreciate the reviewer’s valuable feedback on our study. Regarding the accuracy of beach extraction, we fully agree that beaches are relatively stable landforms that typically do not undergo drastic changes. However, due to the influence of tides, especially in areas with significant tidal fluctuations, the exposure of beaches can change. These tidal variations often result in beach areas being incorrectly identified as water bodies in the yearly extraction results, leading to an underestimation of the actual beach area. The impact of tides not only introduces errors in the single-year beach extraction results but also exacerbates the errors due to the timing differences of satellite image acquisitions, further complicating the tidal effects in different regions.
  To address this issue, we employed a multi-year synthesis approach, combining the yearly beach extraction results to derive the maximum beach extent. This method effectively reduces errors caused by tidal fluctuations, as the beach area extracted at different tidal levels each year varies. The combined result provides a more comprehensive distribution of the beaches, especially reflecting the maximum exposure during low tide. We believe that by synthesizing multi-year beach data, we can more accurately represent the actual distribution of beaches, avoiding errors caused by tidal fluctuations in single-year data, particularly for beach areas that may be covered by water during high tide.
  Furthermore, to further improve the accuracy of the results, we also used land use data from 1990 to 2024 for a compression analysis. By creating a 100-meter buffer zone, we were able to identify the impact of land use on beach areas, particularly how the expansion of human infrastructure affects beach distribution.
  
  Citation: https://doi.org/10.5194/essd-2025-264-AC1
CC1:
'Comment on:Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning', Xuecheng Zhou, 06 Oct 2025
Comment on:Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning
This study has made valuable contributions to coastal resource research by addressing long-term challenges in beach identification and human impact assessment. Its most obvious advantage lies in the innovative integration of multi-source remote sensing data and integrated learning, which effectively overcomes the limitations of traditional single data or single model methods. By combining Sentinel-1/2 images, terrain data, and nighttime light data, and extracting four-dimensional features, this study constructed a stacked ensemble model that integrates RF, SVM, CART, and GBDT. The consistent high accuracy from 2016 to 2023, as well as comparisons with three reference datasets, confirm the robustness of this method, particularly in reducing misclassification of bare land and urban areas, which is a common issue in existing research. The resulting 10 meter resolution beach dataset and regional pattern analysis have filled the gap in long-term beach monitoring nationwide, providing a reliable data foundation for coastal ecological management.
This study also provides practical insights into the risk of human infrastructure crowding. By establishing a 100 meter buffer zone and analyzing impermeable surface data from 1990 to 2023, the increase in risk areas and regional differences were quantified. Linking risk trends with economic factors further reveals the coupling relationship between coastal urbanization and beach degradation, providing targeted guidance for policy-making, such as prioritizing protection in high-risk areas such as Shandong and Guangdong.
However, there are several aspects that deserve improvement. Firstly, tidal disturbances have not been fully resolved. Although years of data can alleviate tidal effects, the lack of tidal phase matching may introduce spatial inconsistency in beach extraction. Future work can integrate tidal prediction models or on-site tidal data to select time series images with consistent tidal conditions, improving spatiotemporal accuracy. Secondly, the assessment of infrastructure squeeze is relatively simple: relying solely on impermeable surface expansion and buffer zone analysis cannot capture dynamic and detailed impacts . Higher resolution data and multi criteria models will improve the granularity of risk attribution.
In addition, it can enhance the interpretability of integrated models. This study did not analyze the importance of features or the collaborative or redundant relationships between underlying models. Adding SHAP values or permutation importance analysis will clarify the contribution of each feature and optimize the model structure, reducing computational costs without sacrificing accuracy.
Overall, this study establishes a solid benchmark for coastal beach research, balancing methodological rigor and practical value. Addressing the aforementioned limitations will further enhance its scientific impact and practicality for sustainable coastal area management.
References
Tao Zhang,Yiqun Zhang & Xiaofei Leng.(2025).Trends in gastric cancer burden in the Western Pacific region from 1990 to 2021 and projections to 2040.Frontiers in Oncology,15,1506479-1506479. https://doi.org/10.3389/FONC.2025.1506479.

Jordan Gacutan,Ibon Galparsoro & Arantza Murillas-Maza.(2019).Towards an understanding of the spatial relationships between natural capital and maritime activities: A Bayesian Belief Network approach.Ecosystem Services,40,101034-101034. https://doi.org/10.1016/j.ecoser.2019.101034.
Citation: https://doi.org/10.5194/essd-2025-264-CC1
- AC3:
  'Reply on CC1', Jie Meng, 30 Oct 2025
  Dear Contributing Commenter:
  Thank you very much for your thoughtful and detailed review. Your suggestions have provided us with important and constructive insights, which have significantly improved the manuscript. We have carefully considered all of your comments and have made substantial revisions to the manuscript based on your feedback. We have done our best to enhance the manuscript and hope that the revised version will meet your approval. A point-by-point response to the outstanding comments is attached to this manuscript. The major revisions are summarized as follows:
  Response to Comments by Contributing Commenter 1:
  
  This study has made valuable contributions to coastal resource research by addressing long-term challenges in beach identification and human impact assessment. Its most obvious advantage lies in the innovative integration of multi-source remote sensing data and integrated learning, which effectively overcomes the limitations of traditional single data or single model methods. By combining Sentinel-1/2 images, terrain data, and nighttime light data, and extracting four-dimensional features, this study constructed a stacked ensemble model that integrates RF, SVM, CART, and GBDT. The consistent high accuracy from 2016 to 2023, as well as comparisons with three reference datasets, confirm the robustness of this method, particularly in reducing misclassification of bare land and urban areas, which is a common issue in existing research. The resulting 10 meter resolution beach dataset and regional pattern analysis have filled the gap in long-term beach monitoring nationwide, providing a reliable data foundation for coastal ecological management. This study also provides practical insights into the risk of human infrastructure crowding. By establishing a 100 meter buffer zone and analyzing impermeable surface data from 1990 to 2023, the increase in risk areas and regional differences were quantified. Linking risk trends with economic factors further reveals the coupling relationship between coastal urbanization and beach degradation, providing targeted guidance for policy-making, such as prioritizing protection in high-risk areas such as Shandong and Guangdong. However, there are several aspects that deserve improvement. Firstly, tidal disturbances have not been fully resolved. Although years of data can alleviate tidal effects, the lack of tidal phase matching may introduce spatial inconsistency in beach extraction. Future work can integrate tidal prediction models or on-site tidal data to select time series images with consistent tidal conditions, improving spatiotemporal accuracy. Secondly, the assessment of infrastructure squeeze is relatively simple: relying solely on impermeable surface expansion and buffer zone analysis cannot capture dynamic and detailed impacts . Higher resolution data and multi criteria models will improve the granularity of risk attribution. In addition, it can enhance the interpretability of integrated models. This study did not analyze the importance of features or the collaborative or redundant relationships between underlying models. Adding SHAP values or permutation importance analysis will clarify the contribution of each feature and optimize the model structure, reducing computational costs without sacrificing accuracy. Overall, this study establishes a solid benchmark for coastal beach research, balancing methodological rigor and practical value. Addressing the aforementioned limitations will further enhance its scientific impact and practicality for sustainable coastal area management.
  
  Response:
  Thank you for the valuable comments. Regarding the model interpretability issue, we have incorporated SHAP value analysis into the paper. Figure 7 presents the importance maps and average importance maps of features from 2016 to 2024. Specifically, we updated the content in L255, adding the analysis of feature importance. The revised text reads: "We evaluated the variable importance of the model for the years 2016-2024 (Fig. 7). Based on the average importance ranking from 2016 to 2024, the top five features were Elevation (0.1270) > B3 (0.1242) > B2 (0.1049) > NDVI (0.0783) > VV (0.0600). When classified by feature category, the highest average importance was found in spectral features (0.3403), followed by index features (0.2458), topographic features (0.1847), texture features (0.1317), and polarization features (0.0975). The most important feature in each category was as follows: Elevation for topographic features, B3 for spectral features, NDVI for index features, VV for polarization features, and VAR for texture features."
  By introducing SHAP value analysis, we are able to gain a clearer understanding of the contribution of different features to the model's results, particularly in terms of model optimization and feature selection. The use of SHAP values enhances the transparency of each feature's contribution, helping us identify key features in the model, optimize the model structure, and reduce computational costs. Furthermore, this analysis helps us better understand the collaboration or redundancy between underlying models, providing more room for optimization in future research.
  Although multi-year data synthesis helps mitigate the tidal effects, the inconsistency in tidal phases remains a significant issue affecting beach extraction. The variability of tidal changes in different time periods and geographic areas often leads to spatial inconsistencies in multi-year data synthesis, resulting in inaccurate identification of beach areas. While the multi-year data synthesis method currently used reduces the impact of tidal fluctuations, the lack of precise tidal phase matching still introduces spatial errors. Therefore, future work can further integrate tidal prediction models and field tidal data to select time series images with consistent tidal conditions. This approach would not only improve spatiotemporal accuracy but also effectively reduce spatial inconsistencies introduced by tidal variations, enhancing the accuracy of beach extraction.
  In addition to tidal effects, the evaluation of infrastructure encroachment is another area for improvement in this study. Although we quantified the impact of infrastructure expansion on beach areas through impervious surface expansion and buffer zone analysis, this method is overly simplified and cannot fully capture the dynamic effects of infrastructure development. Infrastructure development is a complex process involving not only changes in impervious surface area but also land use types, urban expansion rates, and socioeconomic factors, all of which can have varying temporal and regional effects on beaches. To better capture these impacts, future research could consider using higher-resolution remote sensing data, along with time series analysis and multi-criteria models, to enhance the granularity of risk assessment and more accurately quantify the impact of infrastructure expansion on beaches in different regions and time periods.
  Moreover, combining various data sources (such as nighttime light data and land use data) for multi-scale, multi-temporal integrated analysis will help reveal the complex relationship between human activities and beach degradation. This method will not only identify the beach areas most affected by human activities but also provide data support for policy-making, guiding the protection and sustainable development of coastal regions. For example, stricter protection measures could be implemented in high-risk areas (such as Shandong and Guangdong) to limit excessive development and infrastructure construction, thus mitigating beach degradation and ecosystem damage.
  While several improvements have been proposed, there are still some technical and data challenges to overcome. For example, accurate tidal phase matching and the acquisition of high-resolution data may face certain technical limitations. Additionally, future research may need to consider integrating other types of data (e.g., field survey data) and more advanced machine learning methods (such as deep learning and reinforcement learning) to further enhance the model's predictive capability and adaptability.
  In summary, this study provides a solid foundation for coastal beach research and ecological conservation, while also pointing out directions for future research. By addressing issues such as tidal effects and infrastructure encroachment assessment, we believe that we can significantly improve the accuracy of beach extraction and human impact assessment, thereby providing more reliable data support and decision-making for coastal ecosystem protection and sustainable development.
  
  Citation: https://doi.org/10.5194/essd-2025-264-AC3
RC2:
'Comment on essd-2025-264', Anonymous Referee #2, 06 Oct 2025
This study establishes a coastal sandy beach recognition framework for China based on multi-source remote sensing data and a stacking ensemble learning approach, producing a nationwide dataset with 10 m resolution from 2016 to 2023. The research demonstrates significant innovation and practical value in terms of data, methodology, results, and applications: the fusion of multi-source data effectively enhances classification reliability, the ensemble learning framework substantially improves accuracy, and high-resolution national-scale mapping has been achieved, revealing the spatiotemporal evolution of human infrastructure encroachment risks from 1990 to 2023. This would provide robust scientific support for coastal ecological conservation and sustainable management. The overall structure of the manuscript is logical, with adequate support from figures and data, and it meets the requirements for journal publication. In general, the work holds considerable academic value and practical significance, and I recommend acceptance after minor revisions to further improve rigor and readability.
Specific comments:
L161: The window size used for gray-level co-occurrence matrix (GLCM) texture feature extraction (e.g., 3×3, 5×5) is not reported. Please specify and explain the choice, as window size has a substantial impact on metrics such as Entropy and ASM.

L304: The overall English fluency is good; however, some minor errors in singular/plural forms and article usage remain. For example, in Section 5.2, the sentence “Stacking not only ensures competitive accuracy but also offers strong applicability…”—the term applicability could be replaced with generalizability. A final proofreading or use of a grammar-checking tool is recommended.

In Figures 1, 2, 5, 6, 10, and 13, the blue regions represent the ocean. However, adjacent countries bordering China are not labeled, which may cause confusion. It is recommended to label neighboring countries to avoid potential misinterpretation.

In Figure 7, consider including Sentinel-2 true-color imagery for comparison with the classification results, thereby enhancing the intuitiveness and persuasiveness of the figure.

Several figure captions contain redundancies or unclear phrasing. Please review and refine them to ensure clarity and conciseness.

The minimum detectable sandy beach size is not explicitly addressed. Please clarify this aspect in the discussion.

The conclusion primarily emphasizes the value of the study, but provides limited discussion of methodological limitations and future perspectives. Please expand this section accordingly.

Inconsistent use of Chinese and English punctuation marks is observed. Please standardize formatting throughout the manuscript.

The conclusion would benefit from elaborating on the transferability of the method, such as its applicability to other countries or to datasets with different spatial resolutions.

Some English references do not follow the journal’s formatting guidelines regarding author abbreviations (e.g., use of “and” vs. “&”). Please check and revise.

For Table 3, please clarify whether the model parameters were optimized through hyperparameter tuning or adopted as default values.

L35–40: The statement that “traditional field surveys are inefficient” could be strengthened by citing 1–2 recent field-based studies (within the past three years), to demonstrate coverage of the latest research progress.
Citation: https://doi.org/10.5194/essd-2025-264-RC2
- AC2:
  'Reply on RC2', Jie Meng, 30 Oct 2025
  Dear Reviewers:
  Thank you very much for your thoughtful and detailed review. Your suggestions have provided us with important and constructive insights, which have significantly improved the manuscript. We have carefully considered all of your comments and have made substantial revisions to the manuscript based on your feedback. We have done our best to enhance the manuscript and hope that the revised version will meet your approval. A point-by-point response to the outstanding comments is attached to this manuscript. The major revisions are summarized as follows:
  Response to Comments by Reviewer 1:
  
  L161: The window size used for gray-level co-occurrence matrix (GLCM) texture feature extraction (e.g., 3×3, 5×5) is not reported. Please specify and explain the choice, as window size has a substantial impact on metrics such as Entropy and ASM.
  
  Response:
  We Thank you for the valuable comments. Regarding the window size for extracting texture features using the Gray Level Co-occurrence Matrix (GLCM), we have chosen a 3×3 window size, which has been updated in Table 2 (see L179). The choice of a 3×3 window is based on its widespread use in the literature as well as the practical needs of our study. Specifically, a 3×3 window provides a good balance in texture feature extraction, as it captures local detail without introducing excessive noise. Additionally, using a 3×3 window effectively extracts texture features for most land cover types, including beaches, especially when image resolution and spatial scale are moderate.
  We have referred to relevant literature, where many studies also used a 3×3 window size for texture analysis with successful results. In these studies, the 3×3 window was shown to effectively reflect the spatial structure and texture variation of land cover types, especially when dealing with different land cover types such as beaches, urban areas, and forests. In contrast, larger windows (e.g., 5×5 or larger) may capture more spatial information but tend to over-smooth the image, resulting in the loss of local details. Smaller windows, on the other hand, may not capture sufficient texture features. Therefore, a 3×3 window provides an ideal compromise, preserving enough local information while avoiding computational overhead and the loss of detail.
  Specifically, the 3×3 window has significant effects on metrics such as Entropy and Angular Second Moment (ASM), which help us better identify and distinguish different land cover types. In practice, using a smaller window helps reduce noise and interference in the image, enhancing the reliability and usability of the texture features. Moreover, adopting the 3×3 window size ensures the consistency and reproducibility of the algorithm, which is crucial for subsequent experimental validation and result comparison.
  In conclusion, the choice of a 3×3 window size is based on its extensive application and excellent performance in texture feature extraction. We believe this choice is reasonable for our study and will ensure the accuracy and effectiveness of the extracted results.
  
  L304: The overall English fluency is good; however, some minor errors in singular/plural forms and article usage remain. For example, in Section 5.2, the sentence “Stacking not only ensures competitive accuracy but also offers strong applicability…”—the term applicabilitycould be replaced with generalizability. A final proofreading or use of a grammar-checking tool is recommended.
  
  Response:
  Thank you for your valuable feedback. We have reviewed the manuscript and corrected the small errors related to singular/plural forms and article usage. Specifically, in Section 5.2, the term "applicability" has been replaced with "generalization" as suggested. We appreciate your recommendation and have made the necessary adjustments to ensure the accuracy and fluency of the language.
  
  In Figures 1, 2, 5, 6, 10, and 13, the blue regions represent the ocean. However, adjacent countries bordering China are not labeled, which may cause confusion. It is recommended to label neighboring countries to avoid potential misinterpretation.
  
  Response:
  Thank you for your valuable suggestion. We appreciate your careful attention to detail. Based on your feedback, we have updated Figures 1, 2, 5, 6, 10, and 13 by labeling the neighboring countries bordering China. This revision aims to clarify the geographical context and avoid any potential confusion that might arise from the blue areas representing the ocean. We believe this addition will significantly improve the clarity of the maps and ensure that the readers can easily distinguish between China and its neighboring countries. Thank you again for your thoughtful recommendation, and we trust that this update will enhance the overall quality of the manuscript.
  
  In Figure 7, consider including Sentinel-2 true-color imagery for comparison with the classification results, thereby enhancing the intuitiveness and persuasiveness of the figure.
  
  Response:
  Thank you for your valuable suggestion. Based on your feedback, I have added the true-color Sentinel-2 image to Figure 7. This addition allows for a comparison with the classification results, enhancing the visual clarity and persuasiveness of the figure, making it easier for readers to understand the classification outcomes. Thank you again for your thoughtful recommendation; the necessary adjustments have been made.
  
  Several figure captions contain redundancies or unclear phrasing. Please review and refine them to ensure clarity and conciseness.
  
  Response:
  Thank you for pointing that out. I have reviewed and revised the titles of the figures and tables to ensure they are clear and concise, removing any redundancy and improving clarity. The necessary modifications have been made to enhance the overall readability and precision of the manuscript. Thank you again for your helpful suggestion. For example, we have changed "Figure 2: Spatial distribution of Sentinel-1 and Sentinel-2 images used in this study: (a) Image Count of Sentinel-1, (b) Image Count of Sentinel-2" to "Figure 2: Spatial distribution of Sentinel-1 and Sentinel-2 images used in this study from 2016 to 2024: (a) Image Count of Sentinel-1, (b) Image Count of Sentinel-2."
  
  The minimum detectable sandy beach size is not explicitly addressed. Please clarify this aspect in the discussion.
  
  Response:
  Thank you for your valuable feedback. Indeed, the impact of tides on beach extraction is a complex issue, especially when dealing with beach data on a national scale. The differences in tidal times across various regions can cause dynamic changes in beach areas, which in turn affect the accuracy of the extraction. To address this, we have synthesized the results of beach extractions over multiple years to minimize the influence of tidal fluctuations and construct a more representative dataset reflecting the actual beach distribution. Firstly, regarding the minimum beach size, we have set the threshold at 1,000 square meters. This threshold helps avoid retaining small areas below this size during post-processing, as these small areas are often misidentified as beaches due to noise or fluctuations. This setting ensures the reliability of the beach extraction and improves the overall quality of the dataset.
  
  The conclusion primarily emphasizes the value of the study, but provides limited discussion of methodological limitations and future perspectives. Please expand this section accordingly.
  
  Response:
  Thank you for the valuable comments on our research. Indeed, the conclusion section currently focuses mainly on summarizing the value and contributions of the study, with limited discussion on the methodological limitations and future perspectives. We acknowledge this and plan to give more attention to these aspects in future versions. Due to space constraints, we were unable to elaborate on the limitations of the methodology and future outlook in the current paper, but we aim to address these in subsequent versions.
  First, regarding the methodological limitations, although our approach has shown good performance in beach extraction tasks, there are still some limitations. For instance, the exposure of beach areas changes due to tidal fluctuations, which may lead to beach regions being misclassified as water bodies at certain times, affecting the accuracy of extraction results. While we have reduced tidal influences by merging multi-year data, tidal differences still contribute to some errors. Future research could consider incorporating more detailed tidal correction models or combining higher-resolution remote sensing imagery to improve the accuracy of beach extraction. Secondly, the generalizability of the methodology is another important consideration. Currently, our method has been validated mainly in specific regions in China, particularly along coastal beach areas. However, different regions, with varying topographies, climates, and tidal conditions, may result in different beach extraction outcomes. Therefore, future work should aim to extend our approach to other regions and types of land cover, especially those with significant tidal variations, such as tropical and subtropical areas. Additionally, given the complexity of beach extraction, we could further optimize the model to better adapt to other land cover types, such as dunes and coastlines. As for future prospects, we believe that with ongoing advancements in remote sensing technology, particularly the increasing resolution of satellite imagery, there is great potential to further improve the accuracy and reliability of beach extraction. Future research could incorporate multi-source data, such as combining LiDAR data with remote sensing imagery to gain richer land cover information. Additionally, we could explore the application of deep learning methods in beach extraction to automate feature extraction and classification, further improving model precision and efficiency. Finally, we hope to apply our method to broader real-world scenarios in the future, such as beach environmental monitoring, management, and conservation, to enable effective monitoring and management of global beach ecosystems. These will be key directions and goals for our future research. Once again, we appreciate the reviewer’s valuable suggestions, and we will refine our research based on your feedback. We aim to continuously improve the application value and practical effectiveness of the methodology in future work.
  
  Inconsistent use of Chinese and English punctuation marks is observed. Please standardize formatting throughout the manuscript.
  
  Response:
  Thank you for your valuable comment. We have reviewed the manuscript and made the necessary revisions to ensure consistent use of punctuation between Chinese and English throughout the document.
  
  The conclusion would benefit from elaborating on the transferability of the method, such as its applicability to other countries or to datasets with different spatial resolutions.
  
  Response:
  Thank you for your valuable feedback. The point you raised regarding the transferability of the method is indeed an important aspect. In this study, we primarily focused on beach extraction and analysis in China, utilizing existing remote sensing data and algorithms for effective application. However, the transferability of the method to other countries or datasets with different spatial resolutions is an area that requires further discussion. First, our methodology relies primarily on Sentinel-1 and Sentinel-2 remote sensing imagery, which provide accurate spatial information across different land cover types. The high temporal frequency and spatial resolution of Sentinel data give it strong potential for global applications. By utilizing an ensemble learning model, we effectively combined different feature information to address beach area identification tasks. If similar high-resolution remote sensing data, like that from Sentinel-1 and Sentinel-2, are available in other countries or regions, we believe the method could still be feasible and applicable. However, the environmental and geographic differences between countries or regions may influence the performance of the method. For example, beaches in tropical regions might differ from those in temperate zones in terms of texture and morphology, which could affect the applicability of the model in different areas. Moreover, lower-resolution datasets might impact the model's ability to capture fine details, thereby affecting the accuracy of extraction results. Therefore, optimizing the model to handle data of varying resolutions and adjusting the algorithm’s parameters to accommodate different land cover characteristics will be an important direction for future research. Additionally, while we mainly used remote sensing imagery in our study, other countries might employ different types of data, such as UAV imagery or aerial photography. Ensuring the adaptability and consistency of the algorithm across these different data sources is another challenge to consider. For instance, lower-resolution satellite images may not provide the same level of detail as high-resolution imagery, requiring adjustments to the algorithm to handle such data, including applying appropriate data processing and analysis techniques for lower-resolution images. In summary, while our method has shown good results in China, its applicability and transferability to other countries or regions still need further exploration and validation. In future research, we plan to explore how this method can be extended to beach extraction tasks in other countries, and how to optimize it for different resolutions and types of data.
  
  Some English references do not follow the journal’s formatting guidelines regarding author abbreviations (e.g., use of “and” vs. “&”). Please check and revise.
  
  Response:
  Thank you for your valuable feedback. We have carefully reviewed the references and ensured that all English references adhere to the journal's guidelines regarding author abbreviations and formatting, particularly the use of "and" and "&." Any discrepancies have been corrected accordingly.
  
  For Table 3, please clarify whether the model parameters were optimized through hyperparameter tuning or adopted as default values.
  
  Response:
  Thank you for your suggestion. We have used the default values for the model parameters in Table 3, and no hyperparameter tuning was performed. Additionally, we have updated the base learners in the ensemble model to tree-based models that are better suited for handling large datasets and wide geographical areas. These models include Random Forest (RF), Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM).
  
  L35–40: The statement that “traditional field surveys are inefficient” could be strengthened by citing 1–2 recent field-based studies (within the past three years), to demonstrate coverage of the latest research progress.
  
  Response:
  Thank you for your suggestion. We have added references to 1-2 recent field studies (published within the past three years) to strengthen the statement on the "inefficiency of traditional field surveys" and to demonstrate coverage of the latest research advancements.
  
  Citation: https://doi.org/10.5194/essd-2025-264-AC2

Jie Meng, Duanyang Xu, Zexing Tao, and Quansheng Ge

Supplement

https://doi.org/10.5194/essd-2025-264-supplement

Data sets

Spatial Patterns of Sandy Beaches in China and Risk Analysis of Human Infrastructure Squeeze Based on Multi-Source Data and Ensemble Learning Jie Meng, Duanyang Xu, Zexing Tao, Quansheng Ge https://zenodo.org/records/15307240

Jie Meng, Duanyang Xu, Zexing Tao, and Quansheng Ge

Viewed

Total article views: 1,457 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
1,325	84	48	1,457	28	34	48

HTML: 1,325
PDF: 84
XML: 48
Total: 1,457
Supplement: 28
BibTeX: 34
EndNote: 48

Views and downloads (calculated since 24 Jun 2025)

Month	HTML	PDF	XML	Total
Jun 2025	98	14	5	117
Jul 2025	93	20	21	134
Aug 2025	191	20	5	216
Sep 2025	859	11	5	875
Oct 2025	84	19	12	115

Cumulative views and downloads (calculated since 24 Jun 2025)

Month	HTML	PDF	XML	Total
Jun 2025	98	14	5	117
Jul 2025	93	20	21	134
Aug 2025	191	20	5	216
Sep 2025	859	11	5	875
Oct 2025	84	19	12	115

Viewed (geographical distribution)

Total article views: 1,374 (including HTML, PDF, and XML) Thereof 1,374 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 31 Oct 2025

Download

Preprint (2358 KB)
Metadata XML

Short summary

This study used multi-source remote sensing data and ensemble learning methods to map the distribution of sandy beaches in China from 2016 to 2023. A total of 2,984 sandy beaches were identified with high accuracy by integrating Sentinel-1/2 satellite imagery, terrain, and nighttime light data. Since 1990, the area at risk from human infrastructure squeeze has significantly increased. This study provides an updated dataset to support sustainable coastal management.


Total:	0
HTML:	0
PDF:	0
XML:	0