CCWEM and CCWIF: 10 m Datasets of Coastal Wetland Extent and Inundation-Frequency Gradients across China
Abstract. Coastal wetlands are critical ecological spaces for coastal-zone management, with direct implications for wetland conservation, shoreline regulation, and nearshore spatial planning. Here, we present two complementary 10 m datasets for China's coastal wetlands in 2025: the China Coastal Wetland Extent Map (CCWEM) and the China Coastal Wetland Inundation Frequency Map (CCWIF). CCWEM delineates the maximum wetland extent along the Chinese coast, whereas CCWIF further characterizes internal inundation-frequency gradients within this extent. To generate CCWEM, we constructed a four-season semantic segmentation dataset based on multi-season Sentinel-2 observations and developed SeasonMamba, which integrates shared encoding, seasonal adaptive fusion, and pseudo-MSIC representation learning to achieve stable delineation of coastal wetlands under varying seasonal and tidal conditions. To generate CCWIF, the Chinese coast was divided into 58 tidal subregions to account for spatial heterogeneity in tidal environments, and a region-specific inundation-frequency estimation workflow was developed by integrating Sentinel-2 MNDWI time series with tide gauge observations. CCWEM mapped 8906.15 km² of coastal wetlands and achieved an overall accuracy of 94.07 %, with an F1 score of 94.83 % in low-tide boundary validation. Independent validation using Sentinel-1 observed water occurrence showed overall monotonic agreement with the CCWIF classes across all validation regions combined (Spearman's ρ = 0.756), supporting its ability to characterize relative inundation-frequency gradients. To our knowledge, CCWIF provides the first national-scale dataset describing coastal wetland inundation-frequency gradients across China. Together, CCWEM and CCWIF provide a consistent high-resolution spatial dataset for characterizing both the extent and internal hydrological structure of China's coastal wetlands, supporting coastal wetland monitoring, assessment, and management under complex tidal conditions.
I thoroughly enjoyed reading this manuscript. The authors have developed two complementary national-scale datasets at 10 m resolution, capturing the spatial extent of China's coastal wetlands and the tidal inundation frequency gradient within these ecosystems. The research objectives are clearly defined, the technical workflow is comprehensive, and the data organization and validation framework are particularly impressive.
I especially appreciate the considerable effort the authors have invested in developing these datasets. Based on my examination of the data products, their spatial details, and the validation results presented in the manuscript, both datasets appear to be of high overall quality. In particular, CCWEM provides an impressive representation of the spatial extent of complex coastal wetlands and delineates intertidal boundaries with remarkable precision. Given the challenges associated with remote sensing mapping of coastal wetlands at the national scale, this level of data quality and spatial detail is particularly valuable.
The manuscript is also well written and logically structured. The authors provide a systematic presentation of the research background, data development, methodological framework, validation results, and potential applications. The relationship between the two data products is clearly articulated, and the figures and tables effectively support the understanding of the technical workflow and major findings.
While I highly appreciate these contributions, several questions arose during my reading that I believe deserve further consideration and discussion.
Major Comments
1. Mechanism for identifying the landward boundary of CCWEM and comparison with the approach of Zhang et al. (2022)
This is one of the aspects of the manuscript that I found most interesting.
I was particularly impressed by the spatial accuracy of CCWEM, especially the clean delineation of landward wetland boundaries in some areas. Given that many recently developed coastal wetland remote sensing products still suffer from the misclassification of inland vegetation, bare soil, and other spectrally similar land-cover types, I would be very interested to understand how the authors achieved such results.
According to the methodological description, CCWEM primarily relies on multi-season manually annotated samples, SeasonMamba, and MSIC composite imagery for semantic segmentation of coastal wetlands. However, the methodology does not explicitly describe any spatial constraint procedure comparable to the Maximum Seawater Extent (MSE).
This reminds me of the study by Zhang et al. (2022), published in Remote Sensing of Environment, entitled "Sub-continental-scale mapping of tidal wetland composition for East Asia: A novel algorithm integrating satellite tide-level and phenological features."
I was very impressed by this study, particularly its approach to identifying intertidal areas. Although the authors primarily employed a conventional random forest classifier, the entire workflow was specifically designed to account for the geographic and spatial characteristics of intertidal environments. In my view, one of the most valuable aspects of their approach is that tidal information is first used to constrain the spatial extent of the intertidal zone, within which subsequent land-cover classification is performed. This effectively transforms the identification of intertidal wetlands into a classification problem subject to explicit geographical constraints.
In contrast, CCWEM relies primarily on a multi-season deep learning framework for nationwide mapping. I am therefore curious about what information SeasonMamba actually uses to distinguish intertidal wetlands from adjacent terrestrial vegetation, bare soil, and other artificial surfaces.
2. Decomposition of the MNDWI time series and the physical meaning of the constant term (a₀)
The authors apply Fourier fitting to the MNDWI time series and extract the constant term (a₀) as an intermediate variable for subsequent tidal inundation frequency mapping. I appreciate the general idea of using time-series decomposition to extract relatively stable long-term signals, and I believe this approach has considerable potential for further exploration.
However, I have some questions about the underlying scientific rationale of this procedure that I would like to discuss.
First, considering the research objective, tidal inundation frequency is defined as the number of days in a year during which a given location is inundated by water. Assuming that we could obtain a continuous, noise-free MNDWI time series and establish a reliable water detection threshold, a relatively straightforward approach would be to identify a pixel as inundated whenever its MNDWI exceeds the predefined threshold. The proportion of time during which the pixel remains in this state could then be calculated to estimate its water occurrence frequency.
Of course, satellite observations are temporally discontinuous, while cloud cover, satellite overpass timing, and vegetation obstruction can all affect the practical implementation of this approach. Nevertheless, from the perspective of the definition of the target variable, this represents a relatively direct estimation strategy.
If a mathematical decomposition approach is adopted instead, I would expect the decomposition to be designed as closely as possible around the underlying mechanisms governing tidal inundation frequency. Temporal variations in MNDWI at a given location can be understood as the combined outcome of multiple influencing factors. Therefore, if the objective is to extract information about tidal inundation frequency, the resulting components should ideally correspond to the key factors controlling the inundation state.
For example, I would initially consider the following factors:
(1) Ground elevation. Ground elevation is a relatively stable spatial attribute that determines the water level required for tidal water to reach a particular location. Under the same tidal conditions, lower-elevation areas generally experience higher inundation frequencies.
(2) Monthly or spring-neap tidal water-level variations. Tidal water levels exhibit regular variations over the spring-neap tidal cycle, which alter the probability of inundation at locations with different elevations.
(3) Daily tidal variations. Different regions exhibit different tidal regimes, including regular semidiurnal, irregular semidiurnal, regular diurnal, irregular diurnal, and mixed tides. These tidal regimes and their associated intra-day water-level variations also influence the duration of pixel inundation.
If these factors could be identified or constrained using remote sensing observations, it might be possible to further derive relative ground elevation and combine it with tidal variations at different temporal scales to estimate the actual tidal inundation frequency at a given location more directly.
Second, another issue that I consider particularly important is whether MNDWI alone can adequately represent changes in water presence.
MNDWI is calculated using the Green and SWIR bands, both of which are also influenced by vegetation conditions. The reflectance of these two bands can vary with leaf area, canopy structure, vegetation water content, and phenological changes.
Consequently, annual variations in MNDWI within salt marshes may simultaneously contain hydrological signals and vegetation phenological signals.
I noticed that the annual variation curves obtained through Fourier fitting may exhibit certain similarities in shape to typical vegetation growth curves. For example, spring green-up, vigorous summer growth, autumn senescence, and winter dormancy can all produce distinct annual spectral cycles.
This suggests that periodic variations in MNDWI may not originate exclusively from water inundation processes; they may also reflect changes in vegetation growth and canopy water status.
Under these circumstances, the physical meaning of (a₀) becomes more complicated. It may integrate information on long-term hydrological conditions, relative ground elevation, vegetation cover, and other surface spectral characteristics. It therefore seems difficult to interpret this term specifically as a measure of long-term water-related responses based solely on the current Fourier decomposition approach.
3. Extending CCWEM and CCWIF to a multi-year framework
I would be very interested to see these datasets further extended to a multi-year framework.
To the best of my knowledge, existing remote sensing products for China's coastal wetlands still have limitations in consistently characterizing the spatial extent of intertidal areas over multiple years. Studies such as the work by Xu Bing's group published in an ISPRS-related journal ("Global mapping of tidal wetlands and adjacent environments using tidal analysis and multi-source Earth observations") and the GWF-ESSD data product (GWL_FCS30) developed by Liu Liangyun's group have provided important data foundations for wetland remote sensing monitoring in China. Nevertheless, there remains considerable room for improvement in the long-term, continuous, and fine-scale characterization of intertidal areas.
In my view, CCWEM has already demonstrated considerable strengths in spatial resolution, boundary delineation, and overall data quality. If the authors could further extend the current framework to multiple years, or at least provide an additional classification map for 2020 or an earlier year in the present study, the scientific value and broader applicability of this work would be substantially enhanced.
Minor Comments
1. Selection of Sentinel-2 as the primary data source
The study primarily uses Sentinel-2 imagery for nationwide mapping. Given the frequent cloud cover in coastal regions and the sensitivity of intertidal environments to tidal conditions at the time of satellite observation, it can sometimes be difficult to obtain a sufficiently complete set of valid images within a relatively short period.
I would be interested to know whether the authors considered using harmonized remote sensing products such as Harmonized Landsat and Sentinel-2 (HLS). Such products could increase the temporal density of observations, although their use also involves considerations related to spatial resolution and spectral consistency across different sensors. What were the main reasons for ultimately selecting Sentinel-2 as the primary data source?
2. Cloud-cover filtering criteria
The authors applied a cloud-cover threshold of less than 10% when selecting imagery for the four seasonal datasets. I personally consider this threshold relatively strict, as it may exclude many images that still contain substantial areas of usable pixels.
Did the authors consider using more refined cloud and cloud-shadow masking algorithms to perform pixel-level quality control on images with greater overall cloud cover, thereby increasing the number of available observations?
This may be particularly relevant to coastal regions in southern China, where cloud cover is frequent. If image selection is based solely on the overall cloud-cover percentage, a considerable number of observations that remain clear and usable over local areas may be discarded. I would therefore be interested to know whether the authors evaluated how different cloud-cover thresholds affect the number of samples, spatial coverage, and final mapping accuracy.
In addition, the authors appear to have adopted different data processing strategies for selecting images used to construct the four seasonal training datasets and for generating the MSIC prediction imagery. I would appreciate further clarification of the rationale behind these differences.
3. Temporal selection of the four seasonal images
The authors consistently selected January, April, July, and October to represent winter, spring, summer, and autumn, respectively.
Given the considerable latitudinal and climatic gradients across China, vegetation phenology varies substantially among regions. The same calendar month may correspond to different stages of vegetation development in northern and southern China. For example, vegetation in southern China may remain relatively active in October, whereas some regions in northern China may already have entered senescence or dormancy.
Could the use of a uniform monthly classification scheme across the entire country, without regionally adjusting the temporal selection according to latitude, climatic zones, or vegetation phenology, affect the model's ability to learn wetland characteristics across different regions?
4. Spatial partitioning codes
The methodology mentions that the training, validation, and test sets all include regional groups such as T49, T50, and T51. However, the main text does not sufficiently explain what these codes represent or the rationale behind the spatial partitioning. I suggest that the authors clarify their meanings and the basis for this spatial division.
5. Relationships among the different components of SeasonMamba
SeasonMamba consists of two main feature extraction branches: the Seasonal Branch and the Pseudo-MSIC Branch. According to the methodological description, the two branches separately process multi-seasonal stable features and MSIC composite imagery features. These features are subsequently integrated through the main encoder for multi-scale feature fusion, while a high-level feature consistency loss is used to establish a relationship between them.
I find this design interesting and would appreciate further clarification of the specific collaborative mechanisms between the two branches.
Based on the current network architecture, the two branches appear to operate relatively independently during feature extraction, with different types of features eventually being incorporated into the main encoder. Does this design primarily represent the integration of two complementary feature sources, or is there a deeper level of feature interaction between the two branches?
In addition, the CASM module in the Pseudo-MSIC Branch is responsible for reorganizing and synthesizing multi-seasonal imagery to generate pseudo-composite images that approximate the actual MSIC prediction imagery. I am also curious about the specific feature selection and reorganization mechanisms used in this process, as well as how the differences between the generated pseudo-MSIC images and the actual MSIC imagery are controlled.
6. Actual contributions of individual modules in the ablation experiments
According to Table 1, the baseline model E6, which uses a single-season input, already achieves an F1 score of 90.64% for the wetland class, while the complete model E1 achieves 93.43%, representing an improvement of approximately 2.79 percentage points.
Removing the Pseudo-MSIC Branch and the Seasonal Branch results in F1 reductions of approximately 0.65 and 0.91 percentage points, respectively.
This raises an interesting question: given that the baseline model already demonstrates relatively strong wetland classification performance, the additional improvement contributed by the newly introduced modules appears relatively modest. Could there be functional overlap between the different modules? Do the two branches generate synergistic effects when used together?
The current ablation experiments demonstrate the contributions of individual branches within the complete model, but they may not be sufficient to fully distinguish the independent contributions of different modules from their interaction effects.
Furthermore, the comparison between E1 and E6 involves simultaneous changes in several factors, including the number of seasonal inputs, sample organization, and network branches. A further comparison between conventional feature fusion and the complete SeasonMamba framework, while keeping the seasonal inputs and training samples consistent, could help better isolate and explain the actual performance gains attributable to the network architecture itself.
Finally, one of the central objectives of this study is to determine the maximum spatial extent of wetlands, particularly through the precise identification of wetland boundaries. The current ablation experiments primarily rely on pixel-level F1 and mIoU scores for evaluation. It would be valuable to further compare boundary delineation accuracy across different model configurations, using metrics such as boundary F1 scores, mean boundary distances, and separate assessments of seaward and landward boundary performance. Such analyses would help demonstrate more directly how the newly introduced modules contribute to the core mapping objective of the study.