Evaluating Consistency and Accuracy of Public Tidal Flat Datasets in China’s Coastal Zone

Highlights

What are the main findings?

• Systematic evaluation of six tidal flat datasets across China reveals pronounced spatial discrepancies and regional variations in accuracy;
• Independent edge-based validation demonstrates that dataset reliability strongly depends on sensor type and index selection.

What is the implication of the main finding?

• It provides a benchmark for dataset selection and methodological optimization, supporting improved coastal wetland monitoring and management.

Abstract

Tidal flats, as critical transitional ecosystems between land and sea, face significant threats from climate change and human activities, necessitating accurate monitoring for conservation and management. However, publicly available tidal flat datasets exhibit substantial discrepancies due to variations in data sources, spectral indices, and classification methods. This study systematically evaluates six widely used 2020 tidal flat datasets (GTF30, GWL_FCS30, MTWM-TP, DCTF, CTF, and TFMC) across China’s coastal zone, assessing their spatial consistency, area estimation differences, and edge classification accuracy. Using a novel edge validation point set (3150 samples) derived from tide gauge stations and low-tide imagery, we demonstrate that MTWM-TP (OA = 0.85) and TFMC (OA = 0.84) achieve the highest accuracy, while DCTF and GTF30 show systematic underestimation and overestimation, respectively. Spatial agreement is strongest in Jiangsu (49.8% unanimous pixels) but weak in turbid estuaries (e.g., Zhejiang). Key methodological divergences include sensor resolution (Sentinel-2 outperforms Landsat in low-tide coverage), spectral index selection (mNDWI reduces false positives in turbid waters), and boundary constraints (high-tide masks suppress inland misclassification). We propose establishing an automated multi-source framework integrating optical (Sentinel-2, Landsat) and radar (Sentinel-1) observation data to enhance low-tide coverage, constructing region-adaptive spectral indices and improving boundary accuracy through the combination of machine learning and thresholding algorithms. This study provides a critical benchmark for dataset selection and methodological advancements in coastal remote sensing.

1. Introduction

Tidal flats, as dynamic intertidal zones periodically submerged and exposed by tides [1,2], are critical transitional ecosystems bridging land and sea. Globally, tidal flats span approximately 354,600 km² (60°N–60°S) [3], serving as biodiversity hotspots [4], key habitats for coastal species [5], and providers of essential ecosystem services, including coastal stabilization [6], water purification [7], and carbon sequestration [8]. However, these vulnerable ecosystems [9] face escalating threats from climate change (e.g., sea level rise [10]), anthropogenic activities (e.g., industrialization and coastal development [11]), and geomorphological processes (e.g., sediment supply fluctuations [3,12] and land subsidence [13]). Between 1984 and 2016, global tidal flat area declined by 16.02% [3], triggering ecological degradation and biodiversity loss. China, with its 18,000 km coastline [14] and the world’s second largest tidal flat area [2], plays a pivotal role in migratory bird pathways and biodiversity conservation. Systematic monitoring of China’s tidal flats is thus imperative—not only to refine global carbon cycle assessments but also to advance sustainable development goals (SDGs) for coastal resilience.

Satellite remote sensing has emerged as the primary tool for tidal flat monitoring, with four key technical components. (1) Data selection: Landsat series (30 m resolution) dominate global-scale studies [15,16], while Sentinel-2 (10 m) is preferred for regional analyses [10,17,18]. (2) Spectral indices: Enhanced water–land contrast [18,19] is achieved via indices like NDWI [20], mNDWI [21], or specialized tidal flat indices (e.g., LTideI [15], TWDI [17]). (3) Temporal analysis: This involves the characterization of tidal phases through time-series images. A common strategy is extreme value compositing (e.g., using maximum NDVI values [18]) to approximate the maximum exposed tidal flat extent during low tides [22]. (4) Classification and validation: This involves machine learning [15] or thresholding [18] for tidal flat extraction, coupled with accuracy assessments using ground reference or visual interpretation points [15,18,23].

Under this framework, various global and regional tidal flat datasets have been developed. At the global scale, representative products include the UQD (1984–2019, 30 m) [2] and GTF30 (2000–2022, 30 m) [15], both based on Landsat time-series data, as well as the GWL_FCS30 (2020, 30 m) [16], which integrates Sentinel-1 and Landsat data. Regionally, the MTWM-TP (2020, 10 m) [9] based on Sentinel-2 was developed for coastal East Asia, while datasets for China’s coastal areas include the Sentinel-2-derived CTF (2019, 10 m) [18] and TFMC (2020, 10 m) [17], along with the Landsat-based FUDAN/OU (2015, 2018, 30 m) [23] and DCTF (1989–2020, 30 m) [24]. Additionally, regional datasets such as SZU (2015, 30 m) [10] and IGSNRR (2017, 10 m) [25] cover northern and southern coastal China, respectively. Although these datasets provide valuable support for tidal flat research, they exhibit significant spatial heterogeneity due to methodological differences (e.g., sensor resolution, spectral indices, and classification algorithms). While most datasets claim accuracies exceeding 85%, their reliability for scientific and policy applications remains limited in the absence of cross-dataset validation. Specifically, three key gaps currently hinder the effective selection and application of these tidal flat products: (1) the lack of a quantitative understanding of the spatial agreement and disagreement among datasets at the pixel level; (2) the uncertainty in their capability to delineate tidal flat boundaries, which is crucial for dynamic monitoring but has rarely been evaluated through traditional intra-class validation; and (3) the absence of a systematic diagnosis that links observed discrepancies to specific methodological factors, such as sensor resolution, spectral indices, and boundary extraction strategies.

To ensure temporal consistency and comparability, we focus on tidal flat datasets representing the year 2020, for it offers the most comprehensive from multiple global and regional mapping efforts, enabling cross-dataset evaluation under a unified temporal baseline. Moreover, the datasets (GTF30, GWL_FCS30, MTWM-TP, DCTF, CTF, TFMC) released during 2020 adopted distinct methodological frameworks (e.g., different sensors, indices and classification algorithms), making this period particularly suitable for assessing methodological impacts on mapping accuracy. In addition, some regional products such as the SZU (2015) and IGSNRR (2017) datasets were excluded from this analysis, for their limited spatial coverage (focused on the northern or southern coasts of China) and reliance on earlier data sources and coarse classification schemes, which hinder direct comparison with the nationwide 2020 products.

To address these gaps, we conduct a systematic evaluation of six 2020 tidal flat datasets (GTF30, GWL_FCS30, MTWM-TP, DCTF, CTF, TFMC) across China’s coastal zone, with three objectives: (1) quantify spatial consistency among datasets; (2) perform independent edge-based validation (3150 samples) to benchmark boundary delineation accuracy; and (3) diagnose methodological drivers (e.g., data sources, indices, algorithms) of observed discrepancies. Our work provides a critical benchmark for dataset selection and methodological optimization in tidal flat mapping.

The paper is structured as follows. Section 2 describes the study area and datasets. Section 3 details the methodology. Section 4 presents comparative results. Section 5 discusses technical implications and Section 6 concludes with recommendations.

2. Study Area and Dataset

2.1. Study Area

China’s coastal zone, spanning from the Yalu River estuary (40.5°N) to the Beilun River estuary (18.2°N), encompasses 12 provincial-level administrative regions (including Taiwan and Hainan islands) with a total coastline of 18,000 km [14] (Figure 1). This region traverses tropical, subtropical, and temperate climate zones, supporting diverse wetland ecosystems such as tidal flats, salt marshes, and mangroves. To capture the full ecological gradient, we defined the study area as a 5 km landward and 30 km seaward buffer along the provincial coastal boundaries, which were obtained from the National Platform for Common Geospatial Information Services (Tianditu, https://www.tianditu.gov.cn/, accessed on 5 May 2025). This delineation ensures comprehensive coverage of intertidal dynamics while minimizing interference from noncoastal landscapes.

2.2. Publicly Available Tidal Flat Datasets

We evaluated six 2020 tidal flat datasets covering China’s coastal zone (

Table 1. Technical parameter comparison of tidal flat datasets along China’s coast (2020).

Dataset	Time	Range	Data Sources	Resolution	Core Index/Band	Nominal Accuracy	Class
GTF30 [15]	2000–2022	Global	Landsat	30 m	Landsat’s six bands, LTideI, NDVI, mNDWI, LSWI	90.34%	Tidal flat
GWL_FCS30 [16]	2020	Global	Sentinel-1 &Landsat	30 m	Landsat’s six bands, NDVI, mNDWI, EVI, LSWI	86.44%	Tidal flats, Salt marshes, Mangroves, Inland wetlands
MTWM-TP [9]	2020	Esat Asia	Sentinel-2	10 m	Sentinel-2’s twelve bands, NDVI, NDWI	97.02%	Tidal flats, Salt marshes, Mangroves
CTF [18]	2020	China	Sentinel-2	10 m	mNDWI, NDVI	95%	Tidal flats
DCTF [24]	1989–2020	China	Landsat	30 m	NDVI, mNDWI, LSWI, BSI, EVI, MSAVI, NDBI	90.84%	Tidal flats, Salt marshes
TFMC [17]	2020	China	Sentinel-2	10 m	mNDWI, TWDI	97%	Tidal flats

). These include two global-scale datasets: (1) GTF30 (2000–2022, 30 m) (https://doi.org/10.5281/zenodo.7936721 (accessed on 10 May 2025)), developed by Zhang et al. [15] using Landsat imagery and a random forest classifier based on the LTideI index, and (2) GWL_FCS30 (2020, 30 m) (https://doi.org/10.5281/zenodo.7340516 (accessed on 10 May 2025)), a global wetland classification that integrates Sentinel-1 and Landsat data, with tidal flats as a subset category [16]. At the regional scale, we assessed four datasets: (1) MTWM-TP (2020, 10 m) (https://figshare.com/articles/dataset/Fujian_zip/14331785 (accessed on 10 May 2025)), covering East Asia’s coastline [10], which combines Sentinel-2 imagery with tidal-phase and phenological features to classify tidal flats, salt marshes, and mangroves; (2) DCTF (1989–2020, 30 m) (https://doi.org/10.3974/geodb.2021.10.06.V1 (accessed on 10 May 2025)), a China-focused dataset [24] derived from Landsat using a random forest classifier and encompassing tidal flats and salt marshes; and (3–4) CTF (2019, 10 m) (https://doi.org/10.1016/j.rse.2021.112285 (accessed on 10 May 2025)) and TFMC (2020, 10 m) (https://code.earthengine.google.com/a4256918c4c85e863bb52f2e57a631de (accessed on 10 May 2025)), both generated via Google Earth Engine (GEE) using Sentinel-2 data. CTF [18] employs NDVI and mNDWI indices, while TFMC [17] introduces the novel Tidal Wetland Dynamic Index (TWDI); both apply time-series index extremum compositing coupled with Otsu threshold segmentation for tidal flat extraction.

3. Methods

This study systematically evaluates tidal flat datasets through four key steps (Figure 2). (1) Dataset standardization: This involves harmonizing classification systems, temporal baselines, and spatial scales to ensure comparability between datasets. (2) Quantitative comparison: This involves quantifying area discrepancies across datasets using the Coefficient of Variation (CV) and analyzing spatial consistency on a per-pixel basis through a six-level pixel-wise metric. (3) Independent validation: This involves the assessment of edge classification accuracy using edge sample points at the maximum exposure of the tidal flats and quantifying precision metrics via a confusion matrix. (4) Technical linkage: This involves linking methodological differences to performance outcomes, in particular investigating how variations in data sources, spectral indices, time-series analysis, and classification algorithms influence the observed differences in area estimation and spatial accuracy among datasets.

3.1. Dataset Standardization

To ensure comparability [26], we addressed three key sources of heterogeneity in the datasets. Firstly, regarding classification systems, for datasets that contain multiple classes (e.g., GWL_FCS30, DCTF, MTWM-TP), we extracted “tidal flat” pixels using their original classification schemes. In this study, tidal flat refers specifically to non-vegetated intertidal surfaces. Although salt marshes and mangroves are important vegetated components of the broader intertidal zone, they were treated as separate classes in most datasets; for example, MTWM-TP distinguishes tidal flats, salt marshes, and mangroves, whereas CTF and TFMC explicitly remove vegetated areas and retain only bare tidal flats. Therefore, to maintain consistency across all datasets, we excluded the vegetated categories. Secondly, for temporal consistency, we standardized the temporal baselines by retaining only 2020 data from multi-year datasets (e.g., GWL_FCS30, DCTF). Although MWTM-TP incorporated imagery from June 2019 to December 2020 and CTF used data from January 2019 to June 2020, we considered these data to be representative of 2020 due to the demonstrated short-term stability of tidal flats, which exhibit change rates below 5% between 2001 and 2016 [27]. Finally, to address differences in spatial scale, all datasets were converted into raster format and imported into Google Earth Engine (GEE), where datasets with spatial resolutions coarser than 10 m were resampled to 10 m using nearest neighbor interpolation. The resampled products were then clipped and analyzed within the study area, and area calculations were conducted under the Albers Equal-Area projection to maintain geometric accuracy.

3.2. Quantitative Comparison

We assessed the heterogeneity among the six tidal flat products along two dimensions: area discrepancy and spatial consistency.

3.2.1. Area Discrepancy

At the provincial scale, the coefficient of variation (CV) of tidal flat area across the six products was calculated as

$$\mathrm{C}\mathrm{V}=\mathrm{s}/\overline{\mathrm{x}}$$

(1)

$$\mathrm{s}=\sqrt{1/\mathrm{N}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)}^2}$$

(2)

where $\overline{\mathrm{x}}$ is the mean area of the six datasets, $\mathrm{s}$ is the standard deviation, ${\mathrm{x}}_{\mathrm{i}}$ denotes the area of the i-th dataset, and N is the total number of datasets (N = 6). A higher CV indicates larger inter-product differences within the province.

3.2.2. Spatial Consistency

(1) Overall consistency

Each product was resampled to a 10 m binary mask (1 = tidal flat; 0 = other). For every pixel, we counted how many products classified it as tidal flat and assigned an agreement level from 1 to 6, with level 6 indicating unanimous agreement among all products.

(2) Pair-wise consistency

The Spatial Consistency index (SC) [28] was used to quantify the overlap between any two products:

$$\mathrm{S}\mathrm{C}=\mathrm{M}/\left[\left(\mathrm{X}+\mathrm{Y}\right)/2\right]$$

(3)

where M is the number of tidal flat pixels jointly identified by both datasets, and X and Y are the total numbers of tidal flat pixels identified by each individual dataset, respectively. SC ranges from 0 to 1; values closer to 1 denote higher consistency between the two datasets.

3.3. Edge Validation

3.3.1. Sample Collection

To adequately cover the different tidal systems along the coast of China, we selected 21 tide stations according to the tidal regime distribution [18]. These stations are evenly distributed in the coastal provinces of China and cover main tidal types (including regular semidiurnal, irregular semidiurnal, irregular diurnal, and regular diurnal tides) (Figure 3). For each station, low-tide images were selected by matching Landsat-8/9 and Sentinel-2 overpass times with tide records from the National Marine Science Data Center (https://mds.nmdis.org.cn/pages/tidalCurrent.html, accessed on 29 October 2025). The instantaneous tide height (H) at image acquisition was calculated through linear interpolation using tidal harmonic equations [29]:

$$\mathrm{D}\mathrm{u}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}-\mathrm{t}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{p}\mathrm{h}\mathrm{a}\mathrm{s}\mathrm{e}\mathrm{s}:\mathrm{H}={\mathrm{h}}_{\mathrm{l}\mathrm{o}\mathrm{w}}+∆\mathrm{h}\times 1/2\left(1-\cos \mathrm{T}/\mathrm{t}\times 180°\right)$$

(4)

$$\mathrm{D}\mathrm{u}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{l}\mathrm{o}\mathrm{w}-\mathrm{t}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{p}\mathrm{h}\mathrm{a}\mathrm{s}\mathrm{e}\mathrm{s}:\mathrm{H}={\mathrm{h}}_{\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}}+∆\mathrm{h}\times 1/2\left(1-\cos \mathrm{T}/\mathrm{t}\times 180°\right)$$

(5)

where ${\mathrm{h}}_{\mathrm{l}\mathrm{o}\mathrm{w}}$ and ${\mathrm{h}}_{\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}}$ represent adjacent low and high tide heights, $∆\mathrm{h}$ denotes the tidal range, T indicates the time interval between image acquisition and the nearest high/low tide, and t corresponds to the duration between two tide extremes. Scenes acquired at low-tide stages with ≤60% cloud cover were retained, followed by cloud masking; the details for each province are listed in

Table 2. Tide gauge station metadata and low-tide image parameters for validation.

Province	Tidal Station	Image Sources	Overpass Times	Tidal Height (cm)	Chart Datum (cm)	Type of Tide
Liaoning	Laobeihekou	Sentinel-2	6 May 2020 10:56:26	51	−209	Irregular Semidiurnal
	Daludao	Sentinel-2	19 January 2020 10:46:27	152	−332	Regular Semidiurnal
Hebei	Caofeidian	Sentinel-2	11 April 2020 10:56:55	60	−178	Irregular Diurnal
Tianjin	Tanggu	Sentinel-2	24 May 2020 11:06:59	94	−241	Irregular Semidiurnal
Shandong	Wanwangoukou	Sentinel-2	8 July 2020 11:07:13	39	−130	Regular Semidiurnal
	Dongying	Landsat 8	14 March 2020 10:41:49	62	−100	Irregular Semidiurnal
	Zhangjiabu	Sentinel-2	14 January 2020 10:47:28	10	−220	Irregular Semidiurnal
Jiangsu	Lianyungang	Sentinel-2	27 December 2020 10:58:09	131	−290	Regular Semidiurnal
	Jianggang	Sentinel-2	28 April 2020 10:48:31	97	−301	Regular Semidiurnal
	Lvsi	Sentinel-2	14 March 2020 10:48:46	135	−310	Regular Semidiurnal
Shanghai	Zhongjun	Landsat 8	12 May 2020 10:24:24	107	−225	Regular Semidiurnal
Zhejiang	Qimengang	Sentinel-2	13 August 2020 10:49:37	295	−379	Regular Semidiurnal
	Damendao	Sentinel-2	11 November 2020 10:40:11	184	−363	Regular Semidiurnal
Fujian	Minjiangkou	Sentinel-2	26 August 2020 10:50:41	140	−353	Regular Semidiurnal
	Quanzhou	Sentinel-2	26 August 2020 10:50:59	133	−366	Regular Semidiurnal
Taiwan	Magong	Sentinel-2	21 November 2020 10:41:08	52	−160	Regular Semidiurnal
Guangdong	Chaozhougang	Sentinel-2	7 December 2020 11:01:11	59	−101	Irregular Semidiurnal
	Shekou	Sentinel-2	20 November 2020 11:11:45	88	−152	Irregular Diurnal
	Zhanjiang	Sentinel-2	3 January 2020 11:22:02	154	−220	Irregular Semidiurnal
Hainan	Xinying	Sentinel-2	2 May 2020 11:22:29	78	−205	Regular Diurnal
Guangxi	Tieshangang	Sentinel-2	2 May 2020 11:22:17	174	−255	Irregular Diurnal

.

Traditional methods employ intra-class random points for accuracy validation [17,18], reflecting overall classification precision yet often failing to capture boundary classification errors. These errors constitute the primary source of uncertainty in tidal flat extraction. Consequently, we adopt an edge validation approach, focusing on two critical boundaries: the tidal flat–land interface and the tidal flat–water interface. The tidal flat–land boundary was defined using artificial shorelines with distinct spectral characteristics, such as seawalls and aquaculture pond perimeters. Conversely, the tidal flat–water boundary was determined through independent visual interpretation by three domain experts, retaining only segments unanimously agreed upon to ensure objectivity.

During sampling, 50 land points were uniformly distributed along the landward side of the tidal flat–land boundary. For the tidal flat–water boundary, a symmetrical sampling method was employed, where 50 tidal flat points were randomly generated within a 10-pixel buffer zone on the landward side of the interpreted boundary line. Corresponding water body points were then generated by mirroring the boundary line, forming spatially paired points. This method yielded a total of 3150 boundary validation samples for assessing the boundary accuracy of each dataset.

3.3.2. Accuracy Assessment

The collected samples were utilized to build confusion matrices for calculating classification accuracy metrics, including overall accuracy (OA), producer’s accuracy (PA), and user’s accuracy (UA), which, respectively, represent total agreement, omission error, and commission error between classified and reference data [30,31].

$$\mathrm{O}\mathrm{A}=\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{P}}_{\mathrm{i}\mathrm{i}}/\mathrm{N}$$

(6)

$${\mathrm{U}\mathrm{A}}_{\mathrm{i}}={\mathrm{P}}_{\mathrm{i}\mathrm{i}}/\sum _{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{P}}_{\mathrm{i}\mathrm{j}}$$

(7)

$${\mathrm{P}\mathrm{A}}_{\mathrm{i}}={\mathrm{P}}_{\mathrm{i}\mathrm{i}}/\sum _{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{P}}_{\mathrm{j}\mathrm{i}}$$

(8)

Here, N denotes the total number of sampling points, ${\mathrm{P}}_{\mathrm{i}\mathrm{i}}$ represents the number of elements in row i and column i, ${\mathrm{P}}_{\mathrm{i}\mathrm{j}}$ denotes the number of elements in row i and column j, and ${\mathrm{P}}_{\mathrm{j}\mathrm{i}}$ indicates the number of elements in row j and column i.

MTWM-TP lacks data for Taiwan Province; the corresponding results are therefore reported as missing.

4. Results

4.1. Inter Dataset Variability in Tidal Flat Area

Figure 4 summarizes the 2020 provincial-scale estimates derived from the six independent products (TFMC, MTWM-TP, CTF, DCTF, GTF30, GWL_FCS30). The analysis reveals three statistically distinct macro-regional classifications. (1) Extensive mudflats: Jiangsu Province hosts over 2000 km² of tidal flats, accounting for over 35% of the national total. (2) Moderately complex coasts: Liaoning, Shandong, Zhejiang, and Fujian each contain approximately 1000 km² of tidal flats, controlled by convoluted shorelines and heterogeneous sediment regimes. (3) Small-scale areas: Hebei, Tianjin, and five other provinces retain less than 500 km², limited by short fetch, rocky headlands, or artificial shoreline features coasts.

Inter-product agreement, quantified by the coefficient of variation (CV), exhibits a clear pattern (Figure 4b). High-consistency provinces (CV < 15%) are Jiangsu (9.4%) and Shandong (11.8%), whereas Tianjin, Taiwan, and Hainan display high dispersion (CV > 40%). The remaining seven provinces fall within the moderate range (CV = 20–30%).

4.2. Provincial Scale Area Rankings

Figure 5 ranks the six products within each province (1 = largest area). In combination with Figure 4a, three consistent patterns emerge.

GTF30 produced the largest estimates, ranking first in nine of the twelve provinces and second in the remaining three. The largest absolute departures occur in Liaoning, Zhejiang, Fujian, and Guangdong, with Liaoning exhibiting the most pronounced outlier.

DCTF and GWL_FCS30 were consistently low. DCTF is the smallest source in seven provinces; in Fujian, it is approximately 200 km² below the second-smallest estimate. GWL_FCS30 ranks fifth in seven provinces and is the minimum value for Shanghai and Taiwan.

CTF occupies the mid-range, ranking third or fourth in most provinces, but records the lowest value in Jiangsu. TFMC and MTWM-TP exhibit dispersed rankings without a clear directional bias. However, the areal estimates of TFMC, MTWM-TP, and GWL_FCS30 demonstrated consistent agreement (standard deviation is 113 km²), while GTF30, CTF, and DCTF produced significantly larger.

The overestimation of tidal flat area by GTF30 may be due to an unintroduced high-water boundary mask causing misclassification of nearshore aquaculture ponds, and the underestimation of DCTF manifests as the inward contraction of tidal flat boundaries, potentially attributable to data resolution and the classification algorithm’s sensitivity to edge mixed pixels. The relevant mechanisms will be analyzed in depth in the Section 5.

4.3. Spatial Agreement Assessment

While aggregate area metrics reveal overall magnitude discrepancies, they do not capture spatial heterogeneity in dataset performance. We therefore quantified pixel-wise agreement among the six products by stacking their 10 m binary masks (1 = tidal flat; 0 = otherwise) and assigning each pixel an agreement level L ∈ {1, 2, …, 6}, where L = 6 indicates unanimous classification. Provincial frequency distributions of these levels are shown in Figure 6.

Nationwide, spatial agreement is modest; a total of 57.2% of all tidal flat pixels fall within low agreement classes (L = 1–3), whereas only 42.8% reach L ≥ 4. Perfect agreement (L = 6) is achieved for merely 22.8% of pixels, underscoring substantial inter-product divergence. Note that low-agreement pixels (L = 1) do not necessarily denote commission or omission errors; they may also represent true tidal flats detected by some, but not all, algorithms.

Across provinces, three distinct patterns emerge. (1) High agreement: Jiangsu exhibits near-maximum consensus, with 47.8% of pixels at L = 6 and <30% at L = 1–2. (2) Moderate agreement: Liaoning, Shandong, Shanghai, Fujian, and Guangxi show 35–45% of pixels at L ≥ 4 and ≤50% at L = 1–2. (3) Low agreement: Hebei, Tianjin, Zhejiang, Taiwan, Guangdong, and Hainan display <35% of pixels at L ≥ 4; Guangdong and Hainan record only 9.3% and 6.1% at L = 6, respectively, while > 55% of pixels fall within L = 1–2, indicating pronounced spatial inconsistency.

Figure 7 illustrates the spatial pattern of agreement levels among the six datasets. High-consistency pixels (level 6) are concentrated in well-developed tidal flat regions, such as the Liao River estuary (Figure 7f) and the Jiangsu shoals, forming contiguous swaths. In contrast, turbid estuaries, like the Oujiang (Figure 7g), Jiaojiang, and Hangzhou Bay, are dominated by low-consistency pixels (levels 1–3). The disagreement is particularly acute along the Oujiang boundary, where most pixels are assigned levels 1–3. Intermediate agreement dominates elsewhere, exemplified by the Nanliujiang estuary in Guangxi (Figure 7h), where agreement levels increase landward, revealing systematic positional uncertainty at the tidal flat margin.

Taken together, Figure 6 and Figure 7 demonstrate pronounced spatial heterogeneity in tidal flat detection, arising from both methodological differences (data sources, algorithms and processing chains) and the intrinsic complexity of certain coastal environments. To quantify these differences, we computed the pairwise Spatial Consistency index (SC) at national and provincial scales (Figure 8).

National-scale SC values (Figure 8a) range from 0.59 to 0.81. The highest agreement is observed between MTWM-TP and TFMC (SC = 0.81), whereas GTF30 and DCTF exhibit the lowest (SC = 0.59). Five additional pairs exceed 0.7: CTF–TFMC, CTF–MTWM-TP, GWL_FCS30–GTF30, GTF30–TFMC and GWL_FCS30–MTWM-TP. All remaining combinations fall within 0.60–0.70, except GTF30–DCTF. Notably, MTWM-TP, TFMC, and CTF form a coherent cluster, whereas DCTF shows uniformly lower agreement with every other product, consistent with its systematic underestimation, noted in Section 4.1.

Provincial SC patterns (Figure 8b) further reveal strong regional modulation. MTWM-TP versus TFMC remains >0.80 in all provinces, peaking in Jiangsu (0.94) and declining slightly in Zhejiang, Guangdong, and Hainan. CTF–TFMC and GWL_FCS30–GTF30 follow similar spatial trends but register weaker performance in Taiwan. Conversely, DCTF paired with any dataset yields SC <0.40 in Guangdong and Hainan, reaching a minimum of 0.21.

To examine the influence of tidal flat extent on inter-product agreement, we plotted SC against the mean area of each provincial pair (Figure 8c). Provinces with large, continuous flats (green symbols) cluster at high SC values, whereas provinces with fragmented or narrow flats (blue and red symbols) exhibit greater scatter and lower SC. This confirms that extensive, morphologically simple flats enhance mapping stability, while small or complex flats amplify methodological differences.

In summary, Jiangsu displays the highest stability, combining the lowest CV (Section 4.1) with the highest SC values, attributable to its broad, regular flats. Fujian, Guangxi, and Liaoning also achieve high agreement. In contrast, Tianjin, Zhejiang, Taiwan, Guangdong, and Hainan show pronounced discrepancies—high proportions of low-agreement pixels and SC values frequently <0.4. These regional contrasts reflect both the geometric and sedimentological properties of the flats and the divergent sensor configurations, classification algorithms, and thresholding strategies employed by each dataset; the latter are explored in detail in Section 5.

4.4. Accuracy Assessment Using 3150 Edge Validation Points

To quantify classification performance at the tidal flat boundary, we generated an independent set of 3150 validation points adjacent to 21 tide stations. Overall accuracies (OA) ranged from 0.66 to 0.85 (

Table 3. Confusion matrix of edge validation points for tidal flat datasets.

Dataset	Class	TF	Non-TF		UA	OA
			Inland	Water
TFMC	TF	926	175	207	0.71	0.84
	Non-TF	124	875	843	0.93
	PA	0.88	0.83	0.80
MTWM-TP	TF	853	64	226	0.75	0.85
	Non-TF	147	936	774	0.92
	PA	0.85	0.94	0.77
CTF	TF	627	94	485	0.69	0.78
	Non-TF	423	956	865	0.81
	PA	0.6	0.91	0.82
DCTF	TF	254	188	85	0.48	0.66
	Non-TF	796	862	965	0.7
	PA	0.24	0.82	0.92
GTF30	TF	664	368	145	0.56	0.71
	Non-TF	386	682	905	0.8
	PA	0.63	0.65	0.86
GWL_FCS30	TF	349	242	58	0.54	0.68
	Non-TF	701	808	992	0.72
	PA	0.33	0.77	0.94

). MTWM-TP achieves the highest OA (0.85), with high user’s and producer’s accuracies (UA and PA) for non-tidal flat classes. It exceeds 0.80 in ten provinces and attains the top rank in six (Figure 9). TFMC ranks second (OA = 0.84), achieving PA of 0.88 for tidal flat. CTF performs moderately well (OA = 0.78), yet both UA and PA for tidal flat are comparatively low, indicating notable omission and commission errors, especially in Tianjin (OA = 0.61) and Guangxi (OA = 0.63).

DCTF, GWL_FCS30, and GTF30 yield lower accuracy (0.66, 0.68, and 0.71, respectively). DCTF’s PA for tidal flat is only 0.24, whereas PA for inland and water classes reaches 0.82 and 0.92, confirming a systematic underestimation of tidal flat extent, consistent with its low area estimates and SC values (Section 4.1). GWL_FCS30 and GTF30 exhibit markedly lower PA for inland classes, implying widespread misclassification of aquaculture ponds and bare soil as tidal flats, thereby inflating area totals. GTF30 overestimates area in every province, and its reduced OA corroborates severe commission errors.

The edge validation accuracies (Table 3) differ from the nominal accuracies reported in the original datasets (Table 1), with the differences ranging from 12 to 24 percent. This primarily stems from the validation strategy employed. By focusing sampling on the edge of the tidal flat, rather than using traditional random points within class, our method is more sensitive to positional errors along the boundary. Edge validation shows that MTWM-TP and TFMC achieve more robust boundary delineation, with accuracy differences around 12%, whereas the remaining datasets exhibit greater discrepancies, fluctuating around 20%.

5. Discussion

Inter-provincial differences in area, spatial consistency, and edge accuracy (Figure 4, Figure 6 and Figure 9) originate from systematic divergence along the entire processing chain; sensor choice, tidal flat boundary constraints, spectral indices, and classification algorithms jointly determine spatial fidelity and agreement. The following sections dissect these factors and propose actionable improvements.

5.1. Sensor-Specific Impacts on Tidal Flat Extraction

Landsat-8/9 and Sentinel-2 A/B are the most widely used optical sensors for large-area tidal flat mapping. They differ markedly in spatial resolution (30 m vs. 10 m), revisit interval (16 d vs. 5 d; 3–5 d when Sentinel-2 A/B are combined), and spectral band configuration. The shorter revisit interval of Sentinel-2 significantly increases the probability of acquiring cloud-free images at low tide. The impact of sensor revisit frequency is pronounced in regions like the Nanliujiang Estuary, Guangxi, which experiences persistent cloud cover. For 2020, 87 Sentinel-2 scenes met a uniform cloud cover threshold (≤60%), 2.6 times the 34 available Landsat scenes (Figure 10). Synchronizing these acquisitions with tide gauge records shows that Sentinel-2 captured a minimum tide height of 92 cm, compared with 105 cm for Landsat. Figure 11 demonstrates that Sentinel-2-based products (CTF, MTWM-TP, TFMC) fully delineate the exposed tidal flat extent, whereas Landsat-based products (DCTF, GTF30, GWL_FCS30) systematically omit areas. Quantitative assessment reveals a 20–60% larger tidal flat area extracted from Sentinel-2 data compared to Landsat in the Nanliu River estuary. Landsat missed approximately 60% of low-tide exposure windows in Beihai, resulting in systematic underestimation of tidal flat areas. Similarly, in areas with complex, mixed tidal regimes, such as the irregular semidiurnal tides along parts of the Guangdong coast, the ability to sample the tidal cycle more frequently increases the likelihood of capturing the true tidal minimum, reducing the underestimation bias inherent in sparser time series [18,22]. Nevertheless, Sentinel-2’s fixed overpass (around 10:30 in Beijing time) and cloud persistence still hinder annual minimum-tide coverage. Sub-meter imagery (e.g., GF-2, WorldView) with tasking capability offers a complementary source. A multi-sensor fusion strategy is therefore recommended to enhance both temporal completeness and spatial detail.

5.2. Suppression of Inland Interference Through Tidal Flat Boundary Constraints

The spatiotemporal extent of tidal flats is strictly governed by cyclic tidal inundation; therefore, inland water bodies (aquaculture ponds, salt pans, and lakes) must be excluded during mapping to prevent systematic commission errors. TFMC, MTWM-TP, and CTF achieve this by deriving an annual maximum high-water shoreline from Sentinel-2 imagery and applying it as a hard boundary mask, effectively eliminating similar spectra. In contrast, GWL_FCS30 and GTF30 omit this step and exhibit misclassification of aquaculture ponds as tidal flats in the Liaohe River Estuary (Figure 12).

Quantitative assessment shows that incorporating the maximum high-water shoreline mask reduces inland commission errors by 35–60% in the Liaohe Estuary. Conversely, unconstrained approaches (e.g., GTF30) yield a producer’s accuracy below 0.40 for aquaculture ponds, directly explaining their overestimated extents and low edge-validation accuracies. Given the widespread distribution of coastal aquaculture ponds in China [32], integrating a maximum high-water shoreline mask is an essential step toward accurate, large-scale tidal flat mapping.

5.3. Local Adaptability of Spectral Indices

A single water index cannot address the full range of tidal flat remote sensing requirements across complex coastal environments. In estuaries with high sediment loads and turbid waters (e.g., Oujiang Estuary and Hangzhou Bay), NDWI and NDVI frequently misclassify sediment-laden waters as tidal flats. Incorporating the shortwave infrared (SWIR) band to construct mNDWI significantly reduces these false positives [33,34,35] (Figure 13 top). Consequently, datasets (CTF, DCTF, and GTF30) that rely on near-infrared indices exhibit extensive misclassification of turbid water in the Oujiang Estuary, whereas products integrating SWIR information (MWTM-TP and TFMC) provide more accurate and consistent classifications (Figure 14). Similar errors occur in other turbid estuaries along the Zhejiang coast (Hangzhou Bay, Feiyun River Estuary, Aojiang River Estuary), leading to systematic overestimation of tidal flat area and reduced spatial consistency for CTF, DCTF, and GTF30. Conversely, in gently sloping, high-moisture flats such as Laizhou Bay, mNDWI underestimates the tidal flat extent due to strong SWIR absorption, whereas NDWI achieves more accurate land–water separation (Figure 13, bottom).

Quantitative evaluation demonstrates a 35% reduction in mNDWI false positives within the turbid Oujiang Estuary, while NDWI supplementation accounted for 55% of the intertidal zone area undetected by mNDWI in the gentle region of Laizhou Bay. In summary, we recommend dynamically selecting or combining indices according to ambient turbidity and surface moisture levels to establish a scenario-adaptive spectral-index framework, thereby enhancing the robustness and transferability of tidal flat products.

In addition to NDWI and mNDWI, other indices such as the Automated Water Extraction Index (AWEI) [36] and the Land Surface Water Index (LSWI) [37] may also contribute to tidal flat detection, particularly in areas with shadow effects or complex moisture gradients. Future studies will further examine their applicability in multi-index or adaptive frameworks to enhance mapping robustness.

5.4. Robustness of Classification Approaches

Otsu’s method [38] relies on a bimodal histogram assumption [39]; however, mixed pixels or class imbalance at tidal flat margins often violate this premise and introduce threshold bias [40]. Hangzhou Bay exhibits a well-defined bimodal histogram, enabling Otsu’s thresholds to achieve accurate land–water separation (Figure 15a). Conversely, in Caofeidian (Hebei), the mNDWI histogram is unimodal, so the automatically derived threshold fragments the shoreline and leading to erroneous segmentation of aquaculture ponds (Figure 15b). Quantitative comparison indicates that Otsu-based thresholding performs well in regions with clear spectral separation (e.g., sandy flats and well-defined tidal channels), but its accuracy decreases in heterogeneous environments dominated by muddy surfaces or aquaculture ponds.

MTWM-TP’s random forest model maintains high overall accuracy (OA = 0.85) along complex and irregular coastlines, as it integrates multiple spectral and textural features. However, it requires extensive training data and greater computational cost. Although the two approaches yield statistically similar results across the full domain, their applicability is complementary; thresholding is computationally efficient where histograms are well-behaved, while machine learning models are more robust to mixed pixels and class imbalance at the cost of additional training samples.

Consequently, the choice of classification algorithm should be guided by the predominant geomorphological conditions of the study area. A hybrid strategy is recommended: employing Otsu’s method for rapid and effective mapping in regions with simple, bimodal spectral characteristics, while reserving resource-intensive machine learning approaches for complex, heterogeneous tidal systems where their robustness justifies the computational cost.

5.5. Recommendations

5.5.1. Methodological Recommendations

Future tidal flat mapping could adopt a “multi-source—multi-index—multi-algorithm” framework that integrates the following components:

1. Multi-source synergy: Integrate the high temporal density of Sentinel-2 with Landsat’s long-term archive, complemented by high-resolution optical (e.g., GF-2, WorldView) and radar (e.g., Sentinel-1) observations to maximize spatial detail, temporal coverage, and all-weather capability.

2. Adaptive spectral indices: Dynamically select NDWI or mNDWI according to locally derived turbidity thresholds.

3. Hybrid classification: Employ rule-based thresholds for clear land–water interfaces and machine learning (RF/CNN) classifiers for hybrid pixel zones, while constraining the domain with an annual maximum high-water shoreline.

5.5.2. Data Recommendations

Based on the results, we offer the following data-specific recommendations to guide future application selection. For regions with frequent cloud cover and complex tidal conditions, Sentinel-2-based datasets (CTF, MTWM-TP, TFMC) are recommended due to their higher temporal resolution and stronger capability to capture low-tide exposures; in contrast, for broad-scale assessments requiring long-term continuity or data integration with historical studies, Landsat-based products (GTF30, GWL_FCS30, DCTF) remain valuable for ensuring temporal consistency and cross-decadal comparisons. In practice, MTWM-TP demonstrates strong adaptability and stability across diverse coastal environments, making it suitable for general nationwide mapping, while TFMC performs best in regions with high cloud frequency and dynamic tidal conditions. GWL_FCS30 and GTF30 are appropriate for users prioritizing global data consistency and historical comparability. These dataset-specific insights can help guide future coastal wetland research and management toward more targeted and effective data use.

5.6. Limitations and Future Perspectives

Although the present study employed 3150 validation samples covering 21 tide-gauge stations across China’s coastal provinces and representing the full range of tidal regimes (regular and irregular semidiurnal and diurnal tides), certain limitations should still be acknowledged. Despite the comprehensive spatial coverage, the representativeness of some specific geomorphic and sedimentary types (such as sandy and muddy tidal flats) may remain limited due to the natural heterogeneity of coastal morphology and hydrodynamics. Local factors such as sediment composition, slope, and vegetation cover can influence spectral responses, potentially leading to regional variations in classification accuracy. Furthermore, owing to the retrospective nature of this study, in situ measurements from 2020 were not available for direct validation. Consequently, accuracy assessment primarily relied on low-tide satellite imagery. Future work will integrate field surveys, UAV verification, and synchronous satellite data acquisition to further reduce classification uncertainty and enhance the robustness of validation.

The present analysis assumes the short-term temporal stability of tidal flat extents, yet climate change and human activities are continuously reshaping the intertidal zone. These dynamic processes may lead to rapid shoreline shifts and sediment redistribution, thereby influencing the distribution of tidal flats. Future mapping efforts could therefore consider a dynamic updating strategy, ideally on an annual or semi-annual basis, to better capture the evolving morphology of China’s tidal flats and to support long-term coastal monitoring under changing environmental conditions.

6. Conclusions

Tidal flats are critical ecotones between terrestrial and marine systems, underpinning biodiversity, carbon sequestration, and coastal protection. Accurate mapping of these habitats is therefore essential for resource management, ecological conservation and global change research. We conducted a systematic inter-comparison of six 2020 national-scale tidal flat datasets (GTF30, GWL_FCS30, MTWM-TP, DCTF, CTF, and TFMC) using three independent metrics: area discrepancy, spatial consistency, and edge-based classification accuracy.

1. Substantial discrepancies exist among datasets in tidal flat area, spatial patterns, and boundary delineation. While broad agreement is achieved over typical mudflats (e.g., the Jiangsu shoals), pronounced divergence emerges along complex coastlines and turbid estuaries (e.g., Oujiang Estuary), indicating large spatial uncertainty.

2. Performance is dataset-specific, with systematic underestimation (DCTF) or overestimation (GTF30). Edge validation reveals that MTWM-TP (OA = 0.85) and TFMC (OA = 0.84) achieve the highest classification accuracy, while CTF (OA = 0.78) is moderate and DCTF (OA = 0.66), GWL_FCS30 (OA = 0.68), and GTF30 (OA = 0.71) suffer from severe misclassification.

3. Cross-product differences stem primarily from four technical dimensions: (1) spatiotemporal resolution of imagery, (2) boundary control strategy, (3) local adaptability of water indices, and (4) robustness of classification algorithms. Relying on a single sensor, index, or segmentation approach is insufficient for China’s diverse coastal environments.

Future improvements could undertake the following: (1) develop an automated multi-source data framework that integrates optical (Sentinel-2, Landsat), radar (Sentinel-1), UAV, and in situ sensor observations to enhance spatial detail and low-tide coverage while maintaining automation; (2) develop dynamic combinations of spectral indices tailored to regional turbidity and geomorphology; and (3) refine thresholding algorithms by integrating image features and prior knowledge for enhanced boundary delineation. These findings provide a scientific basis for dataset selection and algorithm optimization, supporting precise coastal ecosystem management and sustainable development.