Detailed Classification of Vegetation and Assessment of Carbon Stock in the Liaohe Estuary Wetlands Based on Sentinel-2 Imagery

Abstract

Most remote sensing extraction studies utilizing vegetation indices typically classify and extract land cover features based on the phenological characteristics of the study area or rely on a single vegetation index. When attempting to extract multiple land cover types simultaneously, classification accuracy often declines significantly because a single vegetation index is unsuitable for all features. While some recent studies employ deep learning and neural networks for classification and extraction, their complex mechanisms and “black-box effect” hinder clear explanations for accuracy outcomes. In response to the issues outlined above, this paper proposes a simpler and more intuitive method for the hierarchical extraction of typical land cover features. This approach analyzes the difficulty of separating these features based on spectral reflectance data to determine the following extraction order: first water bodies, followed by reed, then Suaeda salsa, and finally tidal flat. Furthermore, by selecting appropriate parameters and substituting vegetation indices for bands that perform better, high extraction accuracy is achieved. The classification and interpretation results were validated using a combination of field survey data and Google imagery, together with a validation sample. Accuracy assessments using overall accuracy and Kappa coefficient demonstrate the following optimal results for the hierarchical approach: NDWI for water, S2REP for reeds, and MSAVI for Suaeda salsa. Overall accuracy reached 98.5% with a Kappa coefficient of 0.9796, validating the effectiveness of this spectral-feature-based hierarchical extraction method using diverse vegetation indices. Using a hierarchical extraction approach to classify typical land cover features in the study area from 2020 to 2025, accuracy rates exceeded 98% in all cases. Based on these classification results, the INVEST model was employed to simulate carbon stock trends in the Liaohe Estuary region over the past five years. The study found that, although factors such as tides and the date of image acquisition had a certain impact on the study area compared with the problems caused by historical development, the ecological environment in the study area is gradually stabilizing at the present stage.

1. Introduction

The Liaohe Estuary, as a significant estuarine wetland ecosystem in China, holds a vital position in the global ecosystem due to its unique geographical location and rich biodiversity [1,2]. This region features extensive tidal flats and diverse vegetation types, such as Suaeda salsa and common reed (Phragmites australis). These plant communities perform irreplaceable ecological services in maintaining biodiversity, providing vital habitats, reducing stormwater runoff, regulating climate, and sequestering carbon [3]. They play a crucial role in the productivity of wetland ecosystems and serve as primary carbon sinks. In recent years, the ecological environment of the Liaohe Estuary has faced severe challenges due to dual impacts from natural factors (such as climate change and sea level rise) and human activities (such as land reclamation and water resource utilization), leading to significant changes in vegetation community structure and spatial distribution [4]. Therefore, precise and efficient monitoring of typical vegetation in the Liaohe Estuary is crucial for assessing regional ecological health and formulating effective conservation strategies.

The distribution of vegetation types, changes in area coverage, and health status at the Liaohe Estuary serve as core indicators for evaluating wetland functions, succession stages, and ecological restoration outcomes. Remote sensing technology, with its advantages of macro-scale, rapid, and periodic observation, has become a powerful tool for wetland vegetation monitoring. Multispectral satellite imagery, in particular, captures the reflectance characteristics of land features across different wavelength ranges, enabling superior identification and classification of diverse vegetation types and growth conditions. Hyperspectral remote sensing not only images spatial features of targets but also provides continuous spectral coverage across dozens of narrow bands within a broader wavelength range for each pixel. This enables deeper investigation into the spectral response mechanisms and physical processes of vegetation, making it a powerful tool for vegetation monitoring and related fields.

Detailed vegetation classification using remote sensing in Hekou District is a key prerequisite for the high-precision assessment of wetland carbon sequestration capacity and carbon stocks; by accurately distinguishing between different vegetation types and their distribution, the accuracy of carbon sequestration models can be significantly improved. Mainstream methods for fine-scale remote sensing vegetation interpretation in coastal wetlands can be categorized into two types. The first type involves classification models based on vegetation indices. Vegetation indices (VIs) simplify complex remote sensing data into quantifiable indicators of land cover health, coverage, and productivity by combining reflectance values across two or more wavelength bands. Existing VIs are widely applied for large-scale vegetation information processing. For instance, the normalized difference vegetation index (NDVI) has seen extensive use in green vegetation studies. As vegetation index research deepens, scientists have developed a series of improvements to mitigate the influence of soil, topography, and atmospheric factors in study areas, thereby expanding the application scope of these indices. For instance, Wu (2014) proposed the Generalized Difference Vegetation Index (GDVI) based on NDVI formulas, enhancing the sensitivity of leaf area index and vegetation cover retrieval in arid regions with low vegetation coverage [5]. To address these limitations, researchers developed vegetation indices tailored to specific vegetation types. For instance, Ye et al. [6] constructed the estuarine tidal flat vegetation index (ETFVI) for tidal flat ecosystems by integrating humidity indices after accounting for the impact of surface moisture variations on vegetation indices [7]. combined the normalized difference vegetation index (NDVI) with monthly average precipitation—two common drought-impact variables—to construct a novel integrated drought index (Meteorology-Agriculture Drought Index, MADI) that synthesizes precipitation and remote sensing vegetation characteristics. This index demonstrated high accuracy in assessing drought conditions in the Yangtze River Basin. Despite these advances, current research faces several challenges. For instance, some methods for detailed vegetation classification and extraction consider only a limited range of land cover types. Additionally, certain studies rely excessively on multi-temporal imagery, and vegetation indices often require computation from high-precision multispectral or hyperspectral reflectance data, imposing stringent demands on both the quality and quantity of remote sensing imagery. Finally, many fine-scale vegetation indices are often constructed for specific study areas, and their robustness remains unclear. Further validation is needed to assess their effectiveness in different environments.

Machine learning algorithms such as support vector machines, random forests, and deep learning have enabled high-precision land cover classification in local areas. For example, Tang et al. [8] improved rapeseed mapping accuracy by 0.23 using vegetation indices and texture features; Zhang et al. [9] simulated maize yield with LSTM; Fan et al. [10] extracted burned vegetation anomalies using spectral-spatial features; Zheng et al. [11] coupled multi-source imagery to achieve 81.5% overall accuracy for salt marsh vegetation; Li et al. [12] integrated three vegetation indices and three machine learning techniques to optimize winter wheat yield estimation; Ren et al. [13] enhanced hyperspectral interpretation using SVM and RF; and Gu et al. [14] combined Sentinel-1/2, SRTM, and the InVEST model on GEE to assess high-altitude wetlands. However, machine learning models often suffer from poor robustness due to sensitivity to sample quality, sensor parameters, and local environmental conditions.

Moreover, in complex vegetation areas like the Liaohe Estuary (e.g., red Suaeda salsa), spectral characteristics differ significantly from typical green vegetation. Traditional vegetation indices, often constructed based on green vegetation reflectance features, perform poorly in remote sensing extraction of such red vegetation [15]. Wetlands with unique ecological environments (e.g., high-salinity soils, tidal effects) face interference from multiple confounding factors. Different vegetation types exhibit distinct growth stages, varying health conditions, and factors like water and nutrient stress, all of which can lead to pronounced differences in spectral characteristics. Current vegetation index-based land cover studies often rely on a single index for classifying and extracting all features, revealing clear limitations in vegetation information extraction through a single index alone. In complex vegetation areas like the Liaohe Estuary, a single vegetation index often fails to capture all land cover types. For instance, Suaeda salsa exhibits vivid red coloration during specific growth seasons. To enhance extraction accuracy for such vegetation, we must develop or optimize more targeted vegetation index formulas. Although some current studies employ machine learning and deep learning algorithms to aid feature classification and recognition, their complex mechanisms and “black-box effect” hinder clear explanations for accuracy improvements. Furthermore, questions arise regarding the applicability of vegetation indices across different remote sensing satellites—specifically, whether formulas developed for one satellite sensor can be directly applied to data from other sensors [16].

In response to the current difficulties and challenges in remote sensing feature extraction and classification, this study proposes a hierarchical extraction method tailored to complex study areas. The optimal sequence for layer extraction was determined using spectral reflectance data, with priority given to extracting water bodies, followed by reed, then Suaeda salsa, and finally tidal flat [17]. Furthermore, appropriate vegetation indices were selected to achieve better extraction results [18]. At the same time, to improve discrimination, the b8 band used in the NDVI calculation was replaced with b8a [19], and the b7 band in the S2REP was replaced with b8a. With carbon sink assessment as its core application, the study directly supports the InVEST model’s quantitative assessment of carbon stocks in the Liaohe Estuary wetlands by generating high-precision vegetation classification results. Based on preprocessed remote sensing imagery, it extracts spectral information of typical features in the study area and matches the most suitable vegetation indices and calculation bands to the spectral characteristics of each feature type, thereby specifically improving the recognition accuracy for each feature [20]. The core objective of this study is to investigate whether this approach can yield higher-precision classification products for typical wetland features. Additionally, we conducted comparative analyses of different extraction schemes. Specifically, we selected the Liaohe Estuary Delta region in China as the study area and analyzed four typical features (reed, Suaeda salsa, water, and tidal flat) within the study area. Multiple vegetation indices were selected and optimized to achieve hierarchical extraction of these four typical land cover types. To demonstrate effectiveness, commonly used vegetation indices were selected as control variables for comparison with the accuracy and performance of the extraction classification in this study. Having demonstrated that the hierarchical extraction approach can effectively improve the accuracy of land cover classification, we employed this method to classify land cover in a single scene for each year from 2020 to 2025—years with similar environmental conditions—by adjusting dynamic thresholds. Furthermore, based on the land cover classification results, we utilized the INVEST model to assess carbon stocks in the study area. Specific research content includes: (1) spectral reflectance data analysis of typical land cover types in the Liaohe Estuary based on Sentinel-2 satellite remote sensing imagery; (2) selection and matching of vegetation indices for different land cover types based on reflectance data to achieve hierarchical extraction of typical land cover features; (3) performing a hierarchical extraction of Sentinel-2 remote sensing imagery from 2020 to 2025, followed by an accuracy assessment; and (4) using the land cover classification results combined with the InVEST model to assess the carbon stock in the Liaohe Estuary region. The innovation of this study is the application of fine-scale classification based on spectral reflectance curve characteristics to carbon stock modeling.

2. Materials and Methods

2.1. Study Area

This study selected the Liaohe Estuary in China as its study area (Figure 1). Located within Panjin City, Liaoning Province, it is the northernmost coastal estuary in China. The region features a typical temperate semi-humid monsoon climate with a frost-free period of 170–200 days. The annual average temperature is approximately 8.5 °C. January is the coldest month, with an average temperature of −10.2 °C, and July is the hottest month with an average temperature of 23.4 °C. Annual precipitation is approximately 650 mm, predominantly concentrated during summer. The average annual evaporation is 1669.60 mm [21]. The Liaohe River’s annual net flow is 2.75 × 10⁹ m³/a, with an average annual sediment transport of 8.89 × 10⁶ t/a. The Liaohe Estuary exhibits a funnel-shaped configuration, characterized by vast reed marshes, tidal flats, and meadows. The terrain is low-lying with well-developed tidal channels. The study area boasts diverse wetland types and varied landscapes. Its core zone, the Liaoning Liaohe Estuary National Nature Reserve, hosts the world’s largest reed marsh wetland. It serves as a critical habitat for numerous rare and endangered bird species, including the red-crowned crane and Saunders’s gull, and functions as a vital stopover and destination along the East Asian–Australasian Flyway, establishing itself as an internationally renowned biodiversity hotspot. The study area exhibits diverse vegetation types, with 126 recorded herbaceous plant species. Reed (Phragmites australis) and Suaeda salsa dominate the vegetation composition. Extensive reed communities form a vast “reed sea,” constituting the wetland’s foundational landscape. The most iconic vegetation is the Suaeda salsa salt marsh community. Each autumn, these plants transform into a vivid purplish-red hue, stretching across the estuarine tidal flats to create the “Red Beach” landscape, hailed as a “natural wonder of the world.”

2.2. Data Sources and Preprocessing

The European Space Agency’s (ESA) Sentinel-2 satellite, as a key component of the Copernicus program, provides Earth observation data featuring high spatial resolution (10 m, 20 m, and 60 m), multispectral coverage (13 bands), and high temporal resolution (5-day revisit cycle). Sentinel-2’s rich spectral information, particularly its three bands in the red edge region (B5, B6, and B7), provides a research foundation for distinguishing different vegetation types with high precision and monitoring vegetation physiological status. These bands are especially sensitive to capturing changes in vegetation chlorophyll content, internal structure, and moisture conditions. Therefore, the Sentinel-2 remote sensing satellite holds broad application prospects in the detailed classification and health assessment of wetland vegetation.

The vegetation indices used in calculations require selection based on high-precision multispectral or hyperspectral reflectance data. Radiant brightness data without atmospheric correction or dimensionless DN values are unsuitable for vegetation index calculations. Data selection for the study area is restricted to the period from June to September over three consecutive years (2023–2025), with cloud cover below 20%, totaling 103 images. After comparing vegetation data from field surveys of typical areas in the Liaohe River Delta, integrating historical Google Earth imagery, and conducting visual interpretation assessments, one image from 13 September 2023 was ultimately selected for the study area. This image underwent preprocessing, including band resampling, using SNAP 12.0.0software. Pure pixels were identified using a combination of high-resolution UAV ortho mosaics (5 cm resolution) and field survey points; for each class, 500 candidate pixels were initially sampled; then, we calculated the mean, and standard deviation of the reflectance for each spectral band was reported for each band to demonstrate spectral homogeneity within classes. Pure pixels were also selected based on over 100 satellite images and field survey data from the past five years, as well as historical Google Earth imagery. The pixels chosen were those from areas where target features had not undergone significant changes in the past five years and where typical features were concentrated. Following the extraction of these pixels, the expected values of the probability density functions were calculated for each typical feature. Sentinel-2A imagery provides 13 bands, with B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12 comprehensively covering key spectral bands for vegetation reflection/absorption across the visible-near-infrared-shortwave infrared range. These 10 bands achieve a spatial resolution of 10 m either natively or through resampling, meeting the accuracy requirements for remote sensing imagery in this study. This study focuses on spectral analysis of surface vegetation. The discarded B1, B9, and B10 bands primarily serve atmospheric, water vapor, and cirrus cloud corrections. With low signal-to-noise ratios, they contribute minimally to surface vegetation signals. Retaining these bands would not provide additional useful information but would increase variable collinearity and computational load [22,23,24,25].

To facilitate subsequent studies and statistics on typical vegetation in the study area, reflectance samples for four representative land cover types—reed, saltwort, bare soil patches, and water bodies—were extracted using ENVI Classic 5.6. The selection criteria involved choosing areas within the study region where the distribution of these four land cover types showed no significant changes in imagery from the past three years, with the total number of pixels per land cover sample controlled at 500. At the same time, a validation set independent of the training set was selected based on field surveys, drone data and manual visual interpretation, in order to subsequently assess classification accuracy. By calculating the expected value of the probability density function for each sample of each land cover type, the reflectance values for the four typical land covers (reed, saltwort, bare ground, and water body) were obtained across 10 Sentinel-2A bands (B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12). The results are presented in

Table 1. Reflectance of the typical objects in the study area, estimated from the Sentinel-2 images.

Band	Suaeda salsa	Water	Tidal Flat	Reed
B2	0.1288	0.1358	0.1728	0.1323
B3	0.1329	0.1264	0.1905	0.1566
B4	0.1577	0.1302	0.2025	0.1453
B5	0.2264	0.1275	0.2225	0.2050
B6	0.2753	0.1072	0.2264	0.3664
B7	0.2904	0.1069	0.2356	0.4292
B8	0.2989	0.1055	0.2384	0.4668
B8A	0.3027	0.1043	0.2407	0.4831
B11	0.1829	0.1030	0.2359	0.2854
B12	0.1497	0.1029	0.1998	0.1906

. The expected value formula is

$$E\left(x\right)={\int }_a^{b}P\left(x\right)\cdot x\mathrm{d}x$$

(1)

where E(x) denotes the expected value; P(x) represents the reflectance probability density; x indicates the reflectance value; and a and b denote the lower and upper limits of the spectral band, respectively.

The research approach for the entire paper is illustrated in Figure 2.

3. Analysis of Land Surface Spectral Characteristics and Selection of Vegetation Indices

3.1. Analysis of Spectral Reflectance Curve Characteristics of Typical Land Cover Types in the Study

This study visualizes and analyzes multi-band spectral reflectance data from the Sentinel-2 remote sensing satellite for the following four typical land cover types: reed, Suaeda salsa, tidal flat, and water (Table 1). Figure 3 displays the spectral reflectance curves of these four land cover types across Sentinel-2 bands B2 (blue) to B12 (shortwave infrared).

From the perspective of reflectance values, across visible light bands such as B2 (blue), B3 (green), and B4 (red), the reflectance of various land cover types is generally low. Reed exhibits its highest reflectance in the red edge to near-infrared bands, peaking in the near-infrared region and significantly exceeding other land cover types. Water shows the lowest reflectance across all bands, with a continuous decline from the visible to near-infrared and shortwave infrared bands. Tidal flats demonstrate the highest reflectance within the blue to red edge 1 band range.

The spectral characteristics of reed exhibit low reflectance in the blue-to-red wavelength range, with relatively low reflectance in the visible bands (B2, B3, and B4), particularly showing a distinct absorption valley in the red band (B4). However, reflectance sharply increases between the red band (B4) and near-infrared bands (B5 to B8A), rising from 15% to 48% and exhibiting a pronounced “red-edge effect.” In the shortwave infrared bands (B11, B12), reed reflectance decreases significantly but remains higher than that of water and tidal flats.

The spectral curve of Suaeda salsa shares some morphological similarities with that of reed. The reflectance of Suaeda salsa adheres to the following “three-stage pattern”: low-level rise—gentle peak—rapid decline. Suaeda salsa exhibits a gradually increasing trend in the visible bands (B2, B3, and B4), differing from the typical characteristics of many green vegetation types that strongly absorb in the red band (B4) and strongly reflect in the green band (B3) [26]. This phenomenon may be related to the accumulation of red pigments such as anthocyanins in Suaeda salsa leaves during specific growth stages, resulting in reflectance characteristics distinct from typical green vegetation in the visible bands. In the red edge bands (B5, B6, and B7) and near-infrared bands (B8, B8A), the reflectance of Suaeda salsa rises gradually to a gentle peak, indicating its typical vegetation characteristic of high near-infrared reflection. This reflects the strong scattering of near-infrared light by the internal cellular structures within Suaeda salsa leaves. In the shortwave infrared bands (B11, B12), reflectance drops sharply. The reflectance curves also show that the slope characteristics of Suaeda salsa at SWIR1 and SWIR2 are essentially parallel to those of tidal flats.

Water exhibits extremely low reflectance across all Sentinel-2 bands, reflecting the typical spectral characteristics of strong electromagnetic wave absorption by water. The spectral curve of tidal flats shows a continuous upward trend in the visible bands (B2, B3, and B4) and reaches higher reflectance in the red edge and near-infrared bands (B5–B8A). Subsequently, reflectance decreases in the shortwave infrared bands (B11, B12) but remains higher than that of water [27].

3.2. Selection and Optimization of Vegetation Indices for Typical Landforms in the Study Area

Vegetation indices serve as effective parameters in remote sensing for quantifying surface vegetation coverage and growth conditions [28]. Essentially, they mathematically combine reflectance values across different spectral bands to enhance vegetation information while suppressing non-vegetation components (such as soil background and atmospheric effects). Different vegetation indices employ distinct participating bands and calculation formulas. Selecting suitable band indices based on actual reflectance conditions can enhance the differentiation between various land cover types and, to some extent, mitigate the “mixed pixel” issue that arises when using a single index to extract all land cover types. Based on the spectral reflectance characteristics of typical land cover types in the study area analyzed earlier, commonly used vegetation indices most suitable for the region were selected as shown in

Table 2. Common vegetation indices for Sentinel-2A.

Vegetation Index	Formula	Reference
NDVI	(B8 − B4)/(B8 + B4)	[29]
EVI	2.5 ∗ ((B8 − B4)/(B8 + 6 ∗ B4 − 7.5 ∗ B2 + 1))	[30]
SAVI	((B8 − B4)/(B8 + B4 + 0.5)) ∗ 1.5	[31]
MSAVI	${\displaystyle \frac{2\cdot B8+1-\sqrt{{\left(2\cdot B8+1\right)}^2-8\cdot \left(B8-B4\right)}}{2}}$	[32]
GNDVI	(B8 − B3)/(B8 + B3)	[33]
NDRE	(B8 − B5)/(B8 + B5)	[34]
S2REP	705 + 35 ∗ (((B7 + B4)/2) − B5)/(B6 − B5)	[35]
CIg	(B8/B3) − 1	[36]
CIre	(B8A/B5) − 1	[36]
IRECI	(B7 − B4)/(B5/B6)	[36]
NDWI	(B3 − B8)/(B3 + B8)	[37]
NDMI	(B8A − B11)/(B8A + B11)	[38]
NBR	(B8 − B12)/(B8 + B12)	[39]
NDTI	(B11 − B12)/(B11 + B12)	[40]

. This enables effective identification of Suaeda salsa, reed, tidal flat, and water. Employ distinct participating bands and calculation formulas. Selecting suitable band indices based on actual reflectance conditions can enhance the differentiation between various land cover types and, to a certain extent, mitigate the “mixed pixel” issue that arises when using a single index to extract all land cover types. Based on the spectral reflectance characteristics of typical land cover types in the study area analyzed earlier, commonly used vegetation indices most suitable for the region were selected as shown in Table 2. This enables effective identification of Suaeda salsa, reed, tidal flat, and water.

As revealed by spectral feature analysis in Section 3.1, the spectral characteristics of Suaeda salsa in the study area are suitable for the normalized difference vegetation index (NDVI).

$$NDVI={\displaystyle \frac{NNIR-RED}{NNIR+RED}}$$

(2)

${\mathsf{\rho }}_{B8A}-{\mathsf{\rho }}_{B4}>0$; $K_{SS}={\displaystyle \frac{{\mathsf{\rho }}_{B8A,SS}-{\mathsf{\rho }}_{B4,SS}}{{\mathsf{\rho }}_{B4,SS}}}>0$; the same applies to reed and Suaeda salsa. Based on the comprehensive spectral data, it can be concluded that

$$\begin{gathered} K_{\mathrm{Reed}}>K_{\mathrm{SS}} \\ {\displaystyle \frac{{\mathsf{\rho }}_{B8A,\mathrm{Reed}}-{\mathsf{\rho }}_{B4,\mathrm{Reed}}}{{\mathsf{\rho }}_{B4,\mathrm{Reed}}}}>{\displaystyle \frac{{\mathsf{\rho }}_{B8A,\mathrm{SS}}-{\mathsf{\rho }}_{B4,\mathrm{SS}}}{{\mathsf{\rho }}_{B4,\mathrm{SS}}}} \end{gathered}$$

Consequently, a greater near-infrared bandgap is formed between B8A and the red light band, which enhances the ability to distinguish between reed and Suaeda salsa.

Our measurements show that K_Reed > K_SS. This larger NIR jump for reed enhances its spectral separability from Suaeda salsa when the red-edge band (B8A) is used instead of the traditional NIR band (B8). The same principle applies to other band replacements (e.g., replacing red-edge band (B8A) with B7 for S2REP).

To achieve more intuitive classification results, the B8A band (Vegetation red edge) from Sentinel-2 was selected to replace the B8 band (NIR) and the B4 band (red) in the calculation. NDVI performs well in identifying green vegetation, but its effectiveness may be limited for non-green vegetation (such as red Suaeda salsa) or in complex wetland environments. When Suaeda salsa appears red, its red reflectance may be high, leading to low NDVI values and consequently affecting identification accuracy. Additionally, Suaeda salsa and tidal flats exhibit overlapping or similar trends in certain bands. To enhance their distinguishability, the Modified Soil-Adjusted Vegetation Index (MSAVI) can be employed. This index aims to reduce the influence of soil background and atmospheric effects on vegetation indices, yielding superior classification results compared to the SAVI vegetation index.

$$MSAVI={\displaystyle \frac{1}{2}}\left(2\hspace{0.17em}NIR+1-\sqrt{{\left(2\hspace{0.17em}NIR+1\right)}^2-8\hspace{0.17em}\left(NIR-RED\right)}\right)$$

(3)

Reeds exhibit a highly pronounced red-edge effect. Combining this with the Sentinel-2 remote sensing imagery used in this study, we selected the Sentinel-2 modified red-edge position index (S2REP). In band operations involving Sentinel-2, B4 (RED), B5 (Vegetation red edge, red edge 1), B6 (Vegetation red edge, red edge 2), and B7 (Vegetation red edge, red edge 3) are typically employed. To achieve better results in this study area, B7 was replaced with B8A (NIR).

$$S2REP=705+35\times {\displaystyle \frac{\frac{NIR+R}{2}-RE1}{RE2-RE1}}$$

(4)

The water body selection employs the normalized difference water index (NDWI).

$$NDWI={\displaystyle \frac{GREEN-NIR}{GREEN+NIR}}$$

(5)

4. Stratified Extraction and Accuracy Analysis of Vegetation Indices for Typical Land Features in the Study Area

4.1. Development of a Multi-Vegetation Index Stratification Scheme Based on Sentinel-2 Bands

Based on the above analysis, observation of Sentinel-2 remote sensing imagery reveals the distribution of four typical land cover types in the study area on 13 September. By examining reflectance values and spectral curve characteristics, we analyze the differences among the following four land cover types: reed, Suaeda salsa, tidal flat, and water. This assessment evaluates the difficulty of extracting each land cover type. Water exhibits the greatest distinction from the other three land cover types and should be prioritized for extraction in subsequent steps to facilitate high-precision differentiation of Suaeda salsa and other features. Reed displays a pronounced “red edge effect,” warranting its extraction after water. Suaeda salsa distribution appears in patches smaller than tidal flats; thus, extraction indices suitable for Suaeda salsa should be selected for its identification. After all three features are extracted, the remaining area constitutes tidal flat. Thus, the extraction sequence for the study area is as follows: first, extract water from the full-scene imagery; second, extract reed from the water-extracted imagery; finally, extract tidal flat from the reed-extracted imagery. Since the study defines four land cover types, tidal flat is the residual area after these three extraction steps. The NDWI vegetation index is selected for water extraction, the S2REP index for reed extraction, and the MSAVI for Suaeda salsa extraction (Figure 4).

4.2. Extraction of Spatial Distribution of Typical Landforms in the Study Area

Based on the extraction scheme outlined in Section 4.1, and after multiple comparative trials using the reflectance curve data of land cover types from the preceding section, optimal thresholds were established for the three vegetation indices to achieve the best overall performance. For NDWI extraction of water, the maximum threshold was set at 0.25 and the minimum threshold at −0.086. For S2REP extraction of reed, the maximum threshold was set at 725 and the minimum threshold at 722. For MSAVI extraction of Suaeda salsa, the maximum threshold was set at 0.301 and the minimum threshold at 0.08. The extraction results are shown in the figure (Figure 5).

Comparing the left and right images in Figure 5 reveals the distribution of reed, Suaeda salsa, tidal flat, and water. Tidal flat occupies the largest proportion of the study area. Extraction results are better in dense reed and Suaeda salsa zones, with nearly all reed in the intertidal zone and estuary being successfully extracted. Based on the above conclusions, a precision assessment system was established by integrating UAV data, Sentinel-2 imagery, and field survey data (Figure 6 and

Table 3. Interpretation accuracy evaluation system.

Land Cover Classes	Sample	Sentinel-2 Image	UAV Imagery	Field Sampling
Reed
Suaeda salsa
Water
Tidal flats

).

4.3. Accuracy Assessment of Stratified Extraction for Vegetation Indices of Typical Land Features

To further evaluate the effectiveness of the multi-index hierarchical extraction scheme proposed in this study for identifying typical land cover types in the study area, other vegetation indices commonly used in remote sensing extraction and hierarchical extraction schemes composed of these indices were selected. Hierarchical extraction Scheme 1 was defined as follows: first, NDWI was used to extract water; then, S2REP was used to extract reed; finally, MSAVI was used to extract Suaeda salsa. Scheme 2 was defined as follows: first, NDVI was used to extract water; then, S2REP was used to extract reed; finally, SAVI was used to extract Suaeda salsa. Scheme 3 was defined as follows: first, NDWI was used to extract water; then, NDVI was used to extract both reed and Suaeda salsa. Typical land cover types in the study area were extracted using threshold extraction and control variable methods. The extraction accuracy of the three schemes was evaluated based on confusion matrices and Kappa coefficients, while the percentage of misclassified pixels was also calculated (Figure 7 and

Table 4. Confusion matrix and accuracy assessment.

Predicted/Ground Truth	Water	Tidal Flat	Reed	Suaeda salsa	Row Total	User Accuracy (%)
Water	235	0	0	0	235	100
Tidal flat	0	162	1	3	166	97.59
Reed	0	5	123	1	129	95.35
Suaeda salsa	0	0	0	138	138	100
Column total (pixels)	235	167	124	142	668
Producer accuracy (%)	100	97.01	99.19	97.18
Omission error (%)	0.00	2.99	0.81	2.82

and

Table 5. Comparison of the accuracy of typical feature extraction across three methods.

Scheme	Vis	Target	Max Threshold	Min Threshold	Overall Accuracy	Kappa
Scheme 1	NDWI	Water	0.25	−0.086	98.50%	0.98
	S2REP	Reed	725	722
	MSAVI	Suaeda salsa	0.301	0.08
Scheme 2	NDVI	Water	0	−∞	98.02%	0.97
	S2REP	Reed	725	722
	SAVI	Suaeda salsa	0.56	0.25
Scheme 3	NDWI	Water	0.25	−0.086	97.16%	0.96
	NDVI	Reed	0.586	0.429
	NDVI	Suaeda salsa	0.393	0.239

Note: to control variables, if identical vegetation indices extract the same land cover features, the threshold settings remain unchanged. Classification accuracy and Kappa coefficient are each rounded to two decimal places.

).

In practical experiments using the threshold method to extract land cover features, the key challenge lies in determining the optimal threshold values for each vegetation index. In actual implementation, threshold settings significantly impact the classification results of extracted features, requiring a balance between extraction accuracy and the number of misclassified pixels. Comparing the three schemes reveals that while all achieve high discrimination rates below the determined thresholds, Scheme 1 delivers the optimal extraction performance and accuracy among the three. Based on field survey data from the Liaohe Estuary, combined with UAV imagery and supplemented by manual visual interpretation, the final conclusion is that NDWI is best suited for water extraction, S2REP for reed extraction, and MSAVI for Suaeda salsa extraction, which yields the most effective overall extraction scheme.

Using this conclusion, we performed a layered extraction of Sentinel-2 remote sensing imagery from 2020 to the corresponding period in 2025. Except for setting the dynamic extraction threshold, all other operations remained consistent with those described earlier (Figure 8 and Figure 9).

Using this conclusion, we performed a layered extraction of Sentinel-2 remote sensing imagery from 2020 to the corresponding period in 2025 (

Table 6. Annual classification accuracy from 2020 to 2025.

Date	Scheme	Overall Accuracy	Kappa
20 September 2020	1	98.60%	0.98
	2	87.41%	0.82
	3	95.46%	0.93
23 September 2021	1	98.22%	0.98
	2	91.74%	0.89
	3	91.34%	0.88
18 September 2022	1	99.89%	0.99
	2	87.08%	0.81
	3	98.58%	0.98
16 June 2024	1	98.50%	0.98
	2	97.16%	0.96
	3	97.15%	0.96
29 September 2025	1	98.41%	0.98
	2	98.31%	0.98
	3	96.85%	0.96

).

Since it is objectively impossible for the selected images to have completely consistent spectral data, dynamic thresholding is employed while keeping the extraction scheme unchanged.

4.4. Based on the Previous Study on Carbon Stock

Based on Sentinel-2 satellite remote sensing imagery from the past five years that has undergone typical land cover classification, and utilizing carbon pool data obtained from relevant studies of the Liao River Estuary, carbon stocks in the study area were calculated using INVEST 3.18 software. This study used the “Carbon stock and Sequestration” module of the InVEST model to assess carbon sink potential. The model’s ecosystem carbon pools consist of the following four basic carbon pools: aboveground biomass carbon, below-ground biomass carbon, soil carbon, and dead organic carbon. Based on land use classifications, the total amounts of aboveground carbon density, below-ground carbon density, soil carbon density, and dead organic matter carbon density are calculated separately for each land-use class pixel. The carbon pool data used to calculate carbon stocks are sourced from the relevant literature [41] (

Table 7. Carbon-accounting data.

LULC_Name	C_Above (t C/ha)	C_Below (t C/ha)	C_Soil (t C/ha)	C_Dead (t C/ha)
Water	0	0	91	0
Reed	4.46	1.12	72.6	2.5
Suaeda salsa	2.43	0.61	112.4	1.5
Tidal flat	0	0	166.46	0.5

).

$$C_i=C_{i\text{-}\mathrm{above}}+C_{i\text{-}\mathrm{below}}+C_{i\text{-}\mathrm{soil}}+C_{i\text{-}\mathrm{dead}}$$

(6)

$$C_{i\text{-}\mathrm{total}}=\sum _{i=1}^n\left(C_i\times F_i\right)$$

(7)

In the equation, Ci represents the total carbon density of land use type i (t·hm⁻²); Ci-above represents the above-ground carbon density of land use type i (t·hm⁻²); Ci-below represents the below-ground carbon density of land use type i (t·hm⁻²); and Ci-soil represents the soil carbon density of land use type i (t·hm⁻²). Ci-dead represents the density of dead organic carbon (t·hm⁻²) for land type i. Ci-total represents the total carbon stock (t) in the study area; n represents the number of land types; and Fi represents the area (hm²) of land type i.

The results are shown in Figure 10.

Since the carbon pool consists of four components—above-ground biomass, below-ground biomass, soil organic carbon, and carbon in dead organic matter—both carbon pool data and field surveys indicate that soil organic carbon in the Liaohe Estuary region is significantly higher than the other three types of carbon stocks. This is reflected in higher carbon stocks in areas such as tidal flats and exposed mudflats. Interannual variations were calculated based on annual carbon stock data from the study area, as shown in Figure 11.

In Figure 11, (a) shows the change in carbon stocks from 2020 to 2021; (b) shows the change in carbon stocks from 2022 to 2021; (c) shows the change in carbon stocks from 2022 to 2023; and (d) shows the change in carbon stock from 2024 to 2025. Changes in carbon stock at the Liao River Estuary over the past five years have been quite pronounced and have shown a pattern of instability. This reflects the complex ecological environment of the Liao River Estuary region. At the same time, due to limitations in data from the Sentinel-2 remote sensing satellite and the influence of high and low tides at the estuary, the distribution of tidal flats in the study area has fluctuated significantly, though it has generally shown a decreasing trend, indicating enhanced water connectivity in the Liao River Estuary. As can be seen from Figure 11, the area within the study region exhibiting the greatest variation in carbon stock is the tidal flat zone. Among the areas labeled a to d, the zone with the most significant change in tidal flat area is the estuarine tidal flat zone. Verification through field surveys and drone data indicates that changes in tidal flat area within the study region have stabilized over the past five years. The reason for the substantial variation in tidal flat area is that the study region is located at the mouth of the Liao River and is significantly influenced by tidal currents. A review of high and low tide data from nearby tidal monitoring stations reveals that, although the Sentinel-2 remote sensing imagery selected for this study was acquired under controlled conditions regarding the month of acquisition and cloud cover, the tidal data indicate significant variations in water levels at the time of acquisition for the six scenes. In particular, as observed in Figure 10, the water levels at the time of acquisition for the 2022 and 2025 scenes were both greater than 3 m, resulting in the submergence of parts of the tidal flats in the Liao River estuary. The tidal data used in this paper are sourced from the National Oceanic Science Data Center (

Table 8. Tidal depth data table.

Date	Highest	Lowest	Imaging Tide	Tidal Range
20 September 2020	414	45	176	369
23 September 2021	383	56	139	327
18 September 2022	387	80	382	307
13 September 2023	404	67	72	337
16 June 2024	315	92	263	223
29 September 2025	397	79	348	318

Tide level data are referenced to the theoretical depth datum (200 cm below mean sea level). Imaging time is approximately 10:45 Beijing Time (UTC + 8). Tide levels at imaging time are estimated from the tidal curves provided by the National Marine Data and Information Service (NMDIS) of China. Tidal range = highest high tide − lowest low tide.

).

5. Discussion

The approach employed in this study—using NDWI to extract water, S2REP to extract reed, and MSAVI to extract Suaeda salsa—effectively captures land cover information in remote sensing extraction research, though the results still exhibit some degree of error. The band replacement strategy red-edge B8A instead of red for S2REP positively enhanced the separability, as well as between Reed and Suaeda salsa. However, a negative effect was observed in areas with mixed vegetation or transitional zones, where the modified indices produced slightly higher false-positive rates due to spectral similarity at class boundaries [42]. The hierarchical extraction order positively prevented cascading misclassifications; but if a class was incorrectly extracted in an early step (e.g., water misclassified as tidal flat), the error propagated to subsequent classes. This negative effect was minimized by using conservative thresholds.

Remote sensing imagery is susceptible to phenomena such as “same spectral signature, different objects” and “different objects, same spectral signature.” Sentinel-2 imagery, with its bands at 10 m, 20 m, and 60 m resolution, cannot guarantee homogeneity of objects within each pixel even after preprocessing steps like atmospheric correction, radiometric calibration, and band resampling. Additionally, issues arising during the acquisition and production of remote sensing imagery—such as distortion, cloud cover, and shading—inevitably introduce certain effects on subsequent extraction results. Furthermore, the study area is located in a coastal wetland at a river estuary, featuring a complex ecological environment. Vegetation such as Suaeda salsa and reed grows on intertidal tidal flats, exhibiting stronger water absorption than typical conditions. Spectral reflectance curves reveal high similarity between the spectral profiles of Suaeda salsa and tidal flats [22]. Consequently, threshold settings and spectral confusion between tidal flats and Suaeda salsa during subsequent extraction processes compromise land cover classification accuracy. The vegetation indices employed in this study specifically replaced the bands used in calculations based on previously extracted spectral reflectance data. Whether these vegetation indices are universally applicable across all remote sensing satellite sensors, and whether using a vegetation index derived from one satellite sensor on another satellite with different parameters can guarantee computational accuracy, remains open to further discussion and research. The generalizability of conclusions drawn from one satellite sensor to a broader range of remote sensing satellites requires additional investigation. In specific, in detailed studies of land feature interpretation, it has been found that high and low tides have a significant impact on the classification of land features in complex coastal wetland ecosystems [23]. Furthermore, the variation in tidal ranges causes substantial fluctuations in the area of exposed mudflats in coastal wetlands, which in turn leads to complex trends in carbon stock estimates derived from mudflat calculations. The InVEST carbon stock estimates (Section 4.4, Table 7) show that tidal flats contribute the largest soil carbon pool (166.46 t C/ha), followed by Suaeda salsa (112.40 t C/ha) and reed (72.60 t C/ha). Tidal dynamics have a strong positive effect on carbon exposure as follows: low tide exposes carbon-rich soils, potentially increasing aerial carbon flux estimates; high tide submerges tidal flats, reducing the visible carbon pool by up to 12%. Even when using a single remote sensing satellite, such as Sentinel-2 imagery with a revisit cycle of five days, it remains impossible to fully meet the data requirements for study areas spanning multiple years. However, the use of multi-source data compromises data consistency and accuracy. When conducting land cover studies in a specific study area within a coastal wetland ecosystem, it is challenging to simultaneously satisfy the three requirements of single-satellite remote sensing data, high and low tide times, and cloud cover conditions [24]. This poses significant difficulties in further enhancing the scientific rigor and robustness of the research [25]. In subsequent studies, we need to identify better methods to ensure these three factors align as closely as possible, or utilize intermediary tools to improve their compatibility, thereby yielding more accurate carbon stock simulation results [43,44].

Furthermore, the robustness of the hierarchical extraction approach developed in this study to other research areas remains unknown. In hierarchical feature extraction, adjusting vegetation index thresholds simultaneously impacts both classification accuracy and the number of misclassified pixels. Balancing these two factors—accuracy and misclassified pixels—requires extensive research to identify optimal patterns. The hierarchical extraction method proposed in this study achieved an overall accuracy of 98.5% (Kappa 0.9796), which is substantially higher than previous wetland classification studies. Our accuracy improvement can be attributed to the following two key innovations: the explicit band replace such as red edge instead of red for reed vs. Suaeda salsa separation, which directly address the spectral confusion that traditional indices fail to resolve; and the fixed hierarchical extraction order, which prevents misclassification cascades common in single-step machine learning models. Despite the high accuracy, certain errors persisted. The approach—using NDWI to extract water, S2REP to extract reed, and MSAVI to extract Suaeda salsa—effectively captures land cover information, though the results still exhibit some degree of error. Remote sensing imagery is susceptible to “same spectrum, different objects” and “different objects, same spectrum” phenomena [45]. Sentinel-2 imagery, with bands at 10 m, 20 m, and 60 m resolution, cannot guarantee object homogeneity within each pixel even after preprocessing. Additionally, acquisition and production issues (distortion, cloud cover, and shading) inevitably introduce effects on extraction results [30]. The study area is located in a coastal wetland at a river estuary, featuring a complex ecological environment. Vegetation such as Suaeda salsa and reed grows on intertidal tidal flats, exhibiting stronger water absorption than typical conditions. Spectral reflectance curves reveal high similarity between Suaeda salsa and tidal flats. Consequently, threshold settings and spectral confusion between tidal flats and Suaeda salsa during subsequent extraction compromise classification accuracy. The vegetation indices employed in this study specifically replaced the calculation bands based on extracted spectral reflectance data. Whether these vegetation indices are universally applicable across all remote sensing sensors, and whether using an index derived from one satellite on another with different parameters can guarantee computational accuracy, remains open to further discussion. The generalizability of conclusions from one sensor to a broader range requires additional investigation. In detailed interpretation of land features, we found that high and low tides significantly impact classification in complex coastal wetland ecosystems. Variation in tidal ranges causes substantial fluctuations in exposed mudflat area, which in turn leads to complex trends in carbon stock estimates derived from mudflat calculations. Even using a single satellite (e.g., Sentinel-2 with a 5-day revisit cycle), it remains impossible to fully meet data requirements for multi-year studies [46]. However, multi-source data compromise consistency and accuracy. When conducting land cover studies in a coastal wetland ecosystem, it is challenging to simultaneously satisfy the three requirements of single-satellite data, appropriate tidal timing, and cloud-free conditions. This poses difficulties in further enhancing scientific rigor and robustness. In subsequent studies, we need to identify better methods to align these three factors, or use intermediary tools to improve their compatibility, thereby yielding more accurate carbon stock simulations.

Furthermore, the robustness of our hierarchical extraction approach to other study areas remains unknown. In hierarchical feature extraction, adjusting vegetation index thresholds simultaneously impacts both classification accuracy and the number of misclassified pixels. Balancing these two factors requires extensive research to identify optimal patterns.

6. Conclusions

This study selected the Liaohe Estuary Delta region in China as the study area. Based on spectral reflectance characteristics of four typical land cover types (reed, Suaeda salsa, water, and tidal flat) analyzed using Sentinel-2 remote sensing imagery, multiple vegetation indices were selected and optimized to achieve hierarchical extraction of these four land cover types. There were Sentinel-2 spectral reflectance curves (B2–B8A, B11, and B12) for four land cover types. Water and tidal flat exhibit a nearly flat, low-reflectance horizontal line across visible to red-edge bands due to high absorption and minimal spectral variation. Reed and Suaeda salsa show a remarkable decline in B11 and B12 (SWIR1 and SWIR2), caused by strong water absorption in plant tissues, which helps distinguish them from tidal flats. By comparing extraction accuracy and effectiveness through control variable methods, manual visual interpretation, and UAV imagery support, this study demonstrates that using suitable vegetation indices for hierarchical feature extraction in complex land cover areas offers strong targeting capability and high accuracy. This enhances our understanding of the study area and provides valuable information for biodiversity and ecosystem research and conservation, laying a solid foundation for future coastal wetland studies and conservation efforts. Key findings are as follows: a distribution map of typical land features in the Liaohe River estuary region for 2023 was obtained, achieving an overall classification accuracy of 98.5% with a Kappa coefficient of 98%. High classification accuracy was attained through targeted adjustments to spectral characteristics and vegetation indices, coupled with hierarchical extraction using vegetation indices. By utilizing the results of detailed vegetation classification in conjunction with the InVEST model to simulate carbon stocks, it was found that the tidal flats in the study area are significantly influenced by tidal patterns, resulting in complex variations in carbon stocks. Simulation results for above-ground vegetation carbon content and total regional carbon stocks from 2020 to 2025 suggest a possible trend toward stabilization compared with earlier periods. While these findings may be interpreted as indirect evidence that the ecological restoration efforts undertaken in the Liaohe Estuary coastal wetland area in recent years have had a beneficial impact, longer-term monitoring and additional field data are needed to confirm this interpretation and to distinguish restoration effects from natural variability.

Based on the findings and limitations of this study, for future work the following should be done: integrating additional satellite sensors to reduce the dependency on cloud-free optical images and to improve temporal resolution for tidal dynamic monitoring; exploring machine learning-based dynamic threshold optimization to further automate the hierarchical extraction process and minimize user intervention; and extending the methodology to other coastal wetlands with different vegetation compositions to test its generalizability.