Distribution and Long-Term Variation of Wetland Land Cover Types in the Yellow River Delta Remote Sensing Monitoring

Abstract

Wetlands are dubbed the “kidneys of the earth” and are involved in climate regulation, carbon sequestration, ecological balance preservation, and reducing the surface water pollution. Ongoing economic development has introduced pressing challenges to wetland environments. In this context, extracting coastal wetland information and monitoring the dynamic changes are essential. Using long-term sequence Sentinel-2 satellite remote sensing images and field observations, this research proposed a Dynamic Bayesian Network classification model framework based on conjugate gradient updates. We compared the wetland feature extraction effects of the Fletcher–Reeves and the Polak–Ribière–Polyak algorithms of the conjugate gradient. Then, remote sensing combined with the FRDBN classification model was used to extract the information pertinent to wetland feature types and changes in wetland areas and analyze alterations in the distribution characteristics of land cover types. The results showed that the FRDBN model achieved high accuracy (above 96%), and kappa coefficients exceeded 0.96. Long-term monitoring revealed that the area of wetlands increased by 0.85 × 104 hm² from 2016 to 2021. Non-aquatic land cover types exhibited pronounced dynamic changes, with the area of change representing 58–69% of the monitored total. Specifically, the transition between salt marsh vegetation and artificial wetlands was relatively obvious. The FRDBN model provides a new method for extracting wetland feature information. Wetland protection, dynamic monitoring, and carbon sink research can provide robust technology support, facilitating investigations into coastal salt marsh carbon sinks and technological advances in carbon sink assessment.

1. Introduction

The Yellow River Delta is an ecologically important delta estuary in China, sustaining the reproduction and survival of numerous wild animals and plants [1,2,3,4,5]. However, it is facing increasingly severe threats posed by human impact and climate change [6,7,8]. Further monitoring and research are needed to promote the technological development of ecological restoration and carbon sink assessment [9,10].

Remote sensing methods are characterized by large-scale synchronization, high efficiency, and cost-effectiveness. They have found wide applications in resource and environmental monitoring in land, ocean, forestry, and agriculture departments. These techniques offer unique advantages over traditional on-site surveys [11,12]. Research on wetland classification and dynamic changes monitoring based on remote sensing technology has evolved since the 1970s [13,14]. Initially, visual interpretation methods based on staff experience and skills were extensively used. With the development of computer, multispectral imagery, and earth observation satellites technologies, semi-automatic classification methods derived from statistical pattern recognition emerged. These methods primarily leveraged the objects’ spectral characteristics for supervised and unsupervised classification. Examples include the Maximum Likelihood Estimation, Minimum Distance Discriminant Method, K-Nearest Neighbor, K-Means Clustering Algorithm, and random forest [15], which are combined with image pre-processing to improve the feature recognition accuracy. Xin, D.Y. et al. proposed a single-level information extraction method based on superpixel segmentation and multi-condition differential merging [16]. Their results indicated that the multi-scale segmentation accurately extracted multi-level information and addressed limitations such as time-consuming segmentation, classification, and complex genetic relationships. Multi-source remote sensing technology and artificial intelligence algorithms have been updated continuously [17,18]. With such progress, various machine learning-based remote sensing image classification methods have emerged [19]. For example, Decision tree [20], Convolutional Neural Networks [21], Support Vector Machine [22], and Attention Mechanism have achieved favorable results. Moreover, researchers adopt pre-processing methods to extract potential image information as part of improving classification accuracy, thereby fully utilizing the features of multi-source remote sensing data [23]. Normalized Difference Vegetation Index, Normalized Difference Water Index [24], Red Edge Index, texture features, polarization features, etc., are also introduced. Xiang, C.T. et al. developed a method to extract wetland vegetation information under different endmember models [25]. The researchers tested the impact of endmembers on extraction accuracy and demonstrated the feasibility of using potential information for feature information extraction. Researchers have been dedicated to machine learning algorithms to improve the accuracy of wetland feature extraction to ensure the effectiveness of dynamic spatiotemporal analysis [26,27]. In addition, the efforts to identify factors affecting wetland land cover changes have gained many insights [28]. Most natural reserves consist of newly formed and original wetlands with extensive areas and intricate tidal channel networks, hindering traditional on-site operational survey [29]. Numerous studies have confirmed that remote sensing techniques exhibited great potential for operational applications in the long-term monitoring of coastal wetland vegetation resources and the local environments [30].

Artificial intelligence algorithms have been successfully applied to wetland feature classification research; despite such achievements, these applications typically neglect the particularity of extracting land cover types [31]. In addition, network designs for dataset-specific feature extraction are lacking. We developed a Dynamic Bayesian Network (DBN) classification model based on conjugate gradient (CG) updates using long-term sequence Sentinel-2 satellite remote sensing images and field observation data to address these deficiencies. The Fletcher–Reeves (FR) and the Polak–Ribière–Polyak (PRP) algorithms of CG were introduced to obtain the update term for the parameter matrix by calculating the 2-norm of the gradient. Remote sensing interpretation and monitoring were conducted based on the FRDBN classification model, including wetland feature types, changes in feature distribution characteristics, and wetland area alterations. Wetland protection, dynamic monitoring, and carbon sink research can provide robust data support. These efforts will contribute to coastal salt marsh carbon sink research and drive the technological development of carbon sink assessment.

2. Study Area and Data

2.1. Study Area

The Yellow River Delta National Nature Reserve is located on the coastal areas in the vicinity of the mouths of the Yellow River in Dongying City, Shandong Province, China. It is designed to preserve valuable ecological natural resources in this region. This reserve borders the Bohai Sea to the north and Laizhou Bay to the east. Its geographical coordinates span from 118°32.981′ to 119°20.450′ E and from 37°34.768′ to 38°12.310′ N. This region is portioned into the current and the old river course protection zones, which cover approximately 153,000 hectares, comprising about 82,700 hectares of land and about 38,300 hectares of intertidal zones. The Yellow River Delta represents the most intact and pristine wetlands in China, with the largest expanses of land [30]. It is characterized by the warm temperate semi-humid continental monsoon climate [3].

The dominant vegetation in the wetlands consists of aquatic and halophytic plants, with prevalent herbaceous vegetation. This study conducted field surveys and remote sensing monitoring in the Yellow River Delta Nature Reserve (Figure 1). Based on these data, the target features were classified into salt marsh vegetation (such as reeds, Suaeda salsa, Spartina alterniflora, and Tamarix chinensis), cultivated land, harvested cropland or bare land, black locust forests, pit ponds, bare tidal flats, and water areas (rivers and seawater).

2.2. Data

(1) Sentinel-2

The Sentinel-2 satellite carries a Multi-Spectral Instrument with a revisit period of five days (for twin satellites A and B) [12]. The European Space Agency offers publicly available Sentinel-2 Level 2A images. The downloaded images were processed using the Sen2cor plugin, which applies atmospheric bottom correction to provide reflectance values. The resulting images include data from 13 spectral bands across 400–2400 nm wavelengths, and the image parameters are listed in

Table 1. Sentinel-2 image parameters.

Bands	Spectral Range	Wavelength/nm	Spatial Resolution/m	Width/km	Orbital Altitude/km
1	VIS	433–453	60	290	786
2		458–523	10
3		543–578	10
4		650–680	10
5	NIR	698–713	20
6		733–748	20
7		773–793	20
8		785–900	10
8a		848–881	20
9		935–955	60
10	SWIR	1360–1390	60
11		1565–1655	20
12		2100–2280	20

. The Sentinel-2 satellite provides data with high spatial resolution (10 m, 20 m, and 60 m), excellent temporal resolution, and favorable spectral resolution. These abilities ensure the consistency and accuracy for time series analysis, rendering it suitable for monitoring ecological changes over prolonged periods. This study utilized atmospheric-corrected surface reflectance images (Level-2A, Collection 3) from the Sentinel-2 satellite’s MSI for remote sensing monitoring of land cover type distribution and long-term changes in the Yellow River Delta wetlands. The study area was covered by two Sentinel-2 tiles, i.e., T150 and T151 (corresponding to paths/rows 150/30 and 151/30, respectively). Each tile captured two images annually from 2015 to 2021, as shown in Figure 2.

(2) Field data

To coincide with the field observation time, the Sentinel-2 data were acquired in September every year. Accordingly, the reconnaissance routes and sampling locations were established based on high-resolution remote sensing images using the CGCS2000 coordinate system, with 143 field stations (Figure 3). The field survey data covered most coastal wetland geomorphic units of the Yellow River Delta Nature Reserve. Specifically, it included approximately 38 types of land features, dominated by salt marsh vegetation species such as reed, Suaeda salsa, and Spartina alterniflora. The rest of the land features comprised willow trees, pit ponds, bare beaches, etc. These data lays a foundation for constructing a classification system for the coastal wetlands of the Yellow River Delta Nature Reserve, along with the verification and evaluation of classification methods.

(3) Classification system

A classification system for monitoring areas was established based on comprehensiveness, hierarchy, and remote sensing image-based principles. We referred to the “National Standard of the People’s Republic of China: Wetland Classification (GB/T 24708-2009)” [32], “The Coastal Wetland Classification System of China” [33], and the classification systems from related coastal wetland studies [5,34]. On this basis, a monitoring area classification system was constructed to investigate wetland salt marsh vegetation. This classification was based on the vegetation types in field observations and their representation in remote sensing images. The developed classification system parameters are summarized in

Table 2. Classification system of the Yellow River Estuary Wetlands.

Primary Category	Secondary Category	Tertiary Category	Explanation
Natural wetlands	Herbaceous wetlands	Reeds biome	Land types dominated by reeds, suaeda salsa, and spartina alterniflora respectively
		Suaeda salsa plant formation
		Spartina alterniflora plant formation
	Shrub wetlands	Tamarix chinensis plant formation	Land types dominated by tamarix chinensis
	Forest wetlands	Black locust plant formation	Land types dominated by black locust
	Other wetlands	Bare tidal flats	Mudflats or dry beach affected by tides
		Water areas	Yellow river, sea water
Artificial wetlands	Artificially constructed or used water bodies	Pit ponds(Water surface)	Pit ponds, reservoirs, canal, ditches, aquaculture ponds (shrimp ponds, crab ponds), salt fields, salt ponds
	Rice paddy wetlands	Cultivated land	Rice fields, corn, cotton, lotus ponds, etc
		Harvested cropland or bare land	Harvested farmland or bare ground

. This study conducted remote sensing classification of wetland features at the Yellow River estuary using ten features at the tertiary classification category in Table 2.

3. Evaluation Method and Classification Results

A classification model of CG Dynamic Bayesian Network (CGDBN) was proposed based on high-resolution remote sensing images and field survey data. This model was built upon the BN theory [35], with particular emphasis on the idea of processing time series data [36]. During backpropagation (BP), this model utilizes the CG to substitute gradient descent, thereby updating network parameters. Two CG methods (i.e., FR and PRP algorithms) are employed for model updating. These algorithms facilitate the convergence of the classification model and improve the accuracy of remote sensing classification, enabling accuracy evaluation. The overall technical approach for wetland monitoring methods in the Yellow River Delta Reserve is illustrated in Figure 4.

3.1. Model

The structure of the CGDBN model framework is illustrated in Figure 5. For remote sensing classification applications, three modules were designed, including extracting spectral-spatial neighborhood features, creating and inputting joint feature samples, and DBN classification based on CG updates.

(1) Spectral Feature Extraction

Spectral data were extracted from image data. In the early stage, spatial neighborhood features were extracted from images by performing Principal Component Analysis transformation and image enhancement. On this basis, we used a 3 × 3 window to obtain pixel values before the Nth component. Subsequently, these values were expanded into a one-dimensional vector as the spatial neighborhood features of the central pixel.

(2) Sample

The spectral-spatial neighborhood feature vectors of each image pixel were normalized using the range transformation method, as shown in Equation (1).

$$y_i={\displaystyle \frac{x_i-min\left(x\right)}{\mathit{max}\left(x\right)-min\left(x\right)}}$$

(1)

where $x_i$ denotes the ith value of the feature vector $x$ to be standardized, $y_i$ represents the ith value of the standardized feature vector $y$, where $i=\mathrm{1,2},\dots ,\mathrm{m}$, and $\mathrm{m}$ indicates the spectral dimension.

(3) CGDBN

CGDBN consists of a Restricted Boltzmann Machine (RBM) and BP layers. The basic structure corresponding to the pre-training and fine-tuning during model training is schematically shown in Figure 6.

Initially, the feature vector is input. Each layer in the RBM network is trained separately and unsupervisedly. This training approach allows for the mapping of feature vectors to different feature spaces without compromising much feature information, resulting in output feature vectors. The Contrastive Divergence algorithm is introduced in this process to facilitate training.

The next step involves the BP network extracting the output feature vectors from the RBM as its input feature vectors, followed by error calculation and BP to each layer. This network updates the weight matrices of each layer using the CG method, thereby fine-tuning the entire DBN. This enables the supervised training of the entity relation classifier. The layer-wise training in the RBM network can only ensure that the weights within the corresponding layer are optimized for the layer-specific mapping of feature vectors. It falls short of establishing the mapping relationship between the input feature vectors and the output class labels, available in the DBN. To address this limitation, the training of stacked RBMs can serve as an effective initialization method for the weight parameters of a BP network. On this basis, optimizing and fine-tuning the BP network parameters can improve classification efficiency.

(4) Network parameter update

When solving large-scale optimization issues, it is necessary to approximate and identify the optimal solution through iteration. The BP network is based on the error $J_k$ (see Formula (2)) between the DBN predictions and the actual values, and the error is backpropagated to each layer. The $W_k^l$ (see Formula (3)) of each layer can be updated by employing optimization algorithms. After multiple iterations, the resulting trained DBN model is obtained.

$$J_{batch}={\displaystyle \frac{1}{2batch}}{\sum }_{i=1}^{batch}{\left(\widehat{Y_i}\left(W^l,b^l\right)-Y_i\right)}^2$$

(2)

where $J_{batch}$ represents the cost function formula when the output layer is a sigmoid classifier, $batch$ corresponds to the number of batch training samples, $\widehat{Y_i}$ denotes the output layer prediction, $Y_i$ represents the true object label, and $\left(W^l,b^l\right)$ are the weight and bias parameters of layer $l$. In this formula, the multiplication by 1/2 is typically used to cancel out the coefficient of 2 that arises during the differentiation of squares, thereby simplifying the gradient calculation

The CG is a classical unconstrained optimization algorithm. This algorithm distinguishes from the gradient descent method by adjusting its search direction. Specifically, it combines the negative gradient direction and the previous search direction. This approach alleviates the limitations of the gradient descent algorithm in traditional models, such as the zigzag phenomenon near the minimum point and the decrease in convergence speed. The improvement in convergence speed has been confirmed. Moreover, the CG method only calculates first-order derivatives to determine its search orientation. This reduces memory, improves convergence and stability, and eliminates the need for external parameters [37]. Due to these advantages, the DBN model was improved based on the CG algorithm to address the slow convergence arising from the orthogonal search directions during iterative fine-tuning of model parameters using gradient descent. A CG-based update quantity $∆W_k^l$ (Formula (4)) was constructed to act on the weight and bias adjustment. DBN model was improved individually based on the FR and PRP algorithms in the classical CG. Both methods have a negative gradient direction in the first iteration, but from the second iteration, the calculation formula of the direction is different, mainly reflected in the calculation of $\beta$. The two algorithms’ $\beta$ computational formulas are expressed in Equations (6) and (7).

$$W_{k+1}^l=W_k^l-∆W_k^l$$

(3)

where ${\mathrm{W}}_{k+1}^l$ is the parameter matrix (including weights and biases) after updating the $l$th layer at the $k$th iteration, $W_k^l$ is the parameter matrix before the update, and $∆W_k^l$ is the matrix update amount.

$$∆W_k^l=\alpha \ast d{W}_k^l-d{W}_{CG}^l$$

(4)

where the conjugate gradient $d{W}_{CG}^l$ of the $l$th layer is according to Equation (5), ${dW}_k^l$ is the gradient of the $l$th layer at the $k$th iteration, and $\alpha$ is the learning rate.

$$d{W}_{CG}^l={\beta }_{k-1}d{W}_{k-1}^l$$

(5)

$${\beta }_{k-1}^{FR}={\displaystyle \frac{{‖d{W}_k^l‖}^2}{{‖d{W}_{k-1}^l‖}^2}}$$

(6)

$${\beta }_{k-1}^{PRP}={\displaystyle \frac{{‖d{W}_k^l‖}^T(‖d{W}_k^l‖-‖d{W}_{k-1}^l‖)}{{‖d{W}_{k-1}^l‖}^2}}$$

(7)

where $k$ denotes the iteration number. Compared to the optimization of the classical CG algorithm, we calculate ${\beta }_{k-1}$ using use the 2-norm to improve optimization and classification results for the DBN model. The regularization factor based on the 2-norm is represented by Equations (8) and (9).

$${\beta }_{k-1}^{FR\_L2}={\displaystyle \frac{norm\left({‖d{W}_k^l‖}^2\right)}{norm\left({‖d{W}_{k-1}^l‖}^2\right)}}{\beta }_{k-1}^{FR}$$

(8)

$${\beta }_{k-1}^{PRP\_L2}={\displaystyle \frac{{norm(‖d{W}_k^l‖}^T(‖d{W}_k^l‖-‖d{W}_{k-1}^l‖))}{norm\left({‖d{W}_{k-1}^l‖}^2\right)}}{\beta }_{k-1}^{PRP}$$

(9)

The norm( ) function returns the 2-norm of a matrix, approximately equal to the maximum singular value of the matrix. The BP parameter update process based on CG is shown in Algorithm 1. The FR and PRP algorithms based on the 2-norm were applied separately to the DBN model for updating purposes. PRPDBN refers to the DBN based on the PRP algorithm and the 2-norm, and FRDBN is the other DBN based on the FR algorithm and the 2-norm. The CGDBN framework is kept unchanged in both models, and PRP and FR algorithms serve for parameter optimization. Algorithm 1. Algorithm description for CGDBN parameter update.

Algorithm 1 CGDBN parameter update

Input: RBM parameter matrix $W_0^l$. Batch training samples $S=\left\{a_k^l\right\}$.

Learning rate    $\alpha =2$. Number of iterations    epochs = 500.

Output: DBN Parameter matrix $W_k^l$.

1:    for  $t=1,2,\dots$, epochs do

2:        forall the $a_k^l\in S, k=\mathrm{1,2},3,\dots$ do

3:        $Y_k^{perdition }\leftarrow \{a_k^l,{ W}_0^l, sigmoid\}$    % sigmoid is the activation function

4:        $J_k \leftarrow \{ Y_k^{groundtruth}, Y_k^{perdition }, batchsize\}$    % Request the cost function value for the training samples

5:        $d{W}_k^l\leftarrow \{ J_k, sigmoid\}$    % Calculate the gradient

6:        $d{W}_{CG}^l\leftarrow \{ {\beta }_k, d{W}_k^l, d{W}_{k-1}^l\}$    % Compute the FR or PRP conjugate gradient. Different conjugate gradients involve the calculation of ${\beta }_k$ through Equation (8) or Equation (9)

7:        if $t=1$ then $∆W_k^l\leftarrow \{d{W}_k^l, \alpha \}$    % The initial update amount is calculated by the steepest descent method

8:        else  $∆W_k^l\leftarrow \{d{W}_k^l, d{W}_{CG}^l, \alpha \}$

9:        $W_{k+1}^l\leftarrow \{W_k^l, ∆W_k^l\}$

10:      end if

11:   end for

12:   end

3.2. Category and Accuracy Evaluation

The classification experiments on public hyperspectral datasets have proven the better theoretical convergence of the FR algorithm than the PRP algorithm. The improved FRDBN model exhibits reduced parameter sensitivity, enhanced model convergence, and increased classification accuracy [38]. It achieves classification accuracy comparable to existing classification methods. In contrast, the PRP algorithm displays superior numerical performance [38]. The main purpose here is to improve classification accuracy by enhancing model convergence. Given this, subsequent discussions focus exclusively on parameter tuning and accuracy evaluation of the FRDBN classification model.

(1) Parameter settings

We cropped two scenes of Sentinel-2A images covering the Yellow River Estuary wetland monitoring area on 29 September 2019, to obtain multi-spectral images of the Dawenliu and Yiqianer management stations, each containing 11,237,950 and 4,854,174 pixels, respectively. The land cover types were divided into ten categories. The Yiqianer management station was classified as Category 8 based on the low prevalence of Spartina alterniflora and black locust. Using five years of field photos and data, 96,344 and 31,484 samples were labeled for the Dawenliu and Yiqianer monitoring areas, respectively. After multiple parameter adjustments, the optimal FRDBN classification model for the monitoring area was ultimately trained. The model input included the 12 spectral features extracted in the early stage and the 4 × (3 × 3) spatial neighborhood enhancement features. The structures of the two regional classification models are 48-100-100-10 and 48-100-100-8. The setting conditions of the remaining network parameters were as follows: 200 iterations, 100 batch training samples, learning rate at 0.1 for the unsupervised training process, as well as 2500 iterations and 50 batch training samples for BP.

(2) Result evaluation

The labeled samples corresponding to the Dawenliu monitoring area were designed with ten types of land features. Then, 300 samples were selected from each category for model training, and the rest served as test samples for accuracy assessment, as shown in

Table 3. Samples and accuracy values of Sentinel-2A classification in the Dawenliu monitoring area.

Number Sequence	Category	Marked Samples	Trained Samples	Tested Samples	Classification Accuracy (%)
1	Reed	10,403	300	10,103	93.05
2	Suaeda salsa	10,573	300	10,273	97.40
3	Spartina alterniflora	11,075	300	10,775	98.09
4	Tamarix chinensis mixed growth	10,077	300	9777	97.45
5	Bare tidal flats	10,289	300	9989	99.00
6	Water areas	11,701	300	11,401	98.42
7	Pit ponds	10,320	300	10,020	98.09
8	Black locust	5063	300	4763	99.16
9	Cultivated land	11,391	300	11,091	96.04
10	Harvested cropland or bare land	5452	300	5152	98.91
Total		96,344	3000	93,344	——
AA (%)					97.56
OA (%)					97.40
Kappa					0.97

. The Sentinel-2A pseudo-color map is shown in Figure 7a, the labeled sample map is presented in Figure 7b, and the classification result map is given in Figure 7c. The overall classification accuracy of FRDBN is 97.40%, and the Kappa coefficient is 0.97. It can be seen from the last column of Table 3 that the average classification accuracy of all categories exceeds 93%. The category distribution in the resulting classification map is consistent with the field survey results, and several vegetation species can be distinguished, indicating good classification results. The post-processing results are shown in Figure 7d.

In the Yiqianer monitoring area, 31,484 samples were labeled for ten types of land features. Similarly, the model training involved 300 samples from each category, and the rest served as test samples to evaluate the accuracy of the classification model. The labeled samples and FRDBN classification accuracy are listed in

Table 4. Samples and accuracy values of Sentinel-2A classification in the Yiqianer monitoring area.

Number Sequence	Category	Marked Samples	Trained Samples	Tested Samples	Classification Accuracy (%)
1	Reed	2821	300	2521	75.24
2	Suaeda salsa	3012	300	2712	99.00
3	Spartina alterniflora	—	—	—	—
4	Tamarix chinensis mixed growth	2631	300	2331	99.31
5	Bare tidal flats	5472	300	5172	99.81
6	Water areas	5492	300	5192	100
7	Pit ponds	5061	300	4761	98.78
8	Black locust	—	—	—	—
9	Cultivated land	5060	300	4760	94.81
10	Harvested cropland or bare land	2025	300	1725	99.71
Total		31484	2400	29084	——
AA (%)					95.83
OA (%)					96.68
Kappa					0.96

. The Sentinel-2A pseudo-color, labeled sample, and classification result maps are presented in Figure 8a–c. Overall, FRDBN exhibits 96.68% classification accuracy, with a Kappa coefficient of 0.96. From the last column of Table 4, the mean classification accuracy across all categories reaches above 94%, except for reed. The category distribution is consistent with the field survey, enabling the effective distinction of several types of vegetation and demonstrating classification performance. This monitoring area has extensive interlaced areas of reed vegetation and cultivated land. Such a distribution pattern is more complex than the regular block distribution of cultivated land in the Dawenliu monitoring area. It is challenging to distinguish reeds from cultivated land, leading to the misclassification of some reeds, which reduces the classification accuracy. The post-processing results are shown in Figure 8d.

According to the experimental results of the high/multi-spectral classification in the Yellow River Estuary coastal wetland region, the overall classification accuracy for labeled samples in the monitoring area exceeds 96%. The accuracy of wetland extraction according to the CGDBN model is significantly improved compared with the accuracy of wetland extraction based on random forest [39,40]. It is indicated that the proposed classification model has positive application outcomes in the monitoring area. The classification results of unlabeled samples in the monitoring area were verified and corrected on an empirical basis, and field survey site data were continuously supplemented to obtain the resulting classification results for the area of interest.

4. Monitoring and Analysis

The FRDBN classification model was adopted to classify wetland features based on Sentinel-2 multi-spectral remote sensing data covering the Yellow River Estuary Nature Reserve from 2015 to 2021 (included), the established classification system, and the interpretation signs for the protected wetlands. The distribution of feature types was recorded annually throughout the investigation period at the Dawenliu and Yiqianer management stations. Considering the limited data coverage in September 2015, our analysis focused on the spatial-temporal distribution characteristics, area changes, and land type conversion in the monitoring area for results obtained after 2016.

4.1. Analysis of Spatial Patterns of Land Cover Types

The pseudo-color composite image (R8, G4, and B2 bands) from Sentinel-2 remote sensing and the distribution map of land types in the Yellow River Estuary Wetland Reserve are presented in Figure 9. The gray shading indicates the part with missing image data in the Dawenliu management station area in September 2015. It can be preliminarily observed that the types of land features in this reserve are intricate and diverse, with pronounced seasonal and temporal dynamics, and considerable interannual variations may occur [41,42].

The number of pixels was counted for each land type in the classification map of the monitoring area and then multiplied by the ground size represented by each pixel (spatial resolution 10 × 10 m), resulting in the total area for each type of land feature in different periods, as shown in

Table 5. Area calculation results of various land types in different periods (104 hm²).

Time	September 2015	October 2016	September 2017	September 2018	September 2019	September 2020	September 2021
Reed	1.29	1.56	1.67	1.95	1.56	1.88	1.77
Suaeda salsa	0.46	0.80	0.88	1.14	0.93	1.15	1.30
Spartina alterniflora	—	0.34	0.39	0.38	0.40	0.42	0.45
Tamarix chinensis mixed growth	0.34	0.49	0.54	0.53	0.40	0.35	0.62
Bare tidal flats	0.62	1.84	1.98	1.61	2.04	1.79	2.25
Water areas	2.63	6.15	5.92	6.05	6.12	5.91	5.30
Pit ponds	1.22	1.66	1.05	1.50	1.32	1.55	1.40
Black locust	0.09	0.19	0.17	0.15	0.15	0.19	0.12
Cultivated land	1.82	0.85	2.41	1.58	2.15	2.06	1.90
Harvested cropland or bare land	1.10	2.20	1.08	1.21	1.03	0.79	0.99
Herbaceous salt marsh vegetation	1.75	2.70	2.94	3.48	2.89	3.45	3.52
Wetlands	6.96	9.94	10.17	10.04	9.97	10.18	10.79
Total	9.59	16.09	16.09	16.09	16.09	16.09	16.09

Note: Herbaceous salt marsh vegetation includes reed, Suaeda salsa, and Spartina alterniflora. Wetlands refer to the land area characterized by water-saturated soil that excludes water bodies. “Total” denotes the combined areas of all types of land features within the monitoring area.

. It can be seen that the area of herbaceous salt marsh vegetation totals the combined areas of reed, Suaeda salsa, and Spartina alterniflora.

The percentage of each land type area in the reserve from October 2016 to September 2021 was calculated, as summarized in Figure 10. The research area totals 16.09 × 10⁴ hm², with the non-water features primarily located in the intertidal and supratidal zones. The dynamic range of these zones is 9.33–11.06 × 10⁴ hm², accounting for 58–69% of the total area. Wetland vegetation mainly consists of herbaceous salt marsh. This type of vegetation is widespread and serves important ecological functions. The area changes in reeds, Suaeda salsa, and Spartina alterniflora range from 2.36 × 10⁴ to 3.52 × 10⁴ hm², representing 15–22% of the total.

The marshy and salt-tolerant reeds are the foundational plant species in the wetlands at the Yellow River mouth. They are widely distributed in the salt marsh vegetation of the reserve, primarily in the intertidal and supratidal zones of the estuary. These plants constitute a narrow, elongated growth belt stretching along the river course extending into the inland areas, covering 1.11 × 10⁴–1.95 × 10⁴ hm², or 75–12% of the total monitoring area. Due to the dense reed bed, they occupy an absolute advantage over potential plant competitors within their habitats. This land type typically hosts a small variety of species, such as Suaeda salsa, Tamarix chinensis, cattail, and Miscanthus sacchariflorus, growing in close proximity.

Suaeda salsa is concentrated on the tidal flats on both sides of the Yellow River mouth. This area is affected by seawater immersion, and the soil is highly salinized. Despite a tendency for seaward expansion, the distribution of Suaeda salsa along the coastline is often sparse. The area change in this land type ranges from 0.21 × 10⁴ to 1.49 × 10⁴ hm², accounting for 1–9% of the total.

Native to North America, Spartina alterniflora was introduced to the Yellow River Estuary Conservation Area in 1996 to preserve embankments and facilitate land reclamation programs. Due to its favorable tolerance to saline-alkali soils and tidal flooding, this species quickly adapted to the shallow beach and tidal flat environment in the Yellow River Estuary and proliferated. It is mainly distributed in the upper intertidal zone, covering 0.34 × 10⁴–0.45 × 10⁴ hm², accounting for 2–3% of the total area in this reserve. The spread of Spartina alterniflora can improve the soil structure, prevent soil compaction, promote the growth of beach silt, consolidate the coast. It is helpful to purify the air and green the environment, increase the wetland area, its developed root system and salt tolerance characteristics make it have the function of sand suppression, wind prevention, and wave resistance. It can also be used as feed, fuel, food and chemical raw materials, with a wide range of applications. At the same time, its strong environmental adaptability and reproductive ability lead to the occupation of local vegetation space, inhibit the growth of local species, destroy the original ecological balance, governance difficulty and other potential environmental problems. In view of potential problems, the special prevention and control of China began in February 2023, striving to achieve centralized prevention and control and proper management.

Aside from waters, cultivated land and harvested cropland or bare land dominate, primarily in the supratidal zone and extending inland along river courses, exhibit a dense, blocky distribution pattern. It covers an area of 2.66 × 10⁴ to 3.83 × 10⁴ hm², representing 17–24% of the total research area.

Bare tidal flats are prevalent in the monitoring area and are mainly found in the intertidal zone, where the majority are regularly flooded by seawater and exhibit large-scale patchy distribution. They are also present in the supratidal zone, along the riversides, and at the estuaries. Due to the influence of marine dynamics and tidal action, this area experiences dynamic changes with high tide flooding and low tide exposure, ranging from 1.07 × 10⁴ to 2.60 × 10⁴ hm², accounting for 7–16% of the total.

Tamarix chinensis mixed growth refers to the combined area of Tamarix chinensis, reeds, and Suaeda salsa. This land type is concentrated in the upper-middle and upper intertidal zones, i.e., saline-alkali lands, covering 0.35 × 10⁴ to 0.68 × 10⁴ hm² (2–4% of the total monitoring area). These zones have intricate terrains with numerous gullies, and the growth and distribution of plants exhibit complexity and diversity. Reeds and Suaeda salsa are dominant plant spices in the distribution area. In addition, a small number of herbaceous plants grow, such as Miscanthus sacchariflorus, Cynanchum chinense, bitter lettuce, Apocynum venetum, and Limonium sinense. Affected by changing seasons and the growth status of plants, the dominant populations vary. Despite such dynamics, Tamarisk chinensis consistently remains the primary species present in this area.

Pit ponds comprise various artificial wetlands, such as man-made reservoirs, aquaculture ponds, salt ponds, and paddy fields, with a limited presence of reeds growing along their edges. The distribution areas are uniform and relatively concentrated due to artificial planning, with an area of variations ranging from 1.05 × 10⁴ to 1.66 × 10⁴ hm², or 7–10% of the total area of monitoring region.

Black locust forests are distributed in higher inland areas adjacent to cultivated land. These forests are home to a small abundance of herbaceous vegetation such as reeds. In addition, tree communities are relatively concentrated and expensive. The area of change ranges from 0.09 × 10⁴ to 0.19 × 10⁴ hm², making up for 0.5–1.2% of the total monitoring area.

4.2. Analysis of Land Cover Type Conversion

As of September 2021, the total area of wetlands in the monitoring area increased by 0.85 × 10⁴ hm² (increased 8.55%) compared to October 2016 (10.79 × 10⁴ hm² vs. 9.94 × 10⁴ hm²). Apart from factors such as climate and tidal changes, the continuous expansion of wetland areas is primarily attributed to the sediment deposition from the Yellow River’s runoff in the Dawenliu monitoring area, which forms tidal flats. In addition, the extensive proliferation and growth of the invasive species Spartina alterniflora on the seaward side are also responsible.

The shift in land types in the Yellow River Estuary Nature Reserve from mid-October 2016 to late September 2021 is shown in Figure 11. It can be seen that the transition between some salt marsh vegetation (such as reeds, Suaeda salsa, and Tamarix chinensis mixed growth) and artificial wetlands (including pit ponds, cultivated land, and harvested cropland or bare land) is pronounced. Statistical analysis of the spatial distribution and land type transition in various regions is described as follows. Firstly, the transformation between reeds and pit ponds, cultivated land, harvested cropland or bare land is significant. A total of 4050.31 hm² of harvested cropland or bare lands was converted into reeds. The areas transitioned from cultivated lands and pit ponds to reeds were 1855.68 hm² and 942.14 hm², respectively. Conversely, 1350.53 hm² of reeds were converted into cultivated land, 1070.10 hm² into harvested cropland or bare land, and 1043.00 hm² into pit ponds. Secondly, we observe a significant transformation between Suaeda salsa and bare tidal flats. Specifically, 6047.20 hm² of bare tidal flats converted into Suaeda salsa, and 826.12 hm² of Suaeda salsa transitioned into bare tidal flats. The conversion also occurs between reed and black locust forests, with 993.42 hm² area shifted to reeds. A total of 967.76 hm² of Suaeda salsa transformed into Tamarix chinensis mixed growth, whereas 815.70 hm² of Tamarix chinensis mixed growth shifted into Suaeda salsa. The additional Spartina alterniflora area mainly arises from bare tidal flats (693.24 hm²) and water areas (678.71 hm²). This reflects the continuous expansion of Spartina alterniflora at the boundaries of bare tidal flats and water areas on the map.

In addition, the transition between non-saltmarsh vegetation types mainly occurs in harvested cropland or bare land, cultivated land, pit ponds, water areas, and bare tidal flats. Specifically, 9868.03 hm² of harvested cropland or bare lands were converted into cultivated lands, 1291.44 hm² of pit ponds shifted into harvested cropland or bare lands, 980.51 hm² of harvested cropland or bare lands were restored to pit ponds, 8107.63 hm² of water areas changed into bare tidal flats, while 696.28 hm² of bare tidal flats converted into water areas.

The percentage of conversion regarding various land types in the monitoring area is listed in

Table 6. Percentage of land cover type conversion from 2016 to 2021 (%).

	2016	Reed	Suaeda salsa	Spartina alterniflora	Tamarix chinensis Mixed Growth	Bare Tidal Flats	Water Areas	Pit Ponds	Black Locust	Cultivated Land	Harvested Cropland or Bare Land
2021
Reed		53.65	2.23	4.87	10.59	1.24	0.02	5.77	52.63	23.05	19.68
Suaeda salsa		0.74	54.84	0.01	19.07	35.83	0.13	1.51	0	0.01	2.69
Spartina alterniflora		0.36	1.33	87.00	0.19	3.77	1.10	0.006	0.002	0.00	0.16
Tamarix chinensis mixed growth		9.99	12.36	0.22	45.09	2.30	0.02	0.72	0.006	1.74	3.00
Bare tidal flats		1.93	22.62	5.07	6.53	50.04	15.79	3.33	0.008	0.21	1.92
Water areas		0.99	2.46	2.81	1.28	4.14	82.52	3.53	0.006	0.05	1.59
Pit ponds		8.30	0.21	0.0012	1.73	0.51	0.19	64.93	0.58	3.82	5.83
Black locust		1.21	0.0006	0.01	0.0002	0.0002	0.00002	0.10	35.55	1.41	0.82
Cultivated land		12.05	0.04	0.0003	3.66	0.08	0.004	3.74	9.90	60.91	49.41
Harvested cropland or bare land		10.77	3.91	0	11.86	2.08	0.22	16.35	1.32	8.79	14.91
Type conversion		46.35	45.16	13.00	54.91	49.96	17.48	35.07	64.45	39.09	85.09

. In 2016, 46.35% of reeds, 45.16% of Suaeda salsa, 13% of Spartina alterniflora, and 54.91% of Tamarix chinensis mixed growth underwent transitions. Bare tidal flats, black locust forests, and harvested cropland or bare lands experienced area loss of 49.96%, 64.45%, and 85.09%, respectively.

The conversion between various natural wetlands (including reeds, Suaeda salsa, Spartina alterniflora, Tamarix chinensis mixed growth, and bare tidal flats) is also noticeable. Specifically, 9.99% of the reed areas were transformed into Tamarix chinensis mixed growth. Suaeda salsas were transformed into Tamarix chinensis mixed growth (12.36%) and bare tidal flats (22.62%). The Spartina alternifloras had 5.07% and 4.87% of areas lost to bare tidal flats and reeds, respectively. In addition, 19.07%, 10.59%, and 6.53% of Tamarix chinensis mixed growth were transformed into Suaeda salsas, reeds, and bare tidal flats, respectively. The transition areas from the bare tidal flats into reeds, Suaeda salsas, Spartina alternifloras, and Tamarix chinensis mixed growth were 1.24%, 35.83%, 3.77%, and 2.3%, respectively. Furthermore, 49.41% of the harvested cropland or bare lands in 2021 were converted into cultivated lands. The main reason is that mature crops were harvested in most of the cultivated areas by mid-October 2016, while the seedlings were in the vigorous growth period in early September 2021. Water areas are the largest type of land cover across the entire monitoring area, and 15.79% of the area was converted into bare tidal flats. This is mainly due to the sediment-laden waters of the Yellow River, which form estuarine tidal flats. The other reason is associated with the changes in tidal flats area caused by tidal, water level, and climatic conditions between the two imagery periods.

5. Conclusions

This paper proposed a CGDBN classification model framework using Sentinel-2 satellite remote sensing images and field observations. The DBN model network parameters were updated based on the classic FR algorithm and PRP algorithm in CG. Given accuracy comparison results, remote sensing interpretation and monitoring were conducted using the FRDBN classification model. Specifically, we extracted wetland feature types and dynamic changes in wetland areas and analyzed alterations in the distribution characteristics of land cover types. The results are summarized as follows:

(1)

The accuracy of inter-category boundaries is crucial for remote sensing classification of coastal wetlands under complex distribution conditions. The existing algorithms face boundary offset, contraction, or expansion during ground feature classification. These deficiencies are relatively pronounced when dealing with isolated patches, mainly due to the marked influence of spatial correlation features. The proposed FRDBN model can solve the above issues.

(2)

The monitoring of land cover type changes revealed that the wetlands area expanded rapidly from 2016 to 2021, with an increase of 8.55%, especially in the Yellow River Estuary area. The distribution of Spartina alterniflora changes dramatically and grows rapidly towards the seaside. The conversion ratio between natural wetlands and constructed wetlands is the highest within land features. The monitoring results indicate intense mutual transformation of land types in the study area. These changes are primarily caused by factors such as the runoff of the Yellow River, the invasion of exotic species, and human activities.

Based on the findings of this study, the CGDBN model will be extended to other deltaic regions and be integrated with socio-economic data for policy implications research. In the future, we will also explore the remote sensing monitoring research of different vegetation carbon sinks.