Spatial and Temporal Changes of Typical Vegetation in the Yellow River Delta Based on Zhuhai-1 Hyperspectral Data

Abstract

The Yellow River Delta wetland boasts a diverse range of vegetation species and harbors an ecosystem that is both sensitive and fragile, so it is of great practical significance to accurately extract the vegetation information and analyze the spatial and temporal changes of this region. Based on the hyperspectral image data of Zhuhai-1, this study selected the characteristic bands through the continuum removal method. It combined these spectral attributes with index-based characteristics, utilizing the random forest algorithm to classify prevalent vegetation types while subjecting the outcomes to thorough analysis. It was shown that (1) when integrating spectral features, red edge indices, water indices, and vegetation indices to classify five distinct vegetation types in the Yellow River Delta during 2020 and 2022, the random forest classification algorithm showed higher classification accuracy, and the study achieved commendable overall classification accuracy rates and Kappa coefficients of 85.92% and 0.84, and 86.25% and 0.84, respectively. (2) In 2020 and 2022, the distribution of vegetation in the Yellow River Delta exhibited the following order: Suaeda glauca > Phragmites > Spartina alterniflora > Tamarisk > Typha orientalis Presl. With the exception of Spartina alterniflora, all categories of vegetation witnessed an increase in their distribution areas. Phragmites experienced the most significant growth, with an area expansion of 9.42%. (3) The ecological restoration and management measures taken in the Yellow River Delta have proven notably effective. The proportion of Spartina alterniflora within the vegetation decreased by 3.45%, the native vegetation showed a resurgence, the distribution pattern of vegetation communities moved toward stability, and the total area of vegetation in the study area exhibited an upward trajectory.

1. Introduction

The wetland ecosystem is often referred to as the “kidney of the earth”, and it plays an important role in regulating the climate and maintaining biodiversity [1,2]. The Yellow River Delta wetland is the world’s most extensive, well-preserved, and youngest wetland ecosystem in the warm temperate zone. However, it faces challenges to its stability due to both environmental factors and human activities, leading to significant changes in wetland resources [3,4]. The Yellow River Delta wetland is home to a rich variety of vegetation [5]. However, the complex and indistinct boundaries between different vegetation types in this area can lead to reduced classification accuracy on remote sensing images due to the phenomenon of “the same object with different spectra, the same spectrum with different objects” [6]. Accurately extracting wetland vegetation information and monitoring its temporal dynamic changes are of great practical significance for the conservation and restoration of the wetland ecology in the Yellow River Delta.

Hyperspectral image data are characterized by a narrow band, high feature dimensions, and high spectral resolution, which make them well suited to the identification of spectral details and features, and they exhibit greater potential in analyzing the reflectance spectral variability of vegetation [7]. At present, some promising results have been yielded in the use of hyperspectral data for vegetation information extraction and classification. For example, Chen Y B [8] and others analyzed the spectral characteristics of five typical vegetation species based on the measured hyperspectral data of the typical vegetation species in Poyang Lake and extracted the spectral difference bands by comparing the original spectra with the differential spectra, so as to effectively classify the five species of vegetation. Huang J [9] and others, based on the measured hyperspectral data of five typical vegetation species in the Honghe area, compared 38 vegetation indices, demonstrating the applicability of hyperspectral vegetation indices and evaluating the classification of the five vegetation species. Zhang Y C [10] et al. realized the fine classification of the Zalong wetland based on the HJ-1A hyperspectral image by comparing and analyzing the FCLS algorithm and SUFCLS algorithm. Gao C J [11] et al., based on coupled WorldView-2 and Zhuhai-1 images, utilized the random forest algorithm to compare the effects of four different feature selection algorithms and different spatial resolutions on the classification accuracy of mangrove species. Bian F Q [12] et al. utilized Zhuhai-1 hyperspectral data to extract the research threshold range through the NDWI-NDVI index thresholding method. They compared four different supervised classification methods and achieved the high-precision extraction and analysis of mangrove forests in coastal areas of Guangdong Province.

In the exploration of species identification methods for remote sensing data, more and more studies are integrating hyperspectral remote sensing with machine learning algorithms for plant classification. This integration has proven effective in enhancing the accuracy of identification and classification, leading to generally higher classification accuracy [13]. Li C et al. [14] extracted 63 spectral feature variables of eight plants in agricultural regions. They employed three machine learning methods—k-nearest neighbors (KNN), support vector machine (SVM), and random forest (RF)—for the classification and identification of these plants. Liu P [15] conducted research on a plant classification and identification model based on KNN and RF, using the extracted feature bands in urban greening and economic crop detection. Yang H Y [16] et al. constructed a random forest classification model with 24 variables. They compared this model with other classifications of machine learning algorithms, such as SVM, KNN, and a maximum likelihood classifier (MLC), to classify the vegetation of a desert grassland. Their results conclusively demonstrated the superior effectiveness of RF classification.

Currently, great progress has been made in the application of hyperspectral data for vegetation information extraction and classification, but it is essential to recognize that satellite hyperspectral images often have a limited spatial resolution. Moreover, the mixed distribution and high similarity in spectral signatures among different vegetation types in wetland areas underscore the critical importance of feature extraction and selection [17]. This study employs Zhuhai-1 satellite images, which integrate a high spatial resolution with hyperspectral resolution, for the innovative assessment of recent changes in five prevalent wetland vegetation types within the Yellow River Delta. A separability lookup table for the range of differences in the five vegetation bands is established, combining the random forest algorithm, support vector machine, and maximum likelihood estimation to classify the study area. Subsequently, based on the classification results obtained from the superior algorithms, statistical analyses are conducted to scrutinize changes in the area of different vegetation types. This study offers guidance for the safeguarding and surveillance of the wetland flora in the Yellow River Delta.

2. Study Area and Data

2.1. Overview of the Study Area

The Yellow River Delta is located in Dongying City, Shandong Province (37°40′ N~38°10′ N, 118°41′ E~119°16′ E), and bordered by the Bohai Sea in the north and Jiaozhou Bay in the south, encompassing a total area of about 5400 km². The region has an average elevation of less than 15 m and falls under the classification of a temperate continental monsoon climate [18]. The region experiences well-defined seasons with an average annual temperature of 12.2 °C and annual precipitation ranging from 540 to 620 mm, primarily concentrated in July and August. The average annual evaporation is 2200 mm [19]. The soils of the region are dominated by tidal and saline soils, which were formed by the Yellow River sediment deposition [20].

The study area (Figure 1) is situated in the eastern part of the modern Yellow River Delta. It primarily comprises coastal wetlands and contains mudflats, grassy swamps, pits and ponds, etc. The predominant land type in this area is characterized by temperate deciduous broadleaf forest, and the native vegetation is dominated by Phragmites, Suaeda glauca, Tamarik, and a small amount of Typha orientalis Presl. Additionally, there is an invasive vegetation species, Spartina alterniflora [21]. This paper focuses on the above five vegetation types (

Table 1. Typical wetland plants of the Yellow River Delta.

Botanical Name	Latin Name	Subjects	Field Survey Photos
Suaeda glauca	Suaeda glauca(Bunge) Bunge	Chenopodiaceae
Tamarisk	Tamarix chinensis Lour	Tamarixaceae
Spartina alterniflora	Spartina alternifloraLoisel	Poaceae
Typha orientalis Presl	Typha orientalis Presl	Typha orientalis
Phragmites	Phragmites australis (Cav.) Trin. ex Steu	Poaceae

).

2.2. Data Sources and Preprocessing

The “Zhuhai-1” hyperspectral satellite, launched on 26 April 2018, is a commercial satellite launched and operated by Zhuhai Orbit Aerospace Science and Technology Company Limited, and its hyperspectral data contain a total of 32 bands, with spectral wavelengths ranging from 400 nm to 1000 nm. The satellite provides a spatial resolution of 10 m and a spectral resolution of 2.5 nm [12]. The data used in this study are hyperspectral remote sensing images captured by the Zhuhai-1 (OHS) hyperspectral group 02 satellite, with a data processing level of L1 (

Table 2. OHS imaging for research.

Satellite	Treatment Level	Imaging Time	Number of Bands	Spectral Range	Spectral Resolution	Spatial Resolution
OHS-2C	L1B	18 September 2020	32	400~1000 nm	2.5 nm	10 m
OHS-2C	L1B	18 September 2020	32	400~1000 nm	2.5 nm	10 m
OHS-3B	L1B	24 July 2022	32	400~1000 nm	2.5 nm	10 m
OHS-3B	L1B	24 July 2022	32	400~1000 nm	2.5 nm	10 m

). The PIE-Hyp6.3 hyperspectral image data processing software was employed to preprocess the Zhuhai-1 hyperspectral image data, and the preprocessing process included radiometric calibration, 6S atmospheric correction, and orthometric correction.

From September 2020 to September 2022, field research and sampling were conducted for five vegetation species in the study area. The coordinates of the corresponding points were recorded using GPS technology, and, in some instances, multispectral images of some points were captured by a UAV to assist in the selection of sample points during the process of vegetation image classification. The number of sample points obtained varied among different vegetation types based on factors such as their respective growing areas and the level of sampling difficulty, among which there were 22 sample points for Phragmites, 24 for Suaeda glauca, 23 for Typha orientalis Presl, 12 for Tamarisk, and 17 for Spartina alterniflora.

3. Research Methods

3.1. Continuum Removal

In the case of similar spectral curves, direct analysis from the original spectral curves is not conducive to extracting spectral features, so it is necessary to conduct further processing of the original spectral curves. Removing the continuum allows for the normalization of different spectral curves to a uniform background, effectively highlighting the absorption, reflection, and scattering features of the spectral curves, which is convenient for the identification and classification of different vegetation [22,23,24]. In this study, continuum removal was performed using the ENVI5.3 software, and its calculation formula is as follows:

$$K=\frac{R_{end}-R_{start}}{{\lambda }_{end}-{\lambda }_{start}}$$

(1)

$$R_{Hi}=R_{start}+K\cdot \left({\lambda }_i-{\lambda }_{start}\right)$$

(2)

$$C{R}_i=\frac{R_i}{R_{Hi}}$$

(3)

where $K$ is the slope between the absorption start and end bands, R_start and R_end represent the reflectance of the original spectra at the absorption start and end points, λ_start start and λ_end denote the wavelengths of the absorption start and end points, R_Hi corresponds to the “Hull” value for band i, CR_i signifies the value of continuum removal for band i, and R_i is the original spectral reflectance for band i.

3.2. Principal Component Analysis (PCA)

PCA (Figure 2), or principal component analysis, is a basic and commonly used method for the linear dimensionality reduction of hyperspectral image data [25]. Its fundamental idea is to convert a large number of original variable indicators in the data into a small number of principal component indicators that can maximize the retention of the original information while remaining mutually independent of each other; we then sort and filter them according to the amount of information that they encapsulate, thus achieving the purpose of dimensionality reduction for high-dimensional data [26,27].

3.3. Random Forest

The random forest algorithm is a machine learning algorithm based on decision trees, with its core being the bagging algorithm [28]. Its basic principle is as follows. It employs bootstrap resampling from the original dataset to create K training sets, with each training set having a size equal to 63% of the original dataset. K decision tree models are constructed using these training datasets, and these models are used to predict the classification of the original dataset, ultimately yielding the classification results.

3.4. Accuracy Evaluation

The method used for the accuracy evaluation of the classification results is the confusion matrix method based on sample points. In order to ensure the objectivity and randomness of the accuracy verification, the sample points are randomly selected according to the percentage of the map area occupied by different types of features in the classification results. The evaluation indexes used include the overall accuracy, Kappa coefficient, user accuracy, and cartographic accuracy.

$$\mathrm{O}\mathrm{A}=\frac{1}{\mathrm{n}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{x}}_{\mathrm{i}\mathrm{i}}$$

(4)

$$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}=\frac{\mathrm{N}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{x}}_{\mathrm{i}\mathrm{i}}-{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{x}}_{\mathrm{i}+}{\mathrm{x}}_{+\mathrm{i}}\right)}{{\mathrm{N}}^2-{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{x}}_{\mathrm{i}+}{\mathrm{x}}_{+\mathrm{i}}\right)}$$

(5)

$$\mathrm{P}\mathrm{A}=\frac{{\mathrm{x}}_{\mathrm{i}\mathrm{i}}}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{x}}_{+\mathrm{i}}}$$

(6)

$$\mathrm{U}\mathrm{A}=\frac{{\mathrm{x}}_{\mathrm{i}\mathrm{i}}}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{x}}_{\mathrm{i}+}}$$

(7)

In the given formula, “N” stands for the total number of samples in the given formula, “N” stands for the total number of samples, and “n” signifies the total number of columns present in the confusion matrix. ${\mathrm{x}}_{\mathrm{i}\mathrm{i}}$ is the number of samples in the i-th row and column of the confusion matrix, which is the number of correct classifications; ${\mathrm{x}}_{\mathrm{i}+}$ and ${\mathrm{x}}_{+\mathrm{i}}$ represent the total number of samples in the i-th row and i-th column, respectively.

3.5. NDVI, NDWI, NDVIre

In this study, the normalized difference vegetation index (NDVI) [29], normalized difference water index (NDWI) [29], and red edge normalized difference vegetation index (NDVIre) [30] were employed to assist in the selection of classification samples. The specific formulas for these indices are as follows:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R})$$

(8)

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{M}\mathrm{I}\mathrm{R})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{M}\mathrm{I}\mathrm{R})$$

(9)

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e})/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e})$$

(10)

4. Results and Analysis

4.1. Characteristic Band Selection Analysis

With reference to the coordinates of the measured data points, five vegetation samples were selected from the preprocessed OHS hyperspectral images, and the spectral mean curves of the five vegetation samples are shown in Figure 3a. From Figure 3a, it can be seen that the spectral mean curves of the five planted vegetation samples exhibit distinct characteristics of vegetation spectral curves, i.e., there is a typical “red edge” feature within the spectral wavelength range of 680 nm–760 nm. It is worth noting that the spectral mean curve for Suaeda glauca does not show obvious characteristics of vegetation spectral curves, which is due to the fact that Suaeda glauca is more widely distributed in the Yellow River Delta and the growth density of vegetation is small, so its spectral curve is easily affected by the soil substrates. From the figure, it can be seen that the differences in the five vegetation curves were primarily observed in the near-infrared band, with less noticeable differences in the visible band.

Figure 3b displays the spectral mean curve of the vegetation after the removal of the continuum. Here, it is evident that the differences in the spectral characteristics of the vegetation in the near-infrared band have essentially been normalized and eliminated, while the differences in the spectral characteristics of the vegetation in the visible band have been accentuated. This is particularly notable in the green and red band ranges, where the differences in the spectral characteristics between different vegetation types are significantly more pronounced compared to the original spectral curves. Moreover, the gap between the maximum and minimum values of spectral reflectance within these two bands is also further widened compared to the original spectrum. The removal of the continuum further enhanced the separability between Phragmites and Spartina alterniflora compared to the original spectra. However, the separability between Tamarix and Suaeda glauca was reduced after the removal of the continuum.

Figure 4 illustrates the pairwise comparison of spectral features between Phragmites and the other four vegetation species. In the original spectra, the differences between Phragmites and Typha orientalis Presl are not prominent, both in the visible and near-infrared bands. However, after the continuum removal, the distinguishable band between Phragmites and the remaining four vegetation types is shifted from the near-infrared band of the original spectra to the visible band. Within the distinguishable band range, the spectral differences between Phragmites and the other vegetation types are significantly amplified, which demonstrates that the removal of the continuum is advantageous for the extraction of the spectral features of Phragmites.

Table 3. Divisibility lookup table (Unit: nm).

	Phragmites	Suaeda Glauca	Tamarisk	Typha Orientalis Presl	Spartina Alterniflora
Phragmites		466~709	466~730	466~700, 806~833	490~700, 806~833
Suaeda glauca	443~686, 746~940		510~560, 620~709	560~730	443~531, 580~746, 820~850, 896~940
Tamarisk	490~686, 746~940	560~709		490~730	443~550, 580~730, 806~833
Typha orientalis Presl	490~686, 760~896	560~709, 760~926	760~940		443~580, 620~746, 820~850, 896~926
Spartina alterniflora	746~940	510~896	510~686, 746~896	746~926

presents a lookup table of the spectral characteristics of the five vegetation types before and after the continuum removal. In this table, the spectral curves of the five vegetation types after the removal of the continuum are the upper-right blue sections of the two comparative graphs of the distinguishable bands, while the original spectral curves are the lower-left green sections of these comparative graphs. By intersecting these bands, wavelength ranges of 466–709 nm and 806–833 nm were selected as the feature bands. Following the selection of the feature bands, principal component analysis (PCA) was employed for the dimensionality reduction of the OHS hyperspectral remote sensing image data. After this reduction, it was observed that the first three principal components already contained more than 99% of the original information. Consequently, the first three principal components were selected to synthesize the effect map to serve as the reference map for the subsequent vegetation classification phase.

4.2. Vegetation Classification and Accuracy Verification

Hyperspectral image data are characterized by narrowing wavelength bands, high feature dimensions, a high spectral resolution, etc., and they can better reveal the details and features of the spectrum, making hyperspectral images advantageous in the selection of spectral indices. In order to select the optimal classification sample data, this study devised three experimental schemes: spectral features + water index + vegetation index (Scheme 1), spectral features + red edge index (Scheme 2), and spectral features + water index + vegetation index + red edge index (Scheme 3). It is important to note that the OHS hyperspectral image incorporates seven red edge bands. Following the elimination of the maximum and minimum values, five bands within the wavelength range of 686–746 nm are employed as the operational bands for the red edge index. The characterizations are detailed in

Table 4. Description of features.

Characteristic Variable	Short Title	Description of Features
Spectral features	Band	B2–B17, B23–B25
Vegetation index	NDVI	(B27–B13)/(B27 + B13)
Water index	NDWI	(B27–B13)/(B27 + B13)
Red edge index	NDVIred1	(B27–B15)/(B27 + B15)
	NDVIred2	(B27–B16)/(B27 + B16)
	NDVIred3	(B27–B17)/(B27 + B17)
	NDVIred4	(B27–B18)/(B27 + B18)
	NDVIred5	(B27–B19)/(B27 + B19)

.

Referring to the field sampling coordinates, three experimental schemes were employed to select the classification samples. Utilizing the JM distance, the separability of category samples selected by different schemes was analyzed. After multiple trials, it was observed that the separability of the Suaeda glauca vegetation and inland mudflats was comparatively low. Conversely, the separation of other category samples was satisfactory. As a result, we ranked the separation between Suaeda glauca vegetation and inland mudflats, from high to low. Consequently, we determined that the highest separation of sample categories was selected by the combination of band + NDVI + NDWI + NDVIred3 in Scheme 3 (

Table 5. Average JM distance for different programs.

Characterization Schemes	Combination of Highest Separating Features	Average JM Distance
Scheme 1	Band + NDVI + NDWI	1.6040
Scheme 2	Band + NDVIred3	1.7162
Scheme 3	Band + NDVI + NDWI + NDVIred3	1.8178

). The images were categorized using random forest classification, the support vector machine method, and maximum likelihood estimation to obtain the classification results of the five typical vegetation types in the Yellow River Delta in 2020 and 2022 (Figure 5). In the study area, eight distinct land cover types were identified: clear water, turbid water, inland mudflat, Suaeda glauca, Tamarisk, Spartina alterniflora, Typha orientalis Presl, and Phragmites.

According to the classification results, 1200 sampling points were randomly generated in the ENVI software as test samples, and the samples were verified through visual interpretation by combining the data of field sampling points and remote sensing images with a 0.5 m high spatial resolution at the same time. Moreover, the confusion matrix of the classification results was calculated to verify the accuracy of the vegetation classification, as shown in

Table 6. Classification accuracy statistics.

Classification	2020						2022
	RF		SVM		MLE		RF		SVM		MLE
	PA/%	UA/%	PA/%	UA/%	PA/%	UA/%	PA/%	UA/%	PA/%	UA/%	PA/%	UA/%
Phragmites	82.00	79.35	81.33	78.21	76.00	73.08	83.33	89.29	78.67	83.10	76.00	73.08
Suaeda glauca	78.67	83.10	80.00	82.19	74.67	76.71	80.67	80.13	76.00	78.62	74.67	76.71
Typha orientalis Presl	83.33	73.53	73.33	66.67	67.33	61.21	78.00	83.57	73.33	72.85	67.33	61.21
Tamarix	79.33	79.33	78.00	78.52	78.00	78.52	82.67	77.02	80.00	78.43	72.67	73.15
Spartina alterniflora	85.33	85.91	84.67	83.01	70.00	68.63	86.67	82.28	84.67	84.11	70.00	68.63
Inland mudflat	90.00	90.00	80.00	83.92	80.00	83.92	85.33	80.00	82.00	72.35	73.33	76.92
Clear water	94.00	100.00	96.67	100.00	96.67	100.00	99.33	100.00	97.33	100.00	96.67	100.00
Turbid water	95.33	100.00	95.33	100.00	95.33	100.00	94.00	100.00	94.67	100.00	95.33	100.00
Overall accuracy	85.92		83.67		79.75		86.25		83.33		78.25
Kappa coefficient	0.84		0.81		0.76		0.84		0.809		0.75

PA: producer’s accuracy, UA: user’s accuracy.

. From Table 6, it can be seen that the classification accuracy of the random forest-based classification algorithm is significantly higher than that of the support vector machine classification algorithm and the maximum likelihood estimation. The overall classification accuracy and Kappa coefficient of the random forest classification algorithm are 85.27% and 0.82, and 86.59% and 0.84, respectively. Compared with the verification results of the single classification accuracy, the classification effect of Spartina alterniflora is notably superior, with mapping accuracy exceeding 85%. On the contrary, the classification effect of Suaeda glauca is less impressive, and the mapping accuracy is only 78.89%. Meanwhile, the classification effect of Tamarisk, Typha orientalis Presl, and Phragmites is moderate, and the mapping accuracy is between 80% and 85%.

4.3. Analysis of Vegetation Area Changes

The changes in the area and percentages of typical vegetation in the study area from 2020 to 2022 are shown in Figure 6. Notably, except for Spartina alterniflora, the area of the other four vegetation species increased. Among these, Phragmites exhibited the most rapid growth, followed by Suaeda glauca, Tamarisk, and Typha orientalis Presl, which increased by 5.27 km², 5.09 km², 0.75 km², and 0.25 km², respectively. The Suaeda glauca community covers the largest percentage of the mapped area in the study area, which is related to the growth characteristics of the Suaeda glauca community, which is sparsely dispersed. In 2022, the percentage of the area increased by 1.38% compared with that of 2020. The Tamarisk community and Typha orientalis Presl community accounted for a relatively small area, and the overall change trend was not significant. The area of the Spartina alterniflora community decreased by 4.45 km², reducing its share from 30.28% to 26.83%.

Table 7. Main areas of vegetation change.

	2020 Raw Images	2022 Raw Images	Classification Results for 2020	Classification Results for 2022	Major Changes in Vegetation
a					Spartina alterniflora
b					Spartina alterniflora
c					Spartina alterniflora, Suaeda glauca
d					Phragmites

presents the before-and-after comparisons of the four regions with significant vegetation changes between 2020 and 2022 in the study area (Figure 7). Notably, regions a, b, and c saw changes in Spartina alterniflora vegetation. Through a review of relevant information, Shandong Province officially launched the Implementing Measures for Prevention and Control of Spartina Alterniflora in Shandong Province in 2020. Based on the images and the documented measures for the management of huperzia, it can be concluded that areas a, b, and c have adopted the “mowing + flooding” and mowing management measures, respectively, which also confirms the view that the area of Spartina alterniflora has been reduced, as mentioned above. Area d comprises an ecological restoration pool within the study area. The vegetation that has experienced more noticeable changes is Phragmites. As this region is primarily characterized by a water body, it is greatly influenced by precipitation and seasonal variations. When the water level drops, revealing bare land, there is rapid vegetation growth, which leads to a change in the cover type.

In terms of distribution patterns, the Spartina alterniflora community continues to maintain a significant presence on both sides of the estuary. However, since 2020, Dongying City has implemented various management measures for Spartina alterniflora, such as “mowing”, “mowing + flooding”, “mowing + tilling”, and so on, depending on different areas. After adopting the above management measures, the trend of the invasion of Spartina alterniflora communities into the Phragmites communities on both sides of the river channel has slowed down, and it has gradually formed a pattern of staggered distribution with the Phragmites communities, presenting a trend of extension to the sea. The Phragmites communities are mainly distributed on both sides of the Yellow River channel. Overall, their distribution tends to be interconnected and clustered. However, due to the impacts of other vegetation types, such as Spartina alterniflora, and human activities, some Phragmites communities exhibit fragmentation. The Suaeda glauca community is mainly distributed in the southern saline mudflats and some mudflats in the north. The community has a wider distribution area but is characterized by sparse growth. The spectral reflectance of remote sensing images is easily affected by the mudflat substrate, so it shows a larger fragmentation feature in the classification map, particularly in the northern region. The Tamarisk community and Typha orientalis Presl community are niche vegetation communities in the study area, with a relatively small area of vegetation, usually mixed with reed communities growing on both sides of the river, with a stable community distribution trend.

Hyperspectral remote sensing images offer the advantage of a high spectral resolution, which is a great advantage in discriminating different vegetation types. However, hyperspectral remote sensing images have the common problem of a low spatial resolution, causing errors in the classification results. Simultaneously, the random forest classification algorithm exhibits better accuracy and versatility compared with other traditional classification algorithms, but the classification algorithms employed in this study primarily rely on implementation within the ENVI software. Moreover, there are still more advanced algorithms for machine learning and deep learning that can be used in this field of study, and we will continue to explore and test the content of the two areas in the future research.

5. Conclusions

In this study, based on the hyperspectral remote sensing image data of “Zhuhai-1”, the continuum removal method was employed to process the images. It focused on five typical vegetation species in the Yellow River Delta, extracting and selecting characteristic difference bands among the vegetation. Additionally, a feature tuning scheme was designed to select appropriate classification sample categories, and the separability among the categories was verified. Finally, the classification of remote sensing images of five typical vegetation species in the Yellow River Delta was completed by the random forest classification method, and the spatial and temporal changes of vegetation in the study area were monitored and analyzed with the classification results. The following conclusions are drawn.

(1) Combining the spectral features, red edge index, water index, and vegetation index, the classification results obtained from the random forest classification algorithm, support vector machine classification algorithm, and maximum likelihood estimation were compared for accuracy, revealing higher accuracy with the random forest classification. The overall classification accuracy and Kappa coefficient of the five typical vegetation species in the study area in 2020 and 2022 were 85.92% and 0.84, and 86.25% and 0.84, respectively.

(2) In both 2020 and 2022, the ranking of the area share of the five vegetation species in the Yellow River Delta remained unchanged, with Suaeda glauca > Phragmites > Spartina alterniflora > Tamarisk > Typha orientalis Presl. Except for Spartina alterniflora, the distribution area of each type of vegetation increased. Phragmites exhibited the most growth, with a 9.42% increase in area.

(3) The ecological restoration and management measures implemented in the Yellow River Delta have proven effective. The proportion of Spartina alterniflora in the vegetation has decreased, the native vegetation has shown signs of recovery, and the distribution pattern of the vegetation communities is becoming more stable. Moreover, the total area covered by vegetation in the study area is exhibiting an upward trend.