Donghai Island (Figure 1) covers an area of 289 km² (latitude 20°55’–21°05’ N; longitude 110°11’–110°33’ E), making it the largest island in Guangdong Province and the fifth largest island in China. The island experiences a subtropical monsoon climate, regulated year-round by maritime influences, with an average annual temperature range of 22.7°C–23.5°C and annual precipitation of 1,395.5–1,723.1 mm. S. alterniflora grows in mangroves near Beishan Harbor and as abundant independent patches within a large number of tidal flats, which together form an area approximately 4.4 km long, accounting for the largest expanse of S. alterniflora in Zhanjiang, China.

The Jeffries–Matusita (JM) distance is an important metric for assessing the spectral separability between two land-cover classes within a region of interest (Bruzzone et al., 1995). The JM distance, calculated based on the differences between the mean vectors and covariance matrices of the two classes, ranges from zero to two, where values closer to two indicate better separability in the feature space, with classes easier to distinguish using a classifier, and values closer to zero indicating that classes are increasingly difficult to differentiate (Jeffreys, 1997). Generally, JM > 1.9 indicates very good separability, whereas JM < 1 may warrant caution when merging the two samples into a single class. For classes that follow a normal distribution, the JM distance can be calculated based on Formulas 1, 2: J M ij = 2 * 1 − e ‐ B ij (1) B = 1 8 m i − m j T C i + C j 2 − 1 m i − m j + 1 2 ln C i + C j 2 C i * C j 1 / 2 (2) where i and j represent the different land-cover types; m _i and m _j denote the mean values; and C _i and C _j denote the covariance matrices of the classes.

After meeting the separation requirements, the features of different groups were input into the RF, SVM, MLC, MaDC, and MDC models, and the accuracy of the classification results was evaluated using the verification set. In addition, sa and nsa area data classified using various methods were counted. A flowchart of the classification process is shown in Figure 2.

A confusion matrix was used to evaluate the performance of the classification models; this matrix displays the relationships between the actual classes and the classes predicted by the model. Various metrics were then calculated from the confusion matrix, including the overall accuracy (OA), producer accuracy, and user accuracy. The kappa coefficient is a statistical measure used to assess the level of agreement between groups and classifications; a kappa value closer to one indicates high consistency and accuracy, whereas a value closer to zero indicates low accuracy and poor agreement. The kappa coefficient was calculated based on Formula 3: Kappa = N ∑ i = 1 n m i , i − ∑ i = 1 n G i C i N 2 − ∑ i = 1 n G i C i (3) where n represents the total number of land-cover classes; N denotes the total number of pixels in the dataset; m _i,i refers to the number of pixels in the i-th row and i-th column of the confusion matrix, indicating the number of correctly classified pixels for class i; G _i is the total number of predicted samples in the classification results for class i; and C _i is the total number of actual samples in the ground truth for class i.

The original spectral profiles of different land-cover classes, including S. alterniflora, mangrove intertidal zones, water bodies, and mud, are shown in Figure 3 . S. alterniflora showed minimal distinction from water bodies in the visible bands, whereas significant differences (>500 nm) were evident in the red-edge, near-infrared, B9, B11, and B2 bands. The reflectance values of S. alterniflora and intertidal zones were comparable in the visible bands, showing no significant differences in the B5 and B12 bands, whereas the differences in all other bands were >800 nm. Spectral differences between S. alterniflora and mangroves were similar to those between S. alterniflora and wetlands, primarily occurring in the B6–B9 and B11 bands. In summary, the bands that differentiate S. alterniflora from the other land-cover types primarily fell within the non-visible bands. Figure 4 shows that the trend of the sa dataset aligned with the spectral curves, whereas the nsa dataset showed minimal variation with increasing wavelength. The spectral bands exhibiting differences between sa and nsa datasets were B6–B9 and B11, which is similar to the spectral differences between S. alterniflora and mangroves. Regarding the SIs, the mean value for the nsa dataset ranged from 0.01 to 0.10, whereas that for the sa dataset ranged from 0.21 to 0.50 (Figure 5). Significant SI differences were observed between the two datasets, which is beneficial for classification.

The differences in spectral bands among datasets yielded JM distance values of 1.931 and 1.900 for training and validation subsets, respectively, indicating good separability (Table 2). The separability values were lowest for the SI datasets, at 1.897 and 1.893 for training and validation sets, respectively, whereas the GLCM datasets exhibited exceptionally high separability, with both subsets exhibiting JM distances of 2.000, making them well-suited for classification tasks. As for the different feature combinations, combining spectral bands with the other two feature categories significantly enhanced separability, with Group 1 training and validation sets showing separation values of 1.997 and 1.998, respectively, and Group 2 achieving values of 2.000 for both subsets. Thus, a combination of different features is beneficial for enhancing the classification distance and facilitating the classification process. It should be noted that the combination of features in Group 3 could not be calculated using the JM distance because of substantial differences in the numerical range.

The classification results of the different algorithm–feature combinations are shown in Supplementary Figures S1–S30. The mean kappa coefficients for the five classification methods (RF, SVM, MLC, MDC, and MaDC) were 0.87, 0.83, 0.88, 0.60, and 0.82, respectively (Figure 6). The standard deviation of the RF classification method was 0.021. As the number of features increased, the kappa coefficient and OA of RF increased by 0.05% and 2.52%, respectively indicating relatively stable performance across different features, with an OA greater than 90% (Figure 7b). The MLC method achieved the best performance with spectral features, achieving a kappa coefficient of 0.9061 and OA of 95.32%, but showed the worst performance when combined with SIs, with a kappa coefficient of 0.82 and OA of 91.01% (Figure 7a). Other features also performed well in MLC classification. For the feature combinations, classification performance improved slightly when more features were included. In SVM classification, the groups that included GLCM features exhibited better classification performance, with an OA of 93.53% (Figure 7c). The MDC classification method showed the poorest overall performance, with kappa coefficients and OA values less than 0.80% and 90% for all feature classifications, respectively (Figure 7d). MDC classification using GLCM features showed the worst combined performance, with a kappa coefficient of only 0.13, whereas the best performing feature was SIs, with a kappa coefficient of 0.77. Similar to SVM, the MaDC method achieved high classification accuracy for groups that included GLCM features, with OA values for GLCM and Group 3 features of 93.53% and 93.88%, respectively (Figure 7e).

The average area of S. alterniflora extracted by the RF, SVM, MLC, and MaDC methods were 75.66, 76.68, 77.65, and 82.73 ha, respectively (Table 3). After removing the outliers from GLCM features, the average area extracted by the MDC method was 81.64 ha. The area of S. alterniflora on Donghai Island extracted through visual interpretation was 68.69 ha. RF exhibited relatively small fluctuations in the extracted area with different feature combinations, showing a maximum difference of 2.43 ha. Although the area extracted by MLC combined with band and Group 3 features was closest to that obtained from visual interpretation, MLC exhibited relatively large fluctuations with different feature combinations, showing a maximum difference of 12.63 ha.

As well as occupying coastal zones, S. alterniflora commonly coexists with salt marsh plants or mangroves. Therefore, image extraction using only spectral or phenological features alone often yields suboptimal results. Here, we combined spectral and phenological features to enhance identification accuracy and obtain more reliable target samples. The key spectral bands differentiating S. alterniflora from the other three land-cover types were found in the non-visible bands. Green vegetation exhibits rapid changes in the red-edge band, with healthy plants displaying the red-edge phenomenon (Horler et al., 1983). July is a rapid growth period characterized by high chlorophyll contents that may even exceed those of mangroves. This resulted in a red-shift phenomenon in S. alterniflora, facilitating its differentiation from mangroves in terms of NDVI and red-edge NDVI values. During low tide, some moisture remains in intertidal zones, which results in a similar overall spectral curve to water bodies, with spectral differences predominantly limited to the short-wave infrared range. S. alterniflora vegetation exhibits strong reflectance in the infrared band, whereas tidal flats and water bodies demonstrate higher absorption rates, which helps distinguish between tidal flats and water bodies.

In the sample dataset, the spectral bands of sa and nsa datasets differed primarily in the near-infrared and shortwave-infrared bands, which is consistent with the spectral differences between terrestrial objects. The vegetation index features all utilize calculations based on the near-infrared band, and differences in the spectral indices are already present; conversely, differences in the texture features derived from the GLCM increased as the dimensionality increased. The differences among the three types of features indicated that a greater difference in the mean vector values during JM distance calculation led to higher sample separability, ensuring that the samples met the classification requirements. Thus, increasing the number of classification features used in supervised classification can improve classification accuracy to a certain extent, except for unsuitable classification methods.

Among the five supervised classification methods, RF demonstrated good adaptability to high-dimensional data. Typically, as the feature dimensions increase, RF exhibits higher accuracy than other methods (Ye et al., 2013). For example, Yan et al. (2021) used an object-based RF algorithm to quantify the expansion and decline of S. alterniflora in the Dafeng River Elk Protection Area from 1993 to 2020, and achieved an OA of greater than 90%. Moreover, Tian et al. (2020) employed RF to calculate the area and expansion rate of S. alterniflora in the Zhangjiang Estuary from 2016 to 2018, and achieved a kappa coefficient greater than 0.89. In our study, RF classification also achieved kappa coefficients greater than 0.84 and OAs greater than 92%, showing stable results across all six groups of features.

The MLC method requires land-cover features to exhibit differences and conform to a certain probability distribution. Zhou et al. (2017) combined Spot six multispectral imagery and UAV data using MLC and NDVI to estimate the coverage of S. alterniflora in Sansha Bay, and achieved an RMSE of 0.117, indicating good accuracy. Hladik et al. (2013) used multisensor data, DEM, MLC, and decision tree methods to extract different ecological types of S. alterniflora in and around the Duplin River, achieving an OA of 90% and a kappa coefficient of 0.88. Similarly, in this study, the kappa coefficients for MLC classification ranged from 0.82 to 0.91, with OAs exceeding 91%, indicating good classification performance. The band values of the samples varied with the land-cover type, resembling a probability distribution, whereas the SIs and GLCM showed no clear probability distribution, leading to lower classification accuracy.

Compared with machine learning and statistical models, machine learning achieves the best accuracy by iteratively calculating classified data; while the classification of statistical models is based on distribution assumptions, and the rationality of the assumptions affects the accuracy of classification. In reasonable samples, RF and SVM achieve relatively high accuracy through self-iterative learning; among the three statistical models, since the methods and assumptions have different requirements, MLC and MaDC show better classification effects through covariance, while MDC only calculates the distance between category centers, and the possibility of misclassification increases as the sample points become relatively close. The five methods employed in this study showed variable adaptability to the six groups of features, resulting in significant differences in accuracy. At the methodological level, when the data dimensionality was low, RF and MLC demonstrated higher classification accuracy. As the data dimensionality increased, RF exhibited the best adaptability to high-dimensional data and achieved better classification results.

According to previous studies, S. alterniflora was introduced to the eastern coastal areas of Zhanjiang in 2006 (Xin et al., 2018) but is currently distributed in multiple areas, including Gaoqiao Town in Lianjiang City, Caotan Town in Suixi County, Nansan Island in Potou District, the coastal areas of Leizhou City, Jinhe Island in Xuwen County, and Donghai Island. The area of S. alterniflora along the eastern coastline was 79.04 ha in 2016, which rapidly increased to 154.42 ha in 2018, with the area on Donghai Island reaching 65.34 ha in 2018 (Huang et al., 2021). According to our results, compared to the extent of S. alterniflora obtained by visual interpretation (68.69 ha), the area of S. alterniflora on Donghai Island increased by 3.35 ha over a 5-year period, indicating a growth rate of 5.13%. The S. alterniflora distribution in Zhanjiang Bay is primarily divided into an intertidal zone south of Donghai Island and a northern intertidal zone of Nansan Island. Influenced by factors such as tides, shorebirds, and typhoons, seeds or rhizomes can spread to suitable intertidal zones (Li et al., 2023; Zhang et al., 2023). Because of its broad tolerance to flooding and salinity (Liu et al., 2023), S. alterniflora initially occupies a favorable habitat in light sandy areas, then continuously reproduces and expands its population to compete for the ecological niches of other species, thus completing the invasion pattern of introduction–population establishment–reproductive stagnation–expansion. Therefore, we speculate that the relatively sparse mangroves and large area of exposed intertidal zone in Beishan Port, combined with the aforementioned conditions, enabled the expansion of S. alterniflora to the maximum extent in the entire western Guangdong region.

Mangroves typically comprise trees or shrubs that grow relatively slowly during their expansion into light sandy areas, with the gaps often occupied by S. alterniflora (Liu et al., 2023). This makes it difficult for mangrove plants to establish population dominance, leading to a reduction in growth space. Zhanjiang had the largest mangrove protection area and is greatly affected by S. alterniflora invasion. To protect the ecological integrity of the mangrove ecosystem, the government has implemented measures such as cutting, filling, and plastic sheeting to curb the growth of S. alterniflora, with some success (Leizhou Municipal Government, 2021; Zhanjiang Daily Network Dissemination Cente, 2023; Forest Administration of Guangdong Province, 2024). However, localized control efforts only achieve short-term effects against the widespread distribution of S. alterniflora. Over time, S. alterniflora may become reestablished in these regions or even migrate from other regions. In terms of management, simultaneous clearing of the entire region is required to eliminate as many patches as possible, along with enhanced monitoring across the city to prevent overlooked patches from forming populations. Additionally, the introduction of native plants into cleared intertidal zones should be employed to restrict potential S. alterniflora habitats.

The uncertainty of the data, the lack of actual survey data that matches the date of remote sensing images, and the certain deviations in classification due to this. In terms of the model, the classification precision of some models is too low to meet the classification requirements, and the models lack broader validation. In terms of parameters, the automatic calculation parameters of ENVI are used, and the correlation between parameter adjustment and classification accuracy has not been studied in detail.