Monitoring the Invasive Plant Spartina alterniﬂora in Jiangsu Coastal Wetland Using MRCNN and Long-Time Series Landsat Data

: Jiangsu coastal wetland has the largest area of the invasive plant, Spartina alterniﬂora ( S. alterniﬂora ), in China. S. alterniﬂora has been present in the wetland for nearly 40 years and poses a substantial threat to the safety of coastal wetland ecosystems. There is an urgent need to control the distribution of S. alterniﬂora . The biological characteristics of the invasion process of S. alterniﬂora contribute to its multi-scale distribution. However, the current classiﬁcation methods do not deal successfully with multi-scale problems, and it is also difﬁcult to perform high-precision land cover classiﬁcation on multi-temporal remote sensing images. In this study, based on Landsat data from 1990 to 2020, a new deep learning multi-scale residual convolutional neural network (MRCNN) model was developed to identify S. alterniﬂora . In this method, features at different scales are extracted and concatenated to obtain multi-scale information, and residual connections are introduced to ensure gradient propagation. A multi-year data uniﬁed training method was adopted to improve the temporal scalability of the MRCNN. The MRCNN model was able to identify the annual S. alterniﬂora distribution more accurately, overcame the disadvantage that traditional CNNs can only extract feature information at a single scale, and offered signiﬁcant advantages in spatial characterization. A thematic map of S. alterniﬂora distribution was obtained. Since it was introduced in 1982, the distribution of S. alterniﬂora has expanded to approximately 17,400 ha. In Jiangsu, the expansion process of S. alterniﬂora over time was divided into three stages: the growth period (1982–1994), the outbreak period (1995–2004), and the plateau period (2005–2020). The spatial expansion direction was mainly parallel and perpendicular to the coastline. The hydrodynamic conditions and tidal ﬂat environment on the coast of Jiangsu Province are suitable for the growth of S. alterniﬂora . Reclamation of tidal ﬂats is the main factor affecting the expansion of S. alterniﬂora


Introduction
Spartina alterniflora (S. alterniflora), which is a perennial herb, is native to the Atlantic coast of North America and inhabits tidal environments, such as estuaries and bays [1]. With its robust adaptability and reproductive ability, it quickly takes root and spreads in a on identification from small-area or mono-temporal image data [33,34] and the analysis of spectral phenological characteristics of S. alterniflora [35].
The geographical features of China's coastal zone are complicated and diverse. S. alterniflora shows different invasion abilities in different areas, and the competitive relationship between S. alterniflora and local vegetation also varies. The coastal zone of Jiangsu has distinct geographical features, with a straight muddy coast from north to south, and it has the largest tidal flat area in China. In Jiangsu Province, S. alterniflora has invaded and expanded for nearly 40 years, and it has the largest distribution area in China. S. alterniflora appears at various states and scales during the invasion process, such as small discrete patches at the early stage of invasion; connected strips, large broken patches, and mixed patches at the mid-invasion stage; and large monoculture patches at the end of the invasion stage. These different states exist at the same time in any intrusion area. For the identification of ground objects such as S. alterniflora with complex growth states, the classical CNN model causes the gradient to disappear as the number of layers increases. Therefore, in the current study, the MRCNN model and post-processing method were proposed to obtain multi-temporal identification of S. alterniflora. An MRCNN combines the advantages of the Inception [36,37] and ResNet [38,39] CNN models. First, the MRCNN extracts features on different scales and obtains rich multi-scale information [40], and second, the residual structure is used to directly connect the shallow and deep layers, and the larger gradient is directly propagated to the deeper layer, which reduces the computational complexity. Using long time series remote sensing data for the large coastal wetland area in Jiangsu, this study proposed a deep learning MRCNN model suitable for the identification of S. alterniflora in Jiangsu to expand the application potential of deep learning. A unified training method for multi-temporal Landsat data was also adopted to improve the temporal scalability of the deep learning model. Field survey data, Landsat data, and the classification method proposed in this study were used to identify S. alterniflora in a complex coastal wetland environment and a thematic map of S. alterniflora distribution in Jiangsu Province from 1990 to 2020 was created. Finally, this study analyzed the influencing factors and the temporal and spatial rules of S. alterniflora expansion to provide reliable data and technical support for remote sensing monitoring and scientific management of S. alterniflora in Jiangsu Province.

Study Area
The study area is located along the coast of Jiangsu Province, China (between 119 • 21 E and 121 • 55 E, 31 • 33 N and 35 • 07 N) (as shown in Figure 1). The coastline of Jiangsu Province has a total length of 954 km. It stretches from the Xiuzhen estuary in the north to the estuary of the Yangtze River in the south and faces the Yellow Sea in the east. There are three cities in the coastal zone of Jiangsu, namely Lianyungang, Yancheng, and Nantong. Controlled by the monsoon climate, the coastal zone of Jiangsu is located in the transition zone between the north subtropical zone and the warm temperate zone and is influenced by both maritime and continental climates. The average annual temperature is 14 • C, and the average annual rainfall is 1000 mm [41]. The average tidal range along the coast of Jiangsu is 1-4 m. Many rivers enter the sea in this region, and the tidal salinity is low. There is a large muddy tidal flat, and the upper part of the intertidal zone suitable for the growth of S. alterniflora can be up to 4 km wide [42]. Both the climate and the tidal flat sediment provide suitable conditions for the growth and expansion of S. alterniflora [43]. More than 90% of the coastal areas in Jiangsu are muddy, and Jiangsu currently has the largest area of tidal flats in China [44]. The Red-Crowned Crane National Nature Reserve (RCCNNR) and the Elk National Nature Reserve (ENNR) are both located in Yancheng, central Jiangsu Province. The wetland vegetation in the reserves mainly includes S. alterniflora at the forefront of the tidal flat, Phragmites australis in freshwater areas, and Suaeda in the high intertidal zone [45]. Yancheng has become the largest area of S. alterniflora in China [46].
Owing to policy changes, the coastline of Jiangsu Province has been greatly disturbed by human activities. Since 1996, a strategy of Marine Sudong, which is characterized by large area development, has been implemented in the Yancheng coastal wetland, accelerating the pace of reclamation. Large areas of coastal wetlands have been used to build artificial aquaculture ponds. In 2019, the Yancheng Nature Reserve was listed as a World Natural Heritage Site, which restricts and regulates further human activities [45].

Data and Pre-Processing
Jiangsu Province introduced S. alterniflora in many areas in 1982 [4]. In the incubation period of early introduction (1982)(1983)(1984)(1985)(1986), S. alterniflora cannot be identified in Landsat images with a spatial resolution of 30 m. Therefore, this study obtained a total of 21 original Landsat images for seven time periods at 5-year intervals: 1990, 1995, 2000, 2005, 2010, 2015, and 2020 (as shown in Table 1). The collected images were cloud-free in the study area. The time of the image data is very important for accurately distinguishing S. alterniflora from other salt marsh plants. On the basis of the local phenological characteristics of S. alterniflora, the selected remote sensing images were concentrated in August-October, when S. alterniflora was in its most vigorous growth period [47,48], and its spectral features were well differentiated from indigenous vegetation. However, because of the large study area and the need to avoid cloud-cover in the images, it was difficult to obtain images that completely covered the study area and met the monthly conditions within a single year. Owing to policy changes, the coastline of Jiangsu Province has been greatly disturbed by human activities. Since 1996, a strategy of Marine Sudong, which is characterized by large area development, has been implemented in the Yancheng coastal wetland, accelerating the pace of reclamation. Large areas of coastal wetlands have been used to build artificial aquaculture ponds. In 2019, the Yancheng Nature Reserve was listed as a World Natural Heritage Site, which restricts and regulates further human activities [45].

Data and Pre-Processing
Jiangsu Province introduced S. alterniflora in many areas in 1982 [4]. In the incubation period of early introduction (1982)(1983)(1984)(1985)(1986), S. alterniflora cannot be identified in Landsat images with a spatial resolution of 30 m. Therefore, this study obtained a total of 21 original Landsat images for seven time periods at 5-year intervals: 1990, 1995, 2000, 2005, 2010, 2015, and 2020 (as shown in Table 1). The collected images were cloud-free in the study area. The time of the image data is very important for accurately distinguishing S. alterniflora from other salt marsh plants. On the basis of the local phenological characteristics of S. alterniflora, the selected remote sensing images were concentrated in August-October, when S. alterniflora was in its most vigorous growth period [47,48], and its spectral features were well differentiated from indigenous vegetation. However, because of the large study area and the need to avoid cloud-cover in the images, it was difficult to obtain images that completely covered the study area and met the monthly conditions within a single year. Therefore, missing images were taken from adjacent years to minimize the data loss caused Remote Sens. 2022, 14, 2630 5 of 24 by cloud cover. In this paper, a series of pre-processing of Landsat data was performed using ENVI 5.3 software. The digital numbers were converted to radiance by radiometric calibration. The atmospheric correction method used was Quick Atmospheric Correction (QUAC). Finally, band combination, image mosaicking, and clipping were performed. Because the data came from different sensors with different spatial resolutions, to unify the data to a spatial resolution of 30 m, this study did not fuse the panchromatic and multispectral bands in Landsat 8 OLI data. The reason was that compared with Landsat 8 OLI, Landsat 5 TM lacked a panchromatic band with a spatial resolution of 15 m. A total of six original bands of Red, Green, Blue, NIR, SWIR-1, and SWIR-2 were selected. The wavelength ranges of NIR, SWIR-1, and SWIR-2 of Landsat 5 TM were 0.760-0.900, 1.550-1.750, 2.080-2.350, respectively. The wavelength ranges of NIR, SWIR-1, and SWIR-2 of Landsat 8 OLI were 0.845-0.885, 1.560-1.660, and 2.100-2.300.

Field Investigation
A field investigation of S. alterniflora in the coastal wetlands of Jiangsu Province was conducted in October 2020, collecting data from 54 survey sites, 52 drone flights, and hundreds of photographs (as shown in Figure 2). The survey sites were concentrated in Yancheng because Yancheng is a typical distribution area of S. alterniflora in the coastal wetlands of Jiangsu Province. The site distribution is shown in Figure 1. The photographs taken by the drone reflect the distribution characteristics of S. alterniflora at various scales in the coastal wetlands of Jiangsu Province. On the basis of the drone photographs, the spectral characteristics, and growth characteristics of S. alterniflora, this study obtained a ground truth map of S. alterniflora in Jiangsu Province using ArcGIS 10.2 software. This map was used as the label image of the deep learning model. Figure 3 illustrates the spectral characteristics of S. alterniflora and several other ground feature types. It can be seen that the types were well distinguished in the wavelength range of the NIR band.

MRCNN Structure
At present, the main multispectral pixel-level classification methods include support vector machine (SVM), random forest (RF), and neural networks. A convolutional neural network (CNN) is a feedforward neural network that uses convolutional operations and has a deep structure. According to literature and experiments [27][28][29], CNNs are widely used owing to their excellent performance. Thus, a CNN was employed for experimentation and prediction, as shown below.
The performance of the visual geometry group network does not always improve as the number of layers increases, mainly because of the vanishing gradients in classical CNNs. To reduce the computational cost and ensure that the gradient was propagated in

MRCNN Structure
At present, the main multispectral pixel-level classification methods include support vector machine (SVM), random forest (RF), and neural networks. A convolutional neural network (CNN) is a feedforward neural network that uses convolutional operations and has a deep structure. According to literature and experiments [27][28][29], CNNs are widely used owing to their excellent performance. Thus, a CNN was employed for experimentation and prediction, as shown below.
The performance of the visual geometry group network does not always improve as the number of layers increases, mainly because of the vanishing gradients in classical CNNs. To reduce the computational cost and ensure that the gradient was propagated in

MRCNN Structure
At present, the main multispectral pixel-level classification methods include support vector machine (SVM), random forest (RF), and neural networks. A convolutional neural network (CNN) is a feedforward neural network that uses convolutional operations and has a deep structure. According to literature and experiments [27][28][29], CNNs are widely used owing to their excellent performance. Thus, a CNN was employed for experimentation and prediction, as shown below.
The performance of the visual geometry group network does not always improve as the number of layers increases, mainly because of the vanishing gradients in classical CNNs. To reduce the computational cost and ensure that the gradient was propagated in deeper layers, we proposed two representative networks: Inception V1 and ResNet. Inception V1 Remote Sens. 2022, 14, 2630 7 of 24 performs feature extraction and pooling at different scales to obtain information at multiple scales and then superimposes the features for output. Inception V1 can greatly reduce the number of calculations needed to ensure the propagation of gradients, which is its most important advantage [36,37]. ResNet uses residual connections to propagate gradients, that is, the shallower layers are directly connected with the deeper layers. Directly propagating larger gradients into deeper layers is both understandable and efficient [38,39]. The focus of this study was the temporal and spatial distribution of S. alterniflora in Jiangsu Province over the past 30 years. S. alterniflora has been found at different scales, such as small discrete patches, connected strips, large broken patches, and mixed patches from the growth period to the outbreak period and then to the plateau period. Inspired by CD-CNN, Inception V1, and ResNet, a multi-scale residual convolutional neural network (MRCNN) was proposed. The structure is shown in Figure 4. It has the characteristics of deep residual network skip-layer identity mapping and also overcomes the disadvantage that traditional CNNs can only extract feature information at a single scale. deeper layers, we proposed two representative networks: Inception V1 and ResNet. Inception V1 performs feature extraction and pooling at different scales to obtain information at multiple scales and then superimposes the features for output. Inception V1 can greatly reduce the number of calculations needed to ensure the propagation of gradients, which is its most important advantage [36,37]. ResNet uses residual connections to propagate gradients, that is, the shallower layers are directly connected with the deeper layers. Directly propagating larger gradients into deeper layers is both understandable and efficient [38,39]. The focus of this study was the temporal and spatial distribution of S. alterniflora in Jiangsu Province over the past 30 years. S. alterniflora has been found at different scales, such as small discrete patches, connected strips, large broken patches, and mixed patches from the growth period to the outbreak period and then to the plateau period. Inspired by CD-CNN, Inception V1, and ResNet, a multi-scale residual convolutional neural network (MRCNN) was proposed. The structure is shown in Figure 4. It has the characteristics of deep residual network skip-layer identity mapping and also overcomes the disadvantage that traditional CNNs can only extract feature information at a single scale. The MRCNN was divided into three blocks: the inception block, deep residual block, and classification block. The main structure of the inception block used three convolutional layers with different convolution kernels to simultaneously extract the features of the samples. The sizes of convolution kernels set in this study were 1 × 1, 3 × 3, and 5 × 5, respectively. Then, the extracted features of each convolution layer were concatenated. The main role of the inception block was to extract multi-scale features and reduce computational complexity. In this study, a batch normalization layer was applied between the convolution layer and the rectified linear unit (ReLU) activation layer to reduce the drift of the internal covariance and accelerate the convergence of the network. The batch normalization layer normalizes the features to a fixed distribution. The specific operations are: where and are the input and output features of the th kernel, respectively; the parameters and represent the scale displacement; is a constant close to 0. The depth residual block was then used to directly connect the features of the shallow layers to the deep layers to ensure gradient propagation. During the forward propagation The MRCNN was divided into three blocks: the inception block, deep residual block, and classification block. The main structure of the inception block used three convolutional layers with different convolution kernels to simultaneously extract the features of the samples. The sizes of convolution kernels set in this study were 1 × 1, 3 × 3, and 5 × 5, respectively. Then, the extracted features of each convolution layer were concatenated. The main role of the inception block was to extract multi-scale features and reduce computational complexity. In this study, a batch normalization layer was applied between the convolution layer and the rectified linear unit (ReLU) activation layer to reduce the drift of the internal covariance and accelerate the convergence of the network. The batch normalization layer normalizes the features to a fixed distribution. The specific operations are: wherex i and y i are the input and output features of the ith kernel, respectively; the parameters γ and β represent the scale displacement; ε is a constant close to 0. The depth residual block was then used to directly connect the features of the shallow layers to the deep layers to ensure gradient propagation. During the forward propagation of the algorithm, a skip connection is formed between convolutional layers to realize the identity mapping of input and output. The structure of the convolution layer + ReLU Remote Sens. 2022, 14, 2630 8 of 24 activation function + convolution layer was used to better extract features and suppress gradient loss. Each convolution kernel was set to 1 × 1 to reduce computational complexity and save time. Assuming that x represents input and F(x) represents residual mapping, the output of the residual unit is: When the residual F(x) = 0, the function of the residual unit is identity mapping. The output of layer L is: The inverse gradient is: where ∂loss ∂x l is the gradient descent of the loss function. In the final classification block, we added a dropout layer to enhance the generalization of the model. During the training process, the network randomly discards some neurons, which destroys the close interaction between features so that the network does not rely on local features too much. In this block, all three-dimensional features were flattened into one-dimensional features for subsequent classification. The classification block used a fully connected layer to map features to class numbers and then used the Softmax activation function and cross-entropy loss function. The Softmax function is defined as: where the input and output of the Softmax function are x i and f (x i ), respectively, N is the number of categories, and f (x i ) is the probability value predicted to be class i. The activation function helps with the pixel-level classification of S. alterniflora and background data.

Unified Training Method Based on MRCNN
The accuracy and applicability of the dataset directly determine the classification effect of the model. Because the distribution area of S. alterniflora in Jiangsu Province was relatively small in 1990, and the sample distribution was extremely unbalanced, the current study began to create a data set from the data in 1995. To improve the temporal scalability of the deep learning model, the method of training data adopted a unified training of the multiphase data approach instead of training a model for the data of each phase separately.
A random sliding window clipping method was used to obtain small sample blocks with a size of 9 × 9. The 54 field survey sites in 2020 were selected as the sample selection areas for S. alterniflora. Samples were randomly selected in 1995, 2000, 2005, 2010, 2015, and 2020 from locations that covered 60% of the area of S. alterniflora. The number of background samples was 10 times that of the corresponding S. alterniflora samples every year. Figure 5 shows the spatial distribution of training samples in 2020 as an example. To capture the spatial characteristics of S. alterniflora better, a six-year sample was used as the training dataset. The diversity and variability of training data sets were increased through data standardization and data expansion [49]. Table 2 shows the training sets from 1995 to 2020. In this six-year period, we selected a total of 7131 S. alterniflora samples and 71,310 background samples, giving a total of 78,441 samples. Eighty percent of the samples in the training set were used for model training, and 20% of the samples were used for model validation. All the remaining data in the coastal areas of Jiangsu Province were used to test the model, and the identification and evaluation of S. alterniflora in each year were examined. The number of test samples from 1995 to 2020 is 312,527, 303,222, 301,792, 301,044, 300,791, and 300,384 respectively. In this process, different spatially and geographically independent samples were acquired for training and testing to prevent information from leaking from the test dataset to the model. This ensured that the classification algorithm had good generalization ability, which was beneficial for analyzing the temporal and spatial change in S. alterniflora. Table 3 shows the composition of test samples per year, and the ratio of the number of model training samples to the number of test samples per year is about 1:3.8.
test the model, and the identification and evaluation of S. alterniflora in each year were examined. The number of test samples from 1995 to 2020 is 312,527, 303,222, 301,792, 301,044, 300,791, and 300,384 respectively. In this process, different spatially and geographically independent samples were acquired for training and testing to prevent information from leaking from the test dataset to the model. This ensured that the classification algorithm had good generalization ability, which was beneficial for analyzing the temporal and spatial change in S. alterniflora. Table 3 shows the composition of test samples per year, and the ratio of the number of model training samples to the number of test samples per year is about 1:3.8.     The research area of the current study was the coastal wetland on the Jiangsu coastline, and the study aimed to explore changes in the distribution of S. alterniflora in the coastal wetland. Considering the broad coverage of the data used in this study, there were more than 2.5 × 10 7 pixels per year. At the same time, in the test data, the number of S. alterniflora samples differed greatly from the number of background samples, which might have resulted in serious background misclassification in binary classification. To avoid this, the output results of the MRCNN were post-processed in this paper using ArcGIS 10.2 software. According to the complex growth characteristics and environment of S. alterniflora, the classification results were optimized using artificial coastline information. In this study, only the classification results on the seaward side were retained, and the classification results on the landward side were discarded. This helped guarantee the accuracy of the global classification results. The output of the MRCNN was a binary image representing S. alterniflora and the background. The whole process flow is shown in Figure 6. The final distribution map of S. alterniflora in each period was obtained by performing morphological processing such as dilation, erosion, and noise elimination on the post-processing results by MatlabR2020a software.  The research area of the current study was the coastal wetland on the Jiangsu coastline, and the study aimed to explore changes in the distribution of S. alterniflora in the coastal wetland. Considering the broad coverage of the data used in this study, there were more than 2.5 × 10 7 pixels per year. At the same time, in the test data, the number of S. alterniflora samples differed greatly from the number of background samples, which might have resulted in serious background misclassification in binary classification. To avoid this, the output results of the MRCNN were post-processed in this paper using ArcGIS 10.2 software. According to the complex growth characteristics and environment of S. alterniflora, the classification results were optimized using artificial coastline information. In this study, only the classification results on the seaward side were retained, and the classification results on the landward side were discarded. This helped guarantee the accuracy of the global classification results. The output of the MRCNN was a binary image representing S. alterniflora and the background. The whole process flow is shown in Figure  6. The final distribution map of S. alterniflora in each period was obtained by performing morphological processing such as dilation, erosion, and noise elimination on the postprocessing results by MatlabR2020a software. In this set of experiments, accuracy, recall, precision, and F1-score were used as metrics. S. alterniflora was regarded as the positive class, and the others, such as vegetation, tidal flat, buildings, and water bodies, as the negative classes. The pixel value of S. alterniflora in the global truth map was set to 1, and the other pixel values were set to 0. Accuracy is the ratio of the number of correctly predicted samples to the total number of samples, and precision is the ratio of the number of correctly predicted positive samples to the total predicted positive samples. Recall is the ratio of the number of samples that are correctly predicted to be positive to the total number of samples that are actually positive, while the F1-score is a quantitative indicator for data with an unbalanced sample distribution, In this set of experiments, accuracy, recall, precision, and F1-score were used as metrics. S. alterniflora was regarded as the positive class, and the others, such as vegetation, tidal flat, buildings, and water bodies, as the negative classes. The pixel value of S. alterniflora in the global truth map was set to 1, and the other pixel values were set to 0. Accuracy is the ratio of the number of correctly predicted samples to the total number of samples, and precision is the ratio of the number of correctly predicted positive samples to the total predicted positive samples. Recall is the ratio of the number of samples that are correctly predicted to be positive to the total number of samples that are actually positive, while the F1-score is a quantitative indicator for data with an unbalanced sample distribution, which measures the balance between accuracy and recall rate. Accordingly, the F1-score is formulated as follows: where precision and recall are obtained from: Precision = True positives True positives + False positives (8) Recall = True positives True positives + False negatives (9)

Testing of the Model
To better identify the advantages of the proposed MRCNN, a set of comparative experiments between CNNs were conducted. The selected comparison algorithms were CascadeNet and dual-tunnel CNN. CascadeNet [50] is a simple and efficient residual network. The dual-tunnel CNN [51] is a network specially designed for hyperspectral images and multispectral images. It contains a branch dedicated to extracting spatial information and a branch dedicated to extracting spectral information. These two CNNs are proven algorithms for efficient multispectral classification and are more than sufficient as comparison algorithms. The deep learning experiments were performed on a computer equipped with NVIDIA RTX2070S discrete graphics, using Pycharm2021 (Python) as the programming environment. The input patch size of the MRCNN model was 9 × 9, the number of channels was 6, the batch size was 500, the number of iterations was 157, and the number of epochs was 1000. The loss curve and accuracy curve of MRCNN during the training process are shown in Figure 7. The model converged after 1000 epochs.
Remote Sens. 2022, 14, x FOR PEER REVIEW 11 of 24 which measures the balance between accuracy and recall rate. Accordingly, the F1-score is formulated as follows: where precision and recall are obtained from:

Testing of the Model
To better identify the advantages of the proposed MRCNN, a set of comparative experiments between CNNs were conducted. The selected comparison algorithms were CascadeNet and dual-tunnel CNN. CascadeNet [50] is a simple and efficient residual network. The dual-tunnel CNN [51] is a network specially designed for hyperspectral images and multispectral images. It contains a branch dedicated to extracting spatial information and a branch dedicated to extracting spectral information. These two CNNs are proven algorithms for efficient multispectral classification and are more than sufficient as comparison algorithms. The deep learning experiments were performed on a computer equipped with NVIDIA RTX2070S discrete graphics, using Pycharm2021 (Python) as the programming environment. The input patch size of the MRCNN model was 9 × 9, the number of channels was 6, the batch size was 500, the number of iterations was 157, and the number of epochs was 1000. The loss curve and accuracy curve of MRCNN during the training process are shown in Figure 7. The model converged after 1000 epochs.  Table 4 shows that the accuracy and recall reached 99% and 95%, respectively. Although it is not objective to directly compare the two metrics, such high accuracy and recall indicate that the effectiveness of all three algorithms is high. Owing to the substantial difference between the number of samples of S. alterniflora and the number of background samples, the precision was relatively low. Even in the period when the distribution area of S. alterniflora was the largest in Jiangsu, the percentage of S. alterniflora pixels on the satellite images used in this study was only 0.8%. However, the vast majority of S. alterniflora samples were identified by the MRCNN according to the recall and accuracy indicators. To reduce the impact of low precision caused by the large number of image pixels and the small proportion of target pixels, this study improved the precision through post-  Table 4 shows that the accuracy and recall reached 99% and 95%, respectively. Although it is not objective to directly compare the two metrics, such high accuracy and recall indicate that the effectiveness of all three algorithms is high. Owing to the substantial difference between the number of samples of S. alterniflora and the number of background samples, the precision was relatively low. Even in the period when the distribution area of S. alterniflora was the largest in Jiangsu, the percentage of S. alterniflora pixels on the satellite images used in this study was only 0.8%. However, the vast majority of S. alterniflora samples were identified by the MRCNN according to the recall and accuracy indicators. To reduce the impact of low precision caused by the large number of image pixels and the small proportion of target pixels, this study improved the precision through post-processing.
The results of the post-processing served for the subsequent analysis and discussion of the invasion process of S. alterniflora. Meanwhile, the F1-score was used to scientifically and effectively measure the indicators of the algorithms in the binary classification. This study focused on using the F1-score to evaluate the advantages of MRCNN over other comparative algorithms.  Table 4 shows the classification result of the MRCNN. The F1-score of the MRCNN was the highest in 1995, 2005, 2010, and 2020. In 2000 and 2015, there was only a small gap between the F1-score of the MRCNN and that of the highest model. It is normal for results to vary from year to year owing to differences in imaging conditions. However, in general, the performance of the MRCNN was the best among the three algorithms in most years. After post-processing, the classification precision of S. alterniflora was greatly improved, and the annual F1-score was more than 80%, indicating that the combination of the MRCNN and post-processing method was suitable for the identification of S. alterniflora in Jiangsu coastal wetland from long time series data. The results of post-processing are shown in Table 5. Taking 2020 as an example, the identification and comparison of S. alterniflora in several typical local areas are shown in Figure 8. For each group, the left image is the combination of the NIR-R-G bands, the middle image is the classified comparison map, and the right image is the ground truth map. All three models can identify the S. alterniflora samples well, and the contours of the S. alterniflora patches are also relatively consistent. However, the MRCNN model was more accurate, especially in reducing misclassified pixels. On the premise of ensuring the accuracy, the MRCNN method combined with post-processing proposed in this study greatly reduced the number of misclassified pixels, which met the requirements for S. alterniflora identification in this study. right image is the ground truth map. All three models can identify the S. alterniflora samples well, and the contours of the S. alterniflora patches are also relatively consistent. However, the MRCNN model was more accurate, especially in reducing misclassified pixels. On the premise of ensuring the accuracy, the MRCNN method combined with post-processing proposed in this study greatly reduced the number of misclassified pixels, which met the requirements for S. alterniflora identification in this study.

Classification of S. alterniflora
The experiments showed that the MRCNN model was able to identify S. alterniflora in long time series data. In this study, the classification results of S. alterniflora in the coastal wetlands of Jiangsu Province were obtained through the MRCNN plus post-processing method and Landsat data from 1990 to 2020. Before determining the area of S. alterniflora, it was necessary to manually correct the misclassified and missed classification pixels. After morphological processing, the thematic map of S. alterniflora distribution was finally created (as shown in Figure 9). This allowed better analysis of changes in the temporal and spatial distribution of S. alterniflora.
In 1990, S. alterniflora was first detected at the estuaries of the Linhong River, Guan River, Guboshanghou River, Sheyang River, and Dongtai River. Small patches of S. alterniflora also appeared in the south of the RCCNNR, while there was no sign of S. alterniflora invasion in the ENNR. In 1995, the north-south length of the S. alterniflora patch south of the Sheyang estuary was 6.2 km, while the length of the S. alterniflora strip in Lianxinggang reached 19.9 km. By 2000, the coastal areas of the RCCNNR and ENNR had been surrounded by S. alterniflora strips, and the nature reserves protected S. alterniflora as well as protecting indigenous species. S. alterniflora strips were distributed from the Sheyang River estuary to the south of the Dongtai River estuary, spanning 111.5 km from north to south, and S. alterniflora strips approximately 12.4 km long had formed near the Linghong

Temporal and Spatial Distribution of S. alterniflora 4.2.1. Classification of S. alterniflora
The experiments showed that the MRCNN model was able to identify S. alterniflora in long time series data. In this study, the classification results of S. alterniflora in the coastal wetlands of Jiangsu Province were obtained through the MRCNN plus post-processing method and Landsat data from 1990 to 2020. Before determining the area of S. alterniflora, it was necessary to manually correct the misclassified and missed classification pixels. After morphological processing, the thematic map of S. alterniflora distribution was finally created (as shown in Figure 9). This allowed better analysis of changes in the temporal and spatial distribution of S. alterniflora.
In 1990, S. alterniflora was first detected at the estuaries of the Linhong River, Guan River, Guboshanghou River, Sheyang River, and Dongtai River. Small patches of S. alterniflora also appeared in the south of the RCCNNR, while there was no sign of S. alterniflora invasion in the ENNR. In 1995, the north-south length of the S. alterniflora patch south of the Sheyang estuary was 6.2 km, while the length of the S. alterniflora strip in Lianxinggang reached 19.9 km. By 2000, the coastal areas of the RCCNNR and ENNR had been surrounded by S. alterniflora strips, and the nature reserves protected S. alterniflora as well as protecting indigenous species. S. alterniflora strips were distributed from the Sheyang River estuary to the south of the Dongtai River estuary, spanning 111.5 km from north to south, and S. alterniflora strips approximately 12.4 km long had formed near the Linghong River estuary. In 2005, the distribution of S. alterniflora in Yancheng City tended to be saturated. From 2010 to 2020, the strips of S. alterniflora were substantially damaged, especially in the Dafeng port area. River estuary. In 2005, the distribution of S. alterniflora in Yancheng City tended to be saturated. From 2010 to 2020, the strips of S. alterniflora were substantially damaged, especially in the Dafeng port area.

Temporal Changes in S. alterniflora
The distribution statistics of S. alterniflora in Jiangsu Province are shown in Figure 10. In the current study, the expansion of S. alterniflora was divided into three stages: the growth stage, outbreak stage, and plateau stage. S. alterniflora was in the growth stage before 1995. From 1996 to 2005, it entered the outbreak stage, with a rapid growth rate of 1186 ha/year, mainly distributed in Yancheng. After 2005, it gradually entered the plateau stage. By 2020, the distribution area of S. alterniflora in Jiangsu Province had reached 17657.14 ha, an increase of nearly 23.6 times compared with 1990. The changing trend in S. alterniflora in the three coastal cities was roughly consistent with the general trend in Jiangsu Province. S. alterniflora was first introduced in Yancheng (marked in red in Figure 10). After several rounds of trial planting from 1982 to 1986, S. alterniflora entered the outbreak stage earlier in Yancheng than elsewhere. From 1990 to 1995, the expansion rate had reached 496.9 ha/year.

Temporal Changes in S. alterniflora
The distribution statistics of S. alterniflora in Jiangsu Province are shown in Figure 10. In the current study, the expansion of S. alterniflora was divided into three stages: the growth stage, outbreak stage, and plateau stage. S. alterniflora was in the growth stage before 1995. From 1996 to 2005, it entered the outbreak stage, with a rapid growth rate of 1186 ha/year, mainly distributed in Yancheng. After 2005, it gradually entered the plateau stage. By 2020, the distribution area of S. alterniflora in Jiangsu Province had reached 17657.14 ha, an increase of nearly 23.6 times compared with 1990. The changing trend in S. alterniflora in the three coastal cities was roughly consistent with the general trend in Jiangsu Province. S. alterniflora was first introduced in Yancheng (marked in red in Figure  10). After several rounds of trial planting from 1982 to 1986, S. alterniflora entered the outbreak stage earlier in Yancheng than elsewhere. From 1990 to 1995, the expansion rate had reached 496.9 ha/year.

Spatial Changes in S. alterniflora
S. alterniflora patches were defined in this study with a 2-pixel gap (60 m) ( Figure  11h). There were two kinds of spatial variation of the centroid of S. alterniflora patches. The first one was an expansion parallel to the coastline, which covered the whole range of muddy tidal flats in Jiangsu. The second one was expansion in the direction perpendicular to the coastline, with a relatively small expansion range; the farthest distance to the sea is approximately 4 km. Figure 11 shows that the centroid of S. alterniflora patches in Yan-

Spatial Changes in S. alterniflora
S. alterniflora patches were defined in this study with a 2-pixel gap (60 m) (Figure 11h). There were two kinds of spatial variation of the centroid of S. alterniflora patches. The first one was an expansion parallel to the coastline, which covered the whole range of muddy tidal flats in Jiangsu. The second one was expansion in the direction perpendicular to the coastline, with a relatively small expansion range; the farthest distance to the sea is approximately 4 km. Figure 11 shows that the centroid of S. alterniflora patches in Yancheng from 1990 to 2005 expanded from the south and the north to the middle to saturation. During this period, S. alterniflora was in the outbreak period, with active artificial reclamation activities and an obvious increase in the number of broken patches.  In Jiangsu Province, the patches of S. alterniflora were most densely distributed in Yancheng, and the changes were most obvious. There are two national nature reserves in the area from Dawa Port to the estuary of Dongtai River, where the patch distribution of S. alterniflora showed the most obvious changes. The RCCNNR is located between Xinyang Port and Doulong Port. Mudflats were widely distributed in this area, and these were replenished by natural sedimentation. The centroid change of S. alterniflora patches In Jiangsu Province, the patches of S. alterniflora were most densely distributed in Yancheng, and the changes were most obvious. There are two national nature reserves in the area from Dawa Port to the estuary of Dongtai River, where the patch distribution of S. alterniflora showed the most obvious changes. The RCCNNR is located between Xinyang Port and Doulong Port. Mudflats were widely distributed in this area, and these were replenished by natural sedimentation. The centroid change of S. alterniflora patches from Xinyang Port to Doulong Port from 1995 to 2000 is shown in Figure 12a. The number of S. alterniflora patches in the outbreak stage increased significantly, and a large number of ecological niches were suitable for invasion by S. alterniflora. These patches expanded significantly towards the sea, with an average width of 1.1 km. Figure 12b shows the changes in the centroid of S. alterniflora patches from Doulong Port to Dafeng Port from 2000 to 2010. Artificial reclamation led to the seaward expansion of S. alterniflora patches between Doulong Port and Dafeng Port. The number of S. alterniflora patches in this area was clearly reduced, and the expansion distance was relatively small, with an average of 0.8 km to the sea. from Xinyang Port to Doulong Port from 1995 to 2000 is shown in Figure 12a. The number of S. alterniflora patches in the outbreak stage increased significantly, and a large number of ecological niches were suitable for invasion by S. alterniflora. These patches expanded significantly towards the sea, with an average width of 1.1 km. Figure 12b shows the changes in the centroid of S. alterniflora patches from Doulong Port to Dafeng Port from 2000 to 2010. Artificial reclamation led to the seaward expansion of S. alterniflora patches between Doulong Port and Dafeng Port. The number of S. alterniflora patches in this area was clearly reduced, and the expansion distance was relatively small, with an average of 0.8 km to the sea.

The Relationship between Inception Block and Patch Characteristics
The depth and width of the network structure are very important when extracting S. alterniflora patches with complex growth states from remote sensing images. The comparison algorithms, CascadeNet, and dual-tunnel CNN, selected in this study mainly used a 3 × 3 single scale convolutional kernel to extract features. Research [52] showed that using a single convolution kernel can easily lead to filter co-adaptation, resulting in the misclassification of background pixels and target pixels and unclear patch boundaries of S. alterniflora. Therefore, the Inception block of the MRCNN network proposed in this study used three convolutional layers with different convolution kernel sizes to extract sample features simultaneously.
After many experiments and comparisons, this study finally selected three convolution kernels: 1 × 1, 3 × 3, and 5 × 5. When the same receptive field is reached, as the convolution kernel decreases, fewer parameters are needed in the training process. As the convolution kernel increases, the number of parameters increases, the complexity of the model increases, and the convergence of the model becomes more difficult. For images with a 30 m spatial resolution, the recognition ranges of 1 × 1, 3 × 3, and 5 × 5 convolution kernels are equivalent to 9 × 10 2 m 2 , 2.7 × 10 3 m 2 , and 2.25 × 10 4 m 2 , respectively. These scale capacities were sufficient to accommodate the patch feature information of S. alterniflora in various growth states (as shown in Figure 13). The function of the 1 × 1 convolution layer was mainly to increase the depth of the network, add nonlinear information, and also reduce the dimension and parameters in the training process. For the 3 × 3 convolution layer, the detailed information of small patches could be better captured, and the

The Relationship between Inception Block and Patch Characteristics
The depth and width of the network structure are very important when extracting S. alterniflora patches with complex growth states from remote sensing images. The comparison algorithms, CascadeNet, and dual-tunnel CNN, selected in this study mainly used a 3 × 3 single scale convolutional kernel to extract features. Research [52] showed that using a single convolution kernel can easily lead to filter co-adaptation, resulting in the misclassification of background pixels and target pixels and unclear patch boundaries of S. alterniflora. Therefore, the Inception block of the MRCNN network proposed in this study used three convolutional layers with different convolution kernel sizes to extract sample features simultaneously.
After many experiments and comparisons, this study finally selected three convolution kernels: 1 × 1, 3 × 3, and 5 × 5. When the same receptive field is reached, as the convolution kernel decreases, fewer parameters are needed in the training process. As the convolution kernel increases, the number of parameters increases, the complexity of the model increases, and the convergence of the model becomes more difficult. For images with a 30 m spatial resolution, the recognition ranges of 1 × 1, 3 × 3, and 5 × 5 convolution kernels are equivalent to 9 × 10 2 m 2 , 2.7 × 10 3 m 2 , and 2.25 × 10 4 m 2 , respectively. These scale capacities were sufficient to accommodate the patch feature information of S. alterniflora in various growth states (as shown in Figure 13). The function of the 1 × 1 convolution layer was mainly to increase the depth of the network, add nonlinear information, and also reduce the dimension and parameters in the training process. For the 3 × 3 convolution layer, the detailed information of small patches could be better captured, and the ability to distinguish the boundaries of different objects was enhanced. For the boundaries and mixed zones of S. alterniflora patches, a 5 × 5 convolution layer could obtain smoother features than small-sized convolutional kernels. The multi-scale convolution interaction allowed different strategies to be used to serve the feature information at different scales. ability to distinguish the boundaries of different objects was enhanced. For the boundaries and mixed zones of S. alterniflora patches, a 5 × 5 convolution layer could obtain smoother features than small-sized convolutional kernels. The multi-scale convolution interaction allowed different strategies to be used to serve the feature information at different scales.

Expansion Factors of S. alterniflora
According to literature [53,54], the seed yield of S. alterniflora and the silting environment of offshore tidal flats are the main factors that affect the expansion of S. alterniflora. The germination rate of S. alterniflora seeds on the offshore tidal flat and sporadic patches was significantly higher than that inside the large patches. The sedimentation rate of the offshore tidal flat was mainly affected by natural sedimentation and artificial reclamation [55][56][57]. Artificial reclamation of the tidal flat changed its sedimentary environment. The decrease in tidal sediment carrying capacity led to the acceleration of the sedimentation rate, promoting the expansion of S. alterniflora on the offshore tidal flat [58], and eventually connecting the strips with larger spans. At the same time, excessive reclamation destroyed the habitat for S. alterniflora, split S. alterniflora patches, and reduced seed yield, which hindered the expansion of S. alterniflora.
The implementation of artificial reclamation had a great influence on the coastal environment, and the construction of reclamation changed the hydrodynamic conditions and sediment deposition rate in this area [59]. The tidal flat reclamation area located in Haizhou Bay in the north of Jiangsu Province was substantial, resulting in the siltation of the upper reaches of the Linhong estuary. The average velocity of the flood tide and ebb tide in the reclamation area decreased, which led to sediment deposition [44,60,61] and promoted the expansion of S. alterniflora patches toward the sea [62]. The width of the S. alterniflora patch extending to the sea in the upper reaches increased from 0.13 km in 2010 to 0.59 km in 2015. Figure 14a shows that the area of S. alterniflora in the Linghong River estuary was 67.14 ha in 2010 and increased to 281.96 ha after 5 years. After the Sheyang estuary project was built, the sediment concentration at the flood tide was obviously greater than that at the ebb tide. At the flood tide, more sediment was deposited on both sides of the estuary, which facilitated the expansion of S. alterniflora patches to the sea and provided a good niche for the growth of S. alterniflora. In 2015, the area of S. alterniflora in the Sheyang estuary was 356.99 ha. Five years later, the area of S. alterniflora increased to 536.35 ha (Figure 14b), with a remarkable growth rate of 35.87 ha/year.

Expansion Factors of S. alterniflora
According to literature [53,54], the seed yield of S. alterniflora and the silting environment of offshore tidal flats are the main factors that affect the expansion of S. alterniflora. The germination rate of S. alterniflora seeds on the offshore tidal flat and sporadic patches was significantly higher than that inside the large patches. The sedimentation rate of the offshore tidal flat was mainly affected by natural sedimentation and artificial reclamation [55][56][57]. Artificial reclamation of the tidal flat changed its sedimentary environment. The decrease in tidal sediment carrying capacity led to the acceleration of the sedimentation rate, promoting the expansion of S. alterniflora on the offshore tidal flat [58], and eventually connecting the strips with larger spans. At the same time, excessive reclamation destroyed the habitat for S. alterniflora, split S. alterniflora patches, and reduced seed yield, which hindered the expansion of S. alterniflora.
The implementation of artificial reclamation had a great influence on the coastal environment, and the construction of reclamation changed the hydrodynamic conditions and sediment deposition rate in this area [59]. The tidal flat reclamation area located in Haizhou Bay in the north of Jiangsu Province was substantial, resulting in the siltation of the upper reaches of the Linhong estuary. The average velocity of the flood tide and ebb tide in the reclamation area decreased, which led to sediment deposition [44,60,61] and promoted the expansion of S. alterniflora patches toward the sea [62]. The width of the S. alterniflora patch extending to the sea in the upper reaches increased from 0.13 km in 2010 to 0.59 km in 2015. Figure 14a shows that the area of S. alterniflora in the Linghong River estuary was 67.14 ha in 2010 and increased to 281.96 ha after 5 years. After the Sheyang estuary project was built, the sediment concentration at the flood tide was obviously greater than that at the ebb tide. At the flood tide, more sediment was deposited on both sides of the estuary, which facilitated the expansion of S. alterniflora patches to the sea and provided a good niche for the growth of S. alterniflora. In 2015, the area of S. alterniflora in the Sheyang estuary was 356.99 ha. Five years later, the area of S. alterniflora increased to 536.35 ha (Figure 14b), with a remarkable growth rate of 35.87 ha/year.  Figure 15). The removal of aquaculture ponds is an important habitat modification for the reproduction and growth of S. alterniflora seeds. According to research [63,64], the sediment of aquaculture ponds is rich in nitrogen and phosphorus. Owing to the decomposition of excreta from aquaculture species, a large amount of organic nitrogen and phosphorus was released and retained in the sediment, which led to the eutrophication of the aquaculture pond system. With this additional nitrogen, the biomass of S. alterniflora grassland increased significantly. The increased nitrogen input stimulated the growth of S. alterniflora seeds and improved their productivity [42]. In the past 20 years, S. alterniflora in Ganyu Port has increased by nearly 384.71 ha.  Before 2005, S. alterniflora was not found in Ganyu Port. After the aquaculture ponds were dismantled in 2000, S. alterniflora patches were found in 2005 ( Figure 15). The removal of aquaculture ponds is an important habitat modification for the reproduction and growth of S. alterniflora seeds. According to research [63,64], the sediment of aquaculture ponds is rich in nitrogen and phosphorus. Owing to the decomposition of excreta from aquaculture species, a large amount of organic nitrogen and phosphorus was released and retained in the sediment, which led to the eutrophication of the aquaculture pond system. With this additional nitrogen, the biomass of S. alterniflora grassland increased significantly. The increased nitrogen input stimulated the growth of S. alterniflora seeds and improved their productivity [42]. In the past 20 years, S. alterniflora in Ganyu Port has increased by nearly 384.71 ha.  Before 2005, S. alterniflora was not found in Ganyu Port. After the aquaculture ponds were dismantled in 2000, S. alterniflora patches were found in 2005 ( Figure 15). The removal of aquaculture ponds is an important habitat modification for the reproduction and growth of S. alterniflora seeds. According to research [63,64], the sediment of aquaculture ponds is rich in nitrogen and phosphorus. Owing to the decomposition of excreta from aquaculture species, a large amount of organic nitrogen and phosphorus was released and retained in the sediment, which led to the eutrophication of the aquaculture pond system. With this additional nitrogen, the biomass of S. alterniflora grassland increased significantly. The increased nitrogen input stimulated the growth of S. alterniflora seeds and improved their productivity [42]. In the past 20 years, S. alterniflora in Ganyu Port has increased by nearly 384.71 ha.  In summary, the effect of tidal flat reclamation is closely related to the expansion process of S. alterniflora. Reclamation of the tidal flat affects the expansion of S. alterniflora, while the growth of S. alterniflora patches also leads to a process of stagnation and silting [46]. Sporadic patches were formed by vegetative propagation of S. alterniflora in the growth stage, and the planting density gradually increased [3,4]. Subsequently, S. alterniflora gradually expanded to the sea through the natural sedimentation and artificial reclamation of tidal flats. In the outbreak stage, the seed yield of S. alterniflora increased greatly and drifted along the tidal creeks. Tidal flat reclamation led to a decrease in the hydrodynamic force near the estuary, allowing the seeds to easily take root and grow in the muddy environment and gradually connect into large patches. The ecological niche of S. alterniflora was gradually saturated, and finally evolved into a strip pattern with a north-south distribution [62]. S. alterniflora is limited by the elevation and tide during the plateau period, and thus it is necessary for the tidal flat on the seaward side to continue silting up before S. alterniflora patches can continue to expand seaward.

Effects of Mixed Pixels and Tidal Inundation on Classification Results
Pixel-level classification algorithms are dependent on the quality of pixels. The influence of cloud cover needs to be considered when selecting image data. In areas covered by clouds, pixel information will be lost. In the initial growth stage of S. alterniflora, the proportion of S. alterniflora patches in a 30 m pixel was small, and the Landsat images with a medium resolution of 30 m were limited. In fact, the phenomenon of mixed pixels containing other vegetation types or tidal flats is unavoidable. Classifying pixels with mixed vegetation types has always been a challenge in land cover type mapping [65,66]. In the future, the use of high-resolution satellite images or hyperspectral data for fine monitoring can reduce classification errors and improve the accuracy of S. alterniflora distribution mapping.
Finally, S. alterniflora grows in intertidal areas, which are easily flooded by periodic tides. It is difficult to classify S. alterniflora if it is underwater, whether in field surveys or in satellite images. For example, the spectral characteristics and pixel quality of S. alterniflora are affected by flooding [67,68]. Optical data are made up of the reflection spectrum of different ground objects and contain rich spectral information, but the quality of pixels is easily affected by clouds, tides, and other factors. Synthetic aperture radar (SAR) reflects the backscattering ability of different ground objects and is not affected by water bodies. It has the ability to penetrate wetland vegetation and has the characteristics of all-day and allweather data collection, but it lacks spectral information [69]. The different characteristics of optical images and SAR can be used for image fusion [70], which resolves the influence of tides and clouds on pixel quality. These combined data will better serve to identify S. alterniflora in coastal wetlands.

Conclusions
S. alterniflora has the characteristics of a complex growth environment and different scales of growth. In this study, a multi-temporal sample dataset of S. alterniflora in Jiangsu Province was constructed. Through multi-scale feature extraction and depth residual connection, a classification method for S. alterniflora based on an MRCNN and post-processing was designed. The classification results for S. alterniflora in Jiangsu Province for 6 years were obtained using the unified training method for multi-temporal data. The results showed that the test set of S. alterniflora in 1995 had the highest F1-score, which was more than 59%. The performance of the MRCNN was the best among the three algorithms in most years. After post-processing, the global F1-scores of S. alterniflora in Jiangsu were all above 80% in each period. This indicates that the proposed method was effective for the identification of S. alterniflora using long time series data over a large area. In this study, a thematic map of the distribution of S. alterniflora in Jiangsu Province from 1990 to 2020 was created.
After nearly 40 years of expansion, by 2020, the distribution area of S. alterniflora in Jiangsu Province had reached 17657.14 ha, mainly in Yancheng. The expansion process of S. alterniflora in time was divided into three stages: the growth period (1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994), the outbreak period (1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004), and the plateau period (2005-2020). The spatial expansion directions of S. alterniflora were mainly parallel to the coastline and perpendicular to the coastline. The reclamation of coastal tidal flats in Jiangsu was one of the main factors affecting the invasion of S. alterniflora. The tidal flat reclamation provided a new niche and suitable environment for the growth of S. Alterniflora, which accelerated the invasion of S. alterniflora. The patches of S. alterniflora gradually connected into large strips in the north-south direction.
This research provided an effective approach to solving the application of wetland remote sensing using deep learning algorithms. In the future, to better understand the distribution changes of S. alterniflora in different regions, we need to construct a large, accurate, and diverse sample library of S. alterniflora. By combining optical data with SAR, the quality of the pixels can be improved. Using a powerful deep learning model to accurately and dynamically monitor the distribution changes of S. alterniflora can effectively manage and protect the coastal wetland ecosystem.