A semantic change detection method for remote sensing images of typical vegetation in coastal wetlands

Abstract

The present invention discloses a method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands, and belongs to the field of optical remote sensing image change detection. The method comprises the following steps: S1, performing data preprocessing on remote sensing images, and producing a remote sensing semantic change detection data set for typical vegetation in coastal wetlands; S2, calculating the normalized difference vegetation index and the normalized difference water index of the remote sensing image, and extracting a deep semantic multi-layer feature map reflecting the characteristics of the coastal wetland features; S3, sending to a domain alignment module to obtain a semantically guided feature map after domain alignment; S4, sending to a multi-task decoder to obtain a coastal wetland change detection map and a typical vegetation category prediction map of a dual-phase coastal wetland remote sensing image; S5, performing mask processing and vectorization processing to obtain a semantic change detection vector map of typical vegetation in coastal wetlands. The present invention can effectively improve the accuracy and efficiency of change detection, and provide a scientific basis for the protection and management of coastal wetlands.

Description

Technical Field

[0001] The invention relates to the technical field of optical remote sensing image change detection, and in particular to a remote sensing image semantic change detection method for a coastal wetland ecosystem. Background Art

[0002] Coastal wetlands are one of the most biodiverse ecosystems on Earth. They play an important role in regulating climate, protecting coastlines, and maintaining ecological balance. With the impact of global climate change and human activities, coastal wetlands are facing serious threats of degradation and destruction. Therefore, the monitoring and management of coastal wetlands has become particularly important.

[0003] As an effective monitoring method, remote sensing technology can provide large-scale and high-frequency surface cover information and is an important tool for studying coastal wetland changes. Traditional remote sensing image change detection methods mainly include pixel-based comparative analysis, feature-based analysis, and model-based methods. These methods work well when dealing with small-scale or simple scenes, but when faced with complex and changeable coastal wetland environments, they often have problems such as low detection accuracy and low computational efficiency.

[0004] In recent years, deep learning technology has made significant progress in the field of image processing, especially in image recognition, classification and segmentation tasks. Deep learning models can automatically extract image features and learn complex nonlinear mapping relationships through multi-layer neural networks, thereby achieving more accurate and efficient image analysis.

[0005] Although deep learning is increasingly used in remote sensing image analysis, there are still few studies on the specific field of coastal wetlands. Remote sensing images of coastal wetlands have unique spectral characteristics and complex spatial structures, which pose new challenges to the design and training of deep learning models. In addition, change detection in coastal wetlands requires not only identifying the changed areas, but also obtaining the changes in typical vegetation in the changed areas. Summary of the invention

[0006] Based on the above technical problems, the present invention proposes a semantic change detection method for remote sensing images of typical vegetation in coastal wetlands.

[0007] The technical solution adopted by the present invention is:

[0008] A method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands comprises the following steps:

[0009] S1: Obtain remote sensing images containing typical vegetation of coastal wetlands and capable of being used for dual-temporal change detection, perform data preprocessing on them, and produce a semantic change detection dataset of remote sensing images of typical vegetation of coastal wetlands;

[0010] S2: Calculate the brightness, normalized difference vegetation index and normalized difference water index of the remote sensing images in the semantic change detection dataset of typical vegetation remote sensing images of coastal wetlands, and send them and the remote sensing images in the semantic change detection dataset of typical vegetation remote sensing images of coastal wetlands into the feature extractor to extract the deep semantic multi-layer feature map reflecting the characteristics of coastal wetland features;

[0011] S3: The extracted deep semantic multi-layer feature map is sent to the domain alignment module to obtain the semantic guidance feature map of the dual-temporal coastal wetland features after domain alignment;

[0012] S4: Subtract the obtained semantic guidance feature map of the dual-temporal coastal wetland features and send it to the multi-task decoder to obtain the coastal wetland binary change detection map and the typical vegetation category prediction map of the dual-temporal coastal wetland remote sensing image;

[0013] S5: Use the coastal wetland binary change detection map to perform mask processing and vectorization processing on the typical vegetation category prediction map of the dual-temporal coastal wetland remote sensing image to obtain the coastal wetland typical vegetation semantic change detection vector map, and complete the semantic change detection of the coastal wetland typical vegetation remote sensing image.

[0014] The beneficial technical effects of the present invention are:

[0015] The present invention aims to solve the deficiencies of the prior art in the change detection of remote sensing images of coastal wetlands, and proposes a semantic change detection method for remote sensing images of typical vegetation in coastal wetlands. The method can obtain a semantic change detection vector map of typical vegetation in coastal wetlands, from which the growth and degradation trends of key protected vegetation in coastal wetland reserves can be observed, which is helpful for the ecological protection of coastal wetlands. The method of the present invention can effectively improve the accuracy and efficiency of change detection by combining domain alignment processing with powerful feature extraction and classification of deep learning, and provide a scientific basis for the protection and management of coastal wetlands.

[0016] Specifically, the method of the present invention combines multi-temporal data processing with deep learning technology to improve the accuracy and robustness of change detection. The present invention uses the Transformer feature extractor and the domain alignment semantic guidance module to model the semantic topological relationship between coastal wetland vegetation in deep semantics, which can effectively solve the problem of fragmented distribution of typical vegetation in coastal wetlands. At the same time, the multi-task decoder is used to decode the semantic topological relationship, which can effectively utilize the difference relationship between the two phases to obtain more accurate results. Finally, the results are vectorized and implemented, and the model results are converted into products that can be directly used by the business department. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the flow of the semantic change detection method of remote sensing images of typical vegetation in coastal wetlands of the present invention;

[0018] Figure 2 This is a schematic diagram of the remote sensing semantic change detection model of typical vegetation in coastal wetlands used in the method of the present invention;

[0019] Figure 3 It is a schematic diagram of the process of using the domain alignment module in the method of the present invention;

[0020] Figure 4 A schematic diagram of the process of deep learning, masking and vectorization processing in the method of the present invention;

[0021] Figure 5 It is the remote sensing image of 2022 after preprocessing in the specific application example of the present invention;

[0022] Figure 6 It is the remote sensing image of 2024 after preprocessing in the specific application example of the present invention;

[0023] Figure 7 The true value map of typical vegetation change detection in coastal wetlands obtained in a specific application example of the present invention;

[0024] Figure 8 A typical vegetation change detection map of coastal wetlands obtained in a specific application example of the present invention;

[0025] Fig. 9 The first phase of the typical vegetation semantic change detection vector diagram of the coastal wetland obtained in the specific application example of the present invention;

[0026] Fig.10 It is a semantic change detection vector diagram of typical vegetation in coastal wetlands in the second phase obtained in a specific application example of the present invention. DETAILED DESCRIPTION

[0027] like Figure 1 As shown, a method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands includes the following steps:

[0028] S1: Obtain several remote sensing image pairs containing typical vegetation of coastal wetlands that can be used for dual-temporal change detection. , , and preprocess the data to produce a remote sensing semantic change detection dataset of typical vegetation in coastal wetlands .

[0029] Specifically, the steps include:

[0030] S11: Select the acquired remote sensing image pairs containing typical vegetation of coastal wetlands that can be used for dual-temporal change detection. , Radiometric calibration, atmospheric correction, orthorectification and image registration are performed to obtain the preprocessed remote sensing image pair for dual-phase semantic change detection, i.e., the first-phase remote sensing image. and the second phase remote sensing image .

[0031] S12: Combine the pre-processed dual-temporal semantic change detection remote sensing image pairs with the existing professional domain knowledge and corresponding software and the field exploration map. and Visual interpretation of typical vegetation was performed to obtain typical vegetation coverage maps of coastal wetlands at different phases. and .

[0032] S13: Typical vegetation coverage map of coastal wetlands and Perform logical AND-NOT operations to obtain the true value map of typical vegetation change detection in coastal wetlands .

[0033] S14: Typical vegetation coverage map of coastal wetlands and The true value map of typical vegetation change detection in coastal wetlands Perform mask operation to obtain the semantic change detection truth map of typical vegetation in coastal wetlands and .

[0034] S15: Remote sensing image pairs for bi-temporal semantic change detection after preprocessing , And the true value map of semantic change detection of typical vegetation in coastal wetlands and The training dataset is created by cropping the images to 256×256 pixels with a 25% overlap ratio. .

[0035] S2: Calculate the brightness, normalized difference vegetation index (NDVI) and normalized difference water index (NDWI) of the remote sensing image containing typical vegetation of coastal wetlands and capable of being used for dual-temporal change detection, and send it and the preprocessed remote sensing image for dual-temporal semantic change detection into the twin Transformer feature extractor, such as Figure 2 As shown in the figure, a deep semantic multi-layer feature map reflecting the characteristics of coastal wetland features is extracted.

[0036] S2 specifically includes the following steps:

[0037] S21: Calculate the first phase remote sensing image and the second phase remote sensing image The normalized difference vegetation index image is as follows:

[0038] ;

[0039] Wherein, NDVI is the normalized difference vegetation index image, NIR is the reflectance image of the near infrared band, and Red is the reflectance image of the red light band.

[0040] S22: Calculate the normalized difference water index image between the first phase remote sensing image and the second phase remote sensing image. The formula is as follows:

[0041] ;

[0042] Wherein, NDWI is the normalized difference water index image, NIR is the reflectance image of the near infrared band, and Green is the reflectance image of the green band.

[0043] S23: splicing the NDVI image, NDWI image and the pre-processed remote sensing image for dual-temporal semantic change detection along the channel direction, dividing them into image blocks of a fixed size, performing a linear transformation on each image block, and obtaining a semantic labeling vector.

[0044] S24: Add the semantic tag vector to the position embedding vector of the image block to obtain a semantic tag vector with position encoding information.

[0045] S25: Send the semantic tag vector with position encoding information into the feature extractor to obtain a deep semantic multi-layer feature map.

[0046] S3: The extracted deep semantic multi-layer feature map is sent to the domain alignment module to obtain the semantically guided feature map of the dual-temporal coastal wetland features after domain alignment.

[0047] like Figure 3 As shown, S3 specifically includes the following steps:

[0048] S31: Soft clustering of deep semantic multi-layer feature graphs to obtain the corresponding structural feature representation of the i-th layer and j-th phase graph ,in represents the node set of the i-th layer and the j-th phase, represents the set of feature matrices of the i-th layer and the j-th phase, represents the node adjacency matrix of the i-th layer and the j-th phase;

[0049] like .

[0050] S32: Combine three sets of feature matrices and , and , and Transformation is performed to obtain three sets of semantically guided feature map structure information and , and , and .

[0051] The specific steps include:

[0052] S321: Set the feature matrix and Use a multi-layer perceptron to transform the features and obtain a graph representation of the query of the bi-phase feature matrix and , key diagram representation and , graphical representation of values and .

[0053] S322: The query graph representation and the key graph representation of the bi-phase feature matrix are calculated using the softmax function to obtain a similarity matrix. The calculation formula is as follows:

[0054] ;

[0055] S323: The similarity matrix is ​​guided by the bi-temporal semantic information to obtain a set of graph feature matrices after semantic guidance. The calculation formula is as follows:

[0056] ;

[0057] S324: Perform graph convolution reasoning on the semantically guided graph feature matrix set to obtain the first layer of semantically guided feature graph structure information. The calculation formula is as follows:

[0058] ;

[0059] Represents the graph convolution function.

[0060] S33: Re-map the semantically guided feature map structure information back to the original feature space to obtain a domain-aligned bi-temporal feature map of semantically guided features.

[0061] S4: Subtract the obtained dual-temporal coastal wetland feature semantic guidance feature map and send it to the multi-task decoder to obtain the coastal wetland change detection map Typical vegetation category prediction map of dual-temporal coastal wetland remote sensing images and .

[0062] S4 specifically includes the following steps:

[0063] S41: The dual-temporal coastal wetland feature element semantic guidance feature map includes a first-temporal semantic guidance feature map and a second-temporal semantic guidance feature map. The first-temporal semantic guidance feature map and the second-temporal semantic guidance feature map are subtracted to obtain a coastal wetland dual-temporal feature element semantic change feature map.

[0064] S42: Send the coastal wetland dual-temporal feature semantic change feature map and the coastal wetland dual-temporal feature semantic guidance feature map (first-phase semantic guidance feature map and second-phase semantic guidance feature map) into the multi-task decoder to obtain the coastal wetland binary change detection map and the dual-temporal coastal wetland remote sensing image typical vegetation category prediction map.

[0065] S42 specifically includes the following steps:

[0066] S421: Send the semantic change feature map of the coastal wetland dual-temporal feature elements, the first-temporal semantic guidance feature map, and the second-temporal semantic guidance feature map to the neural network CNN1 and the neural network CNN2, such as Figure 4 As shown, a typical vegetation difference enhancement feature map of the first phase and a typical vegetation difference enhancement feature map of the second phase are obtained.

[0067] S422: Subtract the first phase typical vegetation difference enhancement feature map and the second phase typical vegetation difference enhancement feature map and send them to the change detection head to obtain a coastal wetland binary change detection map.

[0068] S423: Send the first-phase typical vegetation difference enhancement feature map and the second-phase typical vegetation difference enhancement feature map to their respective segmentation heads to obtain the first-phase typical vegetation category prediction map and the second-phase typical vegetation category prediction map.

[0069] S424: Define the loss function of each task in the multi-task learning framework. The total loss function is the weighted sum of the loss functions of each task:

[0070] ;

[0071] in, It is The loss function of each task is is the corresponding weight coefficient, is the total number of tasks.

[0072] S425: For the coastal wetland change detection task, a binary cross entropy loss function is used Binary change detection map of coastal wetlands The true value map of typical vegetation change detection in coastal wetlands The difference between:

[0073] ;

[0074] in, It is a truth graph In the figure, the pixel's true label is 0, which indicates no change (this is actually the binary cross entropy loss, which is mainly used to calculate change detection), and 1 indicates change. It is the model prediction graph The value of the corresponding pixel.

[0075] S426: For the first phase ground feature category prediction task and the second phase ground feature category prediction task, a pixel-level cross entropy loss function is used to measure the typical vegetation category prediction map of coastal wetland remote sensing images. and truth value graph The difference between Represents the i-th phase:

[0076] ;

[0077] in, is the true label of category b, is the probability that the model predicts class b, and B is the total number of classes.

[0078] S5: Using coastal wetland change detection maps Prediction map of typical vegetation categories based on dual-temporal coastal wetland remote sensing images and Perform masking and vectorization processing, such as Figure 4 As shown in the figure, the semantic change detection vector diagram of typical vegetation in coastal wetlands is obtained. and .

[0079] S5 specifically includes the following steps:

[0080] S51: Binary change detection map of coastal wetlands The typical vegetation category prediction map of the dual-temporal coastal wetland remote sensing image (the first phase typical vegetation category prediction map And the prediction map of typical vegetation categories in the second phase ) to perform logical AND operation, and obtain the semantic change map of typical vegetation in coastal wetlands in two time phases (the semantic change map of typical vegetation in coastal wetlands in the first time phase Semantic change map of typical vegetation in coastal wetlands in the second phase ).

[0081] S52: Semantic change map of typical vegetation in coastal wetlands in two phases and Perform image vectorization to obtain a semantic change detection vector map of typical vegetation in coastal wetlands and .

[0082] The mask processing in S51 specifically includes the following steps:

[0083] S511: Defining Logic and Operations It is used to convert the coastal wetland binary change detection map (also called coastal wetland semantic change detection map) Typical vegetation category prediction map for the first phase Perform logical AND operations to obtain the semantic change map of typical vegetation in coastal wetlands in the first phase , the expression of this operation is: .

[0084] S512: Similarly, define the logical AND operation Used to convert the coastal wetland binary change detection map Typical vegetation category prediction map for the second phase Perform logical AND operation to obtain the semantic change map of typical vegetation in coastal wetlands in the second phase , the expression of this operation is: .

[0085] The present invention will be further described below in conjunction with specific application examples.

[0086] A method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands comprises the following steps:

[0087] S1: Obtain several remote sensing image pairs containing typical vegetation of coastal wetlands that can be used for dual-temporal semantic change detection, perform data preprocessing, and visually interpret the typical vegetation in the remote sensing images based on existing professional knowledge and corresponding software combined with field survey experience to produce the corresponding remote sensing semantic change detection dataset of typical vegetation in coastal wetlands.

[0088] S2: Calculate the brightness, normalized difference vegetation index and normalized difference water index of the remote sensing image containing typical vegetation of coastal wetlands and capable of being used for dual-temporal change detection, and send it and the preprocessed remote sensing image for dual-temporal semantic change detection into the feature extractor to extract the deep semantic multi-layer feature map reflecting the characteristics of coastal wetland land features.

[0089] S3: The extracted bi-temporal high-level semantic feature map is sent to the domain alignment module to obtain the semantically guided feature map of the bi-temporal coastal wetland features after domain alignment.

[0090] S4: Subtract the obtained coastal wetland dual-temporal coastal wetland feature semantic guidance feature map and send it to the multi-task decoder to obtain the coastal wetland change detection map and the dual-temporal coastal wetland remote sensing image typical vegetation category prediction map.

[0091] S5: Use the coastal wetland change detection map to perform mask processing and vectorization on the typical vegetation category prediction map of the dual-temporal coastal wetland remote sensing image to obtain the semantic change detection vector map of typical vegetation in coastal wetlands.

[0092] In the above steps: the remote sensing images for dual-temporal semantic change detection of coastal wetlands are subjected to radiometric calibration, atmospheric correction, orthorectification and image registration to obtain the pre-processed remote sensing images for dual-temporal semantic change detection, such as Figure 5 and Figure 6 As shown, the shooting time is October 17, 2022 and May 13, 2024, the longitude and latitude are 37°38′41.104′′~ 37°54'12.663′′N, 119°0′18.676′′~ 119°19′35.094′′E, and the shooting satellite and sensor are GF1 WFV 16m resolution.

[0093] The typical vegetation coverage map of coastal wetlands in dual-phase is subjected to logical AND-NOT operations to obtain the true value map of typical vegetation change detection in coastal wetlands, such as Figure 7 shown.

[0094] The preprocessed dual-temporal coastal wetland typical vegetation remote sensing image pair and its semantic change detection truth map are cropped according to the size of 256×256 pixels and the overlap ratio of 25% to produce a training data set; Figure 8 shown.

[0095] The image vectorization of the semantic change map of typical vegetation in coastal wetlands in dual time phases is performed to obtain the semantic change detection vector map of typical vegetation in coastal wetlands; Fig. 9 , Fig.10 As shown. Figure 5-Figure 10 It can be seen that the semantic change detection results can be used to observe the growth and degradation trends of key protected vegetation in coastal wetland reserves, which is helpful for the ecological protection of coastal wetlands.

[0096] The parts not described in the above methods can be realized by adopting or drawing on existing technologies.

[0097] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A semantic change detection method for remote sensing images of typical vegetation in coastal wetlands, characterized in that The following steps are involved: S1: Obtain remote sensing images containing typical vegetation of coastal wetlands and capable of being used for dual-temporal change detection, perform data preprocessing on them, and produce a semantic change detection dataset of remote sensing images of typical vegetation of coastal wetlands; S2: Calculate the brightness, normalized difference vegetation index and normalized difference water index of the remote sensing images in the semantic change detection dataset of typical vegetation remote sensing images of coastal wetlands, and send them and the remote sensing images in the semantic change detection dataset of typical vegetation remote sensing images of coastal wetlands into the feature extractor to extract the deep semantic multi-layer feature map reflecting the characteristics of coastal wetland features; S3: The extracted deep semantic multi-layer feature map is sent to the domain alignment module to obtain the semantic guidance feature map of the dual-temporal coastal wetland features after domain alignment; S4: Subtract the obtained semantic guidance feature map of the dual-temporal coastal wetland features and send it to the multi-task decoder to obtain the coastal wetland binary change detection map and the typical vegetation category prediction map of the dual-temporal coastal wetland remote sensing image; S5: Use the coastal wetland binary change detection map to perform mask processing and vectorization processing on the typical vegetation category prediction map of the dual-temporal coastal wetland remote sensing image to obtain the coastal wetland typical vegetation semantic change detection vector map, and complete the semantic change detection of the coastal wetland typical vegetation remote sensing image.

2. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 1, characterized in that: S1 includes the following steps: S11: Select the acquired remote sensing image pairs containing typical vegetation of coastal wetlands and capable of being used for dual-temporal change detection. , Perform radiometric calibration, atmospheric correction, orthorectification and image registration preprocessing to obtain the preprocessed remote sensing image pair for dual-temporal change detection. , i.e. the first phase remote sensing image and the second phase remote sensing image ; S12: Remote sensing image pair for dual-phase change detection after preprocessing and Visual interpretation of typical vegetation was performed to obtain typical vegetation coverage maps of coastal wetlands at different phases. and ; S13: Typical vegetation coverage map of coastal wetlands and Perform logical AND-NOT operations to obtain the true value map of typical vegetation change detection in coastal wetlands ; S14: Typical vegetation coverage map of coastal wetlands and The true value map of typical vegetation change detection in coastal wetlands Perform mask operation to obtain the semantic change detection truth map of typical vegetation in coastal wetlands and ; S15: Remote sensing image pair for dual-phase change detection after preprocessing , And the true value map of semantic change detection of typical vegetation in coastal wetlands and The training dataset is created by cropping the images to 256×256 pixels with a 25% overlap ratio. .

3. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 1, characterized in that: S2 includes the following steps: S21: Calculate the first phase remote sensing image and the second phase remote sensing image The normalized difference vegetation index image is as follows: ； In the formula, NDVI is the normalized difference vegetation index image, NIR is the reflectance image of the near infrared band, and Red is the reflectance image of the red light band; S22: Calculate the first phase remote sensing image and the second phase remote sensing image The normalized difference water index image is as follows: ； Wherein, NDWI is the normalized difference water index image, NIR is the reflectance image of the near infrared band, and Green is the reflectance image of the green light band; S23: splice NDVI, NDWI and the remote sensing images in the dataset along the channel direction and divide them into image blocks of fixed size, perform linear transformation on each image block and obtain a semantic label vector; S24: adding the semantic tag vector to the position embedding vector of the image block to obtain a semantic tag vector with position encoding information; S25: Send the semantic tag vector with position encoding information into the feature extractor to obtain a deep semantic multi-layer feature map.

4. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 1, characterized in that: S3 includes the following steps: S31: Soft clustering of deep semantic multi-layer feature maps to obtain the structural feature representation of the i-th layer and j-th phase map ,in represents the node set of the i-th layer and the j-th phase, represents the feature matrix set of the i-th layer and the j-th phase, represents the node adjacency matrix of the i-th layer and the j-th phase; S32: Set the feature matrix Transform to obtain semantically guided feature map structure information ; S33: Re-map the semantically guided feature map structure information back to the original feature space to obtain a domain-aligned bi-temporal coastal wetland feature element semantically guided feature map.

5. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 4, characterized in that: The domain alignment modules are provided with three in total; the feature matrix sets are provided with three in total, which are respectively represented as ; The three sets of feature matrices are combined and , and , and Transformation is performed to obtain three sets of semantically guided feature map structure information and , and , and ; The structural information of the three sets of semantically guided feature maps are remapped back to the original feature space to obtain the three sets of domain-aligned bi-temporal feature maps of semantically guided features.

6. A method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 5, characterized in that: S32 includes the following steps: S321: Set the feature matrix and Use a multi-layer perceptron to transform the features and obtain a graph representation of the query of the bi-phase feature matrix and , key diagram representation and , graphical representation of values and ; S322: The query graph representation and the key graph representation of the bi-phase feature matrix are calculated using the softmax function to obtain a similarity matrix. The calculation formula is as follows: ； S323: The similarity matrix is ​​guided by the bi-temporal semantic information to obtain a set of graph feature matrices after semantic guidance. The calculation formula is as follows: ； S324: Perform graph convolution reasoning on the semantically guided graph feature matrix set to obtain the first layer of semantically guided feature graph structure information. The calculation formula is as follows: ； Represents the graph convolution function.

7. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 1, characterized in that: S4 includes the following steps: S41: The binary temporal coastal wetland feature element semantic guidance feature map includes a first temporal semantic guidance feature map and a second temporal semantic guidance feature map, and the first temporal semantic guidance feature map and the second temporal semantic guidance feature map are subtracted to obtain a binary temporal coastal wetland feature element semantic change feature map; S42: Send the coastal wetland dual-temporal feature semantic change feature map, the first-temporal semantic guidance feature map, and the second-temporal semantic guidance feature map into the multi-task decoder to obtain a coastal wetland binary change detection map and a dual-temporal coastal wetland remote sensing image typical vegetation category prediction map.

8. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 7, characterized in that: S42 includes the following steps: S421: Sending the coastal wetland dual-temporal feature semantic change feature map, the first-temporal semantic guidance feature map, and the second-temporal semantic guidance feature map to the neural network CNN1 and the neural network CNN2 to obtain the first-temporal typical vegetation difference enhancement feature map and the second-temporal typical vegetation difference enhancement feature map; S422: Subtract the first phase typical vegetation difference enhancement feature map and the second phase typical vegetation difference enhancement feature map and send them to the change detection head to obtain a coastal wetland binary change detection map; S423: sending the first-phase typical vegetation difference enhancement feature map and the second-phase typical vegetation difference enhancement feature map to their respective segmentation heads to obtain the first-phase typical vegetation category prediction map and the second-phase typical vegetation category prediction map; S424: Define the loss function of each task in the multi-task learning framework. The total loss function is the weighted sum of the loss functions of each task: ； in, It is The loss function of each task is is the corresponding weight coefficient, is the total number of tasks; S425: For the coastal wetland change detection task, a binary cross entropy loss function is used Binary change detection map of coastal wetlands The true value map of typical vegetation change detection in coastal wetlands The difference between: ； in, is the true label of the corresponding pixel in the truth map, It is the model prediction graph The value of the corresponding pixel; S426: For the first phase ground feature category prediction task and the second phase ground feature category prediction task, a pixel-level cross entropy loss function is used to measure the typical vegetation category prediction map of coastal wetland remote sensing images. and truth value graph The difference between Represents the i-th phase: ； in, is the true label of category b, is the probability that the model predicts class b, and B is the total number of classes.

9. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 1, characterized in that: S5 includes the following steps: S51: Binary change detection map of coastal wetlands Respectively with the first phase typical vegetation category prediction map And the prediction map of typical vegetation categories in the second phase Perform logical AND operations to obtain the semantic change map of typical vegetation in coastal wetlands in the first phase Semantic change map of typical vegetation in coastal wetlands in the second phase ; S52: Semantic change map of typical vegetation in coastal wetlands in the first phase Semantic change map of typical vegetation in coastal wetlands in the second phase The image is vectorized to obtain a semantic change detection vector map of typical vegetation in coastal wetlands.

10. The method for detecting semantic changes in remote sensing images of typical vegetation in coastal wetlands according to claim 9, characterized in that: The mask processing in S51 specifically includes the following steps: S511: Defining Logic and Operations Used to convert coastal wetland binary change detection map Prediction map of typical vegetation categories in the first phase Perform logical AND operations to obtain the semantic change map of typical vegetation in coastal wetlands in the first phase , the expression of this operation is: ; S512: Similarly, define the logical AND operation Used to convert coastal wetland binary change detection map Typical vegetation category prediction map for the second phase Perform logical AND operation to obtain the semantic change map of typical vegetation in coastal wetlands in the second phase , the expression of this operation is: .

Images