Spadix spartina remote sensing image cross-domain semantic segmentation method based on ECAM-ResUNet2

Abstract

The application discloses a kind of ECAM-ResUNet2-based Spartina alterniflora remote sensing image cross-domain segmentation method, including the construction of efficient convolution attention module ECAM;With ResUNet model as basic network architecture, ResUNet2 model is constructed, and the context semantic information of ResUNet2 is input into ECAM module, and ECAM-ResUNet2 network model is constructed;According to the region to be predicted and training region is divided into target domain and source domain respectively, the remote sensing image data of source domain is carried out image cutting and image enhancement;The source domain data and target domain data are used as training sample to train ECAM-ResUNet2, and the trained ECAM-ResUNet2 model is obtained;The remote sensing image of the region to be predicted is classified by the trained ECAM-ResUNet2 model.The application is based on the idea of semantic segmentation and cross-domain alignment, can be applied to the prediction of the distribution of invasive species Spartina alterniflora in remote sensing image, and can ensure that the prediction accuracy of Spartina alterniflora in remote sensing image of different time and different scene is high.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a cross-domain semantic segmentation method for remote sensing images of Spartina alterniflora based on ECAM-ResUNet2. Background Technology

[0002] The coastal wetlands of Jiangsu Province are the wetlands with the largest area of ​​the invasive plant Spartina alterniflora in my country. Spartina alterniflora has existed in the wetlands for nearly 40 years, posing a serious threat to the security of the coastal wetland ecosystem and is listed as one of the most serious invasive alien plants. Therefore, it is urgent to effectively describe the invasion dynamics and distribution of Spartina alterniflora in the coastal areas of China.

[0003] Many studies have attempted to investigate the invasion dynamics of Spartina alterniflora in China at regional or local scales, but traditional ground surveys struggle to provide a comprehensive understanding of the extent of its invasion. Compared to traditional field surveys, remote sensing technology offers advantages such as large monitoring areas and long data time series. It can effectively and accurately acquire spatiotemporal distribution information of Spartina alterniflora and is therefore widely used to study the expansion of Spartina alterniflora in China.

[0004] Early methods for extracting Spartina alterniflora from remote sensing images included traditional machine learning approaches such as maximum likelihood estimation, support vector machines, and random forests. Some researchers also used training decision rules to classify Spartina alterniflora. While traditional machine learning methods can quickly extract Spartina alterniflora regions from remote sensing multispectral images, their ability to handle complex functional models is limited, and their robustness and automation levels are low, making it difficult to improve the classification accuracy of complex coastal wetland scenes. Therefore, extracting Spartina alterniflora information from large-scale, long-term remote sensing data remains a challenging task, requiring the development of more robust classification methods.

[0005] Deep learning, in essence, transforms the low-level attributes of raw data into more robust abstract features by learning multi-layered nonlinear network structures. This enables it to learn the essential characteristics of data from samples, representing a new field in machine learning. Currently, some deep learning-based methods for identifying Spartina alterniflora primarily employ classification frameworks based on various improvements to the Convolutional Neural Network (CNN) model. Compared to segmentation, classification requires inputting each pixel of the image into the model for discrimination, resulting in higher time consumption and lower accuracy. Semantic segmentation, on the other hand, classifies each pixel of an image according to its semantic label. This method can quickly classify pixels containing Spartina alterniflora in remote sensing images, achieving the goal of extracting Spartina alterniflora regions from the image.

[0006] Meanwhile, most existing methods for extracting Spartina alterniflora are designed for identifying Spartina alterniflora within the same scene. When transformed into a cross-scene task, the classification accuracy drops significantly. Furthermore, due to the varying species complexity across different scenes, cross-domain segmentation of Spartina alterniflora in remote sensing images based on classic deep learning models remains challenging. Summary of the Invention

[0007] The purpose of this invention is to propose a cross-domain semantic segmentation method for Spartina alterniflora remote sensing images based on ECAM-ResUNet2, which can maintain a fast model training speed, as well as high Spartina alterniflora extraction accuracy and low computational complexity in cross-scene prediction.

[0008] To achieve the above-mentioned objectives, the technical solution of the present invention is as follows:

[0009] A cross-domain semantic segmentation method for remote sensing images of Spartina alterniflora based on ECAM-ResUNet2 includes the following steps:

[0010] S1: Construct an efficient convolutional attention module (ECAM);

[0011] S2: Using the ResUNet model as the basic network architecture, construct the ResUNet2 model, input the ResUNet2 contextual semantic information into the ECAM module, and construct the ECAM-ResUNet2 network model;

[0012] S3: Divide the region to be predicted and the training region into the target region and the source region respectively, and perform image cropping and image enhancement on the remote sensing image data of the source region;

[0013] S4: Use the source domain data and target domain data as training samples to train ECAM-ResUNet2, and obtain the trained ECAM-ResUNet2 model;

[0014] S5: Classify the remote sensing images of the region to be predicted using the trained ECAM-ResUNet2 model, and then stitch together the redundantly segmented prediction results obtained from the classification.

[0015] Further, step S1 includes the following sub-steps:

[0016] Step 1.1: Construct a feature fusion module. The input is two feature maps concatenated from ResUNet2 features. Add the corresponding elements of these two feature maps to obtain the synthesized feature map.

[0017] Step 1.2: Construct the convolutional attention module. This module combines channel attention and spatial attention mechanisms. Through the channel attention mechanism, the network focuses more on the near-infrared band where Spartina alterniflora stands out in the remote sensing image. At the same time, using the spatial attention mechanism, the network pays more attention to the segmentation of the region where Spartina alterniflora is located, avoiding missegmentation or omission of small areas of Spartina alterniflora.

[0018] Step 1.3: Construct an efficient convolutional attention module. This module has two inputs: one is the synthesized feature map obtained from the feature combination module in Step 1.1, and the other is the feature map from the downsampling layer in the same layer. These are then fed into the convolutional attention module in Step 1.2.

[0019] Further, step S2 includes the following sub-steps:

[0020] Step 2.1: Construct the ResUNet2 model. It is constructed by reducing the network depth of ResUNet2. To minimize information loss, the original 4-layer encoder-decoder structure is reduced to a 2-layer encoder-decoder structure.

[0021] Step 2.2: Construct the ECAM-ResUNet2 network model. This network is a deep learning-based semantic segmentation network that can be applied to semantic segmentation of regions containing Spartina alterniflora in remote sensing images to achieve extraction. Before training, the input of the ECAM-ResUNet2 network needs to be set. The input and output sizes of the ECAM-ResUNet2 network are set according to the length, width, and number of spectra in the multispectral image training dataset.

[0022] Step 2.3: Based on step 3.1, set the required parameter values ​​for the network according to the multispectral image training dataset, specifically the learning rate, optimization function, loss function, etc.

[0023] Furthermore, step S3 includes the following sub-steps:

[0024] Step 3.1: Select several multispectral remote sensing images of different regions, the same resolution, containing the same spectral information, and of good image quality;

[0025] Step 3.2: First, delete inconsistent spectral information in the remote sensing image, fill zero values ​​around the original image, and crop the original remote sensing image by using half the length of the sub-remote sensing image as the horizontal cropping step and half the width as the vertical cropping step, to obtain several sub-remote sensing images with the same length, width, number of pixels and spectral dimension.

[0026] Step 3.3: Organize the sub-remote sensing image files and their corresponding label files. Based on the training region, divide the sub-remote sensing image files and their corresponding label files into source regions. Perform data augmentation on source region sub-remote sensing images with a high proportion of Spartina alterniflora area by flipping, rotating, and adding Gaussian white noise. Based on the region to be predicted, divide the sub-remote sensing image files into target regions. Target regions do not contain corresponding label files. The source and target regions form a multispectral image training set.

[0027] Furthermore, step S4 includes the following sub-steps:

[0028] Step 4.1: Set the number of training iterations for the ECAM-ResUNet2 network;

[0029] Step 4.2: Optimize the loss function using the mean difference between the target domain data and the source domain data projected onto the Hilbert space from the model output. The loss function of the ECAM-ResUNet2 model is:

[0030] Loss = Loss IOU +lamda*Loss MMD

[0031] Where Loss represents the total loss, Loss IOU The intersection-union ratio loss used to predict the source domain probability is represented by Loss. MMD The loss represents the difference in projected means between the target and source domains, where lambda represents the regularization parameter controlling the alignment of different contributions of the distributions, and:

[0032]

[0033] Where, n s n t These are the number of samples in the source and target domains, respectively. s z t These are the features extracted from the source domain and the target domain, respectively, and the kernel r(·,·) represents the inner product of the features in the Hilbert space.

[0034] The source and target domain data are input into the ECAM-ResUNet2 network constructed in step 2 to begin training. During this process, the backpropagation loss function includes not only the IOULoss calculated between the source domain data and its corresponding labeled file, but also the mean difference MMDLLoss obtained by projecting the predicted graphs of the target and source domains from the model. During training, the network continuously optimizes its parameters through backpropagation of the loss function, achieving a learning effect.

[0035] Step 4.3: After training is complete, save the network parameters that have been debugged.

[0036] Furthermore, step S5 includes the following sub-steps:

[0037] Step 5.1: In Step 3, the remote sensing image of the area to be predicted containing Spartina alterniflora has been segmented. No processing is required; the target domain data is the test dataset, forming a multispectral image test dataset.

[0038] Step 5.2: Input the multispectral image test dataset into the trained ECAM-ResUNet2 network in sequence to obtain the segmentation results of the multispectral image test dataset, that is, the Spartina alterniflora extraction results of cross-regional multispectral images;

[0039] Step 5.3: Use one-quarter of the length and one-quarter of the width of the sub-remote sensing image as the horizontal segmentation step and the vertical segmentation step, respectively. Crop the four sides of the segmentation result and stitch them together according to the segmentation order of the target domain in Step 3.2 to finally obtain the complete remote sensing image Spartina alterniflora extraction result.

[0040] This invention proposes a deep learning semantic segmentation model for multispectral images extracted from Spartina alterniflora using attention mechanisms and cross-domain learning principles, combined with the feature learning capabilities of deep learning. Compared to existing technologies, the advantages of this invention are:

[0041] (1) This method proposes a complete multispectral image Spartina alterniflora extraction model. Using this method, Spartina alterniflora can be extracted from remote sensing images of different regions, which will help researchers to further understand the distribution characteristics of Spartina alterniflora in wetlands.

[0042] (2) This method uses semantic segmentation to extract Spartina alterniflora. Semantic segmentation can classify each pixel in remote sensing images to achieve the extraction purpose. Through semantic segmentation, researchers can obtain extraction result images with the same length and width as multispectral remote sensing images. Compared with other classification methods, the extraction results obtained by semantic segmentation are more intuitive and faster, greatly improving the extraction efficiency of wooded areas;

[0043] (3) This method proposes a cross-domain segmentation framework, which adds the difference between the mean of the target domain and the source domain after projection into the Hilbert space to the loss function, and at the same time uses the source domain to train the model, transfers the shared knowledge in the source domain to the target domain, and realizes the transfer of the model to different scenarios. Attached Figure Description

[0044] Figure 1 This is a flowchart of the cross-domain semantic segmentation method for Spartina alterniflora remote sensing images based on ECAM-ResUNet2 according to the present invention;

[0045] Figure 2 This is a diagram of the ResUNet2 network structure constructed in this invention;

[0046] Figure 3 This is a diagram of the efficient convolutional attention module (ECAM) network structure constructed in this invention;

[0047] Figure 4 This is a diagram of the ECAM-ResUNet2 network structure constructed in this invention;

[0048] Figure 5 This is a diagram of the cross-domain segmentation network architecture constructed in this invention;

[0049] Figure 6 This is a schematic diagram illustrating the redundant cutting of the segmentation results in the embodiment;

[0050] Figure 7 This is an example of a multispectral image source domain image pair involved in the embodiments (2020);

[0051] Figure 8 This is the result of extracting Spartina alterniflora from remote sensing images of the target domain using the trained ECAM-ResUNet2 in the example (2010). Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0053] like Figure 1 As shown, a cross-domain semantic segmentation method for Spartina alterniflora remote sensing images based on ECAM-ResUNet2 includes the following steps:

[0054] Step 1: Construct an efficient convolutional attention module (ECAM);

[0055] Step 2: Construct the ResUNet2 model based on the ResUNet network, and combine ECAM and ResUNet2 to form the ECAM-ResUNet2 network;

[0056] Step 3: Divide the multispectral remote sensing data into target domain data and source domain data according to the region to be predicted and the training region;

[0057] Step 4: Train ECAM-ResUNet2 using source domain data and target domain data;

[0058] Step 5: Use the trained ECAM-ResUNet2 to extract remote sensing images of the target region containing Spartina alterniflora.

[0059] Specifically, step 1 includes the following steps:

[0060] Step 1.1: Construct the channel attention branch, which includes a 2D average pooling operation, two squeeze operations, a one-dimensional convolution operation, and an activation operation. The operation formula for this branch is as follows:

[0061] I out1 =AdaptiveAvgPool2d(I in (1)

[0062] I out2 =conv(Squeeze(Squeeze(I out1 (2)

[0063] I out3 =sigmoid(I out2 (3)

[0064] In the formula, I in It is the input of this branch, I out3 For the output of this branch, `AdaptiveAvgPool2d` is 2D average pooling, `conv` is a one-dimensional convolution operation with a kernel size of 5×5 and a stride of 1, `Squeeze` is a dimensionality reduction operation, and `Sigmoid` is the sigmoid function, with the formula:

[0065]

[0066] Step 1.2: Construct the spatial attention branch. This branch includes operations for finding the maximum and average channel dimensions, a 3×3 2D convolution operation, an activation operation, and a connection operation that combines the output feature maps of the maximum and average channel operations. The operation formula for this branch is:

[0067] I out1 =mean(I in (5)

[0068] I out2 =max(I in (6)

[0069] I out3 =sigmoid(conv(Concatenate(I out2 ,I out1 )))) (7)

[0070] In the formula, I in It is the input of this branch, I out3 The output of this branch is: mean is the operation of adding all channel dimensions and taking the average, max is the operation of taking the maximum value of the channel dimensions, and conv represents a two-dimensional convolution operation with a kernel size of 3×3 and a stride of 1.

[0071] Step 1.3: Construct an efficient convolutional attention module (ECAM), the specific structure of which is as follows: Figure 1 As shown, this module consists of branches from steps 1.1 and 1.2, and the operation formula for this module is:

[0072] I out =I in1 ×I in2 (8)

[0073] I in0 This is the output of the branch in step 1.1, I in1 These are the outputs of the branch in step 1.2. These two outputs serve as the inputs to this module. out This is the output of the module.

[0074] Specifically, step 2 includes the following steps:

[0075] Step 2.1: The initial network model used in this invention is ResUNet2, such as... Figure 2 As shown, this network is based on ResUNet, reducing the original four encoder-decoder structures to two. The encoder structure includes two convolutional operations, two batch normalization operations, and two activation operations. Except for the deepest encoder structure, each structure includes a downsampling operation. Furthermore, the encoder structure employs a residual structure to extract features from remote sensing images, resulting in better feature extraction performance. The operation formula for the encoder structure including downsampling is as follows:

[0076] I out1 =ReLU(BN(Conv(I in ))) (9)

[0077] I out2 =ReLU(BN(Conv(I out1 ))) (10)

[0078] I out3 =MaxPooling(I out2 (11)

[0079] In the formula, I in For the encoder downsampling input, I out3 The output of the encoder downsamples. `Conv` represents a 2D convolution operation with a kernel size of 3×3 and a stride of 1. `BN` represents batch normalization, and the formula for batch normalization is:

[0080]

[0081] ReLU is the activation operation, and MaxPooling represents downsampling, using the max pooling operation.

[0082] Step 2.2: Construct the ECAM-ResUNet2 network, as follows Figure 4 As shown. The decoding structure of this network, in addition to consisting of a series of upsampling operations, convolution, batch normalization, activation, etc., also includes an ECAM module, such as... Figure 3 Simultaneously, it utilizes connection operations to combine the output of the ECAM module with features from a decoder of the same size, and uses upsampling operations to recover image details. The operational formula for this decoding structure is:

[0083] I out1 =ECAM(I in +I' out )×I in0 (5)

[0084] I out2 =ReLU(BN(Conv(I out1 ))) (6)

[0085] I out3 =ReLU(BN(Conv(I out2 +I out1 ))) (7)

[0086] I out4 =UpSampling(I out3 (8)

[0087] I out5 =ReLU(BN(Conv(I out4 ))) (9)

[0088] In the formula, I′ out Indicates the output of the previous decoding structure (or 0 if there is no previous decoding structure), I in This represents the multispectral remote sensing image features extracted by the coding structure. Concatenate represents the concatenation operation, and UpSampling represents the upsampling operation, which can expand the length and width of the features to twice the size of the input.

[0089] Specifically, step 3 includes the following sub-steps:

[0090] Step 3.1: Acquire high-quality multispectral remote sensing image data from different years;

[0091] Step 3.2: Fill the perimeter of the multispectral remote sensing image data with zero values, and fill the length and width by one-quarter of the length and width of the cropped data, respectively;

[0092] Step 3.3: Divide the multispectral remote sensing image data into multispectral remote sensing image training year data and multispectral remote sensing image prediction year data according to the year;

[0093] Step 3.4: Crop the padded training year multispectral remote sensing image data and its corresponding label file to a standard size, i.e., each image needs to have the same length and width. At the same time, perform data augmentation on the cropped remote sensing image data. The segmented data is the source domain data. Crop the padded prediction year multispectral remote sensing image data (excluding the label file) into sub-remote sensing image data of the same standard size as the source domain data. The segmented data is the target domain data.

[0094] This method suggests that the image reference be 128 pixels long and 128 pixels wide. Square images are easier to learn. It is recommended to fill the image with 32 pixels long and 32 pixels wide. It is also recommended to perform data augmentation on the sub-remote sensing image data with 60% to 100% of the Spartina alterniflora area. Augmentation methods include rotation, flipping, and adding Gaussian white noise.

[0095] Specifically, step 4 includes the following sub-steps:

[0096] Step 4.1: Set the number of training iterations for the ECAM-ResUNet network;

[0097] Step 4.2: Construct a cross-domain segmentation framework, such as... Figure 5 As shown. This framework utilizes the mean difference between the target domain data and the source domain data obtained through the network after being projected onto the Hilbert space to optimize the loss function. The loss function of this cross-domain segmentation framework is:

[0098] Loss = Loss IOU +lamda*Loss MMD (10)

[0099] Where Loss represents the total loss, Loss IOU The intersection-union ratio loss used to predict the source domain probability is represented by Loss. MMD The loss represents the difference in projected means between the target and source domains, where lambda represents the regularization parameter controlling the alignment of different contributions of the distributions, and:

[0100]

[0101] Where, n s n t These are the number of samples in the source and target domains, respectively. s z t These are the features extracted from the source domain and the target domain, respectively, and the kernel r(·,·) represents the inner product of the features in the Hilbert space.

[0102] The source and target domain data obtained in step 3.3 are input into the ECAM-ResUNet2 network constructed in step 2 through the designed cross-domain segmentation framework to begin training the ECAM-ResUNet2 network. The network loss function, such as the cross-entropy loss function, is set according to the specific conditions of the multispectral image training dataset. After preparing the network and the multispectral image training dataset, the number of training iterations is set, typically around 30. However, the number of training iterations can be increased or decreased based on the training progress; for example, training can be stopped if the loss value does not decrease for an extended period.

[0103] Step 4.3: After training begins, all training images are input into the network at the start of each training round. The network adjusts its parameters according to the direction that minimizes the loss value of the loss function until the loss value is minimized, at which point the best training effect can be obtained.

[0104] Step 4.4: Repeat steps 3.1 and 3.2 to perform multiple training iterations. During each training iteration, the network's parameter values ​​can be modified. The network model with the best performance is selected and saved; this is called the optimal segmentation model.

[0105] Specifically, step 5 includes the following steps:

[0106] Step 5.1: The multispectral remote sensing data to be predicted, i.e. the target domain remote sensing data, has been processed in step 3.3. Input the target domain remote sensing data into the trained ECAM-ResUNet2 network in sequence. The network will obtain the Spartina alterniflora segmentation results for all images.

[0107] Step 5.2: Redundantly cut the segmentation results, with the length and width of each cut being one-quarter of the image patch, such as... Figure 6 As shown.

[0108] Step 5.3: The results obtained from the cutting in Step 5.2 are spliced ​​together in order to finally obtain the Spartina alterniflora extraction results from the multispectral test image. Figure 7 Example images show multispectral image source domain data and their labeled images, where black represents non-Spartina alterniflora regions and white represents regions where Spartina alterniflora exists. Figure 8 The results of extracting Spartina alterniflora from remote sensing images generated by the Landsat 8 satellite in 2010 are shown using a trained ECAM-ResUNet2 network. White areas represent regions where Spartina alterniflora is present, and black areas represent regions where Spartina alterniflora is not present.

[0109] This method recommends setting the cut length and width to 32 pixels each.

[0110] Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the essence of this invention, and these modifications and combinations are still within the protection scope of this invention.

Claims

1. A method for ECAM-ResUNet2-based remote sensing image cross-domain semantic segmentation of Spartina alterniflora, characterized in that, Comprising the following steps: S1: constructing an efficient convolution attention module ECAM; Comprising the following sub-steps: Step 1.1: constructing a feature combination module, the input is two feature maps of ResUNet2 feature splicing, adding the corresponding elements of the two feature maps to obtain a synthesized feature map; Step 1.2: constructing a convolution attention module; the combination of channel attention mechanism and spatial attention mechanism, through the channel attention mechanism, the network focuses on the near-infrared band highlighted by the Spartina alterniflora in the remote sensing image, and at the same time, the spatial attention mechanism is used, the network focuses on the segmentation of the region where the Spartina alterniflora is located; Step 1.3: constructing an efficient convolution attention module, having two inputs, one input is the synthesized feature map obtained by the feature combination module in step 1.1, and the other input is the feature of the down-sampling layer at the same layer, which is input into the convolution attention module in step 1.2; S2: constructing a ResUNet2 model based on the ResUNet model as the basic network architecture; input the context semantic information of the ResUNet2 into the ECAM module to construct an ECAM-ResUNet2 network model; Step 2.1: adopting the initial network model as ResUNet2, and based on ResUNet, reducing the original 4 encoding-decoding structures to 2 encoding-decoding structures, wherein the encoding structure contains two convolution operations, two batch normalization operations and two activation operations, and contains a down-sampling operation except for the encoding structure at the deepest part of the network; the encoding structure also adopts a residual structure to extract the features of the remote sensing image; Step 2.2: constructing an ECAM-ResUNet2 network, the decoding structure is composed of a series of up-sampling operations, convolution, batch normalization, activation operations, and also adds an ECAM module, and simultaneously uses a connection operation to combine the output of the ECAM module and the decoder features of the same size, and uses an up-sampling operation to restore the image details; S3: dividing the target domain and the source domain according to the to-be-predicted region and the training region respectively, and performing image cropping and image enhancement on the remote sensing image data of the source domain; S4: taking the source domain data and the target domain data as training samples to train the ECAM-ResUNet2, and obtaining a trained ECAM-ResUNet2 model; S5: using the trained ECAM-ResUNet2 model to classify the remote sensing image of the to-be-predicted region.

2. The ECAM-ResUNet2-based Spartina alterniflora remote sensing image cross-domain semantic segmentation method according to claim 1, characterized in that: Step S3 comprises the following sub-steps: Step 3.1: selecting a plurality of multispectral remote sensing images of different regions, same resolution, containing same spectral information, good image quality; Step 3.2: filling zero values around the multispectral remote sensing image data, and the filling length and width are respectively one quarter of the cropping length and width; Step 3.3: first, delete the inconsistent spectral information in the remote sensing image, fill zero values around the original image, take half of the length of the sub-remote sensing image as the horizontal cropping step, and half of the width as the vertical cropping step, crop the remote sensing image to obtain a plurality of sub-remote sensing images with the same length, width, pixel quantity and spectral dimension; Step 3.4: The sub-remote sensing image files and the corresponding label files are sorted; the sub-remote sensing image files and the corresponding label files are divided into a source domain according to the training region, and data augmentation is performed on the sub-remote sensing image of the source domain with a high Spartina alterniflora area ratio; the sub-remote sensing image files are divided into a target domain according to the region to be predicted, and the target domain does not contain the corresponding label files; the source domain and the target domain form a multispectral image training set.

3. The ECAM-ResUNet2-based Spartina alterniflora remote sensing image cross-domain semantic segmentation method according to claim 2, characterized in that: Step S4 includes the following sub-steps: Step 4.1: Set the number of training times of the ECAM-ResUNet2 network; Step 4.2: Input the source domain data and the target domain data into the ECAM-ResUNet2 network constructed in step 2 to start training, wherein the loss function of the back propagation contains not only the IOULoss calculated by the source domain data and the corresponding label file, but also the MMDLoss obtained by projecting the prediction map of the target domain and the source domain through the model; during the training process, the network continuously optimizes the network parameters through the back propagation of the loss function, and achieves the learning effect; Step 4.3: After the training is completed, the network parameters that have been debugged are saved.

4. The ECAM-ResUNet2-based Spartina alterniflora remote sensing image cross-domain semantic segmentation method according to claim 3, characterized in that: In step S4.2, the mean difference of the source domain data and the target domain data projected into Hilbert space obtained by using the model output is also used to optimize the loss function, and the loss function of the ECAM-ResUNet2 model is: Loss = Loss IOU + lambda * Loss MMD wherein Loss represents a total loss, Loss IOU represents a cross-entropy loss used for predicting the source domain probability, Loss MMD represents a projected mean difference loss of the target domain and the source domain, lamda represents a regularization parameter for controlling different contributions of distribution alignment, wherein: where n s , n t are the number of samples in the source domain and the target domain, respectively, z s , z t are the features extracted from the source domain and the target domain, respectively, and the kernel r(·, ·) represents the inner product of the features in the Hilbert space.

5. The ECAM-ResUNet2-based Spartina alterniflora remote sensing image cross-domain semantic segmentation method according to claim 2, characterized in that: Step S5 includes the following sub-steps: Step 5.1: In step 3, the remote sensing image of the region to be predicted containing Spartina alterniflora is divided, and the target domain data is the test data set, forming a multispectral image test data set; Step 5.2: The multispectral image test data set is input into the ECAM-ResUNet2 network which has completed training according to the order, and the segmentation result of the multispectral image test data set is redundantly cut; Step 5.3: The cutting result, i.e., the Spartina alterniflora extraction result of the cross-regional multispectral image, is obtained, and the result is spliced according to the segmentation order of the target domain in step 3.2, and finally the complete remote sensing image Spartina alterniflora extraction result is obtained.

Images