CN118365882A - Optical remote sensing image segmentation method based on VMamba model - Google Patents

Abstract

The invention discloses an optical remote sensing image segmentation method based on VMamba models, which comprises the following steps: acquiring an optical remote sensing image of a scene to be segmented; constructing a VM-UNet model based on VMamba model; training the constructed VM-UNet model by utilizing the optical remote sensing image to obtain a segmentation model; and (5) utilizing the segmentation model to complete segmentation of the optical remote sensing image. The present invention introduces a Visual State Space (VSS) block as a basic block to capture extensive context information and builds an asymmetric encoder-decoder structure that can capture powerful remote information and preserve linear computational complexity. The defects of the CNN model, such as the limitation of local receptive field, the defect of the capability of capturing remote information, and the high calculation burden caused by the secondary complexity of a self-attention mechanism of a transducer structure in the aspect of image size, are avoided.

Description

Optical remote sensing image segmentation method based on VMamba model

Technical Field

The invention relates to the technical field of optical remote sensing images, and in particular to an optical remote sensing image segmentation method based on a VMamba model.

Background technique

Optical remote sensing images are remote sensing images that use electromagnetic waves such as visible light, infrared light, and ultraviolet light to capture and record the earth's surface. These images can reflect information such as surface features, vegetation coverage, distribution of land objects, and distribution of water bodies. By analyzing and processing these images, people can understand the situation on the earth's surface, monitor environmental changes, conduct resource exploration, and conduct disaster monitoring. Optical remote sensing images are the most widely used form of remote sensing technology, and are also one of the most intuitive means of obtaining information on the earth's surface.

Optical remote sensing image segmentation methods can be mainly divided into two categories: traditional image processing methods and deep learning-based processing methods. Regardless of the method used, its essence is to extract feature representations from remote sensing images to achieve effective segmentation of objects. In high-resolution remote sensing images, the texture structure features are rich, the internal features of the same type of objects are obvious, and the boundaries of different objects are clear. These characteristics provide rich features and topological information, which provide a basis for refined object classification. However, due to the diversity of imaging scenes of optical remote sensing images, the feature representations of the same type of objects under different conditions may vary greatly, such as the performance of vegetation at different latitudes or seasons, while different objects may show similarities, such as grassland and woodland. This phenomenon is described as "same object, different spectrum, same spectrum, different objects", which increases the difficulty of refined segmentation.

At present, there are two deep learning models for optical remote sensing image segmentation, one is based on CNN (Convolutional Neural Network) and the other is based on Transformer structure. Existing studies have shown that the application of CNN models to land use classification and segmentation tasks of high-resolution remote sensing images has effectively improved the accuracy compared with traditional methods, and achieved better classification and segmentation effects. CNN models are now widely used in remote sensing image segmentation tasks. They extract features by sharing convolution kernels, reduce the number of network parameters, and improve model efficiency. Although CNN models have many advantages, due to their limited receptive field, they are not conducive to capturing global features, and the accuracy of ground object classification is still limited and unsatisfactory.

In recent years, the Transformer architecture has gradually become the main deep learning model in the field of computer vision and has been gradually applied to the segmentation of remote sensing images. The attention mechanism is the main component of the Transformer structure, which is superior to the CNN structure in obtaining spatial information and establishing global relationships. The Shifted Windows Transformer (SwinT) proposed by Liu et al. uses Transformer blocks to obtain multi-scale features, has a strong global modeling ability, and has a significantly better classification effect than the traditional CNN model, but its model structure is complex and computationally intensive. Inspired by this, researchers apply Transformer to CNN models to learn from each other's strengths and overcome their weaknesses. For example, Liu et al. proposed a new CNN model: ConvNeXt based on ResNet50 (R50) and borrowed the advantages of SwinT. It achieved the accuracy of ImageNet Top-1 by relying solely on the convolutional structure, and surpassed SwinT in both image classification accuracy and model simplicity. Although the above results have shown the advantages of deep learning in optical remote sensing segmentation, the current optical remote sensing image segmentation models have the following problems:

Both CNN-based models and Transformer-based models have inherent limitations. CNN-based models are limited by their local receptive fields, which greatly hinders their ability to capture long-range information. This often leads to insufficient feature extraction, resulting in suboptimal segmentation results. Although Transformer-based models show excellent performance in global modeling, the self-attention mechanism requires quadratic complexity in terms of image size, resulting in a high computational burden, especially for tasks that require dense predictions. The current shortcomings of these models forced us to develop a new architecture for semantic segmentation of optical remote sensing images that is able to capture strong long-range information and maintain linear computational complexity.

Summary of the invention

To solve the technical problems in the above background, the present invention introduces a visual state space (VSS) block as a basic block to capture extensive contextual information, and constructs an asymmetric encoder-decoder structure to capture powerful long-range information and maintain linear computational complexity.

To achieve the above object, the present invention provides an optical remote sensing image segmentation method based on the VMamba model, the steps comprising:

Acquire an optical remote sensing image of the scene to be segmented;

Based on the VMamba model, the VM-UNet model is constructed;

Using the optical remote sensing image to train the constructed VM-UNet model to obtain a segmentation model;

The segmentation model is used to complete the segmentation of the optical remote sensing image.

Preferably, the VM-UNet model is an asymmetric U-shaped structure network model based on the VMamba model, comprising: a block embedding layer, an encoder, a decoder and a final projection layer; wherein,

The block embedding layer is used to embed the input image to obtain a processed image;

The encoder is used to extract features from the processed image;

The decoder is used to generate a segmentation map based on the features extracted by the encoder;

The final projection layer is used to project the segmentation map into a final segmentation result.

Preferably, the step of embedding the block embedding layer comprises:

Dividing the input optical remote sensing image into non-overlapping blocks of size 4×4;

Performing dimension mapping on the divided optical remote sensing image to obtain a mapping image;

The mapped image is normalized to obtain the processed image.

Preferably, the encoder includes four cascaded VSSLayer layers; wherein the first three VSSLayer layers each include two VSSblock blocks and one PatchMerging2D block; and the fourth VSSLayer layer includes two VSSblock blocks.

Preferably, the decoder comprises four cascaded VSSLayer_up layers; wherein, the first VSSLayer_up layer comprises two VSSblock blocks; and the last three VSSLayer_up layers each comprise two VSSblock blocks and one PatchExpand2D block.

Preferably, the VSSblock block includes a layer normalization layer and a SS2D module; the PatchMerging2D block includes a linear layer and a layer normalization layer; the PatchExpand2D block includes a linear layer and a layer normalization layer; the SS2D module includes: two linear layers, a convolutional layer, a SiLU activation layer and a layer normalization layer.

Preferably, in the encoder, each of the VSSLayer layers corresponds to a stage, and in the first three stages, a block merging operation is used at the end of each stage to reduce the height and width of the input features while increasing the number of channels.

Preferably, in the decoder, each of the VSSLayer_up layers corresponds to a stage, and in the latter three stages, a block expansion operation is used at the beginning of each stage to reduce the number of feature channels and increase the height and width.

Compared with the prior art, the present invention has the following beneficial effects:

The present invention introduces the visual state space (VSS) block as a basic block to capture a wide range of contextual information, and constructs an asymmetric encoder-decoder structure that can capture powerful long-range information and maintain linear computational complexity. This avoids the limitation of the local receptive field of the CNN model, which leads to the lack of ability to capture long-range information, and the high computational burden caused by the self-attention mechanism of the Transformer structure requiring quadratic complexity in terms of image size.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

FIG1 is a schematic diagram of a method flow chart of an embodiment of the present invention.

Detailed ways

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

Embodiment 1

As shown in FIG1 , it is a schematic diagram of the method flow of this embodiment, and the steps include:

S1. Obtain an optical remote sensing image of the scene to be segmented.

An optical remote sensing image of a scene to be segmented and its corresponding real-marked map of ground objects are collected. The scene to be segmented can be ground object extraction in a natural scene or regional planning of a city. In this embodiment, the real-marked map of ground objects is used to provide a reference for the prediction results of the model to calculate the loss function. The model performance can also be evaluated by comparing with the model prediction results.

Afterwards, the optical remote sensing images are divided into training sets and validation sets according to a certain ratio. The specific steps include:

The optical remote sensing image is preprocessed, and the steps include: segmenting the optical remote sensing image into a number of data slices of size 512×512 (1820 in this embodiment), and dividing it into a training set and a verification set according to a ratio of 8:2; wherein the training set includes training images and training labels, and the verification set includes verification images and verification labels, and finally the data set required for VM-UNet network model training and verification is obtained.

S2. Based on the VMamba model, build the VM-UNet model.

The VM-UNet (Vision Mamba UNet) model is an asymmetric U-shaped network model based on the VMamba model. It introduces the visual state space (VSS) block as a basic block to capture a wide range of contextual information and constructs an asymmetric encoder-decoder structure that can capture powerful long-range information and maintain linear computational complexity. The VM-UNet model specifically includes: block embedding layer, encoder, decoder, and final projection layer.

In this embodiment, the block embedding layer of the network model, ie, the input layer, has a feature mapping channel number of 3 for input sample feature data, and generates an output feature of x∈R ^H×W×3 ; wherein H represents the height of the image; and W represents the width of the image.

The block embedding layer is used to embed the input image to obtain the processed image. The specific embedding process includes: dividing the input optical remote sensing image into non-overlapping patches of size 4×4 and mapping the divided optical remote sensing image to obtain a mapped image; finally, normalizing the mapped image to obtain the processed image.

The encoder includes four cascaded VSSLayer layers, which are referred to as the first VSSLayer layer, the second VSSLayer layer, the third VSSLayer layer, and the fourth VSSLayer layer in this embodiment. The first three VSSLayer layers each include two VSSblock blocks and one PatchMergig2D block; the fourth VSSLayer layer includes two VSSblock blocks. In this embodiment, each VSSLayer layer of the encoder corresponds to a stage, and in the first three stages, a block merging operation is used at the end of each stage to reduce the height and width of the input features while increasing the number of channels.

For the convenience of subsequent description, this embodiment makes a detailed distinction, including:

The first VSSLayer layer includes the first VSSblock block, the second VSSblock block and the first PatchMerging2D block; the second VSSLayer layer includes the third VSSblock block, the fourth VSSblock block and the second PatchMerging2D block; the third VSSLayer layer includes the fifth VSSblock block, the sixth VSSblock block and the third PatchMerging2D block; the fourth VSSLayer layer includes the seventh VSSblock block and the eighth VSSblock block.

The decoder includes four cascaded VSSLayer_up layers, which are referred to as the first VSSLayer_up layer, the second VSSLayer_up layer, the third VSSLayer_up layer, and the fourth VSSLayer_up layer in this embodiment. Among them, the first VSSLayer_up layer includes two VSSblock blocks; the last three VSSLayer_up layers each include two VSSblock blocks and a PatchExpand2D block. In this embodiment, each VSSLayer_up layer of the decoder corresponds to a stage, and in the last three stages, a block expansion operation is used at the beginning of each stage to reduce the number of feature channels and increase the height and width.

Similarly, this embodiment also makes a detailed distinction therebetween, specifically including:

The first VSSLayer_up layer includes the ninth VSSblock block and the tenth VSSblock block; the second VSSLayer_up layer includes the first PatchExpand2D block, the eleventh VSSblock block and the twelfth VSSblock block; the third VSSLayer_up layer includes the second PatchExpand2D block, the thirteenth VSSblock block and the fourteenth VSSblock block; the fourth VSSLayer_up layer includes the third PatchExpand2D block and the fifteenth VSSblock block.

The above-mentioned VSSblock block includes a layer normalization layer and a SS2D module connected in sequence; the PatchMerging2D block includes a linear layer and a layer normalization layer connected in sequence; the PatchExpand2D block includes a linear layer and a layer normalization layer connected in sequence; the SS2D module includes: two linear layers, a convolutional layer, a SiLU activation layer and a layer normalization layer, and its connection relationship is linear layer-convolutional layer-SiLU activation layer-layer normalization layer-linear layer.

Finally, the final projection layer includes a linear layer, a layer normalization layer, and a convolutional layer.

S3. Use optical remote sensing images to train the constructed VM-UNet model to obtain the segmentation model.

The VM-UNet network model is trained using the dataset obtained in S1 to obtain a segmentation model.

The specific steps include:

First, the training set batch is input to the block embedding layer; the input image is divided into non-overlapping patches of size 4×4, and the dimension of the image is mapped to C (that is, the number of mapped channels of the image is C), an embedded image is generated, and a feature matrix is generated as the first skip connection layer. Specifically:

The block embedding layer divides the input image x∈R ^H×W×3 into non-overlapping blocks of size 4×4 through a convolutional layer and maps the dimension of the image to C. The embedded image is obtained through this process Next, this embodiment uses the layer normalization layer to normalize x′ and converts the resulting feature matrix Save it as the first skip connection layer and then feed it into the encoder for feature extraction.

After that, the first and second VSSblock blocks extract features and perform downsampling operations through the first PatchMerging2D block to generate a feature matrix for the second jump connection layer; the third and fourth VSSblock blocks extract features and perform downsampling operations through the second PatchMerging2D block to generate a feature matrix for the third jump connection layer; the fifth and sixth VSSblock blocks extract features and perform downsampling operations through the third PatchMerging2D block to generate a feature matrix for the fourth jump connection layer; the seventh and eighth VSSblock blocks extract features as the input of the ninth VSSblock block. Specifically:

The output produced by embedding the block into the layer Input to the first VSSblock block, and generate features through the layer normalization layer in the first VSSblock Input to the SS2D module in the first VSSblock, and then pass through the first linear layer in the SS2D module to generate features Convolutional layers generate features SiLU activation layer, layer normalization layer generates features The second linear layer produces the output And input to the second VSSblock block, after the layer normalization in the second VSSblock to generate features Input to the SS2D module in the second VSSblock, and then pass through the first linear layer in the SS2D module to generate features Convolutional layers generate features SiLU activation layer, layer normalization layer generates features The second linear layer produces the output And input to the first PatchMerging2D block, and then pass through the linear layer and layer normalization layer of the first PatchMerging2D block to generate output And save it as the second jump connection layer; input the output x ₂ generated by the first PatchMerging2D into the third VSSblock block, and generate features through the layer normalization layer in the third VSSblock Input to the SS2D module in the third VSSblock, and then pass through the first linear layer in the SS2D module to generate features Convolutional layers generate features SiLU activation layer, layer normalization layer generates features and the second linear layer produces the output And input to the fourth VSSblock block, the layer normalization layer in the fourth VSSblock block generates features Input to the SS2D module in the fourth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features Convolutional layers generate features SiLU activation layer, layer normalization layer generates features and the second linear layer produces the output And input to the second PatchMerging2D block, and then pass through the linear layer and layer normalization layer of the second PatchMerging2D block to generate output And save it as the third jump connection layer; input the output x ₃ generated by the second PatchMerging2D into the fifth VSSblock block, and generate features through the layer normalization layer in the fifth VSSblock Input to the SS2D module in the fifth VSSblock, and then pass through the first linear layer in the SS2D module to generate characteristics Convolutional layers generate features SiLU activation layer, layer normalization layer generates features and the second linear layer produces the output And input to the sixth VSSblock block, the layer normalization layer in the sixth VSSblock generates features Input to the SS2D module in the sixth VSSblock, and then pass through the first linear layer in the SS2D module to generate features Convolutional layers generate features SiLU activation layer, layer normalization layer generates features and the second linear layer produces the output And input to the third PatchMerging2D block, and then pass through the linear layer and layer normalization layer of the third PatchMerging2D block to generate output And save it as the third jump connection layer; input the output x ₄ generated by the third PatchMerging2D to the seventh VSSblock block, and generate features through the layer normalization layer in the seventh VSSblock block Input to the SS2D module in the seventh VSSblock block, and then pass through the first linear layer in the SS2D module to generate features Convolutional layers generate features SiLU activation layer, layer normalization layer generates features The second linear layer produces the output And input to the eighth VSSblock block, and generate features after layer normalization in the eighth VSSblock block Input to the SS2D module in the eighth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features Convolutional layers generate features SiLU activation layer, layer normalization layer generates features The second linear layer produces the output And serves as the input of the ninth VSSblock block.

Finally, after the tenth VSSblock block, it is fused with the fourth jump connection layer feature as the first PatchExpand2D block input, and the input feature is upsampled through the first PatchExpand2D block; the features are extracted through the eleventh and twelfth VSSblock blocks and fused with the third jump connection layer feature as the second PatchExpand2D block input, and the input feature is upsampled through the second PatchExpand2D block; the features are extracted through the thirteenth and fourteenth VSSblock blocks and fused with the second jump connection layer feature as the third PatchExpand2D block input, and the input feature is upsampled through the third PatchExpand2D block; the features are extracted through the fifteenth VSSblock block and fused with the first jump connection layer feature as the final projection layer input, and the final segmentation result is generated through the final projection layer. Specifically:

The output z"" generated by the eighth VSSblock is input to the ninth VSSblock, and the feature is generated through the layer normalization layer in the ninth VSSblock. Input to the SS2D module in the ninth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features SiLU activation layer, convolutional layer generates features Layer normalization layer produces features The second linear layer produces the output And input to the tenth VSSblock block, the layer normalization layer in the tenth VSSblock block generates features Input to the SS2D module in the tenth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features SiLU activation layer, convolutional layer generates features Layer normalization layer produces features The second linear layer produces the output And perform feature fusion with the fourth jump connection layer, input to the first PatchExpand2D block, and pass through the linear layer and layer normalization layer of the first PatchExpand2D block to generate output The output x ₅ generated by the first PatchExpand2D block is input to the eleventh VSSblock block, and the feature is generated after layer normalization in the eleventh VSSblock block Input to the SS2D module in the eleventh VSSblock block, and then pass through the first linear layer in the SS2D module to generate features SiLU activation layer, convolutional layer generates features Layer normalization layer produces features The second linear layer produces the output And input to the twelfth VSSblock block, the layer normalization layer in the twelfth VSSblock block generates features Input to the SS2D module in the twelfth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features SiLU activation layer, convolutional layer generates features Layer normalization layer produces features The second linear layer produces the output The feature is fused with the third jump connection layer and input to the second PatchExpand2D block, and then the linear layer and layer normalization layer of the second PatchExpand2D block are used to generate the output. The output _x6 generated by the second PatchExpand2D block is input to the thirteenth VSSblock block, and the feature is generated through the layer normalization layer in the thirteenth VSSblock block Input to the SS2D module in the thirteenth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features SiLU activation layer, convolutional layer generates features Layer normalization layer produces features The second linear layer produces the output And input to the fourteenth VSSblock block, the layer normalization layer in the fourteenth VSSblock block generates features Input to the SS2D module in the fourteenth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features SiLU activation layer, convolutional layer generates features Layer normalization layer produces features The second linear layer produces the output The feature is fused with the second jump connection layer and input to the third PatchExpand2D block, and then the linear layer and layer normalization layer of the third PatchExpand2D block are used to generate the output. The output _x7 generated by the third PatchExpand2D block is input to the fifteenth VSSblock block, and the feature is generated through the layer normalization layer in the fifteenth VSSblock block Input to the SS2D module in the fifteenth VSSblock block, and then pass through the first linear layer in the SS2D module to generate features SiLU activation layer, convolutional layer generates features Layer normalization layer produces features The second linear layer produces the output And the features are fused with the first skip connection layer and input to the final projection layer, and the linear layers of the linear projection layer are sequentially passed to generate features Layer normalization layer produces features The convolutional layer produces the final output ^{y∈RH ×W×numclass} , where numclass is the number of classes to be segmented.

S4. Use the segmentation model to complete the segmentation of the optical remote sensing image.

In this embodiment, the block embedding layer of the network model, i.e., the input layer, is set to have 3 feature mapping channels for input sample feature data, 96 feature mapping channels for output feature data (i.e., C is 96), the size of the convolution filter is 4×4, the moving stride is 4, and the generated output feature is x∈R ^128×128×96 . The output feature is input to the above segmentation model, and the parameters of the final projection layer are set as follows:

The fused features are input to the final projection layer, i.e., the output layer, and pass through the linear layer of the linear projection layer in sequence. The number of feature mapping channels of the input sample feature data is 192, the number of feature mapping channels of the output feature data is 6, the size of the convolution filter is 1×1, the moving stride is 1, and the final output is y∈R ^512×512×6 , where 6 represents the number of categories that need to be segmented.

Embodiment 2

In order to verify the segmentation effect of the model of the present invention, a simulation experiment is set up in this embodiment to verify the superiority of the model.

1. Simulation conditions

The hardware platform is: Intel(R) Core(TM) i7-7700@3.20GHZ, 64.0GB RAM;

The software platform is: Pycharm, with TensorFlow as the backend under the PyTorch framework.

2. Simulation method

The method of the present invention is used to obtain optical remote sensing data, wherein the optical remote sensing data includes optical remote sensing image data and its corresponding ground object real marking map; according to the optical remote sensing data, the data is divided into a training set and a validation set according to a certain ratio to train the VM-UNet network model and finally segment the remote sensing optical image;

3. Simulation content and simulation results

The optical remote sensing images selected for the simulation experiment are ten optical remote sensing images with real object marking maps. The optical remote sensing images mainly include 6 types of object coverage, including background, building, farmland, grassland, forest, and water area. The number of each object coverage type is represented by {0, 1, 2, 3, 4, 5}, and the size of the corresponding real object marking map is the same as that of the optical remote sensing image. The values of the pixels at the determined object categories in the real object marking map correspond to black (0, 0, 0) for the background, red (255, 0, 0) for the building, green (0, 255, 0) for the farmland, cyan (0, 255, 255) for the forest, yellow (255, 255, 0) for the grassland, and blue (0, 0, 255) for the water area.

In the simulation experiment, 10 sets of optical remote sensing data were divided into 1820 data slices of size 512×512, and divided into training set and validation set according to the ratio of 8:2. The validation set data was used to train the VM-UN et model to achieve optical remote sensing image segmentation. The segmentation accuracy on the validation set is shown in Table 1:

Table 1

As can be seen from Table 1, in the optical remote sensing segmentation scenario, the VM-UNet provided by the present invention not only maintains linear or near-linear computational complexity in remote sensing optical image segmentation, but also ensures the segmentation accuracy of various categories, especially in water segmentation, and has achieved very good results. It provides a baseline for another segmentation model other than CNN and Transformer in the field of optical remote sensing image segmentation, and provides an effective solution to the existing problems of CNN and Transformer.

The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.

Claims (8)

1. The optical remote sensing image segmentation method based on VMamba model is characterized by comprising the following steps:

Acquiring an optical remote sensing image of a scene to be segmented;

Constructing a VM-UNet model based on VMamba model;

The VM-UNet model built by the optical remote sensing image training is utilized to obtain a segmentation model;

and utilizing the segmentation model to complete the segmentation of the optical remote sensing image.

2. The VMamba model-based optical remote sensing image segmentation method as set forth in claim 1, wherein the VM-UNet model is an asymmetric U-shaped structure network model based on VMamba model, and includes: a block embedding layer, an encoder, a decoder, and a final projection layer; wherein,

The block embedding layer is used for carrying out embedding processing on the input image to obtain a processed image;

the encoder is used for extracting the characteristics of the processed image;

the decoder is used for generating a segmentation map based on the characteristics extracted by the encoder;

the final projection layer is used for projecting the segmentation graph into a final segmentation result.

3. The method for optical remote sensing image segmentation based on VMamba model according to claim 2, wherein the step of embedding the block embedding layer includes:

dividing the input optical remote sensing image into non-overlapping blocks with the size of 4 multiplied by 4;

performing dimension mapping on the divided optical remote sensing image to obtain a mapped image;

and carrying out normalization processing on the mapping image to obtain the processed image.

4. The VMamba model-based optical remote sensing image segmentation method of claim 2, wherein the encoder comprises four cascaded VSSLayer layers; wherein the first three VSSLayer layers each include two VSSblock blocks and one PATCHMERGING2D block; the fourth VSSLayer layer includes two VSSblock blocks.

5. The VMamba model-based optical remote sensing image segmentation method as defined in claim 4, wherein the decoder includes four concatenated VSSLayer _up layers; wherein a first one of said VSSLayer _up layers comprises two VSSblock blocks; the last three VSSLayer up layers each include two VSSblock blocks and one PatchExpand2D block.

6. The VMamba model-based optical remote sensing image segmentation method of claim 5, wherein the VSSblock block comprises a layer normalization layer and an SS2D module; the PATCHMERGING D block includes a linear layer and a layer normalization layer; the PatchExpand D block includes a linear layer and a layer normalization layer; the SS2D module includes: two linear layers, one convolutional layer, one SiLU active layer, and one layer normalization layer.

7. The method of claim 4, wherein in the encoder, each VSSLayer layer corresponds to a stage, and in the first three stages, the end of each stage employs a block merging operation to reduce the height and width of the input features while increasing the number of channels.

8. The method of claim 5, wherein in the decoder, each VSSLayer _up layer corresponds to a stage, and in the last three stages, the start of each stage uses a block expansion operation to reduce the number of feature channels and increase the height and width.