A multi-scale lightweight glacier segmentation method based on global context

Abstract

The application discloses a kind of multi-scale light glacier segmentation method based on global context, it is related to the technical field of deep learning and algorithm improvement, in view of the problems such as complex spatial distribution, serious homogenization and poor real-time in the segmentation of plateau glacier remote sensing image, efficient and accurate segmentation is realized through the generation of adversarial network framework, including sliding cutting and data enhancement to remote sensing image, constructing multi-scale training set, designing layered encoder with Segformer as generator core, combining multi-scale context module and global attention module, reducing parameter quantity through depth separable dilated convolution, fusing local details and global semantic features, using fully convolutional discriminator for pixel-level adversarial learning, restoring image spatial continuity, introducing conditional random field to optimize segmentation boundary, eliminating isolated regions.

Description

Technical Field

[0001] This invention relates to the field of deep learning and algorithm improvement technology, specifically to a multi-scale lightweight glacier segmentation method based on global context. Background Technology

[0002] Glaciers are extremely sensitive to climate change, and glacier changes are closely related to regional ecology, natural disasters, and water resources. The high inter-class similarity and severe homogenization of glacier resources mean that annotating geological element samples requires significant costs and expertise, often resulting in insufficient experimental data accuracy and a limited number of annotated samples. Furthermore, building a highly robust and accurate identification model requires substantial parameters and computational costs. Therefore, a lightweight, multi-scale segmentation algorithm based on generative adversarial networks (GANs) with global context was designed.

[0003] Existing technologies, such as the invention patent application with publication number CN118135239A, disclose a multi-scale, high-resolution remote sensing glacier extraction method using fusion filtering. First, high-resolution glacier remote sensing images are preprocessed to create a deep learning semantic segmentation dataset. Data augmentation techniques are then used to enrich the dataset, ensuring the robustness of model training. To address the insufficient ability to identify scattered and small glaciers, a gated multi-scale filtering layer is designed to provide the model with extraction and feature fusion methods at different scales to capture minute glaciers. The gating mechanism effectively filters out useless feature information. To address the problem of blurred glacier outlines, a parallel dual-channel attention module is designed to encode the rich contextual information of the glacier boundary as local features of the feature map, thereby enhancing its feature representation ability and providing effective assistance for large-area glacier extraction in plateau regions.

[0004] The above solutions have at least the following technical problems: 1. The above solutions lack the introduction of generative adversarial network framework, which makes it difficult for the model to improve its robustness through adversarial learning under limited labeled data conditions. In the existing technology, the traditional U-Net architecture only relies on supervised learning and cannot use unlabeled data to generate pseudo-labels for semi-supervised training. When faced with the common problem of insufficient samples in plateau glacier images, the model's generalization ability will be significantly limited, which in turn leads to more breaks and gaps in traditional methods in complex scenarios.

[0005] 2. The above-mentioned schemes lack a global context multi-scale feature fusion mechanism, which makes it difficult for the model to handle the homogeneity problem between glaciers and background features. Although the existing Glacier-Unet model fuses shallow details and deep semantic features through skip connections, it does not design an object-level context extraction layer and a GAM global attention module similar to the ASP-OLC module, and cannot effectively model long-distance spatial dependencies. Experiments show that this defect will reduce the MIoU index by about 8.7% when the model distinguishes glaciers with similar spectra from clouds or mountain shadows. For example, the MIoU of traditional methods is only 62.41%, while Lmgformer reaches 71.13%.

[0006] 3. The above-mentioned schemes lack lightweight design and computational efficiency optimization, which makes it difficult for the model to meet the real-time processing requirements of high-resolution remote sensing images. The existing methods do not adopt lightweight techniques such as depthwise separable convolution and sequence reduction attention, and the number of parameters is as high as 15.6M, which is 3.3 times that of Lmgformer. This redundant structure not only increases the GPU memory usage, but also reduces the inference speed by more than 40%, making it unsuitable for edge computing devices. In addition, the traditional U-Net does not introduce a conditional random field post-processing module, and its ability to handle isolated noise and edge spikes in the segmentation results is insufficient. Quantitative tests show that its Hausdorff distance is 15.2% higher than the result after CRF optimization, which seriously affects the mapping accuracy. All of these factors limit the application value of the method in scenarios with high timeliness requirements, such as dynamic monitoring of glaciers. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-scale lightweight glacier segmentation method based on global context, which solves the problems existing in the background technology.

[0008] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a multi-scale lightweight glacier segmentation method based on global context, including: S1, segmenting high-resolution remote sensing images by sliding cutting, and performing random flipping, noise addition, blurring and distortion enhancement to construct a training set containing original data and enhanced data.

[0009] S2. Segformer is used as the backbone network, integrating a multi-scale context module and a global attention module. Local to global features are extracted through a hierarchical encoder, and depthwise separable dilated convolution is used to reduce the number of parameters.

[0010] S3. Construct a discriminator network: Design a fully convolutional network as a discriminator to distinguish between the pseudo-labels and real labels output by the generator through a pixel-level confidence map.

[0011] S4. Model Training and Optimization: A frozen training strategy is adopted to accelerate convergence, and the generator and discriminator are jointly optimized using the Softmax loss function and adversarial learning mechanism.

[0012] S5. By modeling the spatial context of pixels using a fully connected conditional random field, holes and isolated regions in the segmentation results are eliminated.

[0013] S6. Evaluate the segmentation performance of the Lmgformer model and the baseline model on the self-built dataset.

[0014] The beneficial effects of the present invention are as follows: 1. The present invention provides a multi-scale lightweight glacier segmentation method based on global context. In the data preprocessing and enhancement process, a sliding window cutting technique is used to segment large-size remote sensing images into standard sub-images of 512×512. Combined with a variety of data augmentation strategies, this method is conducive to significantly improving the diversity and richness of training data. It not only effectively alleviates the problem of limited sample size in plateau glacier datasets, but also enables the model to learn the glacier characteristics under different weather conditions, illumination changes and geometric deformations, greatly enhancing the model's generalization ability in real complex environments.

[0015] 2. In the process of multi-scale feature extraction and fusion, the embodiments of the present invention construct a hierarchical encoder structure with Segformer as the core, and combine it with the multi-input-output ASP-OLC module with object-level context information, which is conducive to realizing multi-level feature capture from local details to global semantics. Among them, the use of depthwise separable dilated convolution reduces the number of parameters of traditional dilated convolution by about 65%, and significantly improves the computational efficiency while ensuring that the receptive field remains unchanged.

[0016] 3. In the global feature optimization process, the present invention introduces the GAM module, which is conducive to realizing cross-dimensional feature interaction of channels and spatial dimensions. The module adopts a 3D arrangement to retain cross-dimensional spatial information, strengthens the inter-channel dependency relationship through a two-layer MLP, and removes the max pooling operation in the traditional spatial attention to avoid information loss. This significantly improves the model's ability to distinguish glaciers from the background, such as clouds, mountain shadows, and glacial lakes, and effectively solves the problem of missegmentation caused by severe homogenization.

[0017] 4. In the adversarial training optimization process, this invention uses a fully convolutional discriminator instead of a traditional classification discriminator, which is beneficial for achieving pixel-level true / false discrimination and feature learning. The FCD contains 5 convolutional layers and uses the LeakyReLU activation function. Finally, it outputs a confidence map through Sigmoid, which is beneficial for learning the spatial distribution pattern of the label data in detail and gradually repairing the spatial discontinuity caused by factor map partitioning. Experiments show that this design improves the generator's segmentation results by about 15.3% in terms of edge smoothness and region consistency, especially when dealing with glacier regions segmented by ridges.

[0018] 5. In the post-processing and result optimization process of this invention, the introduction of a fully connected conditional random field module is beneficial for spatial context modeling and optimization of the preliminary segmentation results. By considering the color similarity and spatial proximity between pixels, CRF classifies adjacent similar pixels into the same category, effectively eliminating about 62% of isolated noise points and segmentation holes, making the final segmentation results more visually coherent and natural. Quantitative evaluation shows that the post-processing step further improves the average intersection-union ratio by 3.2 percentage points, reaching 71.13%, while the kappa coefficient increases to 72.43%, which is significantly better than traditional post-processing methods. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the implementation steps of the present invention.

[0021] Figure 2 This is a flowchart of a fusion-filtered multi-scale high-resolution remote sensing glacier extraction process, as shown in one embodiment.

[0022] Figure 3 This is a structural diagram of a multi-scale, high-resolution remote sensing glacier extraction model based on an embodiment of fusion filtering.

[0023] Figure 4 This is a schematic diagram of a parallel dual-channel attention module in one embodiment.

[0024] Figure 5 This is a schematic diagram of a position attention module in one embodiment.

[0025] Figure 6 This is a schematic diagram of a channel attention module in one embodiment. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1As shown, the present invention provides a multi-scale lightweight glacier segmentation method based on global context. The method includes: S1, segmenting high-resolution remote sensing images by sliding cut, and performing random flipping, noise addition, blurring and distortion enhancement to construct a training set containing original data and enhanced data.

[0028] In a specific embodiment, the construction of a training set containing original data and augmented data is carried out as follows: when the size of the glacier remote sensing image is larger than the preset size, it cannot be directly input into the network framework for training. In this case, the glacier remote sensing image is cropped to a specified size, and the remote sensing image composed of two band combinations of band 1, band 2, band 3 and band 4, band 5, and band 6 is cropped and enlarged.

[0029] Using a sliding cut method with a repetition rate of 0.1, the original large-size glacier remote sensing TIFF image was cut into 512*512 TIFF image maps for dataset construction. The original images were randomly flipped horizontally and vertically, and random noise, random blur, and random distortion were added. The brightness, contrast, and saturation were all set to 0.5. The final training images not only included the original slice images but also the data-enhanced images of the original slice images. The same operation was performed on the label images, which were then paired with the processed original images to form data pairs. When there were original images that did not contain glacier targets during the slicing process, the original images that did not contain glacier targets were removed to simplify the dataset.

[0030] It should be noted that this invention achieves an accuracy of 90.3% on the self-built Anyemaqen Glacier dataset, with an average crossover ratio improved to 71.13% and a parameter reduction of 58%, significantly outperforming traditional models. It is suitable for dynamic glacier monitoring and environmental protection using high-resolution remote sensing imagery.

[0031] In the process of multi-scale feature extraction and fusion, this invention constructs a hierarchical encoder structure with Segformer as the core, and combines it with a multi-input / output ASP-OLC module with object-level context information. This facilitates the capture of multi-level features from local details to global semantics. The use of depthwise separable dilated convolution reduces the number of parameters of traditional dilated convolution by about 65%, and significantly improves computational efficiency while keeping the receptive field unchanged.

[0032] S2. Segformer is used as the backbone network, integrating a multi-scale context module and a global attention module. Local to global features are extracted through a hierarchical encoder, and depthwise separable dilated convolution is used to reduce the number of parameters.

[0033] It's important to note that, unlike the ViT model which only outputs a single-resolution feature map, the hierarchical encoder generates multi-scale feature representations similar to convolutional neural networks. These representations not only contain high-resolution global information from glacier data but also low-resolution local details, thus improving accuracy for glacier identification tasks in complex scenes. In short, given an input image with a resolution of H×W×3, the merged image yields a resolution of... Hierarchical feature map F i Where i∈{1,2,3,4}, C i+1 Compared to C i big.

[0034] In a specific embodiment, the reduction of parameter count using depthwise separable dilated convolution is achieved as follows: A point convolution with a kernel size of 1×1, three depthwise separable dilated convolutions with dilation rates of [6, 12, 18], and a context information fusion layer are used. By calculating the similarity between the current pixel p and pixel i, a pixel-level similarity probability map W is obtained. pi According to the self-attention formula, W p Defined as follows: Where X p and X i These are vector tables for pixels p and i, respectively. fp is the lookup transformation, f k It is a key-value transformation.

[0035] Z p It is the sum of similarities between pixel p and the remaining pixels. Furthermore, it is determined by weighting the distance W between pixel p and the remaining pixels. pi The summation yields the target context feature representation C. p , in As a feature map transformation function, a 1×1 convolution is used instead.

[0036] It should be noted that the role of the information fusion layer described below is to use the information of other pixels belonging to the same semantic object to represent the current pixel and obtain its object context information representation.

[0037] In the global feature optimization process, this invention introduces a GAM module, which facilitates cross-dimensional feature interaction between channels and spatial dimensions. This module uses a 3D arrangement to preserve cross-dimensional spatial information, strengthens the inter-channel dependencies through a two-layer MLP, and removes the max pooling operation in traditional spatial attention to avoid information loss. This significantly improves the model's ability to distinguish glaciers from the background, such as clouds, mountain shadows, and glacial lakes, effectively solving the problem of missegmentation caused by severe homogenization.

[0038] S3. Construct a discriminator network: Design a fully convolutional network as a discriminator to distinguish between the pseudo-labels and real labels output by the generator through a pixel-level confidence map.

[0039] In a specific embodiment, the process of distinguishing between pseudo-labels and real labels output by the generator using a pixel-level confidence map is as follows: Lmgformer improves the discriminator to a fully convolutional network architecture, analyzes the contextual information within the receptive field of each pixel through a pixel-level discrimination mechanism, and completes the input of FCD through the probability map predicted by the generator network and single encoding during adversarial learning. The output of FCD is a confidence map C of size . For unlabeled data, FCD selects pixels with high confidence for semi-supervised learning by setting a threshold. The expression for the confidence map C is as follows:

[0040] The input data is high-resolution remote sensing imagery. sigmoid() is the sigmoid function, where n represents the sample index of the input data, and n takes the value of a positive integer. h represents the height of the feature image and w represents the width of the feature image.

[0041] It should be noted that in the network structure, FCD contains five convolutional layers, each with four convolutional kernels. The kernel size is 2, and the number of kernels is [64, 128, 256, 512, 1]. Each convolutional layer is followed by a LeakyReLU activation function with a coefficient of 0.2.

[0042] In the adversarial training optimization process, this invention employs a fully convolutional discriminator instead of a traditional classification discriminator, which facilitates pixel-level real / false discrimination and feature learning. The FCD contains five convolutional layers, uses the LeakyReLU activation function, and finally outputs a confidence map through the Sigmoid function. This is beneficial for finely learning the spatial distribution patterns of the label data and gradually repairing the spatial discontinuities caused by factor map partitioning. Experiments show that this design improves the generator's segmentation results by approximately 15.3% in terms of edge smoothness and region consistency, especially when dealing with glacier regions segmented by ridges.

[0043] S4. Model Training and Optimization: A frozen training strategy is adopted to accelerate convergence, and the generator and discriminator are jointly optimized using the Softmax loss function and adversarial learning mechanism.

[0044] In a specific embodiment, the process of jointly optimizing the generator and discriminator using the Softmax loss function and adversarial learning mechanism is as follows: The Softmax function is used to determine the class probability of each pixel in the feature, and the Loss value is calculated using the Softmax function. The Loss function, L(θ), is defined as follows: Among them, f u θ represents the category label and has Q possible values. Since the research object is three-class semantic segmentation, the value of Q is 3. θ represents the model parameters. It is θ q transpose of g u The input image pixel observation vector G = {g1, g2, g3, ..., g...} represents the pixel observation vector of the input image. U In the set {}, u represents the number of pixels in each image, and {g} is an explicit function.

[0045] In the post-processing and result optimization process of this invention, the introduction of a fully connected conditional random field module facilitates spatial context modeling and optimization of the initial segmentation results. By considering the color similarity and spatial proximity between pixels, CRF classifies adjacent similar pixels into the same category, effectively eliminating approximately 62% of isolated noise points and segmentation holes, making the final segmentation results more visually coherent and natural. Quantitative evaluation shows that the post-processing step further improves the average intersection-union ratio by 3.2 percentage points, reaching 71.13%, while the kappa coefficient increases to 72.43%, which is significantly better than traditional post-processing methods.

[0046] S5. By modeling the spatial context of pixels using a fully connected conditional random field, holes and isolated regions in the segmentation results are eliminated.

[0047] In a specific embodiment, the process of eliminating holes and isolated regions in the segmentation result is as follows: Based on the generation process of the pseudo data feature map, Segformer is introduced as a generator. Multi-scale feature information is obtained through a multi-input / output lightweight pyramid module with object-level context information. At the same time, depthwise separable dilated convolution reduces the high parameter and computational cost of the original dilated convolution. The global attention module further extracts and fuses channel and spatial interaction correlation feature information in the global scope. A fully convolutional network FCD is introduced as a discriminator to replace the traditional classification network discriminator. The sample features are deeply learned in the process of distinguishing real data and pseudo data based on ground real label samples. The spatial context information of domain pixels is modeled using a fully connected CRF module to classify adjacent pixels into the same category, thereby eliminating holes and isolated regions in the segmentation result.

[0048] It should be noted that, specifically, the experimental parameters for training the improved Lmgformer model include: using a frozen training method to accelerate model training, setting the training epochs to 500, and the batch size to 8. The model uses the Adam optimizer to optimize the learning rate, with a maximum learning rate of 10⁻³ and a minimum learning rate of 10⁻³*0.01. The model uses a weight decay strategy to prevent overfitting, with a weight decay value set to 5*10⁻⁴.

[0049] S6. Evaluate the segmentation performance of the Lmgformer model and the baseline model on the self-built dataset.

[0050] In a specific embodiment, the evaluation of the segmentation performance of the Lmgformer model and the benchmark model on the self-built dataset is carried out as follows: The model accuracy is evaluated using kappa coefficients, pixel accuracy, and mean intersection-over-union (MIoU) ratio. When comparing the extraction results, the evaluation categories of pixel accuracy and MIoU are considered, and the pixel accuracy (PA) and MIoU ratio are directly calculated from the confusion matrix. The calculation process for pixel accuracy (PA) and MIoU ratio is as follows:

[0051] Where j is the number corresponding to each category in the dataset, j = 1, 2, ..., k, k is the total number of categories, and k takes the value of a positive integer. TP means that the target in the test sample can be correctly predicted, TN means that the non-target in the test sample is accurately predicted as a non-target, FP means that the non-target in the test sample is accurately predicted as a target, and FN means that the target in the test sample is accurately predicted as a non-target.

[0052] It should be noted that the improved Lmgformer model remote sensing glacier segmentation method described above acquires high-resolution remote sensing data through the Landsat-9 satellite. The sampled high-resolution remote sensing images are preprocessed, and a glacier dataset and corresponding label data are manually created. The dataset is divided into training, validation, and test sets proportionally. The training and validation sets are input into a supervised network for training and validation. When the epoch is around 300, the loss function reaches a low value and remains relatively stable. All metrics on the validation set achieve their highest values. The optimal model is saved, and training is stopped, resulting in a remote sensing glacier recognition model. Test images from the test set are then input into the model for remote sensing glacier recognition, improving the accuracy of the glacier recognition model.

[0053] Please see Figure 2 As shown, Figure 2The data processing workflow and Lmgformer model framework are defined, including a generator network, a discriminator network, and a post-processing section. The generator network comprises a backbone feature extraction network (Segformer), a multi-scale context module (ASP-OLC), and a global attention module. The discriminator network primarily uses FCD, and post-processing employs a Conditional Random Field (CRF). Optimization is performed. In the basic generator network, both local and global contextual features of high-resolution remote sensing imagery are considered. First, a lightweight Segformer network is used to extract useful spectral and texture structure information. The ASP-OLC module further extracts multi-scale features with object-level contextual information. Then, the global feature fusion module (GAM) performs global cross-dimensional information interaction fusion in the channel and spatial dimensions. Finally, the predicted feature probability map G(X) is output. n ) (h,w,c) Its expression is as follows:

[0054] G(X n ) (h,w,c) =GAM(ASP-OLC(Segformer(X) n ))), c represents the number of feature image channels.

[0055] Please see Figure 3 As shown, high-resolution remote sensing glacier images contain complex ground features, including clouds, glacial lakes, vegetation, and mountain shadows. The experiment only extracts glacier information, while other features are classified as background. The generation process of pseudo-data feature maps is considered first, so Segformer is introduced as the generator.

[0056] Please see Figure 4 , Figure 5 and Figure 6 As shown, Figure 4 This is a schematic diagram of the lightweight backbone network Segformer. Figure 5 and Figure 6 The ASP-OLC module consists of channel and spatial attention submodules. Global structure and spatial feature information distribution are crucial for target extraction from high-resolution remote sensing images. With the fusion of multi-scale and contextual information in the ASP-OLC module, the amount of invalid features and redundant interference information increases, while the proportion of interaction information between channel and spatial features decreases or is even ignored. Therefore, a global attention mechanism is used for global cross-dimensional feature interaction and fusion of channel and spatial dimension feature information. The input to GAM is a multi-scale feature map, denoted as F1. The formulas for the feature map F2 passing through the channel attention mechanism and the final feature map F3 passing through the spatial attention mechanism are as follows: Where M c and M s These represent the channel attention map and the spatial attention map, respectively. The channel attention submodule represents element-wise multiplication operations. It adopts a 3D arrangement to fully preserve important information across three-dimensional space. It uses a two-layer MLP to effectively strengthen the spatial dependencies between cross-dimensional channels. The MLP and BAM modules of the encoding and decoding structure share the same reduction ratio r to achieve feature compression and reconstruction. The spatial attention submodule uses two convolutional layers to fuse spatial information.

[0057] This invention provides a multi-scale lightweight glacier segmentation method based on global context. During data preprocessing and augmentation, a sliding window segmentation technique is used to segment large-size remote sensing images into standard 512×512 sub-images. Combined with various data augmentation strategies, this method significantly improves the diversity and richness of training data. It not only effectively alleviates the problem of limited sample size in plateau glacier datasets, but also enables the model to learn the glacier characteristics under different weather conditions, illumination changes, and geometric deformations, greatly enhancing the model's generalization ability in real complex environments.

[0058] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. A lightweight, multi-scale glacier segmentation method based on global context, characterized in that, include: S1. The high-resolution remote sensing image is segmented by sliding cut, and then randomly flipped, noise added, blurred and distorted to enhance it, and a training set containing the original data and the enhanced data is constructed. S2. Segformer is used as the backbone network, integrating a multi-scale context module and a global attention module. Local to global features are extracted through a hierarchical encoder, and depthwise separable dilated convolution is used to reduce the number of parameters. S3. Construct a discriminator network: Design a fully convolutional network as a discriminator to distinguish between the pseudo-labels and real labels output by the generator through a pixel-level confidence map; S4. Model Training and Optimization: A frozen training strategy is adopted to accelerate convergence, and the generator and discriminator are jointly optimized using the Softmax loss function and adversarial learning mechanism. S5. Model the pixel spatial context relationship through a fully connected conditional random field to eliminate holes and isolated regions in the segmentation results; S6. Evaluate the segmentation performance of the Lmgformer model and the baseline model on the self-built dataset.

2. The multi-scale lightweight glacier segmentation method based on global context according to claim 1, characterized in that, The specific process for constructing a training set containing both original and augmented data is as follows: When the size of the glacier remote sensing image is larger than the preset size, it cannot be directly input into the network framework for training. In this case, the glacier remote sensing image is cropped to the specified size, and the remote sensing image after the combination of two bands, band 1, band 2, band 3 and band 4, band 5, and band 6, is cropped and enlarged. Using a sliding cut method with a repetition rate of 0.1, the original large-size glacier remote sensing TIFF image was cut into 512*512 TIFF image maps for dataset construction. The original images were randomly flipped horizontally and vertically, and random noise, random blur, and random distortion were added. The brightness, contrast, and saturation were all set to 0.

5. The final training images not only included the original slice images but also the data-enhanced images of the original slice images. The same operation was performed on the label images, which were then paired with the processed original images to form data pairs. When there were original images that did not contain glacier targets during the slicing process, the original images that did not contain glacier targets were removed to simplify the dataset.

3. The multi-scale lightweight glacier segmentation method based on global context according to claim 2, characterized in that, The method of reducing the number of parameters using depthwise separable dilated convolution is as follows: The algorithm consists of a 1×1 point convolution, three depthwise separable dilated convolutions with dilation rates of [6, 12, 18], and a context information fusion layer. By calculating the similarity between the current pixel p and pixel i, a pixel-level similarity probability map W is obtained. pi According to the self-attention formula, W p Defined as follows: Where X p and X i These are vector tables for pixels p and i, respectively. fp is the lookup transformation, f k It is a key-value transformation. Z p It is the sum of similarities between pixel p and the remaining pixels. Furthermore, it is determined by weighting the distance W between pixel p and the remaining pixels. pi The summation yields the target context feature representation C. p , Here, θ is used as a feature map transformation function, which is replaced by 1×1 convolution.

4. The multi-scale lightweight glacier segmentation method based on global context according to claim 3, characterized in that, The process of distinguishing between the pseudo-labels and real labels output by the generator using a pixel-level confidence map is as follows: Lmgformer improves the discriminator to a fully convolutional network architecture, analyzing the contextual information within the receptive field of each pixel through a pixel-level discrimination mechanism. During adversarial learning, the input to the FCD is completed by generating a probability map predicted by the network and a single encoding. The output of the FCD is a confidence map C of size . For unlabeled data, the FCD selects pixels with high confidence for semi-supervised learning by setting a threshold. The expression for the confidence map C is as follows: X unlabn The input data is represented as high-resolution remote sensing images. sigmoid() is the sigmoid function, where n represents the sample index of the input data, and n takes the value of a positive integer. h represents the height of the feature image and w represents the width of the feature image.

5. A multi-scale lightweight glacier segmentation method based on global context according to claim 4, characterized in that, The specific process of jointly optimizing the generator and discriminator using the Softmax loss function and adversarial learning mechanism is as follows: The Softmax function is used to determine the class probability of each pixel in the feature, and the Loss value is calculated using the Softmax function. The Loss function is defined by the formula L(θ): Among them, f u θ represents the category label and has Q possible values. Since the research object is three-class semantic segmentation, the value of Q is 3. θ represents the model parameters. It is θ q transpose of g u The input image pixel observation vector G = {g1, g2, g3, ..., g...} represents the pixel observation vector of the input image. U In the set {}, u represents the number of pixels in each image, and {g} is an explicit function.

6. A multi-scale lightweight glacier segmentation method based on global context according to claim 5, characterized in that, The specific process for eliminating holes and isolated regions in the segmentation results is as follows: Based on the generation process of pseudo-data feature maps, Segformer is introduced as a generator. Multi-scale feature information is obtained through a multi-input / output lightweight pyramid module with object-level context information. At the same time, depthwise separable dilated convolution reduces the high number of parameters and computational cost of the original dilated convolution. The global attention module further extracts and fuses channel and spatial interaction correlation feature information in the global scope. Fully convolutional network FCD is introduced as a discriminator to replace the traditional classification network discriminator. In the process of distinguishing real data from pseudo data based on ground real label samples, the sample features are deeply learned. The fully connected CRF module is used to model the spatial context information of domain pixels, and adjacent pixels are grouped into the same category, thereby eliminating holes and isolated regions in the segmentation results.

7. A lightweight glacier segmentation method based on global context according to claim 6, characterized in that, The specific process for evaluating the segmentation performance of the Lmgformer model and the baseline model on the self-built dataset is as follows: The accuracy of the model is evaluated using the kappa coefficient, pixel accuracy, and mean intersection-over-union (MIoU) index. When comparing the extraction results, the evaluation categories of pixel accuracy and MIoU index are considered, and the pixel accuracy (PA) and MIoU index are directly calculated from the confusion matrix. The calculation process of pixel accuracy (PA) and MIoU index is as follows: Where j is the number corresponding to each category in the dataset, j = 1, 2, ..., k, k is the total number of categories, and k takes the value of a positive integer. TP means that the target in the test sample can be correctly predicted, TN means that the non-target in the test sample is accurately predicted as a non-target, FP means that the non-target in the test sample is accurately predicted as a target, and FN means that the target in the test sample is accurately predicted as a non-target.

Images