A semi-supervised remote sensing image semantic segmentation method and device based on multi-stage pseudo label optimization

Abstract

This invention discloses a semi-supervised remote sensing image semantic segmentation method and device based on multi-stage pseudo-label optimization. The method includes: dataset processing and initial pseudo-label generation; image-level pseudo-label stability evaluation; and multi-stage progressive optimization of the pseudo-label training model. The first stage uses labeled data and high-quality pseudo-label data to train the model and optimizes medium-quality pseudo-labels. The second stage introduces high-quality pseudo-labels and optimized medium-quality pseudo-labels to train the model and optimizes low-quality pseudo-labels. The third stage integrates all labeled data and all optimized pseudo-label data to train the final semantic segmentation model. This invention can fully exploit the value of unlabeled data and effectively suppress the interference of pseudo-label noise with a small amount of labeled data, significantly improving the accuracy and robustness of remote sensing image semantic segmentation.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a semi-supervised remote sensing image semantic segmentation method and device based on multi-stage pseudo-label optimization, for semantic segmentation tasks in scenarios with limited labeled data. Background Technology

[0002] With the rapid development of remote sensing technology, remote sensing images are increasingly widely used in urban planning, environmental monitoring, agricultural management, and other fields. In these applications, remote sensing image semantic segmentation technology plays a crucial role, as it can assign each pixel in the image to a corresponding land cover category, thereby enabling accurate identification and management of land cover in the image.

[0003] In recent years, pseudo-label-based self-training methods have effectively expanded the scale of training data by generating pseudo-labels using the model's predictions of unlabeled data. However, in current semi-supervised learning scenarios, a confidence threshold is usually set to filter high-confidence pixel-level pseudo-labels. This strategy has significant limitations: its pixel-level filtering mechanism relies solely on local confidence decisions, disrupting the global semantic consistency that the segmentation task depends on, leading to fragmented segmentation results. Furthermore, setting the confidence threshold too high results in low label utilization, while setting it too low introduces numerous noisy labels that interfere with training. In addition, the model's predictive ability evolves dynamically with the training phase; early low-confidence samples may become more reliable later, but a fixed threshold cannot capture this dynamic characteristic, resulting in the loss of effective information. Ultimately, this leads to limited pseudo-label quality, reliance on manual parameter tuning, and insufficient segmentation robustness. Summary of the Invention

[0004] This invention provides a semi-supervised remote sensing image semantic segmentation method and device based on multi-stage pseudo-label optimization. By jointly training labeled data and hierarchical pseudo-labels in stages, optimizing pseudo-labels and iteratively training the model in multiple stages, error accumulation is mitigated and the robustness of the remote sensing image semantic segmentation model is improved.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions.

[0006] A semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization includes:

[0007] Obtain the publicly available ISPRS Potsdam remote sensing image dataset, and select test images from the dataset as the test set for the model. A subset of test images from the test set is randomly selected as the labeled training dataset. The remaining test images in the test set, after having their labels removed, were used as the unlabeled training dataset. .

[0008] test set The input is fed into a pre-trained remote sensing image semantic segmentation network model to obtain the semantic segmentation output.

[0009] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization includes the following steps in constructing and training the remote sensing image semantic segmentation network model:

[0010] Step 1: Constructing a Teacher Network Model Student Network Model , where k=1,2,3.

[0011] Step 2: Prepare the labeled dataset Input Teacher Network Model The network is trained by using the cross-entropy loss function to quantify the difference between the predicted segmentation map and the real annotation. The network parameters are optimized by the backpropagation algorithm, and the optimal parameters are saved as the trained teacher network model.

[0012] Step 3: Transfer the unlabeled dataset Input the trained teacher network model The initial pseudo-label dataset is obtained; the initial pseudo-label dataset is divided into high-quality pseudo-labels. Medium-sized pseudo-labels and inferior counterfeit labels Level 3 pseudo-labels.

[0013] Step 4: Perform multi-stage progressive optimization training using the three-level pseudo-labels, and provide optimized labeled data to train the student network model. The trained student network model is obtained. This ultimately forms a well-trained remote sensing image semantic segmentation network model.

[0014] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization specifically divides the labeled training data samples and unlabeled training data samples as follows:

[0015] Obtain the ISPRS Potsdam remote sensing image dataset D, and divide D into training sets in an 8:2 ratio. and test set The formula is as follows:

[0016]

[0017] in, The representative model's first One input image, The model is for the first The target value predicted for each image; These represent the sizes of the training set and the test set, respectively. ; Representative dataset The total number of samples.

[0018] training set Select 10% of the data sample As a labeled dataset, the remaining training set As an unlabeled training set, it satisfies:

[0019]

[0020] in, .

[0021] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization, the teacher network model Student Network Model Both adopt an encoder-decoder architecture;

[0022] The encoder uses ResNet-50 as the backbone network to extract deep semantic features from the input image.

[0023] The decoder includes a multi-scale feature enhancement module and a multi-scale feature fusion module. The multi-scale feature enhancement module enhances the deep features extracted by the encoder at multiple levels for ground objects of different scales in remote sensing images. The multi-scale feature fusion module achieves effective fusion of multi-scale semantic features through a cross-level feature interaction mechanism, which is used to improve the model's segmentation accuracy for complex ground object boundaries.

[0024] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization, the teacher network model Student Network Model The specific structure is as follows:

[0025] The encoder uses a ResNet-50 backbone network and includes five stages of downsampling operations on the input image. The encoder outputs a multi-scale feature map. , ;in, , , They represent the first The height, width, and number of channels of the feature map output by each feature extraction layer.

[0026] The multi-scale feature enhancement module performs multi-level feature enhancement on the deep features output by the encoder through a parallel branching structure. Each branch uses convolution operations with different receptive fields to output enhanced multi-scale features:

[0027]

[0028] in, This indicates a concatenation operation, used to merge multiple input tensors along the channel dimension, thereby integrating feature information from different branches or sources; This represents a convolutional layer, used to extract features from input data; This indicates dilated convolution, which obtains contextual information at different scales by adjusting the dilation rate, thereby enhancing the model's ability to distinguish ground features at different scales. This represents a convolution kernel with a height and width of 1. This represents a convolution kernel with a height and width of 3.

[0029] The multi-scale feature fusion module uses a cross-level feature interaction mechanism to combine the shallow features output by the encoder with the enhanced multi-scale features. Effective integration can improve the segmentation accuracy of complex terrain boundaries;

[0030] in This indicates an upsampling operation that matches the spatial dimensions of the feature map, improving the segmentation accuracy of complex feature boundaries through cross-level feature interaction.

[0031] The prediction module gradually restores the spatial resolution of the feature map through upsampling operations. First, it merges the features... Reduce the number of channels using convolutional layers:

[0032]

[0033] in This represents the predicted feature map, which is the output feature after processing by the prediction module.

[0034] Upsampling via bilinear interpolation Gradually restore to the original input image size:

[0035]

[0036] Output segmentation results ,pass The function calculates the probability that each pixel belongs to each category.

[0037] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization, wherein the teacher model The training process includes:

[0038] Step 1: Initialize the teacher network model parameter.

[0039] Step 2: Select remote sensing images from the labeled test set. .

[0040] Step 3: The dataset will have labels. Input Teacher Network Model The difference between the predicted segmentation map and the ground truth annotation is calculated using the cross-entropy loss function:

[0041]

[0042] Step 4: Optimize network parameters using the backpropagation algorithm Minimize the loss function .

[0043] Step 5: Save the network parameters that perform best on the validation set during training as the final teacher network model. .

[0044] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization divides the initial pseudo-label dataset into high-quality pseudo-labels. Medium-sized pseudo-labels Labeled as inferior Level 3 pseudo-tags; specifically including:

[0045] Step 1: Copy the teacher model The parameters are used to initialize the student network parameters.

[0046] Step 2: Select a set of remote sensing images from the unlabeled training set. .

[0047] Step 3: Save the final data during the teacher network model training process. Model checkpoints per training cycle .

[0048] Step 4: Calculate each unlabeled image Stability rating:

[0049]

[0050] in This represents the stability score of the image; the higher the score, the more stable and reliable the model's predictions are. This represents the total number of model checkpoints saved during training. and Representing the first The and the first One model checkpoint; The average crossover ratio (CROR) is used to measure the similarity between predictions made by two different models for the same image.

[0051] Step 5: Set the grading threshold for the stability score calculated from the root stability:

[0052]

[0053] in The quantile parameters are derived based on statistical distribution characteristics and satisfy... ; Represents the stability score set of quantiles; Represents the stability score set of Quantiles.

[0054] Step 6: Divide the unlabeled dataset into three quality levels based on the corresponding grading threshold:

[0055]

[0056]

[0057] .

[0058] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization includes the following multi-stage progressive optimization training:

[0059] The first phase involves training student network models. Using labeled datasets and high-quality pseudo-labels As training data, a composite loss function is used for supervised training:

[0060]

[0061] in For cross-entropy loss, These are the weight coefficients for the first stage; after training, the trained student network model is used. The original unlabeled image corresponding to the medium pseudo-label set Re-predict and generate an optimized set of intermediate pseudo-labels. .

[0062] The second stage involves training student network models. Based on the first stage, an optimized medium-level pseudo-label is introduced. The data used includes , and Supervised training is performed using an improved composite loss function:

[0063]

[0064] in, These are the weighting coefficients for the second stage.

[0065] After training, the trained student network model is used Inferior and counterfeit labels The corresponding original unlabeled image is re-predicted to generate optimized, low-quality pseudo-labels. ;

[0066] The third stage involves training student network models. Based on the second stage, an optimized version of the inferior pseudo-label was introduced. The data used includes , , and At this point, all pseudo-labels are incorporated into the learning process, and the complete conformance loss function is used for final training.

[0067]

[0068] in The weight coefficients are for the third stage; after three rounds of pseudo-label optimization and iterative model training, a student network model that has fully learned the labeled data is finally obtained. .

[0069] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization also includes verifying the performance of the remote sensing image semantic segmentation prediction model by using the difference between the label data and the output of the model prediction.

[0070] A computer device, comprising:

[0071] Memory is used to store computer instructions.

[0072] A processor is configured to execute the computer instructions to implement the semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization.

[0073] The beneficial effects of this invention are as follows: This invention provides a semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label progressive optimization. By constructing an image-level pseudo-label stability evaluation mechanism, it completely abandons the traditional pixel-level screening strategy based on confidence thresholds, effectively avoiding the destruction of the semantic continuity of ground features. A multi-stage progressive optimization training strategy is adopted, following a learning paradigm from easy to difficult, allowing the model to first establish stable feature representations on high-quality pseudo-label data, and then gradually learn and optimize medium and low-quality pseudo-label data, significantly suppressing the accumulation of pseudo-label noise. By designing a progressively enhanced data perturbation strategy, applying differentiated data augmentation intensity at different training stages, it ensures the stability of model training and improves the model's generalization ability. This invention fully explores the value of unlabeled data while effectively improving the model's segmentation accuracy for complex ground feature boundaries, significantly enhancing the performance and robustness of the semi-supervised semantic segmentation model. Attached Figure Description

[0074] Figure 1 A schematic diagram of the process of a semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label hierarchical optimization provided by the present invention;

[0075] Figure 2 This is a schematic diagram illustrating the training process of the remote sensing image semantic segmentation network model constructed in this invention.

[0076] Figure 3 This is a schematic diagram of the multi-stage pseudo-label hierarchical optimization strategy adopted in this invention. Detailed Implementation

[0077] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0078] Example 1: As Figure 1 As shown, this embodiment provides a semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label step-by-step optimization, including:

[0079] Obtain the publicly available ISPRS Potsdam remote sensing image dataset, and select test images from the dataset as the test set for the model. A subset of test images from the test set is randomly selected as the labeled training dataset. The remaining test images in the test set, after having their labels removed, were used as the unlabeled training dataset. .

[0080] test set The input is fed into a pre-trained remote sensing image semantic segmentation network model to obtain the semantic segmentation output.

[0081] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization includes the following steps in constructing and training the remote sensing image semantic segmentation network model:

[0082] Step 1: Constructing a Teacher Network Model Student Network Model , where k=1,2,3.

[0083] Step 2: Prepare the labeled dataset Input Teacher Network Model The network is trained by using the cross-entropy loss function to quantify the difference between the predicted segmentation map and the real annotation. The network parameters are optimized by the backpropagation algorithm, and the optimal parameters are saved as the trained teacher network model.

[0084] Step 3: Transfer the unlabeled dataset Input the trained teacher network model The initial pseudo-label dataset is obtained; the initial pseudo-label dataset is divided into high-quality pseudo-labels. Medium-sized pseudo-labels and inferior counterfeit labels Level 3 pseudo-labels.

[0085] Step 4: Perform multi-stage progressive optimization training using the three-level pseudo-labels, and provide optimized labeled data to train the student network model. The trained student network model is obtained. This ultimately forms a well-trained remote sensing image semantic segmentation network model.

[0086] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization specifically divides the labeled training data samples and unlabeled training data samples as follows:

[0087] Obtain the ISPRS Potsdam remote sensing image dataset D, and divide D into training sets in an 8:2 ratio. and test set The formula is as follows:

[0088]

[0089] in, The representative model's first One input image, The model is for the first The target value predicted for each image; These represent the sizes of the training set and the test set, respectively. ; Representative dataset The total number of samples.

[0090] training set Select 10% of the data sample As a labeled dataset, the remaining training set As an unlabeled training set, it satisfies:

[0091]

[0092] in, .

[0093] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization, the teacher network model Student Network Model Both adopt an encoder-decoder architecture;

[0094] The encoder uses ResNet-50 as the backbone network to extract deep semantic features from the input image.

[0095] The decoder includes a multi-scale feature enhancement module and a multi-scale feature fusion module. The multi-scale feature enhancement module enhances the deep features extracted by the encoder at multiple levels for ground objects of different scales in remote sensing images. The multi-scale feature fusion module achieves effective fusion of multi-scale semantic features through a cross-level feature interaction mechanism, which is used to improve the model's segmentation accuracy for complex ground object boundaries.

[0096] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization, the teacher network model Student Network Model The specific structure is as follows:

[0097] The encoder uses a ResNet-50 backbone network and includes five stages of downsampling operations on the input image. The encoder outputs a multi-scale feature map. , ;in, , , They represent the first The height, width, and number of channels of the feature map output by each feature extraction layer.

[0098] The multi-scale feature enhancement module performs multi-level feature enhancement on the deep features output by the encoder through a parallel branching structure. Each branch uses convolution operations with different receptive fields to output enhanced multi-scale features:

[0099]

[0100] in, This indicates a concatenation operation, used to merge multiple input tensors along the channel dimension, thereby integrating feature information from different branches or sources; This represents a convolutional layer, used to extract features from input data; This indicates dilated convolution, which obtains contextual information at different scales by adjusting the dilation rate, thereby enhancing the model's ability to distinguish ground features at different scales. This represents a convolution kernel with a height and width of 1. This represents a convolution kernel with a height and width of 3.

[0101] The multi-scale feature fusion module uses a cross-level feature interaction mechanism to combine the shallow features output by the encoder with the enhanced multi-scale features. Effective integration can improve the segmentation accuracy of complex terrain boundaries;

[0102] in This indicates an upsampling operation that matches the spatial dimensions of the feature map, improving the segmentation accuracy of complex feature boundaries through cross-level feature interaction.

[0103] The prediction module gradually restores the spatial resolution of the feature map through upsampling operations. First, it merges the features... Reduce the number of channels using convolutional layers:

[0104]

[0105] in This represents the predicted feature map, which is the output feature after processing by the prediction module.

[0106] Upsampling via bilinear interpolation Gradually restore to the original input image size:

[0107]

[0108] Output segmentation results ,pass The function calculates the probability that each pixel belongs to each category.

[0109] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization, wherein the teacher model The training process includes:

[0110] Step 1: Initialize the teacher network model parameter.

[0111] Step 2: Select remote sensing images from the labeled test set. .

[0112] Step 3: The dataset will have labels. Input Teacher Network Model The difference between the predicted segmentation map and the ground truth annotation is calculated using the cross-entropy loss function:

[0113]

[0114] Step 4: Optimize network parameters using the backpropagation algorithm Minimize the loss function .

[0115] Step 5: Save the network parameters that perform best on the validation set during training as the final teacher network model. .

[0116] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization divides the initial pseudo-label dataset into high-quality pseudo-labels. Medium-sized pseudo-labels Labeled as inferior Level 3 pseudo-tags; specifically including:

[0117] Step 1: Copy the teacher model The parameters are used to initialize the student network parameters.

[0118] Step 2: Select a set of remote sensing images from the unlabeled training set. .

[0119] Step 3: Save the final data during the teacher network model training process. Model checkpoints per training cycle .

[0120] Step 4: Calculate each unlabeled image Stability rating:

[0121]

[0122] in This represents the stability score of the image; the higher the score, the more stable and reliable the model's predictions are. This represents the total number of model checkpoints saved during training. and Representing the first The and the first One model checkpoint; The average crossover ratio (CROR) is used to measure the similarity between predictions made by two different models for the same image.

[0123] Step 5: Set the grading threshold for the stability score calculated from the root stability:

[0124]

[0125] in The quantile parameters are derived based on statistical distribution characteristics and satisfy... ; Represents the stability score set of quantiles; Represents the stability score set of Quantiles.

[0126] Step 6: Divide the unlabeled dataset into three quality levels based on the corresponding grading threshold:

[0127]

[0128]

[0129] .

[0130] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization includes the following multi-stage progressive optimization training:

[0131] The first phase involves training student network models. Using labeled datasets and high-quality pseudo-labels As training data, a composite loss function is used for supervised training:

[0132]

[0133] in For cross-entropy loss, These are the weight coefficients for the first stage; after training, the trained student network model is used. The original unlabeled image corresponding to the medium pseudo-label set Re-predict and generate an optimized set of intermediate pseudo-labels. .

[0134] The second stage involves training student network models. Based on the first stage, an optimized medium-level pseudo-label is introduced. The data used includes , and Supervised training is performed using an improved composite loss function:

[0135]

[0136] in, These are the weighting coefficients for the second stage.

[0137] After training, the trained student network model is used Inferior and counterfeit labels The corresponding original unlabeled image is re-predicted to generate optimized, low-quality pseudo-labels. ;

[0138] The third stage involves training student network models. Based on the second stage, an optimized version of the inferior pseudo-label was introduced. The data used includes , , and At this point, all pseudo-labels are incorporated into the learning process, and the complete conformance loss function is used for final training.

[0139]

[0140] in The weight coefficients are for the third stage; after three rounds of pseudo-label optimization and iterative model training, a student network model that has fully learned the labeled data is finally obtained. .

[0141] The aforementioned semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization also includes verifying the performance of the remote sensing image semantic segmentation prediction model by using the difference between the label data and the output of the model prediction.

[0142] A computer device, comprising:

[0143] Memory is used to store computer instructions.

[0144] A processor is configured to execute the computer instructions to implement the semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization.

[0145] Example 2: This example provides a computer device, including:

[0146] Memory is used to store computer instructions.

[0147] A processor is configured to execute the computer instructions to implement the semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization.

Claims

1. A semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization, characterized in that, include: Obtain the publicly available ISPRS Potsdam remote sensing image dataset, and select test images from the dataset as the test set for the model. A subset of test images from the test set is randomly selected as the labeled training dataset. The remaining test images in the test set, after having their labels removed, were used as the unlabeled training dataset. ; test set The input is fed into a pre-trained remote sensing image semantic segmentation network model to obtain the semantic segmentation output; The construction and training of the remote sensing image semantic segmentation network model includes: A teacher network model and a student network model are constructed, both of which adopt an encoder-decoder architecture. The encoder uses ResNet-50 as the backbone network to extract deep semantic features of the input image, and the decoder includes a multi-scale feature enhancement module and a multi-scale feature fusion module to improve the model's segmentation accuracy for complex ground feature boundaries. The labeled training dataset The data is input into the teacher network model for training. The cross-entropy loss function is used to quantify the difference between the predicted segmentation map and the real annotation. The network parameters are optimized through the backpropagation algorithm. Finally, the optimal parameters are saved as the trained teacher network model. The unlabeled training dataset Input the trained teacher network model to obtain the initial pseudo-label dataset; The initial pseudo-label dataset is used for multi-stage progressive optimization training to obtain optimized label data for training. The student network model is trained based on the optimized label data to obtain the trained student network model, and finally a trained remote sensing image semantic segmentation network model is formed.

2. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 1, characterized in that, The initial pseudo-label dataset was divided into high-quality pseudo-labels. Medium-sized pseudo-labels and inferior counterfeit labels Level 3 pseudo-labels.

3. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 1, characterized in that, The specific division between the labeled training dataset and the unlabeled training dataset includes: Obtain the ISPRS Potsdam remote sensing image dataset ,Will The training set was divided into two groups at an 8:2 ratio. and test set The formula is as follows: ； in, and Representing the model's first The j-th and j-th input images, and The model is for the first The target value predicted for the j-th input image, where i and j are index variables; , These represent the sizes of the training set and the test set, respectively. , ; Representative dataset The total number of samples; Select training set 10% of the data samples were used as the labeled training dataset. The remaining data samples are used as the unlabeled training dataset. ,satisfy: ； in, , These represent the sizes of the labeled dataset and the unlabeled training dataset, respectively. , .

4. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 1, characterized in that, The multi-scale feature enhancement module enhances the deep features extracted by the encoder at multiple levels for ground objects at different scales in remote sensing images. The multi-scale feature fusion module achieves effective fusion of multi-scale semantic features through a cross-level feature interaction mechanism, which is used to improve the model's segmentation accuracy for complex ground object boundaries.

5. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 4, characterized in that, The specific structures of the teacher network model and the student network model are as follows: The encoder uses a ResNet-50 backbone network and includes five stages of downsampling operations on the input image. The encoder outputs a multi-scale feature map. , ;in, For the input image, Represents the space of real numbers. These represent the height, width, and number of channels of the input image, respectively. The feature map output by the i-th feature extraction layer. , , They represent the first The height, width, and number of channels of the feature map output by each feature extraction layer; The multi-scale feature enhancement module performs multi-level feature enhancement on the deep features output by the encoder through a parallel branching structure. Each branch uses convolution operations with different receptive fields to output enhanced multi-scale features: ； in, For enhanced multi-scale features, This indicates a concatenation operation, used to merge multiple input tensors along the channel dimension, thereby integrating feature information from different branches or sources; This represents a convolutional layer, used to extract features from input data; This indicates dilated convolution, which obtains contextual information at different scales by adjusting the dilation rate, thereby enhancing the model's ability to distinguish ground features at different scales. This represents a convolution kernel with a height and width of 1. This represents a convolution kernel with a height and width of 3. This represents the feature map output by the 5th feature extraction layer; The multi-scale feature fusion module uses a cross-level feature interaction mechanism to combine the shallow features output by the encoder with the enhanced multi-scale features. Effective fusion can improve the segmentation accuracy of complex terrain boundaries. ； in, Indicates the first Multi-scale features obtained by effectively fusing multiple feature extraction layers This indicates an upsampling operation that matches the spatial dimensions of the feature map, thereby improving the segmentation accuracy of complex terrain boundaries through cross-level feature interaction; The prediction module gradually restores the spatial resolution of the feature map through upsampling operations. First, it fuses the features from the fourth feature extraction layer. Reduce the number of channels using convolutional layers: ； in, This represents the predicted feature map, which is the output feature after processing by the prediction module. Upsampling via bilinear interpolation Gradually restore to the original input image size: ； in, For the segmentation results, ; Output segmentation results ,pass The function calculates the probability that each pixel belongs to each category.

6. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 2, characterized in that, The training process of the teacher network model includes: Step 1: Initialize the parameters of the teacher network model; Step 2: Select remote sensing images from the labeled training set. ,in Let A be the image obtained by weak data augmentation of the i-th input image. Image B is obtained by applying strong data augmentation to the i-th input image; For the model to the first The target value to be predicted from each input image; N is the size of the labeled dataset; Step 3: Add the labeled training dataset Input the teacher network model and use the cross-entropy loss function to calculate the difference between the predicted segmentation map and the ground truth annotation: ； in, The cross-entropy loss function of the teacher network model. For the model's first One input image; For the teacher network model, h is the height position index; w is the width position index; and c is the channel position index. These represent the height, width, and number of channels of the input image, respectively. Step 4: Optimize network parameters using the backpropagation algorithm Minimize the loss function ; Step 5: Save the network parameters that perform best on the validation set during training as the final teacher network model. .

7. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 6, characterized in that, The initial pseudo-label dataset was divided into high-quality pseudo-labels. Medium-sized pseudo-labels and inferior counterfeit labels Level 3 pseudo-tags; specifically including: Step 1: Replicate the teacher network model The parameters are used to initialize the student network parameters; Step 2: Select a set of remote sensing images from the unlabeled training set. ; Step 3: Save the final data during the teacher network model training process. Model check for each training cycle , Teacher network models corresponding to three training cycles; Step 4: Calculate each unlabeled image Stability rating: ； in, This represents the set of stability scores for an image. The higher the score, the more stable and reliable the model's predictions for that image are. This represents the total number of model checkpoints saved during training. and Representing the first The and the first One model checkpoint; The average intersection-union ratio (IUU) is used to measure the similarity between the predictions of two different models for the same image. and Representing the first The and the first A teacher network model with checkpoints; Step 5: Set the grading threshold based on the calculated stability score: ； in, and The quantile parameters are derived based on statistical distribution characteristics and satisfy... ; Represents the stability score set of quantiles; Represents the stability score set of quantiles; It is a quantile function; Step 6: Divide the unlabeled dataset into three quality levels based on the corresponding grading threshold: High-quality fake labels : ; Medium-sized pseudo-labels : ; Labeled as inferior : .

8. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 7, characterized in that, The multi-stage progressive optimization training includes: Training the first phase of the student network model Using labeled datasets and high-quality pseudo-labels As training data, a composite loss function is used for supervised training: ； in, The loss function for the first-stage student network model is... Let cross-entropy be the loss function. These are the weighting coefficients for the first stage; After training is complete, the trained first-stage student network model is used. Re-predict the original unlabeled images corresponding to the medium pseudo-label set to generate an optimized medium pseudo-label set. ; Training the second-stage student network model Based on the first stage, an optimized medium-level pseudo-label is introduced. The data used includes , and Supervised training is performed using an improved composite loss function: ； in, The loss function for the second-stage student network model. These are the weighting coefficients for the second stage; After training, the trained second-order student network model is used Inferior and counterfeit labels The corresponding original unlabeled image is re-predicted to generate optimized, low-quality pseudo-labels. ; Training the third stage student network model Based on the second stage, an optimized version of the inferior pseudo-label was introduced. The data used includes , , and At this point, all pseudo-labels are incorporated into the learning process, and the complete composite loss function is used for final training. ； in, The loss function for the third-stage student network model. The weight coefficients are for the third stage; after three rounds of pseudo-label optimization and iterative model training, a student network model that has fully learned the labeled data is finally obtained. .

9. The semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization according to claim 1, characterized in that, It also includes testing the performance of the remote sensing image semantic segmentation prediction model by using the difference between the label data and the model's predicted output.

10. A computer device, characterized in that, include: Memory, used to store computer instructions; A processor for executing the computer instructions to implement the semi-supervised remote sensing image semantic segmentation method based on multi-stage pseudo-label optimization as described in any one of claims 1-9.

Images