Passive domain noise-aware domain adaptive segmentation method for remote sensing images

Abstract

The application relates to a passive domain noise perception domain adaptive segmentation method for remote sensing images, and the method comprises the following steps: S1, training a source domain model to obtain an initial target domain model; S2, respectively based on an entropy value and a difference between an initial target domain model distribution and a distribution after inputting a target domain image into the initial target domain model, all target domain images are sorted to determine pseudo source domain, easy-to-adapt target domain images and difficult-to-adapt target domain images; S3, an initial classification model is trained by using a course learning method; S4, a segmentation category output by the initial classification model is used as a pseudo label, and the trained initial classification model is trained based on the pseudo label to obtain a trained segmentation model; S5, steps S2 to S4 are repeated until an iteration termination condition is reached; and S6, a target domain image is input into the segmentation model to obtain a remote sensing image segmentation result. Compared with the prior art, the application has the advantages of improving the segmentation accuracy of remote sensing images and the like.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image analysis technology, and in particular to an adaptive segmentation method for passive domain noise perception domain of remote sensing images. Background Technology

[0002] The goal of unsupervised adaptive domain segmentation methods is to address the performance degradation of models when applied to new data domains. This is primarily due to differences in feature distributions between the source and target domains, preventing the model from directly generalizing to the new domain. With the development of adaptive domain segmentation techniques, numerous studies have been proposed by researchers, taking different approaches such as adversarial learning, self-training, knowledge distillation, and contrastive learning, all achieving promising results. This technique has also been introduced into the field of remote sensing, successfully enabling the transfer and application of remote sensing image source domain models to the target domain.

[0003] However, in real-world scenarios, due to the privacy and security concerns surrounding remote sensing images, source domain data is often not easily accessible. Furthermore, the online deployment of some models requires that users have no access to the source domain data. This difficulty in obtaining source domain data in real-world scenarios prevents the direct application of current unsupervised adaptive segmentation methods.

[0004] Currently, when segmenting images in the remote sensing domain, one approach is to assume the availability of labeled source domain data. However, this approach fails to consider the difficulty in obtaining source domain data in real-world industrial scenarios due to privacy and security concerns. Another approach uses a passive domain segmentation model suited for natural images. However, this model is not suitable for remote sensing images, and the passive domain adaptive method lacks a feature alignment process between source and target domain data, resulting in poorer semantic segmentation of the target domain. When using this model for remote sensing image segmentation, the predicted target domain pseudo-labels contain more noise, leading to inaccurate segmentation results. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive segmentation method for passive domain noise perception domain of remote sensing images to solve the problems of excessive pseudo-label noise and inaccurate segmentation results in passive domain remote sensing image segmentation.

[0006] The objective of this invention can be achieved through the following technical solutions: An adaptive segmentation method for passive domain noise perception domain in remote sensing images, the method includes: S1. Obtain the statistical distribution in the pre-trained source domain model, calculate the difference between the statistical distribution and the statistical distribution after the target domain image is input into the source domain model, use the difference as the model loss function, train the source domain model, and obtain the initial target domain model. The target domain image is a remote sensing image. S2. Calculate the entropy value of the prediction result of the target domain image by the initial target domain model, and calculate the difference between the distribution of the initial target domain model and the distribution of the target domain image after inputting into the initial target domain model. Sort all target domain images based on the entropy value and the difference between the distribution of the initial target domain model and the distribution of the target domain image after inputting into the initial target domain model. Based on the sorting results, determine the pseudo source domain, the easily adaptable target domain image and the difficult-to-adapt target domain image. S3. Using the initial target domain model as the skeleton module, and combining it with the cross-attention module and discriminator to generate an initial classification model for adversarial learning, the initial classification model is trained using a curriculum-based learning method. During the training of the initial classification model, adversarial learning from the pseudo-source domain to the easily adaptable target domain is performed first, and then adversarial learning from the pseudo-source domain to the easily adaptable target domain to the difficult-to-adapt target domain is performed. S4. Input the target domain image into the trained initial classification model. Use the segmentation category output by the initial classification model as a pseudo-label. Train the skeleton module in the trained initial classification model based on the pseudo-label. Use the trained skeleton module as the segmentation model. S5. Repeat S2~S4 until the iteration termination condition is met; S6. Input the actual target domain image into the segmentation model to obtain the remote sensing image segmentation result.

[0007] Furthermore, the specific steps of S1 are as follows: Obtain the statistical distribution of the mean and variance data in the batch normalization layer of the pre-trained source domain model. Then, input the target domain image into the pre-trained source domain model and record the mean and variance of the changes in the batch normalization layer of the source domain model. The mean and variance of the changes are the statistical distribution of the target domain image after inputting into the source domain model. Calculate the difference between the two statistical distributions and use the difference as the model loss function to train the source domain model and obtain the initial target domain model.

[0008] Furthermore, the model loss function is: in, L It is the total number of batch normalization layers in the pre-trained model. and Let each represent the t-th image in the target domain image. The first after the pre-trained model l Changes in mean and variance after batch normalization and represents the mean and variance of all batch normalized layers in the source domain model, respectively.

[0009] Furthermore, the specific steps of S2 are as follows: The target domain image is input into the initial target domain model, which outputs a predicted target domain image. The entropy value of the predicted target domain image is calculated, and the difference between the distribution of the initial target domain model and the distribution of the target domain image after input into the initial target domain model is calculated. Based on the entropy value and the difference between the distribution of the initial target domain model and the distribution of the target domain image after input into the initial target domain model, all target domain images are sorted in ascending order, yielding entropy value sorting results and difference sorting results. The top two sorting results are then compared. Images with identical percentages are considered pseudo-source regions. These pseudo-source regions will be removed from both sorting results. The first of the two sorting results after removing the pseudo-source regions will be... Images with the same percentage are considered as easily adaptable target domain images. Then, from the two sorting results excluding the pseudo-source domain, easily adaptable target domain images are removed, and the remaining images are difficult-to-adapt target domain images.

[0010] Furthermore, the entropy value of the target domain image prediction result is the average value of the entropy map of the prediction result, specifically: in, Represents the first in the target domain image i Zhang Image and These represent the height and width of the input target domain image, respectively. The number of segmentation categories representing the input target domain image. An entropy graph representing the prediction results. , and These represent the height and width of the prediction result, as well as the number of segmentation categories, respectively.

[0011] Furthermore, the image The entropy graph of the prediction results is as follows: in, Represents the initial target domain model. This represents the softmax activation function.

[0012] Furthermore, the difference between the initial target domain model distribution and the distribution of the target domain image after inputting into the initial target domain model is as follows: in, It is the total number of batch normalization layers in the initial target domain model. and They represent the first and second parts of the target domain image, respectively. i Zhang Image After the initial target domain model, the first lChanges in mean and variance after batch normalization and Let represent the mean and variance of all batch normalized layers in the initial target domain model, respectively. Represents the first in the target domain image i Zhang Image .

[0013] Furthermore, the specific steps for training the initial classification model are as follows: A1. The pseudo-source domain is taken as the first source domain, and the easily adaptable target domain is taken as the first target domain. The first source domain and the first target domain are input into the skeleton module. The skeleton module outputs a semantic feature map. The semantic feature map is upsampled to obtain the segmentation result. The segmentation result is downsampled and input into the cross-attention module, and then into the feedforward module to obtain the final noise perception features of the target domain branch and the final noise perception features of the source domain branch. The final noise perception features are input into the discriminator for adversarial learning, the loss function is calculated, and the initial classification model is trained based on the loss function. A2. Using the pseudo-source domain and the easily adaptable target domain as the second source domain and the difficult-to-adapt target domain as the second target domain, input the second source domain and the second target domain into the skeleton module, repeat the steps of A1, and train the initial classification model.

[0014] Furthermore, the initial classification model is divided into a target domain branch and a source domain branch. The specific steps of A1 are as follows: The first source domain and the first target domain are input into the skeleton module, which outputs the target domain semantic feature map of the corresponding target domain branch and the source domain semantic feature map of the corresponding source domain branch. The target domain semantic feature map and the source domain semantic feature map are upsampled to obtain the target domain segmentation result and the source domain segmentation result. The target domain segmentation result and the source domain segmentation result are downsampled to obtain the target domain feature and the source domain feature. The target domain feature and the source domain feature are respectively input into the target domain branch and the source domain branch of the cross-attention module. The cross-attention module calculates the cross-attention weights of the source domain branch and the target domain branch. The cross-attention weights are input into the feedforward module to obtain the final noise perception feature of the target domain branch and the final noise perception feature of the source domain branch. The final noise perception feature is input into the discriminator for adversarial learning, and the loss function is calculated. The initial classification model is trained based on the loss function.

[0015] Furthermore, the specific steps for inputting the actual target domain image into the segmentation model to obtain the remote sensing image segmentation result are as follows: The actual target domain image is input into the segmentation model, the segmentation model outputs the actual semantic feature map, and the actual semantic feature map is upsampled to obtain the remote sensing image segmentation result.

[0016] Compared with the prior art, the present invention has the following beneficial effects: To address the privacy and security concerns of remote sensing image data, as well as the need for online model deployment, and the difficulty in obtaining remote sensing source domain data in actual industrial environments, this invention first calculates the difference between the statistical distribution of the source domain image and the statistical distribution of the target domain image after inputting them into the source domain model. This allows for the initial adaptation of the source domain model to the target domain. Then, based on entropy values ​​and distribution differences, the target domain images are categorized into pseudo-source domains, easily adaptable target domain images, and difficult-to-adapt target domain images. Adversarial training is then conducted from easy to difficult to obtain pseudo-labels with less noise, thereby improving the accuracy of the segmentation results. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention; Figure 2 This is a training block diagram for S1 to S4 of the present invention; Figure 3 This is a block diagram of the cross-attention module of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0019] This invention proposes an adaptive segmentation method for passive domain noise perception domain in remote sensing images. The flowchart of the method is as follows: Figure 1 As shown. The method includes the following steps: S1. Obtain the statistical distribution in the pre-trained source domain model, calculate the difference between the statistical distribution and the statistical distribution after the target domain image is input into the source domain model, use the difference as the model loss function, train the source domain model, and obtain the initial target domain model. The target domain image is a remote sensing image. S2. Calculate the entropy value of the prediction result of the target domain image by the initial target domain model, and calculate the difference between the distribution of the initial target domain model and the distribution of the target domain image after inputting into the initial target domain model. Sort all target domain images based on the entropy value and the difference between the distribution of the initial target domain model and the distribution of the target domain image after inputting into the initial target domain model. Based on the sorting results, determine the pseudo source domain, the easily adaptable target domain image and the difficult-to-adapt target domain image. S3. Using the initial target domain model as the skeleton module, and combining it with the cross-attention module and discriminator to generate an initial classification model for adversarial learning, the initial classification model is trained using a curriculum-based learning method. During the training of the initial classification model, adversarial learning from the pseudo-source domain to the easily adaptable target domain is performed first, and then adversarial learning from the pseudo-source domain to the easily adaptable target domain to the difficult-to-adapt target domain is performed. S4. Input the target domain image into the trained initial classification model. Use the segmentation category output by the initial classification model as a pseudo-label. Train the skeleton module in the trained initial classification model based on the pseudo-label. Use the trained skeleton module as the segmentation model. S5. Repeat S2~S4 until the iteration termination condition is met; S6. Input the actual target domain image into the segmentation model to obtain the remote sensing image segmentation result.

[0020] The steps S1 to S4 of this invention can be divided into two stages. The first stage corresponds to steps S1 and S2. The second stage corresponds to steps S3 and S4.

[0021] In S1, the statistical distribution information in the pre-trained source domain model is used to calculate the difference between the statistical distribution of the source domain model and the target domain image, enabling the model to initially adapt to the target domain. S1 corresponds to... Figure 2 The content on the left side of the first stage. The pre-trained source domain model is built based on the RESNET101 network.

[0022] The difference between the statistical distribution of the input target domain image and the statistical distribution contained in the source domain model is calculated, and the model loss function is calculated. This is used to supervise the training of the source domain model, so that the model can initially adapt to the target domain.

[0023] The specific steps of S1 are as follows: First, the statistical distribution information of the mean and variance data in the batch normalization (BN) layers of the source domain model is obtained, which can reflect the distribution of source domain data to a certain extent. Furthermore, the model's mean and variance values ​​are dynamically updated by inputting a target domain image, thereby obtaining the data distribution information of the input image. Therefore, to adapt the source domain model to the target domain, this invention extracts the mean from all batch normalization layers in the source domain model. With variance And record; subsequently, record each remote sensing image of the target domain one by one. In the input source domain model, record the mean of changes in the current batch normalization layer. With variance By leveraging the differences in data distribution between domains, the batch normalization loss for learning the pseudo-source domain is calculated. This loss is used to learn a model that is initially aligned with the target domain, i.e., to initialize the target domain model. As shown below: in, L It is the total number of batch normalization layers in the pre-trained model. and These represent small batches. After the first lThe mean and variance changes after each batch of normalization layers are analyzed. Finally, each input of all target domain images is considered one round, and this process is repeated multiple times until the model obtains the initialized target domain model. .

[0024] In S2, the entropy value and statistical distribution differences of the model's prediction results for the target domain image are comprehensively considered. The target domain images are sorted and a pseudo-source domain is formed, which includes easily adaptable target domain images and difficult-to-adapt target domain images.

[0025] The magnitude of the entropy of the prediction result of the input image using the model The difference between the statistical distribution of this image and the source domain Sort all target domain images, and select the top one from the two sorted results. Percentage of the image, select the intersection image as the pseudo-source domain Similarly, for the remaining target domain images, the first... Percentage of the image, select the intersection and set it as the easily adaptable target domain image. The remaining images are set as images that are difficult to adapt to the target domain. .

[0026] The sorting and field selection methods are as follows: in, This represents the softmax activation function; This represents the initialized target domain model obtained in step 1; Represents each entropy graph The average value; Represents the i-th target domain image patch; Indicates the first i The distribution difference between the target image patch and the source domain; This indicates the first [part of] the entropy sorting results. Percentage of the image set; The ranking results of statistical distribution differences are as follows: Percentage of the image set.

[0027] S1 corresponds to Figure 2 The content on the right side of the first stage. In S2, the entropy value represents the uncertainty of the model's prediction result for the image, and the statistical distribution difference represents the similarity between the distribution of the target domain image and the distribution of the source domain. Taking both into account, the model can select the target domain image that is most similar to the source domain to form a pseudo source domain, and then form an easily adaptable target domain image and a difficult-to-adapt target domain image.

[0028] In the above sorting and domain selection method, in order to obtain the pseudo-source domain, the input model of each target domain image is first calculated. The output results are then calculated, and the entropy graph of the current results is plotted. Its definition is as follows: in, This represents the softmax activation function. The average value of each entropy map is then calculated. : in, Indicates the first i A target domain image patch, and These represent the height and width of the input image, respectively. This represents the number of image segmentation categories. Next, the sorting method defines the... i Distribution differences between target images and source regions : Finally, set a hyperparameter scaling. To obtain two pseudo-source domains with different ranges { }and{ Firstly, based on the previous... Percentage entropy sorting results { }get This includes target image patches with low entropy scores. Secondly, the ranking results are based on distribution differences. }get This includes target image patches whose distribution is relatively similar to that of the source domain. Then, the results of these two sorting methods are intersected: Received The final selected pseudo-source domain is used to calculate the remaining target domain image in the same way, and the remaining target domain is divided into easily adaptable patches. And difficult-to-adapt tiles .

[0029] In S3, the initial target domain model is used as the skeleton module, combined with a cross-attention module and a discriminator to generate an initial classification model for adversarial learning. Based on a convolutional neural network and a cross-attention Transformer, curriculum-based adversarial learning is performed between the pseudo-source domain and the target domain. The curriculum-based learning method first performs adversarial learning from the pseudo-source domain to the easily adaptable target domain, and then performs adversarial learning from the pseudo-source domain and the easily adaptable target domain to the difficult-to-adapt target domain. In each round of learning, model training is supervised simultaneously through model segmentation loss function and adversarial loss function. A noise perception module is designed using a cross-attention Transformer and inserted into the adversarial learning model.

[0030] The specific steps of curriculum-based noise perception adversarial learning are as follows: in Represents the pseudo-source domain input image. This represents the current best segmentation model (the current skeleton module). The pseudo-label represents the pseudo-source domain; s and t represent the pseudo-source domain and the target domain, respectively. It is a vector Dimensions It's a softmax operation; and These are the domain labels for the source and target maps, respectively. This represents a semantic domain discriminator. This represents a Transformer based on cross-attention. Indicates the semantic feature extraction results In the middle position Its characteristics.

[0031] like Figure 2 As shown in Phase Two, this invention performs model training from easy to difficult, starting in the pseudo-source domain. With adaptable target domain Adversarial learning is conducted between different regions, and secondly, in the pseudo-source domain... It is easy to adapt to the target domain. and difficulty adapting to the target domain Adversarial learning is employed to allow the model to learn the features of the target domain step by step. The specific process of adversarial learning is as follows: After extracting semantic features from the input image, the model upsamples it to the original image size and calculates the segmentation result. The model training is supervised by calculating the difference between the segmentation result and the pseudo-label. The loss function... as follows: in Represents the pseudo-source domain input image. This represents the current best segmentation model (the current skeleton module). A pseudo-tag representing a pseudo-source domain.

[0032] While training the model using pseudo-labels, the output segmentation results are downsampled into features, and an inter-domain similarity feature alignment is enhanced through a noise-aware module, namely the cross-attention Transformer module. For example... Figure 3 As shown, the output of the downsampled features in the model is first obtained, with the source domain being... The target domain is Then, through a single-layer Transformer module with two branches, the input features are mapped into three vectors—the query vectors—through convolutional token embedding (the first two layers) and convolutional projection (the first three to five layers). ,key Sum Then, the cross-domain attention is calculated: Where s and t represent the pseudo-source domain and the target domain, respectively. It is a vector Dimensions It's a softmax operation. Then, this invention will calculate the cross-attention weights obtained in the source branch. Multiplying the output of the convolutional token embedding serves as a masking mechanism to enhance the expression of semantically relevant features in the target domain branch. The calculation method is the same. Then, The input is fed into the feedforward module (the last three layers), and the resulting output is the same as... The final noise perception feature is obtained by addition. In the same way Get the target branch .

[0033] Finally, the output features are input into the discriminator for adversarial learning. The loss function for adversarial learning is as follows: in and These are the domain labels for the source and target maps, respectively. This represents a semantic domain discriminator. This represents a Transformer based on cross-attention. Indicates the semantic feature extraction results In the middle position The features of the target domain are analyzed. Adversarial training is conducted between the discriminator's ability to distinguish domain features and the segmentation model's ability to confuse the discriminator, enabling the source domain model to be adapted to the target domain.

[0034] In S4, generalized curriculum learning allows the model to be fine-tuned in the target domain.

[0035] The model trained using S3 is used to predict the segmentation category of all target domain images. This category is then treated as a pseudo-label and used as a supervision signal to fine-tune the model, thus generalizing from curriculum-based learning. This step further utilizes the hard-to-adapt target domain feature information that is not fully leveraged in curriculum-based adversarial learning, promoting the model's domain adaptation to the target domain.

[0036] In S5, the best model obtained in step 4 is used to re-predict the pseudo-label of the target domain image and re-divide the pseudo-source domain. Steps 2 to 4 are repeated. By continuously improving the accuracy of the pseudo-labels, the segmentation performance of the model is improved.

[0037] In S6, the trained network is used to input a remote sensing image of the target domain into the model to obtain the predicted image segmentation result.

[0038] The model in this embodiment was pre-trained on the DeepLab model and trained using training data from the publicly available datasets ISPRS and LoveDA. Test data was used for testing. The domain adaptation test results are as follows: In the ISPRS dataset, the mIoU and F1-score accuracies for domain adaptation from Vaihingen IR-RG to Potsdam IR-RG are 49.12% and 61.38%, respectively; for adaptation from Potsdam IR-RG to Vaihingen IR-RG, the mIoU and F1-score accuracies are 51.25% and 64.98%, respectively; for adaptation from Potsdam RGB to Vaihingen IR-RG, the mIoU and F1-score accuracies are 41.9% and 56.31%, respectively; and for adaptation from Vaihingen IR-RG to Potsdam RGB, the mIoU and F1-score accuracies are 40.3% and 54.01%, respectively. In the LoveDA dataset, the mIoU accuracy reaches 42.98% when adapting from the Rural domain to the Urban domain, and 38.91% when adapting from the Urban domain to the Rural domain.

[0039] The present invention has the following beneficial effects: First, this invention proposes a passive domain adaptive segmentation method for remote sensing images. Addressing the privacy and security concerns of remote sensing image data, as well as the need for online model deployment, and the difficulty in obtaining remote sensing source domain data in real-world industrial environments, this invention solves the passive domain problem during model domain adaptation, thus increasing the feasibility of implementing remote sensing image domain adaptive methods.

[0040] Second, in response to the problem of complex feature distribution and large inter-domain differences in remote sensing images, this invention proposes a course-based learning and multi-level feature perception module, which solves the problem of difficult domain adaptation caused by the wide land coverage and diverse imaging methods of remote sensing images from the perspectives of framework and module.

[0041] Third, to address the negative impact of noisy samples when the target domain is pseudo-labeled during model training, this invention introduces a Transformer module based on cross-attention, which enables the model to strengthen the semantic similarity feature correlation between the source and target domains during training, enhance the model's segmentation ability, and mitigate the negative impact of noisy samples.

[0042] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A passive domain noise-sensing domain adaptive segmentation method for remote sensing images, characterized in that, The methods include: S1. Obtain the statistical distribution in the pre-trained source domain model, calculate the difference between the statistical distribution and the statistical distribution after the target domain image is input into the source domain model, use the difference as the model loss function, train the source domain model, and obtain the initial target domain model. The target domain image is a remote sensing image. S2. Calculate the entropy value of the prediction result of the target domain image by the initial target domain model, and calculate the difference between the distribution of the initial target domain model and the distribution of the target domain image after inputting into the initial target domain model. Sort all target domain images based on the entropy value and the difference between the distribution of the initial target domain model and the distribution of the target domain image after inputting into the initial target domain model. Based on the sorting results, determine the pseudo source domain, the easily adaptable target domain image and the difficult-to-adapt target domain image. S3. Using the initial target domain model as the skeleton module, and combining it with the cross-attention module and discriminator to generate an initial classification model for adversarial learning, the initial classification model is trained using a curriculum-based learning method. During the training of the initial classification model, adversarial learning from the pseudo-source domain to the easily adaptable target domain is performed first, and then adversarial learning from the pseudo-source domain to the easily adaptable target domain to the difficult-to-adapt target domain is performed. S4. Input the target domain image into the trained initial classification model. Use the segmentation category output by the initial classification model as a pseudo-label. Train the skeleton module in the trained initial classification model based on the pseudo-label. Use the trained skeleton module as the segmentation model. S5. Repeat S2~S4 until the iteration termination condition is met; S6. Input the actual target domain image into the segmentation model to obtain the remote sensing image segmentation result; The specific steps of S2 are as follows: The target domain image is input into the initial target domain model, which outputs a predicted target domain image. The entropy value of the predicted target domain image is calculated, and the difference between the distribution of the initial target domain model and the distribution of the target domain image after input into the initial target domain model is calculated. Based on the entropy value and the difference between the distribution of the initial target domain model and the distribution of the target domain image after input into the initial target domain model, all target domain images are sorted in ascending order, yielding entropy value sorting results and difference sorting results. The top two sorting results are then compared. Images with identical percentages are considered pseudo-source regions. These pseudo-source regions will be removed from both sorting results. The first of the two sorting results after removing the pseudo-source regions will be... Images with the same percentage are considered as easily adaptable target domain images. Then, from the two sorting results excluding the pseudo-source domain, easily adaptable target domain images are removed, and the remaining images are difficult-to-adapt target domain images.

2. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 1, characterized in that, The specific steps of S1 are as follows: Obtain the statistical distribution of the mean and variance data in the batch normalization layer of the pre-trained source domain model. Then, input the target domain image into the pre-trained source domain model and record the mean and variance of the changes in the batch normalization layer of the source domain model. The mean and variance of the changes are the statistical distribution of the target domain image after inputting into the source domain model. Calculate the difference between the two statistical distributions and use the difference as the model loss function to train the source domain model and obtain the initial target domain model.

3. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 2, characterized in that, The model loss function is: in, L It is the total number of batch normalization layers in the pre-trained model. and Let each represent the t-th image in the target domain image. The first after the pre-trained model l Changes in mean and variance after batch normalization and represents the mean and variance of all batch normalized layers in the source domain model, respectively.

4. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 1, characterized in that, The entropy value of the target domain image prediction result is the average value of the entropy map of the prediction result, specifically: in, Represents the first in the target domain image i Zhang Image and These represent the height and width of the input target domain image, respectively. The number of segmentation categories representing the input target domain image. An entropy graph representing the prediction results. , and These represent the height and width of the prediction result, as well as the number of segmentation categories, respectively.

5. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 4, characterized in that, image The entropy graph of the prediction results is as follows: in, Represents the initial target domain model. This represents the softmax activation function.

6. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 5, characterized in that, The difference between the initial target domain model distribution and the distribution of the target domain image after inputting into the initial target domain model is as follows: in, It is the total number of batch normalization layers in the initial target domain model. and They represent the first and second parts of the target domain image, respectively. i Zhang Image After the initial target domain model, the first l Changes in mean and variance after batch normalization and Let represent the mean and variance of all batch normalized layers in the initial target domain model, respectively. Represents the first in the target domain image i Zhang Image .

7. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 1, characterized in that, The specific steps for training the initial classification model are as follows: A1. The pseudo-source domain is taken as the first source domain, and the easily adaptable target domain is taken as the first target domain. The first source domain and the first target domain are input into the skeleton module. The skeleton module outputs a semantic feature map. The semantic feature map is upsampled to obtain the segmentation result. The segmentation result is downsampled and input into the cross-attention module, and then into the feedforward module to obtain the final noise perception features of the target domain branch and the final noise perception features of the source domain branch. The final noise perception features are input into the discriminator for adversarial learning, the loss function is calculated, and the initial classification model is trained based on the loss function. A2. Using the pseudo-source domain and the easily adaptable target domain as the second source domain and the difficult-to-adapt target domain as the second target domain, input the second source domain and the second target domain into the skeleton module, repeat the steps of A1, and train the initial classification model.

8. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 7, characterized in that, The initial classification model is divided into a target domain branch and a source domain branch. The specific steps of A1 are as follows: The first source domain and the first target domain are input into the skeleton module, which outputs the target domain semantic feature map of the corresponding target domain branch and the source domain semantic feature map of the corresponding source domain branch. The target domain semantic feature map and the source domain semantic feature map are upsampled to obtain the target domain segmentation result and the source domain segmentation result. The target domain segmentation result and the source domain segmentation result are downsampled to obtain the target domain feature and the source domain feature. The target domain feature and the source domain feature are respectively input into the target domain branch and the source domain branch of the cross-attention module. The cross-attention module calculates the cross-attention weights of the source domain branch and the target domain branch. The cross-attention weights are input into the feedforward module to obtain the final noise perception feature of the target domain branch and the final noise perception feature of the source domain branch. The final noise perception feature is input into the discriminator for adversarial learning, and the loss function is calculated. The initial classification model is trained based on the loss function.

9. The passive domain noise-sensing domain adaptive segmentation method for remote sensing images according to claim 1, characterized in that, The specific steps for inputting the actual target domain image into the segmentation model to obtain the remote sensing image segmentation result are as follows: The actual target domain image is input into the segmentation model, the segmentation model outputs the actual semantic feature map, and the actual semantic feature map is upsampled to obtain the remote sensing image segmentation result.

Images