High-resolution remote sensing image semantic segmentation passive domain adaptive method and device

Abstract

The invention relates to the field of remote sensing image processing, and discloses a high-resolution remote sensing image semantic segmentation passive domain self-adaption method and device, and the method employs a prototype guiding consistency learning module of teacher and student knowledge distillation, optimizes pseudo label generation and quality through knowledge transmission, data enhancement and a prototype guiding mechanism, and improves the semantic segmentation precision of a high-resolution remote sensing image. Feature alignment of a source domain and a target domain is improved, and particularly, the segmentation effect is remarkably improved in small ground feature recognition; through a cross-domain feature enhancement module, the feature extraction capability of complex boundaries and small ground features is enhanced by using channel adaptive weight distribution, so that the segmentation precision of a target domain is improved; a passive domain adaptive strategy is adopted, under the condition that source domain data cannot be accessed, a target domain is optimized based on a pre-trained source domain model, the data privacy and safety problems are solved, and the generalization performance of cross-domain semantic segmentation is improved. According to the method, the condition that source domain data are unavailable is effectively handled, and the segmentation precision of the cross-domain difficult-to-classify region in the remote sensing image is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image processing, and in particular to a high-resolution remote sensing image semantic segmentation passive domain adaptive method and device. Background Art

[0002] Semantic segmentation of remote sensing images has been widely used in urban planning, environmental monitoring, agricultural management, disaster monitoring, military and other fields, and has been a key research topic for decades. In the task of semantic segmentation of remote sensing images, the adequacy of data annotation is crucial to the performance of the semantic segmentation model. Fully supervised methods can achieve high segmentation accuracy by using a large amount of labeled data to train the model.

[0003] However, obtaining sufficient annotated data is often very time-consuming and costly, especially in high-resolution remote sensing images, and the annotation process usually requires the participation of professionals. In addition, fully supervised methods have a significant disadvantage. When the training data and test data exhibit different distribution characteristics, these methods are overly dependent on the characteristics of the training data and are difficult to adapt to new domains (i.e., test data). In practical applications, the difference in feature distribution between the source domain (i.e., the domain where the annotated data is located) and the target domain (i.e., the actual application domain) may cause the performance of the model on the target domain to drop significantly.

[0004] In order to solve the problem of obtaining labeled data and the limitations of cross-domain capabilities, unsupervised learning methods perform domain transfer learning through labeled source domain data and unlabeled target domain data, thereby enhancing the generalization ability of semantic segmentation models in the target domain. However, in some applications where data privacy protection requirements are relatively high or data cannot be obtained, such as military fields and disaster emergency scenarios, it is often impossible to obtain remote sensing images in the source domain and their high-quality labeled samples, and unsupervised cross-domain methods are not well suited for such tasks. When the target domain data is unlabeled, the passive domain adaptation method relies entirely on the pre-trained source domain model for domain adaptation without directly accessing the source domain data.

[0005] In recent years, passive domain adaptation technology has made some progress in the field of natural image processing. However, due to the complex and changeable scenes of remote sensing images, obvious scale differences, rich details and various types of objects, the effect of migrating passive domain adaptation technology based on natural images to the field of remote sensing image processing is often not good. In particular, in the process of generating pseudo labels, complex boundaries and small-scale objects in remote sensing images are prone to noise, resulting in reduced segmentation accuracy.

[0006] To address this problem, this study proposes a new passive domain adaptation method for semantic segmentation of high-resolution remote sensing images. This method effectively improves the segmentation accuracy when source domain data is not available by optimizing the quality of pseudo labels in the target domain. Summary of the invention

[0007] The purpose of the present invention is to propose a high-resolution remote sensing image semantic segmentation passive domain adaptation method and device to solve the technical problem that noise is easily generated in the passive domain adaptive segmentation processing of remote sensing images in the prior art, resulting in inaccurate segmentation accuracy.

[0008] The present invention provides a high-resolution remote sensing image semantic segmentation passive domain adaptive method, comprising the following steps:

[0009] S1: Obtain the trained source domain model and the remote sensing image dataset of the target domain, crop the remote sensing image dataset of the target domain, and obtain remote sensing images and semantic label images;

[0010] S2: Based on the source domain model, a semantic segmentation model adapted to the target domain is constructed. The semantic segmentation model includes: a student model and a teacher model;

[0011] S3: Input a batch of target domain remote sensing images into the teacher model to generate pseudo labels for the batch of images; then, concatenate and fuse the batch of remote sensing images and pseudo labels respectively to generate the same number of mixed images and corresponding mixed pseudo labels as the batch, which serve as input data for subsequent model training;

[0012] S4: Input the mixed image and the enhanced mixed image into the student model to generate the prototype features and student labels of the mixed image and the enhanced prototype features and enhanced student labels. The enhanced student labels are used for comparison and optimization with the pseudo labels to further improve the learning ability of the student model for the target domain data.

[0013] S5: Update the cluster center and corresponding class label of the class based on the prototype features, enhanced prototype features and mixed pseudo labels. The pseudo labels serve as a guide for the categories, helping the model to perform cross-domain consistency learning on the target domain or unlabeled data, ensuring that the updated cluster center can better represent the feature distribution of different categories.

[0014] S6: Calculate the consistency loss of semantic segmentation based on the pseudo-labels and enhanced student labels of the mixed image, and calculate the prototype consistency loss based on the enhanced model output and cluster centers to ensure the alignment between the class prototype and the model prediction;

[0015] S7: Introduce cross-domain feature enhancement modules in student models and teacher models;

[0016] S8: Train the model according to the overall loss and update the model parameters; input the remote sensing image dataset of the target domain into the trained student model to obtain the final semantic segmentation result.

[0017] A storage medium stores instructions and data for implementing a high-resolution remote sensing image semantic segmentation passive domain adaptation method.

[0018] A high-resolution remote sensing image semantic segmentation passive domain adaptation device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a high-resolution remote sensing image semantic segmentation passive domain adaptation method.

[0019] The beneficial effects provided by the present invention are as follows: the present invention adopts a passive domain adaptation strategy, and in the case of being unable to access the source domain data, the target domain is adaptively optimized through a pre-trained source domain model, thereby solving the source domain data privacy and security issues, while ensuring segmentation accuracy. Through the prototype-guided consistency learning module of teacher-student knowledge distillation and the cross-domain feature enhancement module, the pseudo-label generation process is optimized, the feature extraction capability of small objects and complex boundaries is enhanced, and the segmentation effect is effectively improved, especially in complex scenes in remote sensing images. Through the prototype-guided mechanism combined with the data enhancement strategy, the global feature alignment of the source domain and the target domain is achieved, the cross-domain adaptability is significantly improved, and the problem of poor adaptability of traditional methods between the source domain and the target domain with large differences in feature distribution is solved. The cross-domain feature enhancement module effectively improves the ability to extract complex features of the target domain through adaptive channel weight allocation, ensures a more accurate target domain feature representation, and thus improves the semantic segmentation accuracy of the model in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic flow chart of the method of the present invention;

[0021] Figure 2 is a schematic diagram of passive domain adaptation of the present invention;

[0022] Figure 3 It is the backbone network after adding the cross-domain feature enhancement module in the present invention;

[0023] Figure 4 It is a schematic diagram of the structure of the cross-domain feature enhancement module of the present invention;

[0024] Figure 5 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0026] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.

[0027] Please refer to Figure 1 , Figure 1 It is a schematic diagram of the process of the present invention.

[0028] The present invention provides a high-resolution remote sensing image semantic segmentation passive domain adaptation method, comprising the following steps:

[0029] S1: Obtain the trained source domain model and the remote sensing image dataset of the target domain, crop the remote sensing image dataset of the target domain, and obtain remote sensing images and semantic label images;

[0030] It should be noted that in the present invention, the remote sensing image dataset of the source domain is input into deeplabv2 to obtain a trained source domain model.

[0031] As an exemplary embodiment, the remote sensing image datasets used in the source domain and the target domain are ISPRS 2D semantic segmentation benchmark datasets, where the source domain dataset is Potsdam RGB dataset, the target domain dataset is Vaihingen IRRG dataset, and the semantic labels include six types of objects, namely buildings, impervious surfaces, trees, low vegetation, clutter, and cars. Since the feature distribution of the source domain and the target domain in the 'clutter' category is significantly different, and in the absence of source domain data, the model of the target domain is difficult to effectively identify and classify 'clutter'. Therefore, in order to ensure the stability of the model evaluation, the 'clutter' category is removed in the final performance evaluation.

[0032] Crop high-resolution remote sensing images of the source and target domains, and crop semantically labeled images of the source domain;

[0033] According to actual needs, the present invention crops the remote sensing images, elevation images, and semantic label images of the Potsdam RGB dataset with a fixed step size of 512 pixels, and obtains 4598 images with a pixel size of 512×512;

[0034] The present invention crops the remote sensing images, elevation images and semantic label images of the Vaihingen IRRG dataset with a fixed step size of 512 pixels, and its semantic labels are only used to evaluate the model performance.

[0035] S2: Based on the source domain model, a semantic segmentation model adapted to the target domain is constructed. The semantic segmentation model includes: a student model and a teacher model; Figure 2 is a schematic diagram of passive domain adaptation of the present invention;

[0036] It should be noted that step S2 is specifically as follows:

[0037] S21: Use source domain data to pre-train the semantic segmentation model to obtain pre-training weights;

[0038] S22: Based on the pre-trained weights of the source domain model, a student model is constructed; the structure of the student model is the same as the source domain model; a teacher model is constructed, and the teacher model has the same structure as the student model;

[0039] S23: Establish a knowledge distillation mechanism between the student model and the teacher model; the teacher model is responsible for providing stable prediction results and feature representations, while the student model learns how to extract useful knowledge from the teacher model.

[0040] As an exemplary embodiment, in step S2:

[0041] The teacher model is initialized based on the source domain model and remains unchanged to provide stable pseudo-labels. The student model is trained using unlabeled data from the target domain and gradually learns the knowledge of the teacher model through the knowledge distillation mechanism to achieve adaptive optimization. In the passive domain adaptation framework, the model can improve the segmentation performance in the target domain without accessing the source domain data.

[0042] S3: Input a batch of target domain remote sensing images into the teacher model to generate pseudo labels for the batch of images; then, concatenate and fuse the batch of remote sensing images and pseudo labels respectively to generate the same number of mixed images and corresponding mixed pseudo labels as the batch, which serve as input data for subsequent model training;

[0043] It should be noted that step S3 is specifically as follows:

[0044] S31: Multiple remote sensing images of each batch are input into the teacher model to obtain pseudo labels of each batch;

[0045] S32: Use the SaMix method to mix each batch of multiple remote sensing images and their pseudo labels into multiple mixed images and multiple mixed pseudo labels.

[0046] As an exemplary embodiment, each batch of remote sensing images in the present invention is 4, that is, the four remote sensing images in each batch are mixed with their pseudo labels using the SaMix method to form four mixed images and four mixed pseudo labels. Of course, as other optional embodiments, other numbers of images can also be used as a batch.

[0047] S4: Input the mixed image and the enhanced mixed image into the student model to generate the prototype features and student labels of the mixed image and the enhanced prototype features and enhanced student labels. The enhanced student labels are used for comparison and optimization with the pseudo labels to further improve the learning ability of the student model for the target domain data.

[0048] It should be noted that step S4 is specifically as follows:

[0049] S41: performing data enhancement on the mixed image by color jittering and Gaussian blur to obtain an enhanced mixed image;

[0050] S42: Input the mixed image and the enhanced mixed image into the student model respectively to generate the prototype features and student labels of the mixed image and the enhanced prototype features and enhanced student labels of the enhanced mixed image.

[0051] Among them, color dithering is a technology used in image processing, which is mainly used to simulate more colors on devices with a limited number of colors (such as black and white printers or monitors with a limited color palette). By introducing slight color changes in certain areas of the image, the image can be made to look closer to the color of the original image, thereby improving the quality and visual effect of the image.

[0052] Gaussian blur (English: Gaussian Blur), also known as Gaussian smoothing, is often used to reduce image noise and reduce detail levels. The image generated by this blurring technique has a visual effect like observing the image through a frosted glass, which is significantly different from the bokeh effect of the lens out-of-focus imaging and the effect of ordinary lighting shadows. Gaussian smoothing is also used in the pre-processing stage of computer vision algorithms to enhance the image effect at different scales (see scale space representation and scale space implementation).

[0053] S5: Update the cluster center and corresponding class label of the class based on the prototype features, enhanced prototype features and mixed pseudo labels. The pseudo labels serve as a guide for the categories, helping the model to perform cross-domain consistency learning on the target domain or unlabeled data, ensuring that the updated cluster center can better represent the feature distribution of different categories.

[0054] It should be noted that step S5 is specifically as follows:

[0055] S51: updating the cluster center of the class and the corresponding class label according to the prototype feature, the enhanced prototype feature and the mixed pseudo label;

[0056] The calculation formula of the cluster center can be expressed as:

[0057]

[0058] in, represents the cluster center of class c, represents the probability value of the pixel (h, w) belonging to class c in the mixed target domain image without data enhancement, Represents the feature vector of the pixel (h,w) in the mixed image.

[0059] The similarity calculation formula between the output features of the enhanced mixed image and the mixed pseudo-label calculated based on the cluster center can be expressed as:

[0060]

[0061] in, Represents the predicted probability of the pixel at position (h, w) in the enhanced mixed target domain image for category c. Represents the feature vector of the enhanced target domain mixed image, represents the transpose of the above cluster centers, Represents the final predicted category of the enhanced mixed target domain image pixel (h,w).

[0062] S52: The obtained cluster center labels and prototype features are interpolated to adapt to the image size of the target domain and used for subsequent loss calculations. This ensures that the loss function can effectively compare the differences between the model prediction results and the updated cluster centers and pseudo labels.

[0063] S6: Calculate the consistency loss of semantic segmentation based on the pseudo-labels of the mixed image and the enhanced student labels, and calculate the prototype consistency loss based on the enhanced model output and the cluster center (prototype) to ensure the alignment between the class prototype and the model prediction;

[0064] It should be noted that step S6 is specifically as follows:

[0065] S61: Calculate the weight of the unlabeled pixel according to the mixed pseudo-label, apply the calculated unlabeled pixel weight to the entire confidence map, and generate a pixel-level weight matrix for subsequent loss calculation; the calculation formula of the weight matrix is ​​as follows:

[0066]

[0067] Among them, U represents the unlabeled weight, C represents the confidence score of each pixel in the segmentation task, I(C≥α) represents the indicator function, which calculates the number of pixels in C≥α, T represents the total number of pixels, 1 represents the all-1 matrix with the same shape as C, and W represents the weight matrix of all pixels.

[0068] S62: Calculate the cross entropy loss of the mixed pseudo-label of the mixed image and the enhanced mixed student label and apply the pixel weight to the calculated loss to obtain the consistency learning loss of the target domain semantic segmentation; the calculation formula of the consistency learning loss of the target domain semantic segmentation is as follows:

[0069]

[0070] Among them, O represents the enhanced network output, l represents the mixed pseudo-label, c represents the number of categories, h, w represent the pixel coordinates, Wh,w represents the pixel weight matrix, and N is the total number of pixels.

[0071] S63: The cluster center label and the enhanced feature are cross entropy loss calculated to obtain the consistency loss guided by the first prototype; the formula is as follows:

[0072]

[0073] Among them, O represents the enhanced network output, y t represents the label after clustering update, c represents the number of categories, and h, w represent pixel coordinates.

[0074] S64: The enhanced mixed student labels and prototype features are cross-entropy loss calculated to obtain the consistency loss guided by the second prototype; the formula is as follows:

[0075]

[0076] Among them, O represents the enhanced network output, P t represents the predicted probability of each category after clustering update, c represents the number of categories, and h, w represent pixel coordinates.

[0077] S65: After adding the loss guided by the first prototype and the loss guided by the second prototype, and assigning loss weights, the total prototype-guided consistency loss is obtained. The formula is as follows:

[0078] L prototype =L prototype1 +L prototype2 .

[0079] S7: Introduce cross-domain feature enhancement modules in student models and teacher models;

[0080] This module adaptively weights the channels of the feature map, allowing the model to focus more on important features, thereby improving its ability to learn target domain data. During the training process, the module will help the model more accurately align category prototypes and model predictions, improve the calculation results of semantic segmentation consistency loss and prototype consistency loss, and improve the final target domain segmentation performance.

[0081] As an exemplary embodiment, step S7 is specifically as follows:

[0082] S71: First, a cross-domain feature enhancement module is integrated into the feature extraction network (deeplabv2-Resnet101) of the student model and the teacher model. This module enhances the model's ability to focus on important features by adaptively weighting the channels of the feature map. Figure 3 It is the backbone network after the cross-domain feature enhancement module is added in the present invention.

[0083] Figure 4 It is a schematic diagram of the structure of the cross-domain feature enhancement module of the present invention. Figure 4 The structure of the cross-domain feature enhancement module is demonstrated: first, the input feature map is compressed into a channel description vector of size 1×1×C through global average pooling (Fsq) to extract global features; then the weight of each channel is calculated through a fully connected layer; finally, the weight is multiplied by the original feature channel by channel to enhance the influence of important features and improve model performance.

[0084] S72: During training, the feature maps enhanced by the cross-domain feature enhancement module are used to calculate the consistency loss and prototype consistency loss. By improving the loss calculation method, the loss function can more accurately reflect the difference between the model prediction and the actual label or category prototype.

[0085] S73: Verify the improvement of the target domain segmentation performance by the feature enhancement module through training and optimization, and optimize based on the feedback.

[0086] S8: Train the model according to the overall loss and update the model parameters. Input the remote sensing image dataset of the target domain into the trained student model to obtain the final semantic segmentation result. The calculation formula of the overall loss is as follows:

[0087] L total =L consistency +λ·L prototype

[0088] where λ is the prototype-guided consistency loss weight.

[0089] It should be noted that step S8 is specifically as follows:

[0090] S81: Add the consistency learning loss of semantic segmentation and the consistency loss of prototype guidance to get the overall loss.

[0091] S82: Based on the overall loss, use the loss function to update the semantic segmentation model and train the model until convergence;

[0092] S83: Input the remote sensing image of the target domain into the trained model, obtain the semantic segmentation prediction result of the target domain through the final classifier, and achieve accurate object classification.

[0093] See also Figure 5 , Figure 5 40 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a high-resolution remote sensing image semantic segmentation passive domain adaptation device 401, a processor 402 and a storage medium 403.

[0094] A high-resolution remote sensing image semantic segmentation passive domain adaptation device 401: The high-resolution remote sensing image semantic segmentation passive domain adaptation device 401 implements the high-resolution remote sensing image semantic segmentation passive domain adaptation method.

[0095] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the high-resolution remote sensing image semantic segmentation passive domain adaptation method.

[0096] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the high-resolution remote sensing image semantic segmentation passive domain adaptation method.

[0097] The beneficial effects of the present invention are as follows: the present invention adopts a passive domain adaptation strategy, and in the case of being unable to access the source domain data, the target domain is adaptively optimized through a pre-trained source domain model, thereby solving the source domain data privacy and security issues, while ensuring segmentation accuracy. Through the prototype-guided consistency learning module of teacher-student knowledge distillation and the cross-domain feature enhancement module, the pseudo-label generation process is optimized, the feature extraction capability of small objects and complex boundaries is enhanced, and the segmentation effect is effectively improved, especially in complex scenes in remote sensing images. Through the prototype-guided mechanism combined with the data enhancement strategy, the global feature alignment of the source domain and the target domain is achieved, the cross-domain adaptability is significantly improved, and the problem of poor adaptability of traditional methods between the source domain and the target domain with large differences in feature distribution is solved. The cross-domain feature enhancement module effectively improves the ability to extract complex features of the target domain through adaptive channel weight allocation, ensures a more accurate target domain feature representation, and thus improves the semantic segmentation accuracy of the model in the target domain.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A passive domain adaptive method for semantic segmentation of high-resolution remote sensing images, characterized by: The method comprises the following steps: S1: Obtain the trained source domain model and the remote sensing image dataset of the target domain, crop the remote sensing image dataset of the target domain, and obtain remote sensing images and semantic label images; S2: Based on the source domain model, a semantic segmentation model adapted to the target domain is constructed. The semantic segmentation model includes: a student model and a teacher model; S3: Input a batch of target domain remote sensing images into the teacher model to generate pseudo labels for the batch of images; then, concatenate and fuse the batch of remote sensing images and pseudo labels respectively to generate the same number of mixed images and corresponding mixed pseudo labels as the batch, which serve as input data for subsequent model training; S4: Input the mixed image and the enhanced mixed image into the student model to generate the prototype features and student labels of the mixed image and the enhanced prototype features and enhanced student labels. The enhanced student labels are used for comparison and optimization with the pseudo labels to further improve the learning ability of the student model for the target domain data. S5: Update the cluster center and corresponding class label of the class based on the prototype features, enhanced prototype features and mixed pseudo labels. The pseudo labels serve as a guide for the categories, helping the model to perform cross-domain consistency learning on the target domain or unlabeled data, ensuring that the updated cluster center can better represent the feature distribution of different categories. S6: Calculate the consistency loss of semantic segmentation based on the pseudo-labels and enhanced student labels of the mixed image, and calculate the prototype consistency loss based on the enhanced model output and cluster centers to ensure the alignment between the class prototype and the model prediction; S7: Introduce cross-domain feature enhancement modules in student models and teacher models; S8: Train the model according to the overall loss and update the model parameters; input the remote sensing image dataset of the target domain into the trained student model to obtain the final semantic segmentation result.

2. The high-resolution remote sensing image semantic segmentation passive domain adaptation method according to claim 1, characterized in that: Step S2 is specifically as follows: S21: Use source domain data to pre-train the semantic segmentation model to obtain pre-training weights; S22: Based on the pre-trained weights of the source domain model, a student model is constructed; the structure of the student model is the same as the source domain model; a teacher model is constructed, and the teacher model has the same structure as the student model; S23: Establish a knowledge distillation mechanism between the student model and the teacher model; the teacher model is responsible for providing stable prediction results and feature representations, while the student model learns how to extract useful knowledge from the teacher model.

3. The high-resolution remote sensing image semantic segmentation passive domain adaptation method according to claim 1, characterized in that: Step S3 is specifically as follows: S31: Multiple remote sensing images of each batch are input into the teacher model to obtain pseudo labels of each batch; S32: Use the SaMix method to mix each batch of multiple remote sensing images and their pseudo labels into multiple mixed images and multiple mixed pseudo labels.

4. The high-resolution remote sensing image semantic segmentation passive domain adaptation method according to claim 1, characterized in that: Step S4 is specifically as follows: S41: performing data enhancement on the mixed image by color jittering and Gaussian blur to obtain an enhanced mixed image; S42: Input the mixed image and the enhanced mixed image into the student model respectively to generate the prototype features and student labels of the mixed image and the enhanced prototype features and enhanced student labels of the enhanced mixed image.

5. The high-resolution remote sensing image semantic segmentation passive domain adaptation method according to claim 1, characterized in that: Step S5 is specifically as follows: S51: updating the cluster center of the class and the corresponding class label according to the prototype feature, the enhanced prototype feature and the mixed pseudo label; S52: The obtained cluster center labels and prototype features are interpolated to adapt to the image size of the target domain and used for subsequent loss calculation.

6. The high-resolution remote sensing image semantic segmentation passive domain adaptation method according to claim 1, characterized in that: Step S6 is specifically as follows: S61: Calculate the weight of the unlabeled pixel according to the mixed pseudo-label, apply the calculated unlabeled pixel weight to the entire confidence map, and generate a pixel-level weight matrix for subsequent loss calculation; S62: Calculate the cross entropy loss of the mixed pseudo-labels of the mixed image and the enhanced mixed student labels and apply the pixel weights to the calculated loss to obtain the consistency learning loss of the target domain semantic segmentation; S63: Calculate the cross entropy loss of the cluster center label and the enhanced feature to obtain the consistency loss guided by the first prototype; S64: Calculate the cross entropy loss of the enhanced mixed student labels and the prototype features to obtain the consistency loss guided by the second prototype; S65: Add the loss guided by the first prototype and the loss guided by the second prototype, and assign loss weights to obtain the total prototype-guided consistency loss.

7. The high-resolution remote sensing image semantic segmentation passive domain adaptation method according to claim 1, characterized in that: Step S7 is specifically as follows: S71: Integrate a cross-domain feature enhancement module in the feature extraction network of the student model and the teacher model; this module enhances the model's ability to focus on important features by adaptively weighting the channels of the feature map; S72: During the training process, the feature map enhanced by the cross-domain feature enhancement module is used to calculate the consistency loss and the prototype consistency loss; S73: Through training and optimization, verify the improvement of the feature enhancement module on the target domain segmentation performance, and optimize based on the feedback.

8. The high-resolution remote sensing image semantic segmentation passive domain adaptation method according to claim 7, characterized in that: Step S8 is specifically as follows: S81: Add the consistency learning loss of semantic segmentation and the consistency loss of prototype guidance to get the overall loss; S82: Based on the overall loss, use the loss function to update the semantic segmentation model and train the model until convergence; S83: Input the remote sensing image of the target domain into the trained model, obtain the semantic segmentation prediction result of the target domain through the final classifier, and achieve accurate object classification.

9. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a high-resolution remote sensing image semantic segmentation passive domain adaptation method as described in any one of claims 1 to 8.

10. A passive domain adaptive device for semantic segmentation of high-resolution remote sensing images, characterized by: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a high-resolution remote sensing image semantic segmentation passive domain adaptation method as described in any one of claims 1 to 8.

Images