A multi-task cooperative learning remote sensing unsupervised domain adaptive semantic segmentation method

Abstract

The present application relates to the field of remote sensing image processing, and discloses a multi-task collaborative learning remote sensing unsupervised domain self-adaptive semantic segmentation method, which comprises the following steps: sending task-specific semantic features and elevation features into a cross-task feature correlation learning module to learn the correlation between different tasks and obtain enhanced task-specific feature representation; using a pseudo-label guided mixed consistency learning module to improve the quality of pseudo labels and promote global domain alignment; and using an entropy guided class level alignment to improve the separability of difficult-to-classify classes. The present application combines multi-task learning with remote sensing domain adaptive semantic segmentation, fully utilizes elevation information to assist remote sensing cross-domain feature classification, and significantly improves the cross-domain segmentation accuracy of difficult-to-classify areas of remote sensing images.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing, and in particular to a remote sensing unsupervised adaptive semantic segmentation method based on multi-task collaborative learning. Background Technology

[0002] Many tasks in remote sensing use optical imagery as input, such as image classification, object detection, semantic segmentation, elevation estimation, and edge estimation. By setting different tasks and supervision information, deep learning models can learn different types of features to generate specific prediction results. For example, in semantic segmentation, neural networks aim to establish the correlation between semantic features and land cover category information to generate pixel-level land cover classification results. In elevation estimation, neural networks focus more on the correlation between elevation features and elevation information, aiming to predict the height values ​​of land covers. Therefore, the focus of neural networks differs in different tasks. By effectively combining different tasks, cross-task feature representations can be implicitly learned to improve model performance. Existing research often uses multi-task learning techniques to combine different representations from multiple tasks. Simply put, multi-task learning involves training and predicting different visual tasks on a single model to better collaboratively learn the correlations between different tasks. In this way, different tasks can learn from each other's beneficial experiences and knowledge, thereby improving the overall performance of the model.

[0003] Due to the insufficient information provided by optical imagery, phenomena such as homonymous objects and homonymous objects exhibiting different spectra are prevalent in various remote sensing datasets. Neural networks heavily rely on optical imagery when learning semantic features, making these phenomena prone to semantic ambiguity and increasing the likelihood of confused classification of different land features in cross-domain segmentation tasks. Simply adjusting the convolutional layer structure of the network or enhancing the feature learning ability of deep neural networks cannot fundamentally address the limitations inherent in the data itself. To overcome these limitations, two strategies can be considered: first, leveraging additional information to mitigate the negative impact of homonymous objects and homonymous objects exhibiting different spectra; second, promoting semantic category-level alignment of land features to enhance classification guidance. To acquire additional information to compensate for the deficiencies of single datasets, existing semantic segmentation research typically uses multi-task learning techniques to acquire and fuse implicit feature representations from supervised information, further exploring the correlations between different tasks to improve model performance. This provides numerous theoretical foundations and practical examples. However, such methods are rarely applied in cross-domain segmentation of remote sensing data, neglecting challenging scenarios in remote sensing datasets and overlooking the effective combination of multi-task learning techniques and domain adaptation strategies. Summary of the Invention

[0004] The purpose of this invention is to propose a remote sensing unsupervised adaptive semantic segmentation method based on multi-task collaborative learning, which solves the technical problem that existing technologies cannot accurately segment cross-domain scenarios for specific classification scenarios.

[0005] A remote sensing unsupervised adaptive semantic segmentation method based on multi-task collaborative learning includes the following steps:

[0006] S1: Obtain the remote sensing image datasets of the source and target domains, crop the remote sensing image datasets of the source and target domains to obtain elevation images, remote sensing images, and semantically labeled images.

[0007] S2: Construct a semantic segmentation model, which includes a student model and a teacher model; both the student model and the teacher model include an initial feature extraction module, a cross-task feature association learning module, and a final classifier.

[0008] S3: Obtain remote sensing images and elevation images of the mixed samples;

[0009] Input the remote sensing image of the target domain into the teacher model to obtain the labels of the target domain samples and the pseudo-labels of the mixed samples;

[0010] The source domain remote sensing image and the enhanced hybrid remote sensing image are input into the student model to obtain the final elevation estimation results of the source domain samples and the final elevation estimation results of the hybrid samples, as well as the final semantic segmentation results of the source domain samples and the final semantic segmentation results of the hybrid samples.

[0011] S4: Construct a domain adaptation module, which includes a pseudo-label-guided hybrid consistency learning module and an entropy-guided class-level alignment module. The pseudo-label-guided hybrid consistency learning module uses the pseudo-labels generated above to supervise the student model's learning of mixed samples. The entropy-guided class-level alignment module weights the entropy map and the mean entropy of each class into the segmentation result to guide fine-grained class-level alignment.

[0012] S5: Calculate the semantic segmentation loss of the source domain samples based on the source domain sample segmentation results and semantic label images. Use the mixed pseudo-labels as supervision information, calculate the semantic segmentation loss and weighted cross-entropy loss of the mixed samples using the semantic segmentation results of the mixed samples, and calculate the overall semantic segmentation loss.

[0013] S6: Calculate the berHu loss from the elevation estimates and true elevation values ​​of the source domain samples and the mixed samples to obtain the elevation estimation loss of the source domain samples and the mixed samples, and calculate the overall elevation estimation loss;

[0014] S7: Weight the entropy map and the mean entropy of each class into the prediction results of source domain samples and mixed samples to promote semantic alignment of data distribution between the source domain and the target domain; use the source domain ground truth and mixed pseudo-labels as supervision information, and calculate the entropy loss with the semantic segmentation results weighted by the entropy value;

[0015] S8: Train the model based on 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.

[0016] A storage medium storing instructions and data for implementing a remote sensing unsupervised adaptive semantic segmentation method based on multi-task collaborative learning.

[0017] A remote sensing unsupervised adaptive semantic segmentation device based on multi-task collaborative learning includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a remote sensing unsupervised adaptive semantic segmentation method based on multi-task collaborative learning.

[0018] The beneficial effects provided by this invention are as follows: It utilizes cross-task feature association learning, fully mining and optimizing cross-task feature representations through continuous cross-feature enhancement units, and closely associating elevation and semantic features to leverage implicit elevation information and improve the accuracy of cross-domain segmentation tasks. Furthermore, this invention optimizes the quality of pseudo-labels, further using pseudo-label-guided hybrid consistency learning and entropy-guided category-level alignment to narrow the domain gap between the source and target domains, promoting good transfer of source domain knowledge and significantly enhancing the model's cross-domain capability. This invention greatly improves the segmentation accuracy of challenging scenes in remote sensing datasets, such as shaded regions, boundary regions, and similar spectral regions with significant height differences. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0020] Figure 2 This is a schematic diagram of the cross-task feature association learning module structure of the present invention;

[0021] Figure 3 This is a schematic diagram of the attention fusion module in the cross-feature enhancement unit;

[0022] Figure 4 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0024] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0025] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the method of the present invention;

[0026] This invention provides a remote sensing unsupervised adaptive semantic segmentation method based on multi-task collaborative learning, comprising the following steps:

[0027] S1: Obtain the remote sensing image datasets of the source and target domains, crop the remote sensing image datasets of the source and target domains to obtain elevation images, remote sensing images, and semantically labeled images.

[0028] It should be noted that, specifically, the remote sensing image datasets used in the source and target domains are the ISPRS2D semantic segmentation benchmark datasets. The source domain dataset is the Potsdam IRRG dataset, and the target domain dataset is the VaihingenIRRG dataset. The semantic labels include six types of land features: buildings, impervious surfaces, trees, low vegetation, clutter, and vehicles.

[0029] Cropping elevation images and high-resolution remote sensing images of the source and target domains, and cropping semantically labeled images of the source domain;

[0030] Based on actual needs, this invention uses a fixed step size of 512 pixels to crop remote sensing images, elevation images and semantic label images of the Potsdam IRRG dataset, resulting in 4598 images with a pixel size of 512×512.

[0031] This invention uses a fixed step size of 256 pixels to crop remote sensing images, elevation images, and semantically labeled images from the Vaihingen IRRG dataset. The semantic labels are used only to evaluate model performance.

[0032] S2: Construct a semantic segmentation model, which includes a student model and a teacher model; both the student model and the teacher model include an initial feature extraction module, a cross-task feature association learning module, and a final classifier.

[0033] It should be noted that in step S2:

[0034] The initial feature extraction module includes: a parameter-shared encoder, an elevation estimation bottleneck, a semantic segmentation bottleneck, an initial elevation classifier, and an initial semantic classifier; the initial feature extraction module is used to extract initial features and generate initial semantic segmentation results and initial elevation estimation prediction results.

[0035] As one embodiment, the parameter-sharing encoder uses a ResNet101 network. The semantic segmentation bottleneck and the elevation estimation bottleneck are both composed of two consecutive residual blocks, including two 1×1 convolutional layers and one 3×3 convolutional layer. The initial elevation classifier and the initial semantic classifier are composed of 1×1 convolutional layers.

[0036] Specifically, the initial semantic features and initial elevation features are obtained through the initial feature extraction module, and the corresponding initial semantic segmentation prediction results and initial elevation estimation prediction results are generated using the initial classifier.

[0037] Please refer to Figures 2-3 , Figure 2 This is a schematic diagram of the cross-task feature association learning module structure of the present invention. Figure 3 This is a schematic diagram of the attention fusion module in the cross-feature enhancement unit.

[0038] It should be noted that the cross-task feature association learning module consists of two consecutive cross-feature enhancement units and two task-specific gating structures; the cross-task feature association learning module is used to generate adaptive fusion features based on the initial features;

[0039] The initial features are input into two consecutive cross-feature enhancement units to obtain intermediate and final features in sequence. The initial feature F init and intermediate features F mid Through channel merging, and with the final feature F last The original and complementary features are input together into the gating structure to obtain the adaptive fusion feature F. fuse This fusion operation can be represented as follows:

[0040]

[0041] Where, conv 1×1 (g) represents the channel dimensionality reduction operation, which reduces the number of channels to half of the original number of channels. [g] represents the channel merging operation. sigmoid(g) is the activation function.

[0042] The cross-feature enhancement unit consists of two task-specific ASPPs and a cross-attention fusion block. The cross-attention fusion block comprises a channel attention module, a spatial attention module, and a nested fusion gating structure. This module is used to cross-integrate task features, task-specific salient features, and cross-task salient features to fully enhance the feature representations of different tasks. For semantic segmentation tasks, this process can be represented as:

[0043]

[0044] Where conv 1×1 (g) represents the channel dimensionality reduction operation, which reduces the number of channels to half of the original number. [g] represents the channel merging operation. sigmoid(g) is the activation function. Indicates significant elevation characteristics. This represents the initial semantic features.

[0045] The final classifier is used to generate the final semantic segmentation and elevation estimation prediction results based on the adaptive fusion features.

[0046] As one example, the decoder of the Deeplabv2 network is used as the final classifier.

[0047] S3: Obtain remote sensing images and elevation images of the mixed samples;

[0048] Input the target domain remote sensing image into the teacher model to obtain the labels of the target domain samples and the pseudo labels of the mixed samples; input the source domain remote sensing image and the enhanced mixed remote sensing image into the student model to obtain the final elevation estimation results of the source domain samples and the final elevation estimation results of the mixed samples, as well as the final semantic segmentation results of the source domain samples and the final semantic segmentation results of the mixed samples.

[0049] It should be noted that step S3 specifically involves:

[0050] S31: Use the ClassMix method to generate a hybrid mask, and select the corresponding regions of the source and target domain samples based on the mask to generate a remote sensing image of the hybrid samples;

[0051] S32: Input the remote sensing image of the target domain into the teacher model to generate pseudo-labels; combine the hybrid mask, the target domain pseudo-labels and the source domain label ground truth to generate pseudo-labels for the hybrid samples;

[0052] S33: Data augmentation of mixed samples, including random horizontal flipping, vertical flipping, color jittering, and Gaussian noise;

[0053] S34: Input the remote sensing images of the source domain and the target domain into the initial feature extraction module, generate initial features through the parameter sharing encoder, and generate initial semantic segmentation and elevation estimation prediction results using the elevation estimation bottleneck, semantic segmentation bottleneck and initial semantic classifier;

[0054] S35: The initial features are fed into the cross-task feature association learning module, and the final adaptive fusion features of semantics and elevation are generated through the cross feature enhancement unit and the task-specific gating structure.

[0055] S36: Input the adaptive fusion features into the final classifier to generate the final semantic segmentation and elevation estimation prediction results.

[0056] S4: Construct a domain adaptation module, which includes a pseudo-label-guided hybrid consistency learning module and an entropy-guided class-level alignment module. The pseudo-label-guided hybrid consistency learning module uses the pseudo-labels generated above to supervise the student model's learning of mixed samples. The entropy-guided class-level alignment module weights the entropy map and the mean entropy of each class into the segmentation result to guide fine-grained class-level alignment.

[0057] It should be noted that step S4 specifically involves:

[0058] S41: Construct a pseudo-label-guided hybrid consistency learning module to balance the model's feature representation capability and pseudo-label quality; the pseudo-label-guided hybrid consistency learning module adopts an incremental pseudo-label selection strategy, dynamically adjusting the current threshold based on the given minimum threshold, maximum threshold, total number of iterations and current number of iterations, and selecting pseudo-labels that meet the conditions;

[0059] As one embodiment, the specific calculation process of step S41 is as follows:

[0060]

[0061] in These are the filtered pseudo-labels, Y soft It is the original pseudo-label of the target domain, σ(Y) soft ) represents the value used to generate pseudo-label Y soft The predicted probability plot.

[0062] S42: Construct an entropy-guided category-level alignment module; calculate the entropy map based on the final semantic segmentation prediction results and labels, and use the entropy map to calculate the average entropy value of each class, followed by normalization.

[0063] As one embodiment, the calculation process of the entropy map and the mean entropy value for each class in step S42 is as follows:

[0064]

[0065] Where P represents the predicted probability map, W, H, and C are the length, width, and number of channels of the probability map P, respectively, and Y... S Represents semantic tags, num c Representing label Y S The number of pixels belonging to class c.

[0066] S5: Calculate the semantic segmentation loss of the source domain samples based on the source domain sample segmentation results and semantic label images. Use the mixed pseudo-labels as supervision information, calculate the semantic segmentation loss and weighted cross-entropy loss of the mixed samples using the semantic segmentation results of the mixed samples, and calculate the overall semantic segmentation loss.

[0067] It should be noted that step S5 specifically involves:

[0068] S51: Calculate the semantic segmentation loss of the source domain samples based on the initial semantic segmentation result, the final semantic segmentation result, and the source domain semantic label image;

[0069] The semantic segmentation loss for source domain samples is as follows:

[0070]

[0071] in The segmentation loss representing the source domain, and The initial and final segmentation predictions represent the source domain samples.

[0072] S52: Calculate the semantic segmentation loss of the mixed samples based on the initial semantic segmentation results, the final semantic segmentation results, and the mixed pseudo-labels;

[0073] The semantic segmentation loss for mixed samples is as follows:

[0074]

[0075] in, These are pseudo-labels for mixed samples. and The initial and final segmentation predictions represent the mixed samples.

[0076] S53: Calculate the penalty weight based on the difficulty of adapting to the elevation task, and further calculate the weighted cross-entropy loss of the mixed samples.

[0077] The weighted cross-entropy loss for the mixed samples is as follows:

[0078]

[0079] Among them W M It is the highly adaptive weight of the mixed samples, X M It is a mixed sample. It is a pseudo-label for a mixed sample, P M This is the predicted probability map of the mixed sample, where W, H, and C are the probability maps P, H, and C, respectively. M The length, width, and number of channels, G S () and G T () represent elevation estimation classifiers for the source and target domains, respectively. M The true elevation value represents the mixed sample.

[0080] S6: Calculate the berHu loss from the elevation estimates and true elevation values ​​of the source domain samples and the mixed samples to obtain the elevation estimation loss of the source domain samples and the mixed samples, and calculate the overall elevation estimation loss;

[0081] It should be noted that step S6 specifically includes:

[0082] Obtain the elevation estimates of the source domain samples and the mixed samples. Calculate the berHu loss using the elevation estimates and true elevation values ​​of the source domain samples and the mixed samples to obtain the elevation estimation loss of the source domain samples and the mixed samples. Summate the elevation estimation losses of all source domain samples and the mixed samples to obtain the overall elevation estimation loss.

[0083] S7: Weight the entropy map and the mean entropy of each class into the prediction results of source domain samples and mixed samples to promote semantic alignment of data distribution between the source domain and the target domain; use the source domain ground truth and mixed pseudo-labels as supervision information, and calculate the entropy loss with the semantic segmentation results weighted by the entropy value;

[0084] Step S7 is as follows:

[0085] S71: Obtain the prediction results of source domain samples and mixed samples; weight the prediction results of source domain samples and mixed samples with the calculated entropy map and the mean entropy of each class.

[0086] As one example, the weighted calculation process is as follows:

[0087]

[0088] Where P represents the semantic segmentation prediction probability map, W, H, and C are the length, width, and number of channels of the probability map P, respectively, and Y... S Represents the tag value.

[0089] S72: Using the ground truth labels of the source domain and the pseudo labels of the mixed samples as supervision information, compare them with the semantic segmentation results after entropy weighting, and calculate the entropy loss.

[0090] It should be noted that the calculation process for step S72 is as follows:

[0091]

[0092] Where Y S and These are the semantic labels of the source domain samples and the pseudo-labels of the mixed samples, respectively. S and P M These represent semantic segmentation prediction probability maps for source domain samples and mixed samples, respectively.

[0093] S8: Train the model based on 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.

[0094] It should be noted that step S8 specifically includes:

[0095] S81: Based on the overall loss, update the semantic segmentation model using the loss function, and train the model until convergence; the overall loss is as follows:

[0096]

[0097] in and ω represents the Berhu loss for the source domain samples and the mixed samples, respectively. ele and ω ent These are the weighted values ​​of the loss.

[0098] S82: Obtain the trained student model and remove the domain adaptation module;

[0099] S83: Input the remote sensing image of the target domain into the trained model, and obtain the semantic segmentation prediction result of the target domain through the final classifier to achieve accurate land cover classification.

[0100] As an example, this invention provides five comparative examples for comparison with this invention. These include the SourceOnly method and four state-of-the-art multi-task learning-based UDA methods: Deep Domain Aware Adaptive Network (DADA), Cross-Task Relation Network (CTRL), Mutual Information Domain Adaptive Network (MIDA), and Correlation Aware Domain Adaptive Network (CorDA). See Table 1 for reference.

[0101] Table 1 shows the quantization results using Potsdam IRRG as the source domain and Vaihingen IRRG as the target domain.

[0102]

[0103]

[0104] The evaluation metrics are F1 score and IoU, and the detailed formulas are as follows:

[0105]

[0106] Where TP represents true positives, FP represents false positives, and FN represents false negatives. Precision represents the proportion of correctly predicted positives out of all predicted positives, and recall represents the proportion of correctly predicted positives out of all true positives. F1_Score represents the weighted average of precision and recall. P represents the prediction mask, G represents the ground truth mask, and IoU represents the intersection-union ratio of the predicted and ground truth values.

[0107] Table 1 compares the experimental results of this invention with other domain-adaptive semantic segmentation methods based on multi-task learning on the Potsdam IRRG->VaihingenIRRG, evaluating the performance of the method under different geographical environments and large differences in the appearance of similar land features. Compared with Source Only, all domain-adaptive methods achieved higher accuracy. Overall, the average F1 score and mIoU of this invention are 35.22% and 33.76% higher than Source Only, respectively. Notably, this invention achieved the best accuracy in all land feature categories, significantly outperforming other state-of-the-art methods, with an average F1 score and mIoU 15.33% and 13.95% higher than the best state-of-the-art method "CorDA," respectively. Furthermore, compared with all methods, this invention significantly improves the prediction accuracy of clutter / background classes and reduces the confusion between land features with large inter-class differences and other land features. The results show that the method proposed in this invention is an effective combination of existing multi-task learning methods and domain-adaptive semantic segmentation tasks, verifying that the proposed method has stronger feature learning and land feature classification capabilities.

[0108] Please see Figure 4 , Figure 4 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a remote sensing unsupervised domain adaptive semantic segmentation device 401 with multi-task collaborative learning, a processor 402, and a storage medium 403.

[0109] A remote sensing unsupervised adaptive semantic segmentation device 401 based on multi-task collaborative learning: The remote sensing unsupervised adaptive semantic segmentation device 401 based on multi-task collaborative learning implements the remote sensing unsupervised adaptive semantic segmentation method based on multi-task collaborative learning.

[0110] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the remote sensing unsupervised domain adaptive semantic segmentation method of multi-task collaborative learning.

[0111] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the remote sensing unsupervised domain adaptive semantic segmentation method of multi-task collaborative learning.

[0112] The beneficial effects of this invention are as follows: It utilizes cross-task feature association learning, fully mining and optimizing cross-task feature representations through continuous cross-feature enhancement units, and closely associating elevation and semantic features to leverage implicit elevation information and improve the accuracy of cross-domain segmentation tasks. Furthermore, this invention optimizes the quality of pseudo-labels, further using pseudo-label-guided hybrid consistency learning and entropy-guided category-level alignment to narrow the domain gap between the source and target domains, promoting good transfer of source domain knowledge and significantly enhancing the model's cross-domain capability. This invention greatly improves the segmentation accuracy of challenging scenes in remote sensing datasets, such as shaded regions, boundary regions, and similar spectral regions with significant height differences.

[0113] 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 principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-task co-learning remote sensing unsupervised domain adaptation semantic segmentation, characterized in that: The method comprises the following steps: S1: Obtain remote sensing image datasets of source and target domains, crop the remote sensing image datasets of the source and target domains, and obtain an elevation image, a remote sensing image, and a semantic label image; S2: Construct a semantic segmentation model, which comprises a student model and a teacher model; both the student model and the teacher model comprise an initial feature extraction module, a cross-task feature correlation learning module, and a final classifier; S3: Obtain remote sensing images and elevation images of mixed samples; input the target domain remote sensing image into the teacher model to obtain the label of the target domain sample and the pseudo label of the mixed sample; input the source domain remote sensing image and the enhanced mixed remote sensing image into the student model to obtain the final elevation estimation result of the source domain sample and the final elevation estimation result of the mixed sample, the final semantic segmentation result of the source domain sample and the final semantic segmentation result of the mixed sample; S4: Construct a domain adaptive module, which comprises a pseudo label guided mixed consistency learning module and an entropy guided class level alignment module; the pseudo label guided mixed consistency learning module is used to supervise the learning of the student model on the mixed sample by using the generated pseudo label; the entropy guided class level alignment module is used to guide the fine-grained class level alignment by weighting the entropy graph and the entropy mean value of each class to the segmentation result; S5: Calculate the semantic segmentation loss of the source domain sample according to the segmentation result of the source domain sample and the semantic label image; use the mixed pseudo label as the supervision information, calculate the semantic segmentation loss and the weighted cross-entropy loss of the mixed sample by using the semantic segmentation result of the mixed sample, and calculate the overall semantic segmentation loss; S6: Calculate the berHu loss by using the elevation estimation value and the elevation true value of the source domain sample and the mixed sample to obtain the elevation estimation loss of the source domain sample and the mixed sample, and calculate the overall elevation estimation loss; S7: Weight the entropy graph and the entropy mean value of each class to the prediction result of the source domain sample and the mixed sample to promote the semantic alignment of the data distribution of the source domain and the target domain; use the source domain true value and the mixed pseudo label as the supervision information, and calculate the entropy loss by using the entropy value weighted semantic segmentation result; S8: Train the model according to the overall loss, update the model parameters, and input the remote sensing image dataset of the target domain into the trained student model to obtain the final semantic segmentation result; Step S3 specifically comprises: S31: Use the ClassMix method to generate a mixed mask, and select the corresponding regions of the source domain and target domain samples according to the mask to generate the remote sensing image of the mixed sample; S32: input the target domain remote sensing image into the teacher model to generate the pseudo label; combine the mixed mask, the target domain pseudo label, and the source domain label true value to generate the pseudo label of the mixed sample; S33: perform data enhancement on the mixed sample, including random horizontal flipping, vertical flipping, color jittering, and Gaussian noise; S34: input the remote sensing images of the source and target domains into the initial feature extraction module to generate initial features through a parameter sharing encoder, and generate initial semantic segmentation and elevation estimation prediction results by using an elevation estimation bottleneck, a semantic segmentation bottleneck, and an initial semantic classifier; S35: The initial features are input into the cross-task feature association learning module, and adaptive fusion features of final semantics and height are generated through the cross-feature enhancement unit and the task-specific gating structure; S36: The adaptive fusion features are input into the final classifier to generate the final semantic segmentation and height estimation prediction results. 2.The multi-task co-learning remote sensing unsupervised domain adaptation semantic segmentation method of claim 1, wherein: In step S2: The initial feature extraction module includes a parameter-shared encoder, a height estimation bottleneck, a semantic segmentation bottleneck, an initial height classifier, and an initial semantic classifier; the initial feature extraction module is configured to extract initial features and generate initial semantic segmentation results and initial height estimation prediction results; The cross-task feature association learning module is composed of two consecutive cross-feature enhancement units and two task-specific gating structures; the cross-task feature association learning module is configured to generate adaptive fusion features according to the initial features; The final classifier is configured to generate final semantic segmentation and height estimation prediction results according to the adaptive fusion features. 3.The multi-task co-learning remote sensing unsupervised domain adaptation semantic segmentation method of claim 1, wherein: Step S4 is specifically: S41: A pseudo-label-guided hybrid consistency learning module is constructed to balance the feature representation capability of the model and the quality of the pseudo labels; the pseudo-label-guided hybrid consistency learning module adopts an incremental pseudo-label screening strategy, dynamically adjusts the current threshold according to the given minimum threshold, maximum threshold, total number of iterations, and current number of iterations, and screens the pseudo labels that meet the conditions; S42: An entropy-guided class-level alignment module is constructed; An entropy map is calculated according to the final semantic segmentation prediction results and the labels, and the average entropy value of each class is calculated using the entropy map, followed by normalization processing. 4.The multi-task co-learning remote sensing unsupervised domain adaptation semantic segmentation method of claim 1, wherein: Step S5 is specifically: S51: The semantic segmentation loss of the source domain sample is calculated according to the initial semantic segmentation result, the final semantic segmentation result, and the source domain semantic label image of the source domain sample; S52: The semantic segmentation loss of the mixed sample is calculated according to the initial semantic segmentation result, the final semantic segmentation result, and the mixed pseudo label of the mixed sample; S53: The penalty weight is calculated according to the adaptive difficulty of the height task, and the weighted cross-entropy loss of the mixed sample is further calculated.

5. The multi-task co-learning remote sensing unsupervised domain adaptation semantic segmentation method of claim 1, wherein: Step S6 is specifically: The height estimation values of the source domain samples and the mixed samples are obtained, the berHu loss is calculated using the height estimation values and the height true values of the source domain samples and the mixed samples, the height estimation loss of the source domain samples and the mixed samples is obtained, and the total height estimation loss is obtained by summing the height estimation loss of all source domain samples and mixed samples.

6. The multi-task co-learning remote sensing unsupervised domain adaptation semantic segmentation method of claim 1, wherein: Step S7 is specifically: S71: The prediction results of the source domain samples and the mixed samples are obtained; the prediction results of the source domain samples and the mixed samples are weighted with the calculated entropy map and the average entropy value of each class; S72: The true value labels of the source domain and the pseudo labels of the mixed samples are used as supervision information, and are compared with the semantic segmentation results weighted by the entropy value to calculate the entropy loss.

7. The method of claim 1, wherein the method further comprises: Step S8 is specifically: S81: The semantic segmentation model is updated using the loss function according to the total loss, and the model is trained until convergence; S82: The trained student model is obtained, and the domain adaptation module is removed; S83: The target domain remote sensing image is input into the trained model, and the semantic segmentation prediction result of the target domain is obtained through the final classifier to realize accurate feature classification.

8. A storage medium characterized by: The storage medium stores instructions and data for implementing the multi-task cooperative learning remote sensing unsupervised domain adaptive semantic segmentation method in any one of claims 1-7.

9. A multi-task co-learning remote sensing unsupervised domain adaptation semantic segmentation device, characterized by: Comprise: A processor and a storage medium; the processor loads and executes the instructions and data in the storage medium for implementing the multi-task cooperative learning remote sensing unsupervised domain adaptive semantic segmentation method in any one of claims 1-7.

Images