A remote sensing image cross-domain migration classification method for a label-free target region

Abstract

The application relates to a remote sensing image cross-domain migration classification method for a label-free target area, comprising the following steps: acquiring multi-source remote sensing data covering a region; constructing an adversarial domain self-adaptive network; calculating a classification loss and a domain discrimination loss; constructing an adversarial training total loss function, and reversely propagating network parameters after dynamically adjusting a weight coefficient to update the network parameters; inputting target domain images into a trained feature extractor and a classifier, and finally outputting a refined classification result map. The application has the beneficial effects that the application enhances the multi-dimensional representation capability of a model for a complex geographical environment, overcomes the limitation of a single data source, and provides a more robust feature basis for cross-domain migration classification.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image classification and extraction technology, and more specifically, to a method for cross-domain migration classification of remote sensing images for unlabeled target areas. Background Technology

[0002] Remote sensing image classification, as a crucial technology for geospatial information extraction, is widely applied in various fields such as land use monitoring, resource surveys, environmental assessments, disaster response, urban planning, and agricultural monitoring. With the rapid development of remote sensing Earth observation technology, multi-source remote sensing image data (including hyperspectral, multispectral, SAR, and LiDAR data) has grown exponentially, providing a rich data foundation for land cover identification and classification at regional scales. Hyperspectral images capture the spectral characteristics of land cover through multiple narrow bands, enabling precise differentiation of various land cover types and are widely used in vegetation monitoring, mineral exploration, and soil type identification. Multispectral images, typically acquired by satellite or airborne platforms, cover a wide spectral band and effectively distinguish land cover categories through spectral information from different bands, making them widely used in agricultural monitoring, vegetation health assessment, and urban expansion monitoring. SAR images are all-weather and all-time compatible, capable of penetrating clouds and fog, making them particularly suitable for post-disaster monitoring, soil moisture analysis, and forest resource monitoring. LiDAR (Light Detection and Ranging) data acquires high-precision 3D ground information through laser scanning technology, accurately depicting surface elevation, topographic features, and vegetation structure. It is widely used in high-precision terrain modeling, forest structure analysis, and urban modeling. The fusion of different types of remote sensing imagery can significantly improve the accuracy and reliability of land cover classification, especially in complex geographical environments and situations with varying land cover, providing more comprehensive land cover identification capabilities. By effectively fusing data from different sensors, the complementary advantages of various remote sensing data can be fully utilized, improving classification accuracy and supporting large-scale, multi-dimensional geographic information analysis, thus promoting widespread applications in environmental protection, resource management, urban development, and disaster response.

[0003] However, in practical applications, remote sensing image classification methods still face many technical challenges, mainly reflected in the following aspects: (1) Insufficient model generalization ability: When the classification model trained on the labeled source domain (i.e. the region where the training samples are located) is directly transferred to the unlabeled target domain (i.e. the region to be predicted), due to the combined effects of sensor parameter differences, imaging condition changes, geographical environment diversity, seasonal fluctuations and other factors, a significant "domain shift" phenomenon often occurs, resulting in a significant decrease in classification accuracy. The cross-domain transfer ability of the model needs to be improved. (2) High labeling cost: For classification tasks in new regions, a large number of manually labeled samples are usually required for model fine-tuning or retraining. This process is not only time-consuming and laborious, but also, with the continuous growth of the number of regions and the scale of data, the labeling workload increases exponentially, which seriously restricts the large-scale promotion and application of the technology. (3) Limited feature expression ability: Traditional classification methods mostly rely on manually designed features, which have bottlenecks in extracting deep features with cross-domain invariance and semantic discrimination ability. Especially in regions with complex land boundary and similar spectral features, problems such as classification ambiguity and boundary mixing are easy to occur, affecting the final classification accuracy and application effect.

[0004] To address the cross-domain problem in remote sensing image classification, Domain Adaptation (DA) technology has been introduced and has gradually developed into a core technology for solving this challenge. This technology improves the model's generalization ability in the target domain by reducing the feature distribution differences between the source domain (i.e., the area containing labeled data) and the target domain (i.e., the area with no or little labeled data). This reduces the cost of manual annotation while ensuring high-accuracy cross-region classification. With the rapid advancement of deep learning technology, domain adaptation methods based on Generative Adversarial Networks (GANs) and Domain Adversarial Neural Networks (DANNs) have been widely applied to remote sensing image classification. GANs generate pseudo-samples consistent with the style of the target domain through adversarial training, thereby enhancing the model's adaptability; while DANNs directly align the feature distributions of the source and target domains by introducing a domain adversarial loss function. Furthermore, multimodal data fusion has become an important direction for the development of domain adaptation technology. By fusing different types of remote sensing data (such as high-resolution visible light imagery, multispectral imagery, SAR imagery, etc.), more comprehensive ground feature information can be obtained, significantly improving the model's representation ability in complex geographical environments and thus enhancing its generalization performance.

[0005] While these domain-adaptive methods have improved the accuracy of cross-domain classification of remote sensing images to some extent, they still face several challenges. In complex geographical environments, especially in cross-regional applications, the differences in land cover types, spatial layouts, and spectral characteristics across different regions are significant, leading to unstable domain alignment results. Existing methods mainly focus on global feature alignment, neglecting fine-grained alignment for specific categories (such as vegetation, water bodies, and bare land), resulting in insufficient accuracy in fine-grained classification tasks and failing to meet the demands of high-precision classification. Furthermore, how to effectively utilize information from unlabeled data in the target domain and gradually improve model performance through pseudo-labels and iterative optimization strategies remains a hot topic and a challenge in current research. Therefore, how to achieve high-precision cross-regional and cross-scene remote sensing image classification in complex geographical environments, reduce dependence on labeled data in the target domain, and improve the robustness and generalization ability of the model across different regions remains a key technical problem that urgently needs to be solved in the field of remote sensing image classification. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for cross-domain migration classification of remote sensing images for unlabeled target areas.

[0007] Firstly, a method for cross-domain migration classification of remote sensing images for unlabeled target areas is provided, including:

[0008] S1. Acquire multi-source remote sensing data of the coverage area, perform preprocessing on each data source, and then spatially register all preprocessed images to complete feature-level fusion of multi-source data.

[0009] S2. Construct an adversarial domain adaptive network, which includes a feature extractor, a classifier, and a domain discriminator with shared weights, and insert a gradient inversion layer between the feature extractor and the domain discriminator;

[0010] S3. Simultaneously input the images of the source domain and the target domain into the feature extractor to extract depth features, and use the classifier to classify the source domain features to obtain the classification loss; use the domain discriminator to perform domain classification calculation on the source domain features and the target domain features to obtain the domain discrimination loss;

[0011] S4. Combine the classification loss and the domain discrimination loss to construct the total loss function for adversarial training. After dynamically adjusting the weight coefficients, backpropagate to update the network parameters so that the feature extractor learns domain-invariant features.

[0012] S5. Input the target domain image into the trained feature extractor and classifier, and iteratively optimize the classification results by introducing confidence learning to correct the false labels with low confidence, and finally output a refined classification result map.

[0013] Preferably, in S1, the multi-source remote sensing data includes at least two of hyperspectral images, multispectral images, and SAR data; the preprocessing includes orthorectification, radiometric calibration, and atmospheric correction; the feature-level fusion is used to form fused feature data containing multi-source information, and to construct a fused feature sample set with labels in the source domain and a fused feature sample set without labels in the target domain.

[0014] Preferably, in S2, the feature extractor uses a convolutional neural network, the classifier consists of a fully connected layer and a softmax activation function, and the domain discriminator consists of a fully connected layer and a sigmoid activation function; the gradient reversal layer implements an identity mapping during forward propagation and reverses the gradient direction during backward propagation.

[0015] Preferably, in S3, the classification loss is calculated using the cross-entropy function, and the domain discrimination loss is calculated using the binary cross-entropy function.

[0016] Preferably, in S4, the mathematical form of the total loss function for adversarial training is: the sum of classification loss and dynamic weight coefficient multiplied by domain discrimination loss; the dynamic weight coefficient increases non-linearly with the training progress.

[0017] As a preferred embodiment, S5 includes:

[0018] Input the target domain image into the trained feature extractor and classifier to generate pseudo-labels and their confidence scores;

[0019] Set a confidence threshold and select high-confidence samples to construct a self-training loss;

[0020] The self-training loss is combined with the classification loss and the domain discrimination loss to construct the overall loss function;

[0021] By iteratively updating network parameters through backpropagation, and repeatedly executing the processes of pseudo-label generation, confidence filtering, and parameter updating, the classification results of the target domain are gradually corrected.

[0022] Secondly, a remote sensing image cross-domain migration classification system for unlabeled target areas is provided for performing any of the methods described in the first aspect, including:

[0023] The acquisition module is used to acquire multi-source remote sensing data of the coverage area, perform preprocessing on each data source, and then spatially register all the preprocessed images to complete the feature-level fusion of the multi-source data.

[0024] The first construction module is used to construct an adversarial domain adaptive network, which includes a feature extractor, a classifier, and a domain discriminator with shared weights, and inserts a gradient inversion layer between the feature extractor and the domain discriminator.

[0025] The first input module is used to simultaneously input images of the source domain and the target domain into the feature extractor to extract depth features, and to perform classification calculation on the source domain features through the classifier to obtain classification loss; and to perform domain classification calculation on the source domain features and the target domain features through the domain discriminator to obtain domain discrimination loss.

[0026] The second construction module is used to construct an adversarial training total loss function by combining the classification loss and the domain discrimination loss, dynamically adjust the weight coefficients and then backpropagate to update the network parameters, so that the feature extractor learns domain-invariant features.

[0027] The second input module is used to input the target domain image into the trained feature extractor and classifier. By introducing confidence learning, the classification results are iteratively optimized to correct low-confidence pseudo-labels, and finally output a refined classification result image.

[0028] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.

[0029] Fourthly, an electronic device is provided, comprising:

[0030] Memory, used to store computer programs;

[0031] A processor for executing the computer program to implement the method as described in any of the first aspects.

[0032] The beneficial effects of this invention are:

[0033] 1. This invention introduces a feature-level fusion strategy for multi-source remote sensing data, fully leveraging the complementary advantages of different data sources at the spectral, spatial, and semantic levels to construct a fusion feature set rich in geographic semantic information. This strategy enhances the model's ability to represent complex geographic environments in multiple dimensions, overcomes the limitations of single data sources, provides a more robust feature foundation for cross-domain transfer classification, and effectively improves the model's generalization ability across different regions.

[0034] 2. This invention introduces a confidence learning mechanism into the pseudo-label optimization process. Through confidence quantification and dynamic threshold screening, high-confidence samples are accurately selected for iterative training, effectively suppressing classification bias caused by low-quality pseudo-labels. This mechanism gradually corrects the classification boundary during iteration, achieving a progressive improvement from coarse classification to refined feature recognition. This solves the error accumulation problem in traditional self-training, ensuring the convergence stability of the iterative process and the continuous improvement of classification accuracy.

[0035] 3. This invention constructs a joint learning framework that combines adversarial training and confidence learning for collaborative optimization. Based on domain-adaptive extracted domain-invariant features, it utilizes high-confidence pseudo-labels to introduce target domain supervision signals, achieving multi-objective joint optimization of classification loss, domain discrimination loss, and self-training loss. This framework achieves high-precision classification through knowledge transfer and self-supervised iteration even when the target domain is completely unlabeled, significantly reducing manual annotation costs and promoting an efficient application model of "one-time labeled training, multiple unlabeled reuses." Attached Figure Description

[0036] Figure 1 A flowchart of a remote sensing image cross-domain migration classification method for unlabeled target areas provided by the present invention;

[0037] Figure 2 The flowchart for multi-source remote sensing data preprocessing and registration in the Hangzhou Bay area provided by this invention;

[0038] Figure 3 This is a schematic diagram of the Hangzhou Bay-Yancheng domain adaptation dataset provided by the present invention. Detailed Implementation

[0039] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0040] Example 1:

[0041] To address the problems of existing technologies, Embodiment 1 of this application provides a method for cross-domain migration classification of remote sensing images for unlabeled target areas, including:

[0042] S1. Acquire multi-source remote sensing data of the coverage area, perform preprocessing on each data source, and then spatially register all preprocessed images to complete feature-level fusion of the multi-source data.

[0043] In S1, the multi-source remote sensing data includes at least two of the following: hyperspectral images, multispectral images, and SAR data; the preprocessing includes orthorectification, radiometric calibration, and atmospheric correction; the feature-level fusion is used to form fused feature data containing multi-source information, and to construct a fused feature sample set with labels in the source domain and a fused feature sample set without labels in the target domain.

[0044] Specifically, in S1, the feature-level fusion adopts the following strategy:

[0045] First, using the spatial resolution of the multispectral image as a benchmark, bicubic interpolation is used to resample the hyperspectral image and SAR data to this benchmark resolution, ensuring that all images are perfectly aligned on the spatial grid. Then, for each pixel, the reflectance values ​​of all effective bands are extracted from the hyperspectral image to form a complete image. 3D spectral eigenvectors The reflectance values ​​of each discrete band are extracted from the multispectral image to form... 3D feature vector The backscattering coefficients are extracted from SAR images after adaptive Lee filtering, and an N-dimensional feature vector is constructed based on four texture features: contrast, correlation, energy, and homogeneity, derived from the gray-level co-occurrence matrix. Next, the three elements are concatenated end-to-end to form the original fused feature vector. Finally, the mean value of each feature channel is calculated using the labeled data from the source domain. with standard deviation Z-score normalization is performed on the original fusion features of all source and target domain samples. ,in To prevent division by zero constants, the standardized fused feature data are arranged into a three-dimensional tensor according to their spatial location. ,in For image height and width, The number of channels is used to construct a fusion feature sample set with labels in the source domain and a fusion feature sample set with no labels in the target domain.

[0046] S2. Construct an adversarial domain adaptive network, which includes a feature extractor, a classifier, and a domain discriminator with shared weights, and insert a gradient inversion layer between the feature extractor and the domain discriminator.

[0047] In S2, the feature extractor uses a convolutional neural network, the classifier consists of a fully connected layer and a softmax activation function, and the domain discriminator consists of a fully connected layer and a sigmoid activation function; the gradient reversal layer implements an identity mapping during forward propagation and reverses the gradient direction during backward propagation.

[0048] Specifically, in S2, the feature extractor uses a ResNet-18 network, and its input is a multi-source fusion feature tensor. Firstly, through a The convolutional layer reduces the number of channels to 3, resulting in Then The data is fed into the ResNet-18 backbone network (containing multiple residual blocks and a total of 18 convolutional layers), and after global average pooling, it outputs 512-dimensional deep features. Its mathematical expression is:

[0049]

[0050] in, The first layer corresponds to Convolution dimensionality reduction, output ; Level 2 to Level 1 =19 layers correspond to the 18 convolutional layers inside ResNet-18 (including convolutions in residual connections), that is for =2,…,19, final features It is obtained by global average pooling from the output of the last layer. and The first Layer weights and biases Represents the ReLU activation function. This represents a convolution operation. The classifier consists of a fully connected layer and a softmax activation function, and its mathematical expression is:

[0051]

[0052] in, and These are the parameters for the fully connected layer. Represents matrix multiplication. To predict probabilities for each category, a gradient reversal layer (GRL) is inserted between the feature extractor and the domain discriminator. This GRL layer performs an identity mapping during forward propagation and reverses the gradient direction during backward propagation. The domain discriminator consists of a fully connected layer and a sigmoid activation function, using the GRL-processed features... For input, the mathematical expression is:

[0053]

[0054] in, and These are the parameters for the fully connected layer. This represents the probability that a sample belongs to the target domain.

[0055] S3. Simultaneously input the images of the source domain and the target domain into the feature extractor to extract depth features, and use the classifier to classify the source domain features to obtain the classification loss; use the domain discriminator to perform domain classification calculation on the source domain features and the target domain features to obtain the domain discrimination loss.

[0056] Specifically, in S3, the classification loss is calculated using the cross-entropy function:

[0057]

[0058] in, The number of samples in the source domain. For the number of categories, For real labels, To predict probabilities, the domain discrimination loss is calculated using a binary cross-entropy function:

[0059]

[0060] in, The number of samples in the target domain. For the domain label (source domain is 0, target domain is 1). Predict the probability for the domain.

[0061] S4. Combine the classification loss and the domain discrimination loss to construct the total loss function for adversarial training. After dynamically adjusting the weight coefficients, backpropagate to update the network parameters so that the feature extractor learns domain-invariant features.

[0062] In S4, the mathematical expression for the total loss function of adversarial training is:

[0063]

[0064] in, The weights are inverted through a gradient reversal layer, which minimizes the gradient of the feature extractor. Maximize at the same time This allows them to learn domain-invariant features. The weight coefficients... Adopt a dynamic adjustment strategy:

[0065]

[0066] in, Training progress (current iteration number / total iteration number), Hyperparameters for controlling the growth rate.

[0067] S5. Input the target domain image into the trained feature extractor and classifier, and iteratively optimize the classification results by introducing confidence learning to correct the false labels with low confidence, and finally output a refined classification result map.

[0068] S5 includes:

[0069] Input the target domain image into the trained feature extractor and classifier to generate pseudo-labels and their confidence scores;

[0070] Set a confidence threshold and select high-confidence samples to construct a self-training loss;

[0071] The self-training loss is combined with the classification loss and the domain discrimination loss to construct the overall loss function;

[0072] By iteratively updating network parameters through backpropagation, and repeatedly executing the processes of pseudo-label generation, confidence filtering, and parameter updating, the classification results of the target domain are gradually corrected.

[0073] Specifically, in S5, the target domain image patch is input into the trained feature extractor. and classifier The classification probability distribution is obtained through forward propagation. , Generate pseudo tags Its confidence level is Set a confidence threshold. A linear heating strategy is adopted: an initial threshold is set. final threshold The total number of iterations is (For example =20), in the Threshold selection in round iteration Filtering high-confidence sample sets Construct a self-training loss function:

[0074]

[0075] During the iterative optimization phase, the self-training loss and adversarial training loss are combined to construct the overall loss function:

[0076]

[0077] in These are the weight coefficients for the self-training loss. The network parameters are updated iteratively through backpropagation, repeatedly executing the "pseudo-label generation-confidence screening-parameter update" process. While maintaining domain adaptability, the target domain classification results are gradually corrected using high-confidence pseudo-labels.

[0078] Example 2:

[0079] Based on Example 1, Example 2 of this application provides a more specific method for cross-domain migration classification of remote sensing images for unlabeled target areas, including:

[0080] S1. Acquire multi-source remote sensing data of the coverage area, perform preprocessing on each data source, and then spatially register all preprocessed images to complete feature-level fusion of the multi-source data.

[0081] For example, in S1, ZY1-02D ​​hyperspectral imagery, Sentinel-2 multispectral imagery, and Sentinel-1 SAR data are acquired from the source domain (labeled areas) and the target domain (unlabeled areas). These images undergo preprocessing operations such as orthorectification, radiometric calibration, and atmospheric correction to ensure accurate geolocation and eliminate spatial distortions caused by different sensors and observation conditions. Next, all preprocessed images are spatially registered to ensure complete alignment of spatial coordinates and resolution between different data sources, further improving spatial consistency. Finally, a multi-source data feature fusion strategy is used to fuse the registered multi-source images, forming fused feature data containing hyperspectral, multispectral, and SAR information. Labeled fused feature sample sets from the source domain and unlabeled fused feature sample sets from the target domain are constructed to provide high-quality input for subsequent classification and analysis.

[0082] S2. Construct an adversarial domain adaptive network, which includes a feature extractor, a classifier, and a domain discriminator with shared weights, and insert a gradient inversion layer between the feature extractor and the domain discriminator.

[0083] In this step, an adversarial domain adaptation network framework is designed and constructed, utilizing a shared feature extractor, classifier, and domain discriminator to jointly optimize the feature representations of the source and target domains. The feature extractor extracts shared deep features from the source and target domain images, the classifier performs classification prediction of source domain features, and the domain discriminator distinguishes the feature differences between the source and target domains. By inserting a gradient inversion layer between the feature extractor and the domain discriminator, the network is prompted to perform adversarial optimization during training, thereby learning feature representations that are effective for both the source and target domains, providing strong structural support for subsequent domain adaptation.

[0084] S3. Simultaneously input the images of the source domain and the target domain into the feature extractor to extract depth features, and use the classifier to classify the source domain features to obtain the classification loss; use the domain discriminator to perform domain classification calculation on the source domain features and the target domain features to obtain the domain discrimination loss.

[0085] The main goal of this step is to train the feature extractor, classifier, and domain discriminator by simultaneously inputting images from both the source and target domains into the feature extractor, thereby jointly optimizing source domain classification and inter-domain adversarial processing. The classifier calculates the classification loss for the source domain features to ensure that the features in the source domain are separable by category; simultaneously, the domain discriminator calculates the domain discrimination loss to measure the distributional differences between the source and target domain features. This process enables the network to recognize and adversarially process feature differences between the source and target domains, laying the foundation for subsequent learning of domain-invariant features.

[0086] S4. Combine the classification loss and the domain discrimination loss to construct the total loss function for adversarial training. After dynamically adjusting the weight coefficients, backpropagate to update the network parameters so that the feature extractor learns domain-invariant features.

[0087] The purpose of this step is to combine the classification loss and domain discriminant loss of the source domain to construct a total loss function for adversarial training. By dynamically adjusting the weight coefficients of the classification loss and domain discriminant loss during training, we ensure that the feature extractor minimizes the classification loss while maximizing the domain discriminant loss through backpropagation. This forces the feature extractor to learn domain-invariant features that are robust to both the source and target domains. This process effectively reduces the feature distribution differences between the source and target domains, improving the model's cross-domain adaptability and generalization ability.

[0088] S5. Input the target domain image into the trained feature extractor and classifier, and iteratively optimize the classification results by introducing confidence learning to correct the false labels with low confidence, and finally output a refined classification result map.

[0089] The purpose of this step is to introduce a confidence learning mechanism after training to further optimize the classification results of the target domain. By performing forward propagation on the target domain image, initial pseudo-labels are generated and their confidence scores are calculated. High-confidence samples are then selected for further optimization. By combining the self-training loss with the original classification loss and domain discrimination loss, iterative correction of classification results for low-confidence regions is achieved, suppressing the accumulation of errors and ultimately outputting refined, high-precision classification results. This process effectively improves the classification accuracy of unlabeled target domain data, ensuring efficient and accurate classification of target domain images even without manual annotation.

[0090] These steps effectively achieve feature alignment and knowledge transfer between the source and target domains through adversarial training mechanisms, enabling high-precision classification results even when labeled data in the target domain is scarce, significantly improving the model's generalization ability and cross-domain application performance.

[0091] Finally, the classification results of the proposed method were compared with those of four other mainstream methods using a self-constructed multi-source remote sensing dataset (including ZY1-02D ​​hyperspectral imagery, Sentinel-2 multispectral imagery, and Sentinel-1 SAR data). Specific results are shown in Table 1. The results show that on the Hangzhou Bay-Yancheng domain adaptation dataset, the overall accuracy (OA) of the proposed method is 92.12%, and the Kappa coefficient is 0.8171. These results demonstrate that the proposed method significantly outperforms other comparative methods on this domain adaptation dataset, exhibiting cross-domain migration adaptation capabilities for unlabeled target areas, further validating its effectiveness in cross-domain migration classification methods for remote sensing images targeting unlabeled target areas.

[0092] Table 1. Experimental results of Hangzhou Bay-Yancheng domain adaptation data (January 3, 2022 → February 24, 2022) on different domain adaptation methods.

[0093]

[0094] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0095] Example 3:

[0096] Based on Example 2, Example 3 of this application provides a remote sensing image cross-domain migration classification system for unlabeled target areas, including:

[0097] The acquisition module is used to acquire multi-source remote sensing data of the coverage area, perform preprocessing on each data source, and then spatially register all the preprocessed images to complete the feature-level fusion of the multi-source data.

[0098] The first construction module is used to construct an adversarial domain adaptive network, which includes a feature extractor, a classifier, and a domain discriminator with shared weights, and inserts a gradient inversion layer between the feature extractor and the domain discriminator.

[0099] The first input module is used to simultaneously input images of the source domain and the target domain into the feature extractor to extract depth features, and to perform classification calculation on the source domain features through the classifier to obtain classification loss; and to perform domain classification calculation on the source domain features and the target domain features through the domain discriminator to obtain domain discrimination loss.

[0100] The second construction module is used to construct an adversarial training total loss function by combining the classification loss and the domain discrimination loss, dynamically adjust the weight coefficients and then backpropagate to update the network parameters, so that the feature extractor learns domain-invariant features.

[0101] The second input module is used to input the target domain image into the trained feature extractor and classifier. By introducing confidence learning, the classification results are iteratively optimized to correct low-confidence pseudo-labels, and finally output a refined classification result image.

[0102] It should be noted that the system provided in this embodiment is the system corresponding to the method provided in embodiment 2. Therefore, the parts in this embodiment that are the same as or similar to those in embodiment 2 can be referred to each other, and will not be described again in this application.

Claims

1. A method for cross-domain migration classification of remote sensing images for unlabeled target areas, characterized in that, include: S1. Acquire multi-source remote sensing data of the coverage area, perform preprocessing on each data source, and then spatially register all preprocessed images to complete feature-level fusion of multi-source data. S2. Construct an adversarial domain adaptive network, which includes a feature extractor, a classifier, and a domain discriminator with shared weights, and insert a gradient inversion layer between the feature extractor and the domain discriminator; S3. Simultaneously input the images of the source domain and the target domain into the feature extractor to extract depth features, and use the classifier to classify the source domain features to obtain the classification loss; use the domain discriminator to perform domain classification calculation on the source domain features and the target domain features to obtain the domain discrimination loss; S4. Combine the classification loss and the domain discrimination loss to construct the total loss function for adversarial training. After dynamically adjusting the weight coefficients, backpropagate to update the network parameters so that the feature extractor learns domain-invariant features. S5. Input the target domain image into the trained feature extractor and classifier, and iteratively optimize the classification results by introducing confidence learning to correct the false labels with low confidence, and finally output a refined classification result map.

2. The remote sensing image cross-domain migration classification method for unlabeled target areas according to claim 1, characterized in that, In S1, the multi-source remote sensing data includes at least two of the following: hyperspectral images, multispectral images, and SAR data; the preprocessing includes orthorectification, radiometric calibration, and atmospheric correction; the feature-level fusion is used to form fused feature data containing multi-source information, and to construct a fused feature sample set with labels in the source domain and a fused feature sample set without labels in the target domain.

3. The remote sensing image cross-domain migration classification method for unlabeled target areas according to claim 2, characterized in that, In S2, the feature extractor uses a convolutional neural network, the classifier consists of a fully connected layer and a softmax activation function, and the domain discriminator consists of a fully connected layer and a sigmoid activation function; the gradient reversal layer implements an identity mapping during forward propagation and reverses the gradient direction during backward propagation.

4. The remote sensing image cross-domain migration classification method for unlabeled target areas according to claim 3, characterized in that, In S3, the classification loss is calculated using the cross-entropy function, and the domain discrimination loss is calculated using the binary cross-entropy function.

5. The remote sensing image cross-domain migration classification method for unlabeled target areas according to claim 4, characterized in that, In S4, the mathematical form of the total loss function for adversarial training is: the sum of classification loss and dynamic weight coefficient multiplied by domain discrimination loss; the dynamic weight coefficient increases non-linearly with the training progress.

6. The remote sensing image cross-domain migration classification method for unlabeled target areas according to claim 5, characterized in that, S5 include: Input the target domain image into the trained feature extractor and classifier to generate pseudo-labels and their confidence scores; Set a confidence threshold and select high-confidence samples to construct a self-training loss; The self-training loss is combined with the classification loss and the domain discrimination loss to construct the overall loss function; By iteratively updating network parameters through backpropagation, and repeatedly executing the processes of pseudo-label generation, confidence filtering, and parameter updating, the classification results of the target domain are gradually corrected.

7. A remote sensing image cross-domain migration classification system for unlabeled target areas, characterized in that, For performing the method according to any one of claims 1 to 6, comprising: The acquisition module is used to acquire multi-source remote sensing data of the coverage area, perform preprocessing on each data source, and then spatially register all the preprocessed images to complete the feature-level fusion of the multi-source data. The first construction module is used to construct an adversarial domain adaptive network, which includes a feature extractor, a classifier, and a domain discriminator with shared weights, and inserts a gradient inversion layer between the feature extractor and the domain discriminator. The first input module is used to simultaneously input images of the source domain and the target domain into the feature extractor to extract depth features, and to perform classification calculation on the source domain features through the classifier to obtain classification loss; and to perform domain classification calculation on the source domain features and the target domain features through the domain discriminator to obtain domain discrimination loss. The second construction module is used to construct an adversarial training total loss function by combining the classification loss and the domain discrimination loss, dynamically adjust the weight coefficients and then backpropagate to update the network parameters, so that the feature extractor learns domain-invariant features. The second input module is used to input the target domain image into the trained feature extractor and classifier. By introducing confidence learning, the classification results are iteratively optimized to correct low-confidence pseudo-labels, and finally output a refined classification result image.

8. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 6.

Images