Remote sensing image change detection method and system based on large model domain self-adaption

Abstract

The invention discloses a remote sensing image change detection method and system based on large model domain self-adaption. The method comprises the following steps: S1, inputting a pair of to-be-detected dual-time remote sensing images into a trained remote sensing image change detection network; and S2, outputting a change detection graph. The invention belongs to the technical field of remote sensing image change detection, and solves the problems of domain difference and boundary displacement when an SAM2 model is used in a change detection task at present. According to the method, the knowledge difference between the SAM2 and a change detection task is overcome through a hierarchical low-rank self-adaption strategy, and domain self-adaption from the SAM2 to remote sensing is realized by introducing a low-rank matrix into a key layer of the SAM2 and guiding a model to learn knowledge in the remote sensing field.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image change detection, and in particular relates to a remote sensing image change detection method and system based on large model domain adaptation. Background Art

[0002] Remote sensing image change detection aims to obtain the change information between two images taken at different times at the same geographical location. It plays an important role in many fields such as disaster monitoring, urban planning and environmental investigation, and has become an important research hotspot in remote sensing earth observation technology. Traditional change detection methods rely on manual feature extraction, which makes it difficult to achieve efficient and accurate change recognition in complex scenes.

[0003] Recently, visual base models have driven a paradigm shift in change detection by virtue of their task-independent advantages in different downstream tasks. Different from other models, visual base models have become increasingly attractive for change detection tasks due to their good generalization performance and few-shot learning capabilities. SegmentAnything Model 2 (SAM2) is a segmentation model recently developed by Meta. Its core idea is to predict the segmentation mask of an object based on user-provided cues. After high-quality training on large-scale natural datasets, SAM2 demonstrates strong zero-shot generalization and feature representation capabilities in various visual tasks. However, there are the following problems when using the SAM2 model for change detection tasks: (1) Domain difference: At the data level, remote sensing images differ significantly from natural images in spatial resolution, imaging conditions, and object scale. Remote sensing images usually have lower resolution and large spectral overlap between ground objects, which makes it difficult for the model to distinguish foreground from background. When directly applying SAM2 to remote sensing change detection tasks, the network often relies on its inherent knowledge in natural scenes, which limits its sensitivity to small changes in remote sensing scenes. (2) Boundary displacement: At the feature level, the SAM2 model has difficulty integrating the different spatial levels and semantic granularity features of bi-temporal images. Simple feature fusion methods (such as splicing) fail to fully exploit the temporal differences of bi-temporal features. Summary of the invention

[0004] The purpose of the present invention is to provide a remote sensing image change detection method and system based on large model domain adaptation, which solves the problems of domain difference and boundary displacement when the SAM2 model is used in the change detection task in the prior art.

[0005] The present application provides a technical solution: a remote sensing image change detection method based on large model domain adaptation, comprising the following steps: S1: Input a pair of dual-temporal remote sensing images to be detected into the trained remote sensing image change detection network; The change detection network consists of the following steps: S11: Hierarchical low-rank adaptive encoder, first introduces the low-rank matrix into the self-attention layer and linear layer of the SAM2 image encoder, and then extracts the multi-scale features of a pair of dual-temporal remote sensing images ; Extracting multi-scale features of a pair of dual-temporal remote sensing images through a hierarchical low-rank adaptive encoder , where i=1, 2, 3, 4.

[0006] S12: Difference adaptive enhancement module, first S11 dual-time remote sensing image multi-scale features , i=2,3,4 upsampled to The size of each group Adaptive fusion to generate 4 sets of difference features; through the difference adaptive enhancement module, the multi-scale features of a pair of dual-time remote sensing images in S11 are fused , i=1,2,3,4.

[0007] S13: Residual convolution decoder, decodes the four sets of difference features generated by S12 and outputs a change map; the output change map contains two channels, the first channel corresponds to the probability of the no-change class, and the second channel corresponds to the probability of the change class; the argmax operation is used to obtain a binary change map along the channel dimension.

[0008] S2: Output change detection map.

[0009] Preferably, the size of each set of difference features is same.

[0010] Preferably, the low-rank adaptive encoder in S11 specifically includes the following steps: The SAM2 image encoder is selected as the feature extractor for dual-temporal remote sensing images.

[0011] In the Transformer module of every two SAM2 image encoders, the low-rank adaptive strategy is introduced into its self-attention layer and multi-layer perceptron layer.

[0012] Preferably, the low-rank adaptive strategy includes: Given the weight matrix of a layer in the Transformer module , and Represents the weight matrix Dimensions; A branch road is added on one side. Decomposed two low-rank matrices , Composition, r represents the weight matrix and The rank and ; Freeze during training The weight of , Alternative Update of; given the input feature of this layer is , the output sequence , this process is expressed as: (1) (2) In formula 1 and formula 2, The weight matrix after introducing LoRA to this layer, Replacement during training Updated weight matrix.

[0013] Preferably, the S12 difference adaptive enhancement module includes the following steps: The difference adaptive enhancement module captures difference information through two branches: global difference perception and local detail optimization. The specific steps of the two branches are as follows: In the global difference perception branch, firstly, and Add the elements together to get , and then use the global average pooling operation to Each channel feature element is averaged. The size of is changed from [C, H, W] to [C, 1, 1], where C represents the number of feature channels, and H and W represent the height and width of the feature map, respectively. Global attention feature Calculated as: (3) (4) In formula 3, represents element-by-element addition. In Formula 4, GAP represents the global average pooling operation. represents a 1×1 convolution operation, BN represents a batch normalization operation, ; represents the global attention feature, Represents the input features of the global difference perception branch; In the local detail optimization branch, firstly, and Subtract the absolute value of the elements , the maximum pooling operation takes the maximum value of the local receptive field, and the features processed by each layer are Keep the same size, local attention features Calculated as: (5) (6) In formula 5 represents element-by-element subtraction, represents the absolute value operation, and MAP in Formula 6 represents the maximum pooling operation; ; Then, and The features are added and fed into the Sigmoid function to obtain the attention weight matrix and ; at last, and As a measure of global and local attention.

[0014] Preferably, the attention weight matrix and The specific calculation is as follows: (7) (8) The extracted spatial detail features and category discriminant features are multiplied by their relevant weights at the pixel level, and then weighted summed at the pixel level. The enhanced features of each layer Calculated by the following formula: (9)

[0015] In formula 9, represents the feature aggregation strategy with adaptive weights, represents element-wise multiplication; The samples containing labels are divided into training set, validation set and test set. At the beginning of training, the parameters of the SAM2 image encoder itself remain frozen and are not updated during the training process. Initialized to 0 and 1 respectively. Represented by The low-rank matrix obtained by decomposition is Represents the weight matrix of a layer in the transformer. And The rank value of is set to 16; the parameters of the difference adaptive enhancement module and the residual convolution decoder are randomly initialized; and the data enhancement strategy of random flipping and random rotation is adopted during training.

[0016] Preferably, the network optimization strategy is as follows: During the training process, the performance of the network is optimized by minimizing the cross entropy loss; for the binary change detection problem, the mathematical expression of the cross entropy loss function is: (10) In formula 10, represents the category label of the real sample, Represents the predicted category label, and N represents the total number of pixels of each sample; It is an indicator to measure the difference between the model prediction value and the true value. The smaller the value of the loss function, the closer the model prediction value is to the true value, and the better the performance of the model.

[0017] The present invention also provides another technical solution: A remote sensing image change detection system based on large model domain adaptation includes: an image input module, a network composition and processing module and an image output module.

[0018] Image input module: by inputting dual-time remote sensing images into the change detection network.

[0019] Network composition and processing modules: including: hierarchical low-rank adaptive encoder, difference adaptive enhancement module and residual convolution decoder.

[0020] The hierarchical low-rank adaptive encoder extracts multi-scale features (i = 1, 2, 3, 4) by introducing low-rank matrices into the self-attention and linear layers of the SAM2 image encoder. The difference adaptive enhancement module generates four sets of difference features by upsampling and adaptively fusing the multi-scale features (i = 2, 3, 4). The residual convolutional decoder obtains a change map containing two types of probabilities by decoding the difference features, and converts the change map into a binary change map using the argmax operation.

[0021] The image output module is used to output the change detection map.

[0022] The beneficial effects of the present invention are as follows: through the collaborative work of three main modules, namely, a hierarchical low-rank adaptive encoder, a difference adaptive enhancement module and a residual convolution decoder, the remote sensing image change detection network can realize change detection of dual-time remote sensing images, and utilize multi-scale feature extraction, feature fusion and residual convolution decoding technologies to realize effective extraction and visualization of image change information. This method combines low-rank matrix, adaptive enhancement and residual structure, and has certain innovation and practicality in the field of remote sensing image analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of the process framework of the remote sensing image change detection method based on large model domain adaptation in this application; Figure 2 This is a schematic diagram of the low-rank adaptive strategy process framework for this application level; Figure 3 This is a schematic diagram of the process framework of the difference adaptive enhancement module of this application; Figure 4 This is the detection result diagram obtained by the remote sensing image change detection method based on large model domain adaptation in this application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0025] like Figure 1 As shown, the present application provides a remote sensing image change detection method based on large model domain adaptation, comprising the following steps: S1: Input a pair of dual-temporal remote sensing images to be detected into the trained remote sensing image change detection network.

[0026] S2: Output change detection map.

[0027] Among them, the change detection network specifically includes: S11, hierarchical low-rank adaptive encoder, extracts multi-scale features of a pair of dual-time remote sensing images (i=1, 2, 3, 4); pre represents the previous moment, and post represents the later moment; the low-rank matrix is ​​first introduced into the self-attention layer and linear layer of the SAM2 image encoder, and then feature extraction is performed.

[0028] S12, difference adaptive enhancement module, by fusing a pair of multi-scale features of dual-time remote sensing images (i=1, 2, 3, 4); first (i=2, 3, 4) upsampled to size, and then each group Adaptive fusion is used to generate four sets of difference features with rich semantic information. The size of each set of difference features is same.

[0029] S13, residual convolution decoder, decodes 4 sets of difference features and outputs a change map. The first channel corresponds to the probability of the no-change class, and the second channel corresponds to the probability of the change class. The binary change map is obtained by using the argmax operation along the channel dimension.

[0030] The training process of the change detection network is as follows: The labeled samples are divided into training set, validation set and test set.

[0031] At the beginning of training, the parameters of the SAM2 image encoder itself remain frozen and are not updated during the training process; Initialized to 0 and 1 respectively. Represents the weight matrix of a layer in the transformer, Represented by The low-rank matrix obtained by decomposition, and The rank of is set to 16. The parameters of the difference adaptive enhancement module and the residual convolution decoder are randomly initialized. The data enhancement strategy of random flipping and random rotation is adopted during training.

[0032] Example The remote sensing image change detection method based on large model domain adaptation in this application specifically includes the following steps: Step 1: Construct training samples for the change detection network, including training set, validation set and test set. Before training, normalize the dual-time images and labels of the training set to between 0 and 1.

[0033] Step 2: Select the SAM2 image encoder as the feature extractor for dual-time images.

[0034] Step 3: Construct a remote sensing image change detection network based on large model domain adaptation, including a hierarchical low-rank adaptive encoder, a difference adaptive enhancement module, and a residual convolution decoder. The corresponding processing of the hierarchical low-rank adaptive encoder, the difference adaptive enhancement module, and the residual convolution decoder is as follows: The SYSU-CD, WHU-CD and LEVIR-CD datasets are selected to provide change labels (i.e., labeled maps of change areas and non-change areas) for each pair of dual-temporal images. The remote sensing image dataset is divided into three parts: training set, validation set and test set.

[0035] The training set is used for model training, and the validation set is used to adjust hyperparameters and prevent overfitting. The test set is used to evaluate the final model performance. The dual-time images and their labels in the training set, validation set, and test set are normalized. The pixel values ​​of the dual-time images are converted from the original numerical range to values ​​between -0.5 and 0.5 to improve the stability and convergence speed of training.

[0036] Construct a hierarchical low-rank adaptive encoder. SAM2 has been trained on large-scale natural datasets with high quality and has performed well in various downstream tasks. However, the domain differences between natural images and remote sensing images limit its performance in change detection tasks. In order to extract the inductive bias of remote sensing images and further enhance the generalization ability of SMA2, this application sets a hierarchical low-rank adaptive strategy based on parameter efficient transfer learning method to fine-tune the encoder network of SAM2.

[0037] like Figure 2As shown, the present application introduces low-rank adaptation in the self-attention layer and the perceptron layer in the Transformer module of the SAM2 image encoder: in the self-attention layer, the weight matrix is ​​adjusted by low rank to optimize the adaptive weight allocation capability of the attention mechanism, so that the model can more accurately capture the global relationship between input features; in the multi-layer perceptron layer, the features are transformed nonlinearly, and after the introduction of low-rank adaptation, the feature expression can be more effectively adjusted during the transformation process.

[0038] Given the weight matrix of a layer of the Transformer module , R represents the feature space, and Represents the weight matrix Dimensions. A branch road is added on one side. Decomposed two low-rank matrices , Composition, r represents the weight matrix and of rank, and ; Freeze during training The weight of , Alternative Update of; given the input feature of this layer is , the output sequence is , this process is expressed as: (1) (2) In formula 1~2, The weight matrix after introducing LoRA to this layer, Replacement during training Updated weight matrix. In the multi-head self-attention mechanism of the Transformer module, the cosine similarity between different positions is calculated to determine which features should be weighted. This application applies the low-rank adaptive strategy to the query ( ) and value ( ) affects the attention score calculated in the self-attention mechanism, thereby adjusting the model's attention to different regions. Therefore, the attention calculation process after introducing the low-rank adaptive strategy is as follows: (3) (4) (5) (6) In formula 3 to formula 6, , and are the query, key, and value matrices respectively. , , Frozen projection layer for SAM2 , The weight matrix of , , , is a trainable low-rank matrix, and F is the multi-scale contextual feature extracted by the SAM2 image encoder.

[0039] In the MLP layer of the Transformer module, nonlinear transformation allows the model to learn more complex feature representations. This application introduces low-rank adaptation in the first hidden layer of MLP to capture the complex relationship between input features. The linear layer calculation process after the introduction of LoRA is as follows: (7) (8) in, is the input sequence of the MLP layer, Represents the weight matrix after the low-rank matrix is ​​introduced into the first linear layer in the MLP layer, , are the weight matrices of linear layer 1 and linear layer 2 in the MLP layer, , for The decomposed low-rank matrix, , Represents the bias term.

[0040] Construct a difference adaptive enhancement module: This application proposes a difference adaptive enhancement module. Different from the single-branch feature fusion strategy, the difference adaptive enhancement module achieves accurate capture of dual-time difference information through two branches: global difference perception and local detail optimization.

[0041] The two branches of the difference adaptive enhancement module use global attention and local attention to capture category discriminative features and spatial detail features respectively, output their attention weights and adaptively fuse them.

[0042] like Figure 3 As shown, the multi-scale features extracted by the encoder are recorded as (i=1, 2, 3, 4), the difference features enhanced by the difference adaptive enhancement module are recorded as (i=1,2,3,4).

[0043] The operation steps of the difference adaptive enhancement module are as follows: Will and Add the elements together to get ,Will and Subtract the elements to get , then, and They are input into the global difference-aware and local detail-refined branches respectively, taking both local and global contextual information into account.

[0044] In the global difference perception branch, a global average pooling operation is first performed on Each channel feature element is averaged. The size of is changed from [C, H, W] to [C, 1, 1], where C represents the number of feature channels, and H and W represent the height and width of the feature map, respectively. Global attention feature Calculated as: (9) (10) In formula 9 and formula 10, represents element-by-element addition, GAP represents the global average pooling operation, represents a 1×1 convolution operation, BN represents a batch normalization operation, , Represents the global attention feature. Represents the input features of the global difference-aware branch.

[0045] In the local detail optimization branch, firstly, and Subtract the absolute value of the elements , the maximum pooling operation takes the maximum value of the local receptive field, and the features processed by each layer are Keep the same size, local attention features Calculated as: (11) (12) In formula 11 and formula 12, represents element-by-element subtraction, represents the absolute value operation, MAP represents the maximum pooling operation, Represents the input features of the local detail optimization branch.

[0046] Then, and The features are added and fed into the Sigmoid function to obtain the attention weight matrix and The details are as follows: (13) (14) at last, and As a measurement coefficient of global attention and local attention. Specifically, the extracted spatial detail features and category discriminant features are multiplied by their relevant weights on the pixels, and then weighted summed at the pixel level. The enhanced features of each layer It can be calculated by the following formula: (15)

[0047] in, Feature aggregation strategy representing adaptive weights Represents element-wise multiplication.

[0048] The network optimization strategy is as follows: During the training process, the performance of the network is optimized by minimizing the cross entropy loss. For the binary change detection problem, the mathematical expression of the cross entropy loss function is: (16) in, is the category label of the real sample, is the predicted class label and N is the total number of pixels for each sample. It is an indicator to measure the difference between the model prediction value and the true value. The smaller the value of the loss function, the closer the model prediction value is to the true value, and the better the performance of the model.

[0049] The details of the training process are as follows: The SAM2-large image encoder is used as the feature extractor, and multi-scale features are output in the 2nd, 8th, 44th, and 48th transformer blocks to ensure the integrity of the feature space and semantic information. In order to obtain the best performance of the model, a hierarchical low-rank adaptation is introduced every two Transformer Blocks in the encoder, and the rank of the decomposition matrix is ​​set to 16. During the training process, random flipping and random rotation data augmentation strategies are used, and the backbone of the network is kept frozen. The network is trained on a single Nvidia 4090 GPU, and the low-rank matrices are initialized to 0 and 1 respectively, and the remaining trainable parameters are randomly initialized. The learning rate is set to 2.1e-4 based on experience, and the model is trained for 150 rounds. The learning rate is linearly decayed until the last epoch, using the AdamW optimizer with a weight decay of 0.01 and a beta value of (0.9, 0.999).

[0050] In order to more intuitively illustrate the effectiveness of the remote sensing image change detection method based on large model domain adaptation provided by this application, a fully supervised binary change detection experiment was conducted. The evaluation indicators include: F1 score (F1), precision (Pre.), recall rate (Rec.), overall accuracy (OA) and intersection over union (IoU).

[0051] Among them, true positives (TP) represent the number of pixels that are correctly detected as changed, false positives (FP) represent the number of pixels that are actually not changed but are mistakenly detected as changed, true negatives (TN) represent the number of pixels that are not changed and are correctly detected as not changed, and false negatives (FN) represent the number of pixels that are actually changed but are mistakenly detected as not changed.

[0052] (17) (18) (19) (20) (twenty one) As shown in Table 1 below, the performance comparison of the remote sensing image change detection method based on large model domain adaptation in this application and other detection methods on the SYSU-CD dataset is given. Higher IoU and F1 indicate better detection effect. The best result in each column is in bold.

[0053] Table 1 Quantitative comparison of remote sensing image change detection method based on large model domain adaptation in this application and other methods on SYSU-CD and WHU-CD datasets

[0054] The remote sensing image change detection method based on large model domain adaptation in this application performs best on the SYSU-CD dataset. Compared with BAN, Reacll, IoU, F1 score and OA are improved by 7.73%, 5.50%, 3.80% and 0.91% respectively. The above experimental results show that the remote sensing image change detection method based on large model domain adaptation provided by this application has the best performance compared with other change detection networks, verifying that the detection method provided by this application is the most effective.

[0055] In order to further verify the superiority of the remote sensing image change detection method based on large model domain adaptation in this application, the following Figure 4From the visualization analysis shown, it can be observed that most methods have boundary offset and intra-class inconsistency in detecting large-scale objects, while the remote sensing image change detection method based on large model domain adaptation provided by this application shows internal integrity and smooth boundaries. In areas with densely changing objects, the remote sensing image change detection method based on large model domain adaptation in this application detects more complete results than other methods, with basically no missed detection or false detection.

[0056] The present application provides a remote sensing image change detection system based on large model domain adaptation, which includes: an image input module, a network composition and processing module, and an image output module.

[0057] Image input module: by inputting dual-time remote sensing images into the change detection network.

[0058] Network composition and processing modules: including: hierarchical low-rank adaptive encoder, difference adaptive enhancement module and residual convolution decoder.

[0059] The hierarchical low-rank adaptive encoder extracts multi-scale features (i = 1, 2, 3, 4) by introducing low-rank matrices into the self-attention and linear layers of the SAM2 image encoder. The difference adaptive enhancement module generates four sets of difference features by upsampling and adaptively fusing the multi-scale features (i = 2, 3, 4). The residual convolutional decoder obtains a change map containing two types of probabilities by decoding the difference features, and converts the change map into a binary change map using the argmax operation.

[0060] The image output module is used to output the change detection map.

[0061] The hierarchical low-rank adaptation strategy of this application overcomes the knowledge gap between SAM2 and change detection tasks. By introducing low-rank matrices in the key layers of SAM2, the model is guided to learn remote sensing domain knowledge and achieve domain adaptation of SAM2 to remote sensing.

[0062] The difference adaptive enhancement module of this application effectively coordinates spatial detail information and category discrimination information. By generating attention weights for the two, it adaptively enhances the difference information and suppresses boundary displacement and intra-class inconsistency.

[0063] Experimental results on multiple benchmark datasets show that the remote sensing image change detection method based on large model domain adaptation in this application outperforms other state-of-the-art change detection networks in multiple indicators. In addition, the dual-branch structure of the remote sensing image change detection method based on large model domain adaptation in this application can be easily extended to visual tasks such as RGB-T / D semantic segmentation and MRI-CT image fusion.

[0064] Although the content of the present application has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present application. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present application can be made. Therefore, the protection scope of the present application should be limited by the appended claims.

Claims

1. A remote sensing image change detection method based on large model domain adaptation, characterized in that: The steps include: S1: Input a pair of dual-temporal remote sensing images to be detected into the trained remote sensing image change detection network; The change detection network comprises the following steps: S11: Hierarchical low-rank adaptive encoder, first introduces the low-rank matrix into the self-attention layer and linear layer of the SAM2 image encoder, and then extracts the multi-scale features of a pair of dual-temporal remote sensing images ; The multi-scale features of a pair of dual-temporal remote sensing images are extracted by the hierarchical low-rank adaptive encoder , where i=1, 2, 3, 4; S12: a difference adaptive enhancement module firstly transforms the multi-scale features of the dual-time remote sensing image into , i=2,3,4 upsampled to The size of each group Adaptive fusion to generate 4 sets of difference features; Through the difference adaptive enhancement module, the multi-scale features of a pair of dual-time remote sensing images in S11 are integrated. , i=1,2,3,4; S13: residual convolution decoder, decoding the four groups of difference features generated by S12 and outputting a change map; the output change map includes two channels, the first channel corresponds to the probability of the no-change class, and the second channel corresponds to the probability of the change class; using argmax operation along the channel dimension to obtain a binary change map; S2: Output change detection map.

2. The remote sensing image change detection method based on large model domain adaptation according to claim 1 is characterized in that: The size of each of the four groups of difference features is same.

3. The remote sensing image change detection method based on large model domain adaptation according to claim 1 is characterized in that: The S11 mid-level low-rank adaptive encoder specifically includes the following steps: The SAM2 image encoder is selected as the feature extractor for dual-temporal remote sensing images; In the Transformer module of every two SAM2 image encoders, the low-rank adaptation strategy is introduced into its self-attention layer and multi-layer perceptron layer.

4. The remote sensing image change detection method based on large model domain adaptation according to claim 3 is characterized in that: The low-rank adaptive strategy includes: Given the weight matrix of a layer in the Transformer module , and Represents the weight matrix Dimensions; A branch road is added on one side. Decomposed two low-rank matrices , Composition, r represents the weight matrix and The rank and ; Freeze during training The weight of , Alternative Update of; given the input feature of this layer is , the output sequence , this process is expressed as: （1） （2） In formula 1 and formula 2, The weight matrix after introducing LoRA to this layer, Replacement during training Updated weight matrix.

5. The remote sensing image change detection method based on large model domain adaptation according to claim 1 is characterized in that: The S12 difference adaptive enhancement module includes the following steps: The difference adaptive enhancement module captures difference information through two branches: global difference perception and local detail optimization. The specific steps of the two branches are as follows: In the global difference perception branch, firstly, and Add the elements together to get , and then use the global average pooling operation to Each channel feature element is averaged. The size of is changed from [C, H, W] to [C, 1, 1], where C represents the number of feature channels, H and W represent the height and width of the feature map respectively; global attention feature Calculated as: （3） （4） In formula 3, represents element-by-element addition. In Formula 4, GAP represents the global average pooling operation. represents a 1×1 convolution operation, BN represents a batch normalization operation, ; represents the global attention feature, Represents the input features of the global difference perception branch; In the local detail optimization branch, firstly, and Subtract the absolute value of the elements , the maximum pooling operation takes the maximum value of the local receptive field, and the features processed by each layer are Keep the same size, local attention features Calculated as: （5） （6） In formula 5 represents element-by-element subtraction, represents the absolute value operation, and MAP in Formula 6 represents the maximum pooling operation; ; Then, and The features are added and fed into the Sigmoid function to obtain the attention weight matrix and ; at last, and As a measure of global and local attention.

6. The remote sensing image change detection method based on large model domain adaptation according to claim 5 is characterized in that: The attention weight matrix and The specific calculation is as follows: （7） （8） The extracted spatial detail features and category discriminant features are multiplied by their relevant weights at the pixel level, and then weighted summed at the pixel level. The enhanced features of each layer Calculated by the following formula: （9） In formula 9, represents the feature aggregation strategy with adaptive weights, represents element-wise multiplication; The samples containing labels are divided into training set, validation set and test set. At the beginning of training, the parameters of the SAM2 image encoder itself remain frozen and are not updated during the training process. They are initialized to 1 and 0 respectively, and the rank value is set to 16; the difference adaptive enhancement module and the residual convolution decoder are randomly initialized with parameters; and the data enhancement strategy of random flipping and random rotation is adopted during the training process.

7. The remote sensing image change detection method based on large model domain adaptation according to claim 6 is characterized in that: The optimization strategy of the network is as follows: During the training process, the performance of the network is optimized by minimizing the cross entropy loss; for the binary change detection problem, the mathematical expression of the cross entropy loss function is: （10） In formula 10, represents the category label of the real sample, Represents the predicted category label, and N represents the total number of pixels of each sample; It is an indicator to measure the difference between the model prediction value and the true value. The smaller the value of the loss function, the closer the model prediction value is to the true value, and the better the performance of the model.

8. A system using the remote sensing image change detection method based on large model domain adaptation according to any one of claims 1 to 7, characterized in that: The system includes: an image input module, a network composition and processing module and an image output module; The image input module is: inputting dual-time remote sensing images into the change detection network; The network composition and processing module include: a hierarchical low-rank adaptive encoder, a difference adaptive enhancement module and a residual convolution decoder; The hierarchical low-rank adaptive encoder extracts multi-scale features (i = 1, 2, 3, 4) by introducing a low-rank matrix into the self-attention and linear layers of the SAM2 image encoder; the difference adaptive enhancement module generates four sets of difference features by upsampling and adaptively fusing the multi-scale features (i = 2, 3, 4); the residual convolution decoder obtains a change map containing two types of probabilities by decoding the difference features, and converts the change map into a binary change map using an argmax operation; The image output module is used to output a change detection image.

Images