A dual-branch change detection method and system based on a visual base model

Abstract

This invention discloses a bi-branch change detection method and system based on a visual fundamental model, belonging to the field of remote sensing image detection technology. The method proposes a bi-branch change detection network model based on a visual fundamental model, using the frozen visual fundamental model FastSAM as the encoder, which greatly improves the feature extraction capability of the network model. To achieve the transfer of features extracted by the visual fundamental model to the remote sensing image domain, this method designs a change detection adapter module, which achieves effective conversion from general visual features to remote sensing image features by fine-tuning a small number of model parameters. Furthermore, this method proposes a bi-branch change detection module, utilizing a single-temporal semantic feature decoder and a dual-temporal change feature decoder respectively. By mining the semantic information of single-temporal remote sensing images and fusing dual-temporal multi-scale features, the discriminative ability of the features is further improved, thereby enhancing the detection performance of the model.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image detection technology, specifically relating to a dual-branch change detection method and system based on a visual fundamental model. Background Technology

[0002] Remote sensing image change detection is an important technology applied to the interpretation and analysis of high-resolution remote sensing image data. This technology aims to utilize two or more remote sensing images covering the same surface area at different times, along with related geospatial data, combining relevant ground feature characteristics and remote sensing imaging mechanisms. Using image processing algorithms and mathematical models, it determines and analyzes changes in ground features within a specific period, which has significant value and meaning for practical applications such as natural resource management and land spatial planning. Since the 1980s, remote sensing image change detection methods have experienced rapid development, especially benefiting from the rapid advancement of deep learning technology. Remote sensing image change detection technology has achieved significant improvements in detection accuracy and has been widely applied. However, deep learning-based remote sensing image change detection methods still face some challenges. Specifically, existing deep learning models still need improvement in feature extraction capabilities. Besides requiring accurate spatial positioning capabilities, they also need to understand the semantic information corresponding to pixels. The limitations of existing methods in feature extraction capabilities lead to less than ideal performance when handling complex scenes and understanding different types of changed areas, significantly limiting their application in real-world scenarios. Therefore, it is urgent to improve the feature extraction capabilities of deep learning models.

[0003] Recently, visual foundational models, represented by SAM (Signal-Based Aspects), have demonstrated powerful feature extraction and zero-shot generalization capabilities, and are considered a new milestone in the development of computer vision technology. By pre-training on massive amounts of natural scene image data, SAM can identify and segment arbitrary objects using cues such as points, boxes, and text, demonstrating a broad understanding of various objects and scenes. Following the introduction of SAM, more visual foundational models, such as EfficientSAM, FastSAM, and MobileSAM, have emerged. These models not only rival SAM in performance but also significantly improve inference efficiency. Inspired by these methods, leveraging the powerful feature extraction capabilities of advanced visual foundational models to improve existing change detection methods has shown great potential. However, despite the excellent performance of visual foundational models in various natural scenes, directly transferring them to downstream tasks in specific domains, especially in highly specialized fields such as medical imaging and remote sensing imagery, still presents significant challenges. Specifically, visual foundation models are typically pre-trained on large scales of natural images. However, natural images and remote sensing images differ significantly in their image features and distributions, and visual foundation models lack a deep understanding of remote sensing image features and remote sensing knowledge. Therefore, directly applying them to remote sensing image change detection tasks often fails to achieve ideal results.

[0004] In summary, although deep learning-based change detection methods have made some progress, they still suffer from insufficient feature extraction capabilities. While visual basic models possess powerful feature extraction capabilities, they also face many challenges when directly applied to change detection in remote sensing images due to domain differences. Summary of the Invention

[0005] To address these issues, this invention proposes a bi-branch change detection method based on a visual fundamental model. This method uses the visual fundamental model as the encoder, fully leveraging its advantages in feature extraction. Next, a change detection adapter module is designed to transfer features from general visual features to remote sensing image features, and a bi-branch change detection module is introduced to further improve change detection performance. This design not only alleviates the insufficient feature extraction problem of existing deep learning methods in remote sensing image change detection but also effectively transfers features extracted by the visual fundamental model to the remote sensing image domain, thereby improving the accuracy of identifying changed areas.

[0006] In view of this, the purpose of this invention is to provide a two-branch change detection method based on a visual fundamental model. This method proposes a two-branch change detection network model based on a visual fundamental model, using the frozen visual fundamental model FastSAM as the encoder. Leveraging its superior feature extraction capabilities, it extracts general features from dual-temporal remote sensing images, significantly improving the network model's feature extraction ability. To achieve the transfer of features extracted by the visual fundamental model to the remote sensing image domain, this method designs a change detection adapter module. By fine-tuning a small number of model parameters, it achieves an effective conversion from general visual features to remote sensing image features. Furthermore, this method proposes a two-branch change detection module, utilizing a single-temporal semantic feature decoder and a dual-temporal change feature decoder respectively. Through the mining of semantic information from single-temporal remote sensing images and the fusion of dual-temporal multi-scale features, it further improves the feature discrimination ability, thereby enhancing the model's detection performance. This invention effectively addresses the shortcomings of deep learning-based remote sensing image change detection algorithms in feature extraction, while simultaneously achieving effective transfer of the visual fundamental model to the remote sensing image domain, further improving the accuracy and robustness of change detection.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for detecting changes in remote sensing images based on a visual fundamental model, comprising the following steps:

[0008] Step 1) Collect and process the remote sensing image change detection dataset required for training the model;

[0009] Step 2), use a dual-branch change detection network based on a visual base model to train a remote sensing image change detection model on the dataset in Step 1).

[0010] The processing procedure of the remote sensing image change detection model is as follows: A visual base model is used as an encoder to extract general visual features from the dual-temporal images, which are then used as input to the change detection adapter module. The change detection adapter module adjusts the input features to transfer the general visual features to remote sensing image features, and uses the adjusted features as input to the dual-branch change detection module. The dual-branch change detection module uses both the single-temporal semantic feature decoder branch and the dual-temporal change feature decoder branch to output change detection results and supervises the results. Finally, the results output by the dual-branch change detection module are weighted and fused to output the final detection result.

[0011] Step 3) Based on the final remote sensing image change detection model, identify and visualize the changed areas in the dual-temporal remote sensing images to be detected.

[0012] Furthermore, the specific implementation method of step 1) is as follows:

[0013] First, collect the dual-temporal remote sensing image data required for training the remote sensing image change detection model, and perform correction and registration preprocessing steps on the dual-temporal images; second, manually annotate the collected data according to the predetermined land cover change types to obtain an annotation file containing the changed areas and change types in the dual-temporal remote sensing images.

[0014] Furthermore, the visual foundation model is the FastSAM model.

[0015] Furthermore, the specific processing procedure of the change detection adapter module is as follows:

[0016] Using the current feature map as input, the dimensionality of the input feature is first reduced to one-quarter of its original number through a 1×1 convolution, while keeping the spatial dimension unchanged. Then, the adjusted feature is input into the normalization layer. Finally, the ReLU activation function is used to perform a nonlinear transformation on the normalized feature, outputting the final adjusted feature that is more suitable for the field of remote sensing image change detection. The feature adjustment method is shown in formula (1):

[0017]

[0018] Specifically, P i This represents the features extracted by the FastSAM encoder. Indicates the change detection adapter module, F i This represents the features adjusted by the adapter module, where Conv is a 1×1 convolutional layer and BN is the batch normalization function.

[0019] Furthermore, the single-temporal semantic feature decoder branch includes a single-temporal semantic feature decoder module and a change detection module, with the following specific structure:

[0020] 33) The single-phase semantic feature decoder module contains several decoding units. Each decoding unit consists of one upsampling layer, two convolutional layers with a stride of 1 and a kernel size of 3×3, two batch normalization (BN) layers, and two ReLU activation functions. Each decoding unit receives input from high-level features, first performs deconvolution on the current input features, and then concatenates the upsampled features with the feature maps from lower levels to achieve multi-scale feature fusion. Through layer-by-layer upsampling and feature fusion, the semantic features D1 of the multi-scale information of each phase are output. The process is shown in formula (2).

[0021] D i =F i D i-1 =Conv[F i-1 Upsample(D) i (2)

[0022] Specifically, D i F represents the i-th layer feature of the single-temporal semantic feature decoder. i This represents the features at different scales after adjustment by the adapter module. Upsample represents the upsampling operation, and Conv represents a convolutional layer with a kernel size of 3×3.

[0023] Similarly, semantic features D′1, which are obtained by fusing multi-scale information from each phase in the subsequent time period, are obtained.

[0024] 34) Next, a change detection module consisting of two convolutional layers, two batch normalization layers, two ReLU activation functions, and a sigmoid function is used. The image features fused from the preceding and following single-temporal images are stitched together and used as input. The change detection results are then used to predict the change detection results of the single-temporal semantic feature decoder branch. In summary, the calculation process for this branch can be represented by formula (3):

[0025]

[0026] Specifically, The subscripts 1 and 2 refer to the time phase, with 1 representing the previous time phase and 2 representing the subsequent time phase. M represents the feature maps of the before and after temporal remote sensing images output after the change detection adapter is adjusted. θ (·) indicates a branch of the single-phase semantic feature decoder.

[0027] Furthermore, cross-entropy loss is used to evaluate the prediction results. Supervision is performed, and the loss function is calculated as shown in formula (4):

[0028]

[0029] Specifically, N is the number of samples, C is the number of classes, and y i c represents the indicator function of the true label of the i-th sample under category c, which is 1 if it is the true label of the i-th sample and 0 otherwise, while p i .c represents the predicted probability of the i-th sample in class c.

[0030] Furthermore, the specific structure of the dual-temporal change feature decoder is as follows:

[0031] 52) The dual-temporal feature decoder consists of several decoding units within its branches. Each decoding unit comprises two 3×3 convolutional layers, two batch normalization (BN) layers, and two ReLU activation functions. During the decoding of dual-temporal features, each decoding unit concatenates and fuses the extracted features at the same scale and a low-scale feature after upsampling. The features at the same scale come from the outputs of the preceding and following temporal features and other decoding units, while the low-scale features come from the outputs of the following temporal features and other decoding units. The fused feature map undergoes further feature extraction through convolutional blocks to enhance the network's feature representation capability. After decoding operations by all decoding units, four features R containing shallow and deep information are finally output. 1，1 R 1，2 R 1，3 and R 1，4 The decoding process is shown in formula (5):

[0032]

[0033] Specifically, R i，j F represents the output of the decoding unit at the corresponding position. i F′ represents a feature from a previous phase. i This indicates features from the later time phase, Upsample represents the upsampling operation, and Conv represents a convolutional layer with a kernel size of 3×3;

[0034] 52) Next, the obtained features are input into a change detection module consisting of a channel attention module (CAM), a convolutional layer, and a sigmoid function. CAM obtains global information for each channel through global average pooling and global max pooling, then generates channel weights using a fully connected layer, and finally applies these weights to the original feature map to achieve adaptive adjustment of the features. This change detection module receives the features from the dual-temporal change feature decoder (four features concatenated along the channel dimension) as input and outputs the change detection results of the dual-temporal change feature decoder branch. In summary, the calculation process for this branch can be expressed by formula (6) as follows:

[0035]

[0036] Specifically, The subscripts 1 and 2 refer to the time phase, with 1 representing the previous time phase and 2 representing the subsequent time phase. This represents the feature maps of the before and after temporal remote sensing images output after the change detection adapter has been adjusted. This represents the decoder branch for dual-temporal change features.

[0037] Furthermore, the weighted fusion calculation formula for the change detection results output by the dual-branch change detection module is as follows:

[0038]

[0039] Specifically, This represents the change detection results of a single-phase semantic feature decoder branch. For the change detection results of the dual-temporal change feature decoder branch, Conv 1×1 (·) represents a 1×1 convolution operation, and Cat(·) represents channel concatenation.

[0040] Furthermore, during the training process, cross-entropy loss is used to evaluate the prediction results. or To conduct oversight.

[0041] The present invention also provides a dual-branch change detection system based on a visual fundamental model, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a dual-branch change detection method based on a visual fundamental model as described in the above technical solution.

[0042] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0043] 1) This invention utilizes the visual basic model FastSAM to extract feature information from images, which has a stronger feature extraction capability than existing technologies and effectively improves the recognition accuracy of changing areas.

[0044] 2) This invention successfully transferred the general visual features extracted by the visual basic model FastSAM to the field of remote sensing imagery through the change detection adapter module. In addition, the dual-branch change detection module in this invention can utilize the single-temporal semantic feature decoder and the dual-temporal change feature decoder respectively. By mining the semantic information of single-temporal remote sensing images and fusing dual-temporal multi-scale features, the discriminative ability of features is further improved, thereby enhancing the detection performance of the model. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating an implementation of the present invention.

[0046] Figure 2 This is a branch diagram of the single-temporal semantic feature decoder implemented in this invention.

[0047] Figure 3 This is a branch diagram of the dual-temporal change feature decoder implemented in this invention.

[0048] Figure 4This paper compares the performance of this invention with other change detection methods on the WHU-CD dataset. Detailed Implementation

[0049] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] like Figure 1 The technical solution adopted in this invention is a remote sensing image change detection method based on a visual fundamental model, comprising the following steps:

[0051] (1) In the encoder module, the visual basic model FastSAM is used to perform general feature extraction on the input dual-temporal remote sensing images to obtain multi-scale feature maps of the preceding and following temporal phases, and these maps are used as input to the change detection adapter module.

[0052] (2) The change detection adapter module fine-tunes the general visual features extracted by the visual basic model FastSAM, thereby realizing the effective conversion of general visual features to remote sensing image features, and uses the converted features as the input of the dual-branch change detection module.

[0053] (3) The dual-branch change detection module uses a single-temporal semantic feature decoder and a dual-temporal change feature decoder respectively. By mining the semantic information of the single-temporal remote sensing image and fusing the dual-temporal multi-scale features, the detection results of the two branches are obtained. Finally, the results of the two branches are fused to output the final change detection result.

[0054] The specific implementation steps are as follows:

[0055] 1) Input dual-temporal remote sensing images and use the visual basic model FastSAM to extract convolutional feature maps at different spatial resolutions. In this embodiment, convolutional feature maps at four scales were obtained for each temporal image.

[0056] 2) The change detection adapter module is used to adjust each convolutional feature map to obtain convolutional feature maps of the preceding and following time phases that are more suitable for the field of change detection in remote sensing images, and these maps are used as input features for the dual-branch change detection module.

[0057] This change detection adapter module takes the feature map at each scale as input, and adjusts the number of channels of the input feature map to one-quarter of the original number through a 1×1 convolutional layer. Then, the adjusted features are input into a normalization layer to accelerate model training and alleviate overfitting. Finally, the ReLU activation function is used to perform a nonlinear transformation on the normalized features, enabling the model to better fit complex functional relationships and output the final adjusted features that are more suitable for the field of change detection in remote sensing images.

[0058] The specific processing procedure of the change detection adapter module in step 2) is as follows: taking the current feature map as input, the dimensionality of the input feature is first reduced to one-quarter of its original number through a 1×1 convolution, while keeping the spatial dimension unchanged. Then, the adjusted feature is input into the normalization layer to accelerate model training and alleviate overfitting. Finally, the ReLU activation function is used to perform a nonlinear transformation on the normalized feature, enabling the model to better fit complex functional relationships and output the final adjusted feature that is more suitable for the field of change detection in remote sensing images. The feature adjustment method is shown in formula (1):

[0059]

[0060] Specifically, P i This represents the features extracted by the FastSAM encoder. Indicates the change detection adapter module, F i This represents the features adjusted by the adapter module, where Conv is a 1×1 convolutional layer and BN is the batch normalization function.

[0061] 3) The convolutional feature maps of the preceding and following time phases adjusted using the change detection adapter in 2) are used as input to the single-temporal semantic feature decoder branch. A single-temporal semantic feature decoder unit is established using progressive upsampling and fusion. This unit decodes the features of the preceding and following time phases separately to obtain the semantic feature maps of the preceding and following time phases respectively, and uses them as input to the subsequent change detection unit.

[0062] like Figure 2 As shown, the specific details of the single-temporal semantic feature decoder branch are as follows:

[0063] 1) This branch first includes a single-temporal semantic feature decoder module, which contains three decoding units. Each decoding unit consists of one upsampling layer, two convolutional layers with a stride of 1 and a kernel size of 3×3, two batch normalization (BN) layers, and two ReLU activation functions. Each decoding unit receives input from high-level features, first performs deconvolution on the current input features, and then concatenates the upsampled features with the feature maps from lower levels to achieve multi-scale feature fusion. Then, the above steps are repeated to gradually recover spatial information and improve feature representation. The decoding method is shown in formula (2):

[0064] D4 = F4, D i-1 =Conv[F i- 1, Upsample(D i (2)

[0065] Specifically, D i F represents the i-th layer feature of the decoder. i Representing features at different scales, F4 is the feature map at a scale of 1 / 32 of the original image size, Upsample represents the upsampling operation, and Conv represents a convolutional layer with a kernel size of 3×3. Through layer-by-layer upsampling and feature fusion, the semantic features D1, which integrates multi-scale information from each phase, are output.

[0066] Similarly, semantic features D′1 of the multi-scale information of each phase are obtained by fusing the above methods.

[0067] 2) Following the single-temporal semantic feature decoder module is a change detection module consisting of two convolutional layers, two batch normalization layers, two ReLU activation functions, and a sigmoid function. This module takes the image features fused from the preceding and following single-temporal images, concatenates them, and uses the result as input. It then predicts the change detection result to obtain the change detection result of the single-temporal semantic feature decoder branch. The calculation process for this branch can be represented by formula (3):

[0068]

[0069] Specifically, The subscripts 1 and 2 refer to the time phase, with 1 representing the previous time phase and 2 representing the subsequent time phase. M represents the feature maps of the before and after temporal remote sensing images output after the change detection adapter is adjusted. θ (·) denotes a branch of the single-phase semantic feature decoder. Cross-entropy loss is used during training to adjust the prediction results. Supervision is performed, and the loss function is calculated as shown in formula (4).

[0070]

[0071] Specifically, N is the number of samples, C is the number of classes, and y i , c represents the indicator function of the true label of the i-th sample under category c (1 if it is the category, 0 otherwise), and p i .c represents the predicted probability of the i-th sample in class c, as predicted by the model.

[0072] 4) Perform change detection on the semantic feature maps of the preceding and following time phases, and supervise the obtained change detection results to enhance the model's deep semantic understanding of the changed regions, thereby improving the model's ability to capture change details and its robustness in different scenarios.

[0073] 5) The convolutional feature maps of the preceding and following time phases adjusted using the change detection adapter in 2) are used as input to the dual-temporal change feature decoder branch. The change information of all scales of the preceding and following time phases is integrated by layer-by-layer concatenation and fusion to obtain the fused change feature map of all scales of the preceding and following time phases, which is then used as input to the subsequent change detection unit.

[0074] like Figure 3 As shown, the specific details of the dual-temporal change feature decoder branch are as follows:

[0075] 1) This branch contains a dual-temporal feature decoder module, which comprises 10 decoding units. Each decoding unit consists of two 3×3 convolutional layers, two batch normalization (BN) layers, and two ReLU activation functions. During the decoding of dual-temporal features, each decoding unit concatenates and fuses the extracted features at the same scale and a low-scale feature (upsampled). The features at the same scale come from the outputs of the preceding and following temporal features and other decoding units, while the low-scale features come from the outputs of the following temporal features and other decoding units. The fused feature map undergoes further feature extraction through convolutional blocks to enhance the network's feature representation capability. After decoding operations by all decoding units, four features R containing both shallow and deep information are finally output. 1，1 R 1，2 R 1，3 and R 1，4 The decoding process is shown in formula (5):

[0076]

[0077] Specifically, R i，j F represents the output of the decoding unit at the corresponding position. i F′ represents a feature from a previous phase. iThis indicates features from the later phase, Upsample represents the upsampling operation, and Conv represents a convolutional layer with a kernel size of 3×3.

[0078] 2) Following the dual-temporal feature decoder is a change detection module consisting of a Channel Attention Module (CAM), a convolutional layer, and a sigmoid function. CAM acquires global information for each channel through global average pooling and global max pooling, then generates channel weights using a fully connected layer, and finally applies these weights to the original feature map, thus achieving adaptive adjustment of the features. In this way, CAM effectively improves the model's attention to key features, enhancing its expressive power and performance. The change detection module receives the concatenated features from the four features output by the dual-temporal feature decoder along the channel dimension as input and outputs the change detection results from the dual-temporal feature decoder branch. The calculation process for this branch can be expressed by formula (6) as follows:

[0079]

[0080] Specifically, The subscripts 1 and 2 refer to the time phase, with 1 representing the previous time phase and 2 representing the subsequent time phase. This represents the feature maps of the before and after temporal remote sensing images output after the change detection adapter has been adjusted. This represents the decoder branch for dual-temporal variation features. Cross-entropy loss is used during training to adjust the prediction results. Supervision is performed, and the loss function is calculated as shown in formula (4).

[0081]

[0082] Specifically, N is the number of samples, C is the number of classes, and y i .c represents the indicator function of the true label of the i-th sample under category c (1 if it is the category, 0 otherwise), while p i .c represents the predicted probability of the i-th sample in class c, as predicted by the model.

[0083] 6) Perform change detection on the change feature map that integrates information from all scales of the preceding and following time phases, and supervise the obtained change detection results to improve the model's perception of the relationship between detailed changes and the global context;

[0084] 7) The change detection results output from 4) and 6) are weighted and fused to obtain the final change detection result as the final output of the entire method. This result demonstrates excellent performance in accurately identifying changed regions.

[0085] In practical applications, to improve the performance and robustness of change detection results, this method introduces a weighted fusion module after the dual-branch change detection module, which combines the detection results of the single-temporal semantic feature change detection branch and the dual-temporal change feature detection branch. and Perform channel splicing and weighted fusion, and then combine the calculated results. The final change detection result of this method is output. The calculation formula for the weighted fusion step is as follows:

[0086]

[0087] Specifically, Conv 1×1 (·) represents a 1×1 convolution operation, and Cat(·) represents channel concatenation. Cross-entropy loss is used during training to adjust the prediction results. Supervision is performed, and the loss function is calculated as shown in formula (4).

[0088]

[0089] Specifically, N is the number of samples, C is the number of classes, and y i .c represents the indicator function of the true label of the i-th sample under category c (1 if it is the category, 0 otherwise), while p i .c represents the predicted probability of the i-th sample in class c, as predicted by the model.

[0090] Figure 4 To compare the performance of this invention with other change detection methods on the WHU-CD dataset, the proposed dual-branch change detection model based on the visual foundation model, FastSAM, fully leverages the model's feature extraction capabilities. Furthermore, the dual-branch change detection module enhances the discriminative and representational abilities of features, fully mining semantic and change region information contained in multi-scale features, resulting in higher detection accuracy. Table 1 presents a quantitative comparison of the proposed method with other change detection methods on the WHU-CD dataset. The evaluation metrics used in the table are accuracy (P), recall (R), and F1 score (F1), with the F1 score serving as the final reference metric for model performance.

[0091] Table 1. Quantitative comparison results of the present invention with other change detection methods on the WHU-CD dataset.

[0092]

[0093] As shown in Table 1, the model trained by the method of this invention achieves an F1 score of 92.91%, surpassing the highest F1 score FTN model by 0.7%. Experimental results demonstrate that the proposed method exhibits superior performance on the WHU-CD dataset.

[0094] On the other hand, embodiments of the present invention also provide a dual-branch change detection system based on a visual fundamental model, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a dual-branch change detection method based on a visual fundamental model as described in the above technical solution.

[0095] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A two-branch change detection method based on a visual fundamental model, characterized in that, Includes the following steps: Step 1) Collect and process the remote sensing image change detection dataset required for training the model; Step 2) Use a visual base model-based dual-branch change detection network to train a remote sensing image change detection model on the dataset in Step 1). The processing procedure of the remote sensing image change detection model is as follows: A visual base model is used as an encoder to extract general visual features from the dual-temporal images, which are then used as input to the change detection adapter module. The change detection adapter module adjusts the input features to transfer the general visual features to remote sensing image features, and uses the adjusted features as input to the dual-branch change detection module. The dual-branch change detection module uses both the single-temporal semantic feature decoder branch and the dual-temporal change feature decoder branch to output change detection results and supervises the results. Finally, the results output by the dual-branch change detection module are weighted and fused to output the final detection result. The single-temporal semantic feature decoder branch includes a single-temporal semantic feature decoder module and a change detection module, with the following specific structure: 211) The single-temporal semantic feature decoder module contains several decoding units. Each decoding unit consists of one upsampling layer, two convolutional layers with a stride of 1 and a kernel size of 3×3, two batch normalization (BN) layers, and two ReLU activation functions. Each decoding unit receives input from high-level features, first performs deconvolution on the current input features, and then concatenates the upsampled features with the feature maps from lower levels to achieve multi-scale feature fusion. Through layer-by-layer upsampling and feature fusion, the semantic features of the previous temporal fusion, which integrates the multi-scale information of each phase, are output. The process is shown in formula (2): （2） Specifically, This represents the i-th layer features of the single-temporal semantic feature decoder. This represents the features at different scales after the adapter module has been adjusted. This represents an upsampling operation. This represents a convolutional layer with a kernel size of 3×3; Similarly, semantic features obtained from the fusion of multi-scale information in each phase are obtained. ; 212) Next, a change detection module consisting of two convolutional layers, two batch normalization layers, two ReLU activation functions, and a sigmoid function is used. The image features fused from the preceding and following single-temporal images are concatenated and used as input. The change detection results are then used to predict the change detection results of the single-temporal semantic feature decoder branch. In summary, the calculation process for this branch can be represented by formula (3): （3） Specifically, The subscripts 1 and 2 refer to the time phase, with 1 representing the previous time phase and 2 representing the subsequent time phase. This represents the feature maps of the before and after temporal remote sensing images output after the change detection adapter has been adjusted. This represents a branch of the single-temporal semantic feature decoder; The specific structure of the dual-temporal change feature decoder is as follows: 221) The dual-temporal feature decoder comprises several decoding units within its branches. Each decoding unit consists of two 3×3 convolutional layers, two batch normalization (BN) layers, and two ReLU activation functions. During the decoding of dual-temporal features, each decoding unit concatenates and fuses the extracted features at the same scale and a low-scale feature after upsampling. The features at the same scale come from the outputs of the preceding and following temporal features and other decoding units, while the low-scale features come from the outputs of the following temporal features and other decoding units. The fused feature map undergoes further feature extraction through convolutional blocks to enhance the network's feature representation capability. After decoding operations by all decoding units, four features containing both shallow and deep information are finally output. , , and The decoding process is shown in formula (5): （5） Specifically, This indicates the output of the decoding unit at the corresponding position. Indicates features from previous time periods. Indicates features from later phases, This represents an upsampling operation. This represents a convolutional layer with a kernel size of 3×3; 222) Next, the obtained features are input into a change detection module consisting of a channel attention module (CAM), a convolutional layer, and a sigmoid function. CAM obtains global information for each channel through global average pooling and global max pooling, then generates channel weights using a fully connected layer, and finally applies these weights to the original feature map to achieve adaptive adjustment of the features. This change detection module receives the features from the dual-temporal change feature decoder (four features concatenated along the channel dimension) as input and outputs the change detection results of the dual-temporal change feature decoder branch. In summary, the calculation process for this branch can be expressed by formula (6) as follows: （6） Specifically, The subscripts 1 and 2 refer to the time phase, with 1 representing the previous time phase and 2 representing the subsequent time phase. This represents the feature maps of the before and after temporal remote sensing images output after the change detection adapter has been adjusted. This represents the decoder branch representing the dual-temporal variation characteristics; Step 3) Based on the final remote sensing image change detection model, identify and visualize the changed areas in the dual-temporal remote sensing images to be detected.

2. The bi-branch change detection method based on a visual fundamental model as described in claim 1, characterized in that: The specific implementation method for step 1) is as follows: First, collect the dual-temporal remote sensing image data required for training the remote sensing image change detection model, and perform correction and registration preprocessing steps on the dual-temporal images; second, manually annotate the collected data according to the predetermined land cover change types to obtain an annotation file containing the changed areas and change types in the dual-temporal remote sensing images.

3. The bi-branch change detection method based on a visual fundamental model as described in claim 1, characterized in that: The visual foundation model is the FastSAM model.

4. The bi-branch change detection method based on a visual fundamental model as described in claim 1, characterized in that: The specific processing procedure of the change detection adapter module is as follows: The current feature map is used as input. First, the number of channels of the input feature is reduced to one-quarter of the original number through 1×1 convolution, while keeping the spatial dimension unchanged. Then, the adjusted feature is input into the normalization layer. Finally, the ReLU activation function is used to perform a nonlinear transformation on the normalized feature, and the final adjusted feature is output, which is more suitable for the field of remote sensing image change detection. The feature adjustment method is shown in formula (1): （1） Specifically, This represents the features extracted by the FastSAM encoder. This indicates a change detection adapter module. This indicates the adjusted features of the adapter module. It is a 1×1 convolutional layer. This is a batch processing normalization function.

5. The bi-branch change detection method based on a visual fundamental model as described in claim 1, characterized in that: Using cross-entropy loss to evaluate prediction results Supervision is performed, and the loss function is calculated as shown in formula (4): （4） Specifically, For the sample size, For the number of categories, Indicates the first The true label of each sample in the category The indicator function below returns 1 if it belongs to that category, and 0 otherwise. Then it is the predicted number Each sample in category The predicted probability is as follows.

6. The bi-branch change detection method based on a visual fundamental model as described in claim 1, characterized in that: The weighted fusion calculation formula for the change detection results output by the dual-branch change detection module is as follows: （7） Specifically, This represents the change detection results of a single-phase semantic feature decoder branch. This represents the change detection results of the dual-temporal change feature decoder branch. This represents a 1×1 convolution operation. This indicates channel splicing.

7. The bi-branch change detection method based on a visual fundamental model as described in claim 6, characterized in that: During training, cross-entropy loss is used to evaluate the prediction results. or To conduct oversight.

8. A dual-branch change detection system based on a visual fundamental model, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute a two-branch change detection method based on a visual fundamental model as described in any one of claims 1-7.

Images