A remote sensing image change detection method based on Token difference guidance

Abstract

The application discloses a remote sensing image change detection method based on Token difference guidance, comprising the following steps: acquiring double-time-phase remote sensing images and spatially aligned change labels; constructing and training a change detection network ToDi-Net; extracting double-time-phase multi-scale features and fusing the features to obtain a fused feature map; Tokenizing the fused feature map to obtain double-time-phase Token sequences, and generating Token difference regulation coefficients through a Token difference regulation module to adaptively regulate double-time-phase Token semantic responses; modeling cross-time-phase interaction of the regulated Token sequences and mapping the Token sequences back to a pixel space; generating a difference-semantic collaborative weight map through a difference-semantic double-branch collaborative decoder, fusing difference features and semantic reference features, and outputting a change prediction map. The application can enhance remote sensing image change-related expression and suppress non-change redundant interference, and improve the consistency and robustness of change detection results in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection technology, and discloses a remote sensing image change detection method based on token difference guidance, which is used to extract change information and determine the change area from dual-temporal remote sensing images of the same area. Background Technology

[0002] Remote sensing image change detection aims to compare and analyze two-temporal remote sensing images of the same area to output change results, and is widely used in urban sprawl monitoring and disaster assessment. In existing research, deep learning methods have been used to learn feature representations of two-temporal images, and on this basis, a representation of the differences between temporal phases is constructed to obtain change detection results. However, existing remote sensing image change detection networks still have the following problems:

[0003] 1. Change regions in remote sensing images are typically sparse, making change cues easily obscured during bi-temporal feature modeling, and lacking a targeted token response adjustment mechanism based on bi-temporal differences. In real-world scenarios, non-change regions often dominate, leading to insufficient model response to minor or localized changes. Especially at the token semantic level, without a mechanism to adjust the strength of the token response based on bi-temporal differences, change-related cues are difficult to maintain prominence. Existing token compression or pruning strategies, such as patent CN202511846468.7 (application number CN202511846468.X) entitled "Visual Token Compression Method and Device Based on Dual-Path Fusion," primarily focus on computational efficiency optimization; general token adjustment schemes, such as patent CN202511891618.6 (application number CN202511891618.X) entitled "An Application Method of Transformer Architecture Based on Biological Regulation Mechanism," primarily derive their adjustment signals from the general context, failing to establish a targeted token control mechanism around bi-temporal differences.

[0004] 2. Existing methods for extracting change information from remote sensing images employ relatively simplistic construction methods for both differential and semantic information, making it difficult to simultaneously provide both differential representations and semantic references for change detection. Some methods use simple differencing and concatenation to construct change representations, but these approaches often focus on constructing differential information along a single path, lacking parallel modeling of differential and semantic information for change detection tasks. This results in a lack of complementary information support in the subsequent decoding stage.

[0005] 3. Existing remote sensing image change detection methods lack an adaptive collaborative fusion mechanism for difference information and semantic information, affecting the integrity and stability of the results. When relying solely on a single remote sensing image difference representation for decoding, it is difficult to simultaneously consider the semantic consistency between the response of changed areas and those of unchanged areas. Related remote sensing image semantic segmentation schemes, such as patent CN202511924681.5 and application number CN202511924681.X, entitled "A Semantic Segmentation Method, Terminal, and Storage Medium for Remote Sensing Burnt Areas Based on Multi-Scale Fusion and Attention Mechanism," can improve semantic segmentation performance in complex backgrounds, but they are geared towards single-temporal semantic segmentation tasks and still differ from the dual-temporal difference construction and difference-semantic collaborative decoding mechanism in change detection.

[0006] Based on the above problems, there is an urgent need for a remote sensing image change detection method that can respond and adjust the differences between two-temporal remote sensing images at the token semantic level, and collaboratively utilize difference information and semantic information for change prediction during the decoding stage. Summary of the Invention

[0007] The purpose of this invention is to provide a remote sensing image change detection method based on token difference guidance, so as to solve the problems that existing change detection is easily affected by non-change perturbations in the process of dual-temporal semantic representation and fusion, and that it is difficult to simultaneously take into account the response consistency of changed areas and the suppression of non-change areas in the fusion stage.

[0008] This invention provides a token-guided method for detecting changes in remote sensing images, comprising the following steps:

[0009] S1: Obtain change detection samples from dual-temporal remote sensing images. These samples are derived from publicly available remote sensing change detection datasets, or are constructed independently based on the first and second temporal remote sensing images of the study area and experimental area, and spatially aligned change annotations are obtained. The first temporal remote sensing image is denoted as... The second phase remote sensing image is denoted as Changes are marked as ;

[0010] S2: Construct a remote sensing image change detection network ToDi-Net based on token difference guidance; ToDi-Net consists of a CNN backbone network, a multi-scale feature fusion module, a token difference regulation module, a Transformer cross-temporal interaction module, and a difference-semantic dual-branch collaborative decoder;

[0011] The Transformer cross-temporal interaction module includes a Transformer encoder, a Transformer decoder, and an intra-layer feedforward neural network.

[0012] The multi-scale feature fusion module includes a lateral alignment unit, a scale alignment unit, a feature splicing unit, a fusion convolutional unit, and an SE channel attention unit. The lateral alignment unit includes three 1×1 convolutional layers with a stride of 1 and padding of 0, and its input and output channels are (64, 32), (128, 32), and (256, 32), respectively. The scale alignment unit includes an interpolation alignment layer used to adjust each feature to the target spatial size. The feature splicing unit splices three features in the channel dimension to obtain 96-channel features. The fusion convolutional unit includes a 1×1 convolutional layer, a BN layer, and a ReLU layer, with the input and output channels of the convolutional layer being (96, 32). The SE channel attention unit performs channel recalibration on the fused features to obtain the output features.

[0013] The Token Difference Control Module includes a Token Difference Calculation Unit, a Change Probability Prediction Unit, and a Token Difference Modulation Unit. The Token Difference Calculation Unit performs element-wise difference operations on the two-phase Token sequences at corresponding Token positions to obtain the difference Token. The Change Probability Prediction Unit includes a first fully connected layer, a ReLU layer, a second fully connected layer, and a Sigmoid layer. The input-output dimension of the first fully connected layer is (32, 8), and the input-output dimension of the second fully connected layer is (8, 1). The Sigmoid layer normalizes the output of the Change Probability Prediction Unit to obtain the Token Difference Control Coefficient. The Token differential modulation unit adjusts the token differential coefficient accordingly. Perform element-wise weighted modulation on the two-phase token sequences;

[0014] The differential-semantic dual-branch collaborative decoder includes a differential enhancement branch, a semantic reference branch, a differential-semantic collaborative unit, and an output prediction unit. The differential enhancement branch includes a 3×3 convolutional layer, a BN layer, and a ReLU layer, with the convolutional layer having an input / output channel of (32, 32). The semantic reference branch includes a 1×1 convolutional layer, a BN layer, and a ReLU layer, with the convolutional layer having an input / output channel of (32, 32). The differential-semantic collaborative unit includes a 3×3 convolutional layer, a BN layer, and a Sigmoid layer, with the convolutional layer having an input / output channel of (64, 32). It also includes an upsampling layer... Convolutional layers, BN layers, ReLU layers and Convolutional layer; wherein, the The convolutional layer has 32 input and 32 output channels. The convolutional layer has 32 input channels and 2 output channels.

[0015] S3: Will and The features are respectively input into the CNN backbone network for feature encoding to obtain a multi-scale feature set. and and the deepest features and The first phase feature map is obtained through convolution mapping. Compared with the second phase feature map ;in, Represents scale index and ;

[0016] S4: The multi-scale feature set and Input to the multi-scale feature fusion module to obtain the first temporal feature map. Compared with the second phase feature map ;

[0017] S5: Regarding the above and Tokenization is performed to obtain the first phase token sequence. With the second phase token sequence ;

[0018] Specifically, the tokenization process includes: processing the first phase feature map. Second phase feature map Perform separately Convolutional mapping yields the first temporal spatial attention map. Second temporal spatial attention map The spatial attention map is then normalized using softmax in the spatial dimension to obtain the corresponding spatial attention weights. Finally, the input feature maps are weighted and converged using these spatial attention weights to form a semantic token sequence, the calculation formula of which is as follows:

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] in, This represents the spatial attention map of the first phase. This represents the second-phase spatial attention map. Represents the spatial attention weights of the first phase. Indicates the spatial attention weights of the second temporal phase. express Convolution operation, This represents the softmax normalization operation. This represents tensor transformation operations. This indicates a weighted aggregation operation;

[0026] S6: [The sentence is incomplete and requires further context.] and Input the token difference control module to obtain the first phase token sequence. With the second phase token sequence The Token difference control module is used to enhance the difference response between the two temporal features and suppress interference from the non-changing background.

[0027] Furthermore, in step S6, the Token difference control module includes a Token difference calculation unit, a change probability prediction unit, and a Token difference modulation unit.

[0028] The token difference calculation unit calculates the first phase token sequence. With the second phase token sequence Perform element-wise absolute difference operations at the corresponding token position to obtain the difference token. The calculation formula is as follows:

[0029]

[0030] in, This represents element-wise absolute value operation;

[0031] The change probability prediction unit for the difference token Perform a fully connected mapping and normalize using the Sigmoid function to obtain the differential modulation coefficient. The calculation formula is as follows:

[0032]

[0033] in, This represents the mapping operation consisting of a first fully connected layer, a ReLU layer, and a second fully connected layer. This represents the Sigmoid activation function;

[0034] The Token difference modulation unit adjusts the token difference coefficient. right and Perform element-wise weighted modulation to obtain and The formula is as follows:

[0035]

[0036] in, This represents element-wise multiplication.

[0037] S7: [The sentence is incomplete and likely refers to a specific item or instruction.] and Input the Transformer cross-temporal interaction module and output the first temporal decoding feature map. With the second phase decoding feature map ;

[0038] S8: [The sentence is incomplete and likely refers to a specific method or procedure.] and Input difference – semantic dual-branch co-decoder, output dual-temporal change prediction graph The difference-semantic dual-branch collaborative decoder is used to collaboratively fuse and decode difference information and semantic information.

[0039] Furthermore, in step S8, the difference-semantic dual-branch collaborative decoder includes a difference enhancement branch, a semantic reference branch, a difference-semantic collaborative unit, and an output prediction unit;

[0040] The difference enhancement branch includes Convolutional layers, BN layers, and ReLU layers; The convolutional layer has 32 input and 32 output channels; the difference enhancement branch is used to construct and enhance difference features element-wise. The calculation formula is as follows:

[0041]

[0042]

[0043] in, Indicates the difference representation; This represents a temporal decoding feature map; This represents the second phase decoding feature map; Indicates the kernel size as Convolution operation; Indicates BN layer operations; Represents the ReLU activation function;

[0044] The semantic reference branch includes Convolutional layers, BN layers, and ReLU layers; The convolutional layer has 32 input and 32 output channels; the semantic reference branch is used to construct and enhance semantic reference features on an element-wise average basis. The calculation formula is as follows:

[0045]

[0046]

[0047] in, Indicates semantic reference representation;

[0048] The difference-semantic collaboration unit includes a splicing layer, Convolutional layers, BN layers, and Sigmoid layers; The convolutional layer has 64 input channels and 32 output channels; the differential-semantic co-operation unit is used to obtain fused features. The steps are as follows:

[0049] Difference characteristics With semantic reference features By splicing, we obtain joint features. The calculation formula is as follows:

[0050]

[0051] in, This indicates a concatenation operation;

[0052] joint features enter Convolutional layers, Batch Normalization (BN) layers, and Sigmoid layers yield a difference-semantic co-weight graph. The calculation formula is as follows:

[0053]

[0054] in, This represents the Sigmoid activation function;

[0055] The difference-semantic co-weight graph Differential characteristics With semantic reference features Weighted fusion is performed to obtain fusion features The calculation formula is as follows:

[0056]

[0057] in, This represents element-wise multiplication.

[0058] The output prediction unit includes an upsampling layer, Convolutional layers, BN layers, ReLU layers and Convolutional layer; wherein, the The convolutional layer has 32 input and 32 output channels. The convolutional layer has 32 input channels and 2 output channels.

[0059] S9: Utilizing the above , , Train the ToDi-Net network according to Change labeling Calculate the loss function and update the network parameters to obtain the trained ToDi-Net network;

[0060] S10: In the inference phase, the dual-temporal remote sensing image to be tested is input into the trained ToDi-Net network, which outputs a remote sensing image change probability map to complete the remote sensing image change detection.

[0061] Compared with the prior art, the present invention has at least the following beneficial effects:

[0062] 1. By introducing a token difference control mechanism based on dual temporal differences, this invention can specifically enhance change-related information at the token semantic level, enabling change cues to maintain a strong response during cross-temporal feature interaction, thereby improving the situation in the prior art where change information is not easily highlighted under sparse change regions.

[0063] 2. By constructing difference enhancement branches and semantic reference branches, this invention can provide difference information and semantic information respectively, so that the decoding stage no longer relies on a single change representation, thereby improving the problem of the single construction method of difference information and semantic information in the prior art.

[0064] 3. By setting up a difference-semantic coordination unit and an output prediction unit, this invention can adaptively and collaboratively utilize difference information and semantic information, thereby improving the problem of the lack of an adaptive collaborative fusion mechanism between difference information and semantic information in the decoding stage of the prior art, which is conducive to improving the detection effect of change regions. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a ToDi-Net network structure diagram of a remote sensing image change detection method based on token difference guidance according to an embodiment of the present invention;

[0067] Figure 2 This is a structural diagram of a multi-scale feature fusion module for a remote sensing image change detection method based on token difference guidance according to an embodiment of the present invention.

[0068] Figure 3 This is a structural diagram of the difference token control module in a remote sensing image change detection method based on difference token guidance according to an embodiment of the present invention.

[0069] Figure 4 This is a diagram of the differential-semantic dual-branch collaborative decoder structure of a remote sensing image change detection method based on token difference guidance according to an embodiment of the present invention.

[0070] Figure 5 This is a comparison chart of detection results for a remote sensing image change detection method based on token difference guidance according to an embodiment of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] like Figure 1 As shown, this invention discloses a remote sensing image change detection method based on token difference guidance, comprising the following steps:

[0073] S1: Obtain change detection samples from dual-temporal remote sensing images. These samples are derived from publicly available remote sensing change detection datasets, or are constructed independently based on the first and second temporal remote sensing images of the study area and experimental area, and spatially aligned change annotations are obtained. The first temporal remote sensing image is denoted as... The second phase remote sensing image is denoted as Changes are marked as ;

[0074] S2: Construct a remote sensing image change detection network ToDi-Net based on token difference guidance; ToDi-Net consists of a CNN backbone network, a multi-scale feature fusion module, a token difference regulation module, a Transformer cross-temporal interaction module, and a difference-semantic dual-branch collaborative decoder;

[0075] The Transformer cross-temporal interaction module includes a Transformer encoder, a Transformer decoder, and an intra-layer feedforward neural network.

[0076] Furthermore, the multi-scale feature fusion module, as shown in the example... Figure 2 As shown, the system includes a lateral alignment unit, a scale alignment unit, a feature concatenation unit, a fusion convolution unit, and an SE channel attention unit. The lateral alignment unit comprises three 1×1 convolutional layers with a stride of 1 and padding of 0, and its input and output channels are (64, 32), (128, 32), and (256, 32), respectively. The scale alignment unit includes an interpolation alignment layer used to adjust the features of each path to the target spatial size. The feature concatenation unit concatenates three features in the channel dimension to obtain 96-channel features. The fusion convolution unit includes a 1×1 convolutional layer, a BN layer, and a ReLU layer, with the input and output channels of the convolutional layer being (96, 32). The SE channel attention unit performs channel recalibration on the fused features to obtain the output features.

[0077] The Token Difference Control Module, as follows Figure 3 As shown, the system includes a Token Difference Calculation Unit, a Change Probability Prediction Unit, and a Token Difference Modulation Unit. The Token Difference Calculation Unit performs element-wise difference operations on the corresponding Token positions of two temporal Token sequences to obtain the difference Token. The Change Probability Prediction Unit includes a first fully connected layer, a ReLU layer, a second fully connected layer, and a Sigmoid layer. The input-output dimension of the first fully connected layer is (32, 8), and the input-output dimension of the second fully connected layer is (8, 1). The Sigmoid layer normalizes the output of the Change Probability Prediction Unit to obtain the Token Difference Modulation Coefficient. The Token differential modulation unit adjusts the token differential coefficient accordingly. Perform element-wise weighted modulation on the two-phase token sequences;

[0078] The differential-semantic dual-branch collaborative decoder includes a differential enhancement branch, a semantic reference branch, a differential-semantic collaborative unit, and an output prediction unit. The differential enhancement branch includes a 3×3 convolutional layer, a BN layer, and a ReLU layer, with the convolutional layer having an input / output channel of (32, 32). The semantic reference branch includes a 1×1 convolutional layer, a BN layer, and a ReLU layer, with the convolutional layer having an input / output channel of (32, 32). The differential-semantic collaborative unit includes a 3×3 convolutional layer, a BN layer, and a Sigmoid layer, with the convolutional layer having an input / output channel of (64, 32). It also includes an upsampling layer... Convolutional layers, BN layers, ReLU layers and Convolutional layer; wherein, the The convolutional layer has 32 input and 32 output channels. The convolutional layer has 32 input channels and 2 output channels.

[0079] S3: Will and The features are respectively input into the CNN backbone network for feature encoding to obtain a multi-scale feature set. and The deepest features are then convolved to obtain the first phase feature map. Compared with the second phase feature map ;in, Represents scale index and ;

[0080] S4: The multi-scale feature set and Input to the multi-scale feature fusion module to obtain the first temporal feature map. Compared with the second phase feature map ;

[0081] S5: Regarding the above and Tokenization is performed to obtain the first phase token sequence. With the second phase token sequence ;

[0082] Specifically, the tokenization process includes: processing the first phase feature map. Second phase feature map Perform separately Convolutional mapping yields the first temporal spatial attention map. Second temporal spatial attention map The spatial attention map is then normalized using softmax in its spatial dimension to obtain the corresponding spatial attention weights. Finally, the input feature maps are weighted and converged using these spatial attention weights to form a semantic token sequence. The calculation formula is as follows:

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] in, This represents the spatial attention map of the first phase. This represents the second-phase spatial attention map. Represents the spatial attention weights of the first phase. Indicates the spatial attention weights of the second temporal phase. express Convolution operation, This represents the softmax normalization operation. This represents tensor transformation operations. This indicates a weighted aggregation operation;

[0090] S6: [The sentence is incomplete and requires further context.] and Input the token difference control module to obtain the first phase token sequence. With the second phase token sequence The Token difference control module is used to enhance the difference response between the two temporal features and suppress interference from the non-changing background.

[0091] Furthermore, in step S6, the Token difference control module includes a Token difference calculation unit, a change probability prediction unit, and a Token difference modulation unit.

[0092] The token difference calculation unit calculates the first phase token sequence. With the second phase token sequence Perform element-wise absolute difference operation at the corresponding Token position to obtain the remote sensing image difference Token. The formula is as follows:

[0093]

[0094] in, This represents element-wise absolute value operation;

[0095] The change probability prediction unit analyzes the difference token in remote sensing images. Perform a fully connected mapping and normalize using the Sigmoid function to obtain the differential modulation coefficient. The calculation formula is as follows:

[0096]

[0097] in, This represents the mapping operation consisting of a first fully connected layer, a ReLU layer, and a second fully connected layer. This represents the Sigmoid activation function;

[0098] The Token difference modulation unit adjusts the token difference coefficient. right and Perform element-wise weighted modulation to obtain and The formula is as follows:

[0099]

[0100] This indicates element-wise multiplication.

[0101] S7: [The sentence is incomplete and likely refers to a specific item or instruction.] and Input the Transformer cross-temporal interaction module and output the first temporal decoding feature map. With the second phase decoding feature map ;

[0102] S8: [The sentence is incomplete and likely refers to a specific method or procedure.] and Input difference – semantic dual-branch co-decoder, output dual-temporal change prediction graph The difference-semantic dual-branch collaborative decoder is used to collaboratively fuse and decode difference information and semantic information.

[0103] Furthermore, in step S8, the difference-semantic dual-branch collaborative decoder includes a difference enhancement branch, a semantic reference branch, a difference-semantic collaborative unit, and an output prediction unit;

[0104] Among them, the difference enhancement branch includes Convolutional layers, BN layers, and ReLU layers, the The convolutional layer has 32 input and 32 output channels; the difference enhancement branch is used to decode the feature map based on the first temporal phase. With the second phase decoding feature map Element-wise absolute difference construction difference characterization and through Differential features are obtained by processing convolutional layers, BN layers, and ReLU layers. The calculation formula is as follows:

[0105]

[0106]

[0107] in, Indicates the kernel size as Convolution operation; Indicates BN layer operations; Represents the ReLU activation function;

[0108] Semantic reference branches include Convolutional layers, BN layers, and ReLU layers, the The convolutional layer has 32 input and 32 output channels. The semantic reference branch is used to decode the feature map based on the first temporal phase. With the second phase decoding feature map Element-wise average construction of semantic reference representation and through Semantic reference features are obtained through convolutional layers, batch normalization (BN) layers, and ReLU layers. The calculation formula is as follows:

[0109]

[0110]

[0111] The difference-semantic collaboration unit includes a splicing layer, Convolutional layers, BN layers, and Sigmoid layers; The convolutional layer has 64 input channels and 32 output channels; the differential-semantic co-operation unit is used to obtain fused features. The steps are as follows:

[0112] Difference characteristics With semantic reference features By splicing, we obtain joint features. The calculation formula is as follows:

[0113]

[0114] in, This indicates a concatenation operation;

[0115] joint features enter Convolutional layers, Batch Normalization (BN) layers, and Sigmoid layers yield a difference-semantic co-weight graph. The calculation formula is as follows:

[0116]

[0117] in, This represents the Sigmoid activation function;

[0118] The difference-semantic co-weight graph Differential characteristics With semantic reference features Weighted fusion is performed to obtain fusion features The calculation formula is as follows:

[0119]

[0120] in, This represents element-wise multiplication.

[0121] The output prediction unit includes an upsampling layer, Convolutional layers, BN layers, ReLU layers and Convolutional layer; wherein, the The convolutional layer has 32 input and 32 output channels. The convolutional layer has 32 input channels and 2 output channels.

[0122] S9: Supervised training of the ToDi-Net network is performed using the training samples, based on... Change labeling Calculate the loss function and update the network parameters to obtain the trained change detection model;

[0123] S10: In the inference phase, the dual-temporal remote sensing image to be tested is input into the trained ToDi-Net network, which outputs a remote sensing image change probability map to complete the remote sensing image change detection.

[0124] To verify the effectiveness of the method of the present invention, comparative experiments were conducted on the LEVIR-CD dataset. A unified training and testing partition was used in the experiments, and common change detection evaluation metrics were employed to quantitatively analyze the detection results. The experimental results are summarized in Table 1.

[0125] Table 1 - Comparative experimental results of ToDi-Net on the LEVIR-CD dataset

[0126]

[0127] Table 1 shows that different methods, from early CNN fusion structures to the introduction of cross-temporal interaction modeling, show an overall improvement in detection performance. The ToDi-Net of this invention achieves the best results across all metrics: Precision 91.00%, Recall 90.07%, F1 90.53%, IoU 82.71%, and OA 99.04%, indicating that it can effectively suppress non-changing interference while enhancing the response in changing regions, thereby improving the consistency and stability of detection results.

[0128] Figure 5Visual comparison results on the LEVIR-CD dataset are presented to intuitively demonstrate the differences between different methods in terms of changing region localization and contour recovery. Combined with the experimental results on the LEVIR-CD dataset shown in Table 1, the present invention achieves superior performance in Precision, Recall, F1, IoU, and OA metrics on the LEVIR-CD dataset, with an F1 score of 90.53%, an IoU of 82.71%, and an OA of 99.04%. It can be seen that the prediction results of the present invention's ToDi-Net are more consistent with GT, with more complete boundaries and fewer holes and noise, further validating the performance advantages shown in Table 1.

Claims

1. A method for detecting changes in remote sensing images based on token difference guidance, characterized in that, Includes the following steps: S1: Obtain change detection samples from dual-temporal remote sensing images, wherein the first temporal remote sensing image is denoted as... The second phase remote sensing image is denoted as Changes are marked as ; S2: Construct a remote sensing image change detection network ToDi-Net based on token difference guidance; ToDi-Net consists of a CNN backbone network, a multi-scale feature fusion module, a token difference regulation module, a Transformer cross-temporal interaction module, and a difference-semantic dual-branch collaborative decoder; S3: Will and The features are respectively input into the CNN backbone network for feature encoding to obtain a multi-scale feature set. and and the deepest features and The first phase feature map is obtained through convolution mapping. Compared with the second phase feature map ;in, Represents scale index and ; S4: The multi-scale feature set and Input to the multi-scale feature fusion module to obtain the first temporal feature map. Compared with the second phase feature map ; S5: Regarding the above and Tokenization is performed to obtain the first phase token sequence. With the second phase token sequence ; S6: [The sentence is incomplete and requires further context.] and Input the token difference control module to obtain the first phase token sequence. With the second phase token sequence The Token difference control module is used to enhance the difference response between the two temporal features and suppress interference from the non-changing background. S7: [The sentence is incomplete and likely refers to a specific item or instruction.] and Input the Transformer cross-temporal interaction module and output the first temporal decoding feature map. With the second phase decoding feature map ; S8: [The sentence is incomplete and likely refers to a specific method or procedure.] and Input difference – semantic dual-branch co-decoder, output dual-temporal change prediction graph The difference-semantic dual-branch collaborative decoder is used to collaboratively fuse and decode difference information and semantic information. S9: Utilizing the above , , Train the ToDi-Net network according to Change labeling Calculate the loss function and update the network parameters to obtain the trained ToDi-Net network; S10: In the inference phase, the dual-temporal remote sensing image to be tested is input into the trained ToDi-Net network, which outputs a remote sensing image change probability map to complete the remote sensing image change detection.

2. The remote sensing image change detection method based on token difference guidance according to claim 1, characterized in that, In step S6, the Token difference control module includes a Token difference calculation unit, a change probability prediction unit, and a Token difference modulation unit. The token difference calculation unit calculates the first phase token sequence. With the second phase token sequence Perform element-wise absolute difference operation at the corresponding Token position to obtain the remote sensing image difference Token. The calculation formula is as follows: in, This represents element-wise absolute value operation; The change probability prediction unit for the difference token Perform a fully connected mapping and normalize using the Sigmoid function to obtain the differential modulation coefficient. The calculation formula is as follows: in, This represents the mapping operation consisting of a first fully connected layer, a ReLU layer, and a second fully connected layer. This represents the Sigmoid activation function; The Token difference modulation unit adjusts the token difference coefficient. right and Perform element-wise weighted modulation to obtain and The calculation formula is as follows: in, This indicates element-wise multiplication.

3. The Token Difference Control Module according to claim 2, characterized in that, The probability prediction unit includes a first fully connected layer, a ReLU layer, a second fully connected layer, and a Sigmoid layer; wherein the input-output dimension of the first fully connected layer is (32, 8), and the input-output dimension of the second fully connected layer is (8, 1).

4. The remote sensing image change detection method based on token difference guidance according to claim 1, characterized in that, In step S8, the difference-semantic dual-branch collaborative decoder includes a difference enhancement branch, a semantic reference branch, a difference-semantic collaborative unit, and an output prediction unit.

5. The difference-semantic dual-branch collaborative decoder according to claim 4, characterized in that, The difference enhancement branch includes Convolutional layers, BN layers, and ReLU layers; The number of input channels and output channels of the convolutional layer are both 32; The difference enhancement branch is used to construct element-wise differences and enhance difference features. The calculation formula is as follows: in, Indicates the difference representation; This represents a temporal decoding feature map; This represents the second-phase decoding feature map; Indicates the kernel size as Convolution operation; Indicates BN layer operations; This represents the ReLU activation function.

6. The difference-semantic dual-branch collaborative decoder according to claim 4, characterized in that, The semantic reference branch includes Convolutional layers, BN layers, and ReLU layers; The convolutional layer has 32 input and 32 output channels; the semantic reference branch is used to construct and enhance semantic reference features on an element-wise average basis. The calculation formula is as follows: in, It represents a semantic reference representation.

7. The difference-semantic dual-branch collaborative decoder according to claim 4, characterized in that, The difference-semantic collaboration unit includes a splicing layer, Convolutional layers, BN layers, and Sigmoid layers; The convolutional layer has 64 input channels and 32 output channels; the differential-semantic co-operation unit is used to obtain fused features. The steps are as follows: Difference characteristics With semantic reference features By splicing, we obtain joint features. The calculation formula is as follows: in, This indicates a concatenation operation; joint features enter Convolutional layers, Batch Normalization (BN) layers, and Sigmoid layers yield a difference-semantic co-weight graph. The calculation formula is as follows: in, This represents the Sigmoid activation function; The difference-semantic co-weight graph Differential characteristics With semantic reference features Weighted fusion is performed to obtain fusion features The calculation formula is as follows: in, This indicates element-wise multiplication.

8. The difference-semantic dual-branch collaborative decoder according to claim 4, characterized in that, The output prediction unit includes an upsampling layer, Convolutional layers, BN layers, ReLU layers and Convolutional layer; wherein, the The convolutional layer has 32 input channels and 32 output channels; The convolutional layer has 32 input channels and 2 output channels.

9. The remote sensing image change detection method based on token difference guidance according to claim 1, characterized in that: The changed areas include one or more of the following: building change areas, road change areas, vegetation change areas, water body change areas, bare land change areas, and disaster-damaged areas.

Images