Change detection method and system based on multi-scale fusion and integrated kernel adaptive learning

Abstract

The invention discloses a change detection method and system based on multi-scale fusion and integrated kernel adaptive learning. The method comprises the following steps: constructing a Siamese sensitivity enhanced feature extraction network for extracting dual-tense features; establishing a multi-scale dual-time fusion attention module to fuse dual-time images of all branches; defining that a nested decoder is formed by a jump connection mode to reserve high-dimensional semantic details and fine-grained positioning details; constructing an integrated kernel adaptive learning module to fuse output features of four levels, and fusing multi-scale semantic information; and extracting an optimization result, and outputting a change region. According to the method, the false change phenomenon caused by feature redundancy in the feature extraction process can be well relieved, important components in the dual-time features can be more effectively recognized, the feature expression diversity is enhanced, and the omission ratio in a small change area is reduced by capturing subtle changes in the image.

Description

Technical Field

[0001] The present invention belongs to the field of change detection, and in particular relates to a change detection method and system based on multi-scale fusion and integrated kernel adaptive learning. Background Art

[0002] Change detection is a key technology in the field of remote sensing. It determines the changed area by analyzing dual-phase images of the same area obtained at different time points. It plays an important role in many practical applications such as land resource monitoring, natural disaster assessment, urban expansion survey, forest monitoring and environmental monitoring. With the advancement of satellite and radar sensor technology, the spatiotemporal resolution of images has been continuously improved, providing richer image features and ground information, but also increasing the complexity and difficulty of change detection tasks. Therefore, accurate and effective change detection is of great significance to protecting earth resources.

[0003] Traditional change detection methods can be divided into pixel-based change detection algorithms and object-based change detection algorithms according to the different detection units. The pixel-based change detection algorithm compares the pixel differences between the dual-time images to obtain the change map, and then performs threshold segmentation or clustering methods to obtain the detection results; the object-based algorithm segments the image, and then compares the classification results to obtain the change detection results. However, these methods also have high requirements for the registration of images at different times, have high computational costs, and often fall into the accumulation of multiple classification errors, affecting the detection accuracy. With the emergence of deep learning technology, many neural networks and components have been used in the field of change detection. Compared with traditional methods, deep learning methods have better modeling capabilities.

[0004] Although deep learning-based methods have made great progress, they still have some problems. First, although convolutional neural networks have achieved excellent performance in change detection, they consume huge computing resources, and the problem of extracting redundant features in high-resolution image processing is more significant, which increases the computational complexity and reduces the recognition accuracy. In addition, when fusing bi-phase features, existing methods only focus on the important parts of the features at each moment, while ignoring the important parts of the features across the bi-phases, which affects the accuracy of change detection. Secondly, there is a complementary relationship between deep features and shallow features. Current methods tend to use ordinary convolution to fuse deep and shallow features, and cannot discover the internal relationship between them well. Summary of the invention

[0005] The object of the present invention is to provide a change detection method and system based on multi-scale fusion and integrated kernel adaptive learning.

[0006] The technical solution to achieve the purpose of the present invention is as follows: In a first aspect, the present invention provides a change detection method based on multi-scale fusion and integrated kernel adaptive learning, comprising the following steps:

[0007] The first step is to construct a Siamese sensitivity enhanced feature extraction network to extract bi-temporal features;

[0008] In the second step, a multi-scale dual-temporal fusion attention module is established to fuse the dual-temporal images of each branch;

[0009] The third step is to define a nested decoder consisting of skip connections to preserve high-dimensional semantic details and fine-grained localization details;

[0010] The fourth step is to construct an integrated kernel adaptive learning module to fuse the output features of the four levels at the end of the decoder and integrate multi-scale semantic information;

[0011] In the fifth step, the contrast loss function is introduced as a measure between the change map and the true label map, and the neural network is trained by minimizing the loss.

[0012] In a second aspect, the present invention provides a change detection system based on multi-scale fusion and integrated kernel adaptive learning, for implementing the method described in the first aspect, the system comprising:

[0013] Sensitivity enhancement feature extraction module, constructs Siamese sensitivity enhancement feature extraction network to extract bi-temporal features. The inputs at both ends of the network are two registered images, which are respectively passed through the sensitivity enhancement feature extraction module to obtain feature maps of different levels.

[0014] A multi-scale bi-temporal fusion attention module is used to fuse the bi-temporal images of each branch;

[0015] An integrated kernel adaptive learning module is used to fuse the output features of four levels and integrate multi-scale semantic information.

[0016] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the program.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0018] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which implements the steps of the method described in the first aspect when executed by a processor.

[0019] Compared with the prior art, the present invention has the following significant features: (1) constructing a Siamese sensitivity enhanced feature extraction network can reduce redundant features in the feature extraction process through spatial reconstruction and channel reconstruction, capture subtle changes in the image, and suppress pseudo changes caused by redundant features; (2) defining a multi-scale dual-temporal fusion attention module to compare dual-temporal features in the time dimension, and using channel and spatial attention to determine the important parts of the features; (3) establishing an integrated kernel adaptive learning module to fuse multi-scale semantic information in the feature fusion stage, enhance feature diversity, and locally filter and enhance features in space by selecting convolution cores of different sizes, thereby capturing subtle changes in the image and reducing the missed detection rate in small change areas.

[0020] The present invention is further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the change detection method based on multi-scale fusion and integrated kernel adaptive learning of the present invention.

[0022] Figure 2 It is the change detection network structure of the present invention, wherein (a) is the integrated kernel adaptive learning module of the present invention.

[0023] Figure 3 It is the sensitivity enhancement feature extraction module of the present invention.

[0024] Figure 4 It is the multi-scale dual-temporal fusion attention module of the present invention.

[0025] Figure 5 It is the kernel adaptive learning module of the present invention.

[0026] Figure 6 This is a result effect diagram of the method of the present invention on a remote sensing building change detection data set. DETAILED DESCRIPTION

[0027] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Combination Figure 1 , the implementation process of the present invention is described in detail below, the steps are as follows:

[0029] The first step is to construct a Siamese sensitivity enhanced feature extraction network to extract bi-temporal features, such as Figure 2 As shown in the figure, the inputs at both ends of the network are two registered images, which are respectively passed through the sensitivity enhancement feature extraction module to obtain feature maps of different levels. The specific process is:

[0030] (1) Construct a sensitivity enhancement feature extraction module, which consists of a standard convolution kernel, a sensitivity enhancement convolution, and a batch normalization module BN. The above parts form a 6-layer structure through residual connections, as shown in Figure 3 As shown in the figure, the specific process is that the remote sensing image passes through the standard convolution kernel and batch normalization module BN in turn, is processed by the activation function ReLU, then passes through the sensitivity enhancement convolution and batch normalization module BN, and is superimposed with the output of the standard convolution kernel, and is processed again by the activation function ReLU;

[0031] The standard convolution kernel is a 3×3×C convolution kernel, that is, a convolution kernel with the same number of channels as the original data and a length and width of 3×3; the sensitivity enhancement convolution (SC) includes two units, namely the spatial reconstruction unit (SRU) and the channel reconstruction unit (CRU). First, the SRU operation generates a spatial refinement feature x w , then the CRU operation generates the channel refinement feature Y;

[0032] (2) Construct a Siamese sensitivity enhancement feature extraction network, that is, expand the constructed sensitivity enhancement feature extraction module into a Siamese structure, define the network input as two aligned dual-temporal remote sensing images, and obtain feature maps of different levels by passing through multiple sensitivity enhancement feature extraction modules.

[0033] The second step is to build a multi-scale dual-temporal fusion attention module model, such as Figure 4 As shown, the specific process is:

[0034] (1) Establish a multi-scale dual-time fusion attention module model, define the input data as T1, T2, and then define the feature function as follows:

[0035] Output = MDFA (T1, T2)

[0036] Where T1 and T2 represent the input dual temporary features, and MDFA(T1, T2) represents the obtained feature map;

[0037] (2) Introduce a multi-scale feature extraction strategy, use 3×3, 5×5 and 7×7 convolutions to process the input dual-phase features T1 and T2 to obtain feature information at different scales, and then add the different convolution results to form a multi-scale feature map T1 ′ , T2 ′ , thereby enhancing the feature expression ability of the model;

[0038] (3) Spatial information is aggregated through global aggregation across spatial dimensions. The aggregation process can be expressed as:

[0039] F c =Concat(Avg(T1 ′ ),Max(T1′ ),Avg(T2 ′ ),Max(T2 ′ ))

[0040] Among them, F c represents the aggregated spatial features, Avg(·) and Max(·) represent the global average pooling and global maximum pooling across spatial dimensions, respectively;

[0041] (4) The fused channel features are adaptively determined by the number of kernels (K) through one-dimensional convolution, effectively capturing the correlation between channels. Then, the nonlinear activation function Softmax is used to generate the attention weight W for the channel. c1 and W c2 :

[0042]

[0043] Where Conv1(·) and Conv2(·) represent one-dimensional convolution;

[0044] (5) Determine the weights in the spatial dimension, that is, use global pooling in the spatial dimension and use two-dimensional convolution and nonlinear activation function Softmax to generate attention weights W for the space s1 and W s2 ;

[0045] (6) The bi-phase weights are obtained by summarizing the bi-phase channel weights and the bi-phase spatial weights, thereby determining the important parts between the bi-phase features. Finally, the bi-phase weights are multiplied by the bi-phase features and summarized to effectively fuse the bi-phase features:

[0046] Output=(W c1 +W s1 )*T1+(W c2 +W s2 )*T2

[0047] When the sum of the bi-phase weights is equal to 1, the useful part between the bi-phase features is retained, while the useless part is discarded, thus achieving effective feature fusion.

[0048] In the third step, a nested decoder is defined consisting of skip connections to preserve high-dimensional semantic details and fine-grained localization details.

[0049] The decoder consists of multiple layers, each of which contains multiple decoding units. These decoding units process the features extracted from different levels and gradually decode the feature maps. The high-level decoding units are located in the later stage of the decoder, responsible for processing large-scale, low-resolution features and gradually decoding them into higher-resolution outputs. The deeper decoding units are located in the early stage of the decoder, processing high-resolution, fine-grained features from the encoder, focusing on the restoration of image details. During the decoding process, the high-level decoding units not only receive input from the decoding units of the previous level, but also receive input from the deeper decoding units (i.e., the previous level) through jump connections. This jump connection allows information of different scales to be fused with each other, enhancing the expressiveness of the features; through multi-level decoding and jump connections, the network can better integrate features from different scales, thereby effectively retaining the high-dimensional semantic information and fine-grained positioning information of the image.

[0050] The fourth step is to construct an integrated kernel adaptive learning module to fuse the output features of the four levels and fuse the multi-scale semantic information. The specific process is:

[0051] (1) Construct a kernel adaptive learning module, such as Figure 5 As shown, the input data is defined as F, then the characteristic function is defined as follows:

[0052] F out =KAL(F)

[0053] Where F represents the input image features, and KAL(F) represents the obtained feature map. The steps are as follows:

[0054] ① Through 1×1, 3×3, 5×5 and 7×7 convolutions, four new feature sets F with different spatial information are obtained a 、F v 、F c and F d , splicing F a 、F v 、F c and F d Get feature F u ;

[0055] ② For F u Perform global average pooling to obtain global information to generate M, and M is passed through the first fully connected layer to reduce the number of channels to obtain N;

[0056] ③ Pass N to the second fully connected layer, increase its number of channels, and then divide it into four blocks in the channel dimension, perform Softmax operation at the corresponding position, and obtain weights a, b, c, and d from convolution kernel branches of different sizes. Where a+b+c+d=1, note that weights a, b, c, and d can be adaptively adjusted according to the different sizes and shapes of objects, allowing different spatial information to be captured;

[0057] ④F a 、F b 、F c and F d After being input into the CBAM module for feature enhancement, they are multiplied by their corresponding attention weights a, b, c and d to obtain the adaptive spatial feature map F. a ′ 、F b ′ 、F c ′ and F d ′ ;

[0058] ⑤Concatenate the above four adaptive feature maps to get the final result:

[0059] F out =F a ′ +F b ′ +F c ′ +F d ′

[0060] (2) Construct an integrated kernel adaptive learning module. The steps are as follows:

[0061] First, the four groups of outputs of the backbone network are summed, and then the intra-group relations are extracted through the kernel adaptive learning module;

[0062] M intra =KAL(x 0,1 +x 0,2 +x 0,3 +x 0,4 )

[0063] At the same time, the four groups of outputs are connected and then connected to another group of kernel adaptive learning modules to extract the intra-group relations;

[0064] F ensemble =[x 0,1 ,x 0,2 ,x 0,3 ,x 0,4 ]

[0065] M inter =KAL(F ensemble )

[0066] Where [] represents the concatenation of multiple sets of feature maps;

[0067] The final integrated kernel adaptive learning module output is:

[0068]

[0069] The function repeat (n) (·) represents the operation of repeating the attention map n times and concatenating it in the channel dimension, Represents element-wise product.

[0070] In the fifth step, the contrast loss function is introduced as a measure between the change map and the true label map, and the neural network is trained by minimizing the loss. The loss function is defined as follows:

[0071]

[0072] in is the weighted cross entropy loss, is the overlap combined loss;

[0073]

[0074] H and W represent the change diagram The height and width of the class, weight[class] represents the weight corresponding to the class, Indicates that the predicted value at position k belongs to the score of category class, Indicates the score that the predicted value at position k belongs to class l, and the value of "class" is 0 or 1, corresponding to "unchanged" and "changed" pixels respectively;

[0075]

[0076] in The overlap loss is calculated through the softmax() layer, and Y is the true value.

[0077] The present invention defines a Siamese sensitivity enhanced feature extraction network for extracting dual-temporal features, alleviating the pseudo-change phenomenon caused by feature redundancy in the feature extraction process; establishes a multi-scale dual-temporal fusion attention module to fuse the dual-temporal images of each branch, and more effectively identifies the important components in the dual-temporal features; constructs an integrated kernel adaptive learning module to fuse the output features of four levels, enhance the diversity of feature expression, and reduce the missed detection rate in small change areas.

[0078] The effect of the present invention can be further illustrated by the following simulation experiment:

[0079] (1) Simulation conditions

[0080] The simulation experiment uses a remote sensing building change detection dataset. The dataset consists of 637 pairs of 1024×1024 pixel high-resolution images, focusing on the change information of buildings such as high-rise apartments, villas, large warehouses, and small garages. In the experiment, the 1024×1024 pixel images were cropped into non-overlapping images of 256×256 pixels and classified into 7120 / 1024 / 2048 pairs of images for training, validation, and testing, respectively. We compared the proposed method with some fully supervised change detection methods. The simulation experiments were all configured with an Inteli7-8700KCPU (3.7GHz) and NVIDARTX3090GPU under the Linux operating system, and the program was written in Python and PyCharm 2023.

[0081] The evaluation indicators used in the present invention are precision, recall, F1-Score, and average intersection-over-Union (IOU).

[0082] (2) Simulation content

[0083] The present invention uses a remote sensing building change detection data set to test the performance of the algorithm. In order to test the performance of the algorithm of the present invention, the proposed change detection method based on multi-scale fusion and integrated kernel adaptive learning is compared with the currently popular change detection algorithms in the world. The comparison methods include: FC-EF, FC-Siam-diff, FC-Siam-conc, BITNet, and SNUNet.

[0084] (3) Analysis of simulation results

[0085] Table 1 shows the comparison results of different evaluation indicators under different change detection algorithms under the remote sensing building change detection dataset.

[0086] Table 1 Quantitative evaluation of different algorithms on the dataset

[0087]

[0088]

[0089] It can be seen from Table 1 that in the remote sensing building change detection dataset, the change detection method based on multi-scale fusion and integrated kernel adaptive learning proposed in the present invention has strong adaptability, can highlight the changed area, and avoid the influence of unchanged buildings. Compared with FC-EF, FC-Siam-diff, FC-Siam-conc, BITNet, and SNUNet, the accuracy in different evaluation indicators is significantly improved. The results of the method of the present invention and the comparative method on the remote sensing building change detection dataset are shown in the figure below. Figure 6 The simulation experimental results of the above real data sets demonstrate the effectiveness of the method of the present invention.

[0090] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A change detection method based on multi-scale fusion and integrated kernel adaptive learning, characterized in that: The following steps are involved: The first step is to construct a Siamese sensitivity enhancement feature extraction network to extract bi-temporal features. The inputs at both ends of the network are two registered images, which are respectively passed through the sensitivity enhancement feature extraction module to obtain feature maps of different levels. In the second step, a multi-scale dual-temporal fusion attention module is established to fuse the dual-temporal images of each branch; The third step is to define a nested decoder consisting of skip connections to preserve high-dimensional semantic details and fine-grained localization details; The fourth step is to construct an integrated kernel adaptive learning module to fuse the output features of the four levels at the end of the decoder and integrate multi-scale semantic information; In the fifth step, the contrast loss function is introduced as a measure between the change map and the true label map, and the neural network is trained by minimizing the loss.

2. The change detection method based on multi-scale fusion and integrated kernel adaptive learning according to claim 1, characterized in that: The first step is as follows: Construct a sensitivity enhancement feature extraction module, which consists of a standard convolution kernel, a sensitivity enhancement convolution and a batch normalization module BN. The above parts are connected through residual connections to form a 6-layer structure. The specific process is that the remote sensing image passes through the standard convolution kernel and the batch normalization module BN in turn, is processed by the activation function ReLU, then passes through the sensitivity enhancement convolution and batch normalization module BN, and is superimposed with the output of the standard convolution kernel, and is processed again by the activation function ReLU; The standard convolution kernel is a 3×3×C convolution kernel, that is, a convolution kernel with the same number of channels as the original data and a length and width of 3×3; the sensitivity enhancement convolution includes two units, namely the spatial reconstruction unit SRU and the channel reconstruction unit CRU; first, the SRU operation generates a spatial refinement feature X w , then the CRU operation generates the channel refinement feature Y; A Siamese sensitivity enhancement feature extraction network is constructed, that is, the constructed sensitivity enhancement feature extraction module is expanded into a Siamese structure, and the input of the network is defined as two aligned dual-temporal remote sensing images, which are respectively passed through multiple sensitivity enhancement feature extraction modules to obtain feature maps of different levels.

3. The change detection method based on multi-scale fusion and integrated kernel adaptive learning according to claim 1, characterized in that: The specific steps of the second step are as follows: (1) Establish a multi-scale dual-time fusion attention module model, define the input data as T1, T2, and then define the feature function as follows: Output = MDFA (T1, T2) Where T1 and T2 represent the input dual temporary features, and MDFA(T1, T2) represents the obtained feature map; (2) A multi-scale feature extraction strategy is introduced. 3×3, 5×5 and 7×7 convolutions are used to process the input dual-phase features T1 and T2 to obtain feature information at different scales. Then, the results of different convolutions are added together to form multi-scale feature maps T′1 and T′2, thereby enhancing the feature expression ability of the model. (3) Aggregate spatial information through global aggregation across spatial dimensions. The aggregation process is expressed as: F c =Concat(Avg(T′1),Max(T′1),Avg(T′2),Max(T′2)) Among them, F c represents the aggregated spatial features, Avg(·) and Max(·) represent the global average pooling and global maximum pooling across spatial dimensions, respectively; (4) The fused channel features are adaptively determined by the number of kernels through one-dimensional convolution to capture the correlation between channels; then the nonlinear activation function Softmax is used to generate the attention weight W for the channel c1 and W c2 : Where Conv1(·) and Conv2(·) represent one-dimensional convolution; (5) Determine the weights in the spatial dimension, that is, use global pooling in the spatial dimension and use two-dimensional convolution and nonlinear activation function Softmax to generate attention weights W for the space s1 and W s2 ; (6) The bi-phase channel weights and the bi-phase spatial weights are summed up to obtain the bi-phase weights, thereby determining the important parts between the bi-phase features; finally, the bi-phase weights are multiplied by the bi-phase features and summed up to effectively fuse the bi-phase features: Output=(W c1 +W s1 )*T1+(W c2 +W s2 )*T2 When the sum of the bi-phase weights is equal to 1, the useful part between the bi-phase features is retained, while the useless part is discarded, thus achieving effective feature fusion.

4. The change detection method based on multi-scale fusion and integrated kernel adaptive learning according to claim 1, characterized in that: The specific steps of the third step are as follows: The decoder consists of multiple layers, each of which contains multiple decoding units. These decoding units process the features extracted from different levels respectively and gradually decode the feature maps. During the decoding process, the high-level decoding units not only receive input from the decoding units at the previous level, but also receive input from the deeper decoding units through jump connections. Through multi-level decoding and jump connections, features of different scales are integrated to retain the high-dimensional semantic information and fine-grained positioning information of the image.

5. The change detection method based on multi-scale fusion and integrated kernel adaptive learning according to claim 1, characterized in that: The specific steps of the fourth step are as follows: (1) Construct a kernel adaptive learning module and define the input data as F. Then define the feature function as follows: <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> out <h2 style=";text-align:left;direction:ltr"> =KAL(F) Where F represents the input image features, KAL(F) represents the obtained feature map, and the steps are as follows: Four new feature sets F with different spatial information are obtained through 1×1, 3×3, 5×5 and 7×7 convolutions. a 、F b 、F c and F d , splicing F a 、F b 、F c and F d Get feature F u ; F u Perform global average pooling to obtain global information to generate M, and M is passed through the first fully connected layer to reduce the number of channels to obtain N; Pass N to the second fully connected layer to increase its number of channels, and then divide it into four blocks in the channel dimension, perform Softmax operation at the corresponding position, and obtain weights a, b, c, and d from convolution kernel branches of different sizes; where a+b+c+d=1, note that weights a, b, c, and d are adaptively adjusted according to the different sizes and shapes of objects, allowing different spatial information to be captured; F a 、F b 、F c and F d After being input into the CBAM module for feature enhancement, they are multiplied by their corresponding attention weights a, b, c and d to obtain the adaptive spatial feature map F′ a , F′ b , F′ c and F′ d ; The above four adaptive feature maps are concatenated to get the final result: F out =F′ a +F′ b +F′ c +F′ d (2) Construct an integrated kernel adaptive learning module. The steps are as follows: First, the four groups of outputs of the backbone network are summed, and then the intra-group relations are extracted through the kernel adaptive learning module; M intra =KAL(x 0,1 +x 0,2 +x 0,3 +x 0,4 ) The four groups of outputs are connected together and then connected to another group of kernel adaptive learning modules to extract the intra-group relations; F ensemble =[x 0,1 ,x 0,2 ,x 0,3 ,x 0,4 ] M inter =KAL(F ensemble ) Where [] represents the concatenation of multiple sets of feature maps; The final integrated kernel adaptive learning module output is: The function repeat (n) (·) represents the operation of repeating the attention map n times and concatenating it in the channel dimension, Represents element-wise product.

6. The change detection method based on multi-scale fusion and integrated kernel adaptive learning according to claim 1, characterized in that: The specific steps of the fifth step are as follows: The contrast loss function is introduced as a measure between the change map and the true label map, and the neural network is trained by minimizing the loss; the loss function is defined as follows: in is the weighted cross entropy loss, is the overlap combined loss; H and W represent the change diagram The height and width of the class, weight[class] represents the weight corresponding to the class, Indicates that the predicted value at position k belongs to the score of category class, Indicates the score that the predicted value at position k belongs to category l, and the value of "class" is 0 or 1, corresponding to "unchanged" and "changed" pixels respectively; in The overlap loss is calculated through the softmax() layer, and Y is the true value.

7. A change detection system based on multi-scale fusion and integrated kernel adaptive learning, characterized in that: For implementing the method described in any one of claims 1 to 6, the system comprises: Sensitivity enhancement feature extraction module, constructs Siamese sensitivity enhancement feature extraction network to extract bi-temporal features. The inputs at both ends of the network are two registered images, which are respectively passed through the sensitivity enhancement feature extraction module to obtain feature maps of different levels. A multi-scale bi-temporal fusion attention module is used to fuse the bi-temporal images of each branch; An integrated kernel adaptive learning module is used to fuse the output features of four levels and integrate multi-scale semantic information.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of any method described in claims 1-6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 6 are implemented.

Images