Remote sensing image change detection system based on content feature reinforcement learning

Abstract

The invention discloses a remote sensing image change detection system based on content feature reinforcement learning, which is divided into four data processing stages, namely a multi-scale feature coding stage, a content feature focusing decoding stage, a multi-branch change difference decoding stage and a loss calculation stage based on content reinforcement. The method can effectively eliminate the interference on the change detection result caused by the difference of the image acquisition conditions, including the parameter difference of the sensor for acquiring the image data and the difference of cloud and mist and illumination when the image data is acquired, and can enhance the segmentation capability of the system on the valuable content change area. By combining multi-scale feature extraction, content feature focusing, multi-branch difference decoding and loss calculation based on content enhancement, the invention provides an efficient and robust remote sensing image change detection method.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image change detection, and in particular relates to a remote sensing image change detection system based on content feature reinforcement learning. Background Art

[0002] Remote sensing change detection is the process of identifying and quantifying changes in objects, phenomena or landscapes by comparing remote sensing images acquired at different times. This technology is important for understanding land cover changes, urban expansion, environmental transformation and natural disasters. The process is usually carried out by comparing dual-temporal or multi-temporal satellite images or aerial photos, and with the help of advanced algorithms, it can detect phenomena such as deforestation, urban expansion or coastline changes.

[0003] In the past decade, the advancement of remote sensing technology has led to a rapid increase in the number of remote sensing platforms. More importantly, the types of remote sensing platforms have become increasingly diverse, ranging from common space platforms (such as observation satellites and spacecraft) to aviation platforms (such as drones and high-altitude balloons) and even ground platforms (such as remote sensing towers and sensing vehicles). The diversity of remote sensing image sources has significantly increased the complexity of image style features. In addition to the differences between dual-phase remote sensing images caused by the diversity of remote sensing platforms, even observations from different phases of the same platform are usually conducted under different conditions. Factors such as atmospheric conditions (such as cloud cover and humidity), changes in illumination (such as sun angle and seasonal changes), sensor calibration, and platform height or trajectory offsets can have a significant impact on the observation environment. These changes in observation conditions further exacerbate the style differences between images, making it more complicated to separate meaningful content changes. How to reduce the interference of intrinsic style differences in images and focus on more valuable changes in image content is a challenging problem that needs to be solved urgently.

[0004] In the current change detection algorithm, the feature splicing method is a common means of fusing the features of dual-temporal remote sensing images. This method directly splices the features of the two images along the channel dimension and passes them to the deep neural network (DNN) for processing, relying on its strong generalization ability to learn the mapping relationship between the fused features and the change labels. This method can achieve relatively good performance, mainly due to the advantages of DNN in extracting and integrating high-dimensional features. However, the splicing method itself does not explicitly calculate and model the change area between the dual-temporal images directly, but only relies on the automatic learning ability of the network to complete the identification and matching of the change area. This indirect processing method may cause the model to pay insufficient attention to the change area, especially in complex scenes or when the style features are very different, thus affecting the accuracy of change detection. Therefore, how to guide the model to pay more attention to the change area of ​​the dual-temporal image more accurately in the process of feature fusion so as to align it with the change label more effectively is still a core problem that needs to be solved in the change detection task.

[0005] In summary, the existing remote sensing image change detection algorithms have the following problems:

[0006] (1) The diversity of remote sensing platforms and changes in observation conditions (such as atmosphere, lighting, sensor calibration, etc.) lead to significant differences in the style features of dual-temporal images. This difference increases the difficulty of distinguishing between content changes and style changes in the image. The interference of style differences reduces the model's ability to fit content features.

[0007] (2) Existing feature splicing methods fail to explicitly model the change areas of dual-temporal images and rely too much on the fitting ability of deep neural networks. This indirect processing method may cause the model to pay insufficient attention to the change areas, especially in complex scenes or when the style differences are large, which affects the accuracy of change detection. Summary of the invention

[0008] The present invention proposes a remote sensing image change detection system based on content feature reinforcement learning, aiming to solve the problem of insufficient detection accuracy of changed areas due to differences in style features in the prior art.

[0009] The technical scheme adopted by the present invention is as follows: a remote sensing image change detection system based on content feature reinforcement learning, comprising: an image acquisition module, a multi-scale feature encoding module, a content feature focus decoding module and a multi-branch change difference decoding module;

[0010] Image acquisition module, used to acquire dual-temporal remote sensing images;

[0011] The multi-scale feature encoding module is used to input dual-temporal remote sensing images, use a weight-sharing encoder to extract features from the dual-temporal remote sensing images at multiple scales, and output the encoded features at each scale;

[0012] The content feature focused decoding module is used to input the coded features at each scale, use two content feature decoders with unshared weights to perform multi-scale decoding on the coded features at each scale, and decode and output the multi-scale decoding features of the dual-temporal remote sensing image;

[0013] The multi-branch change difference decoding module is used to input the multi-scale decoding features of the dual-phase remote sensing image; the multi-scale decoding features of the dual-phase remote sensing image are fused at each scale, the fused multi-scale decoding features are decoded using the change difference decoder, and the predicted change area segmentation map is output.

[0014] Furthermore, the shared weight encoder is the first 5 stages of the pre-trained EfficientNet-B5 network, and the output of stages 2-5 is taken as the output of the encoder, where:

[0015] The 1st to 5th stages of the EfficientNet network downsample the 256*256 resolution image of the input system to 128*128, 128*128, 64*64, 32*32, and 16*16 resolutions respectively. The outputs of the dual-temporal remote sensing images in the 2nd to 5th stages are defined as { , , , }and{ , , , }.

[0016] Furthermore, the content feature decoder generates the minimum scale feature of the encoder output No processing is performed and the output is directly , , , the features of adjacent scales output by the encoder and , Input feature aggregation module to obtain content features of corresponding scale , represent or ; The content features of the dual-temporal remote sensing image output by the content feature decoder are defined as { , , , }and{ , , , };

[0017] The focuser calculates the weighted graph of the corresponding scale through the content features at each scale output by the content feature decoder .

[0018] Furthermore, the feature aggregation module is used to:

[0019] Small-scale features After transposed convolution with larger scale features Perform scale alignment;

[0020] Concatenate the two adjacent features after scale alignment in the channel dimension;

[0021] The concatenated feature map is obtained by feature fusion through a 1*1 two-dimensional convolutional layer and a residual module. .

[0022] Furthermore, the weighted graph The calculation process is:

[0023] Calculating content features of dual-temporal remote sensing images at the same scale and Cosine similarity between ;

[0024] along Dimension Pair Sum and get the content features and Cosine distance plot ;

[0025] right use The function is normalized;

[0026] The calculation formula is as follows:

[0027] ;

[0028] ;

[0029] .

[0030] Further, the change difference decoder will and Each of them passes through a CBAM module, expands to a new dimension, and then splices on this new dimension, and then passes through a 2*1*1 three-dimensional convolution layer for feature fusion, and then the fused features are combined with the weighted graph at the same scale. , get the change difference characteristics of the dual-phase image ;

[0031] The change difference decoder is , and Do the same thing: and change difference characteristics Input feature fusion module, output is Variation characteristics of the same scale ;

[0032] The change difference decoder uses a change region segmentation header to Upsample by 2 times to 256*256 through a 3*3 2D transposed convolution, and then apply the upsampled feature map The function is used to normalize the feature map, and finally the normalized feature map is binarized, with values ​​greater than 0.5 set to 1 and values ​​less than 0.5 set to 0, and the final segmentation map of the changed area is output.

[0033] Furthermore, the feature blending module is specifically used for:

[0034] Will Through a CBAM module respectively;

[0035] Will Expand a new dimension respectively and splice on this new dimension;

[0036] The concatenated feature maps are fused through a 2*1*1 three-dimensional convolutional layer;

[0037] The feature map after feature fusion and the input change difference feature Input the feature aggregation module;

[0038] Compare the feature map output by the feature aggregation module with the weighted map of the corresponding scale Multiply and get Variation characteristics of the same scale .

[0039] Furthermore, the content feature focused decoding module is provided with a focuser;

[0040] The focuser is used to utilize Focusing on the changed area and unchanged area of ​​the dual-temporal remote sensing image, respectively, we obtain the changed content feature set and unchanged content feature set of image one and image two.

[0041] Furthermore, it also includes a loss calculation module based on content enhancement, which is used to calculate the main loss using the change region segmentation map and the change region label on the main branch ; Calculate the change content feature loss using the change content feature set of image one and image two ; Calculate the unchanged content feature loss using the unchanged content feature set of image one and image two ; Combining the above three losses, multi-dimensional constraints are imposed on the model parameters of the system;

[0042] The main loss It is obtained by calculating the mean square loss function MSELoss of the change region segmentation map and the change region label;

[0043] The change content feature loss The calculation process is:

[0044] Use a random sampler to sample the first Scale and Perform the same sampling to obtain N sampling point pairs in the change area;

[0045] Calculate the cosine similarity internally at N sampling points to obtain a 1*N matrix and ;

[0046] use and calculate , the formula is as follows:

[0047] ;

[0048] For each scale Taking the average value, we get , the formula is as follows:

[0049] ;

[0050] The unchanged content feature loss The calculation process is:

[0051] Use a random sampler to sample the first Scale and Perform the same sampling to obtain N pairs of sampling points in the unchanged area;

[0052] Calculate the cosine similarity internally at N sampling points to obtain a 1*N matrix and ;

[0053] use and calculate , the formula is as follows:

[0054] ;

[0055] For each scale Taking the average value, we get , the formula is as follows:

[0056] .

[0057] Furthermore, the comprehensive , and The method of multivariate constraints on the model parameters of the system is to set variable parameters α and β, which can be dynamically adjusted during the convergence process of the system model. The final total system loss is defined as:

[0058] .

[0059] The beneficial effects achieved by the present invention are:

[0060] The present invention provides an efficient and robust remote sensing image change detection method by combining multi-scale feature extraction, content feature focusing, multi-branch difference decoding and loss calculation based on content enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the overall structure of a remote sensing image change detection system based on content feature reinforcement learning provided by an example of the present invention;

[0062] Figure 2 is a schematic diagram of the structure of a content feature decoder provided by an example of the present invention;

[0063] Figure 3 is a schematic diagram of the structure of a focuser provided by an example of the present invention;

[0064] Figure 4 is a schematic diagram of the structure of a variation difference decoder provided by an example of the present invention;

[0065] Figure 5 It is a schematic diagram of the structure of the changed content feature loss calculation and the unchanged content feature loss calculation provided by the example of the present invention;

[0066] Figure 6 is a schematic diagram of visualization results of comparing the examples of the present invention with other methods on the CLCD data set;

[0067] Figure 7 is a schematic diagram of visualization results of the comparison between the example of the present invention and other methods on the LEVIR-CD dataset;

[0068] Figure 8 It is a schematic diagram of the visualization results of comparing the example of the present invention with other methods on the SYSU-CD dataset.

[0069] Specific implementation methods In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the remote sensing image change detection system based on content feature reinforcement learning proposed by the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0070] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0071] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0072] The embodiment of the present invention proposes a remote sensing image change detection system based on collaborative similarity learning, which mainly solves two problems: first, due to the common differences in imaging conditions and style changes between dual-phase remote sensing images, these unpredictable interferences will significantly weaken the model's ability to accurately identify the changed area, thereby reducing the overall detection effect; second, although the feature splicing method widely used in existing change detection systems has good performance, it cannot directly calculate the changed area between dual-phase images, which makes it difficult for the model to accurately focus on the changed area during feature fusion, affecting the detection accuracy.

[0073] The embodiment of the present invention provides a remote sensing image change detection system based on collaborative similarity learning, see Figure 1. The remote sensing image change detection system includes: a multi-scale feature encoding stage, a content feature focused decoding stage, a multi-branch change difference decoding stage, and a content enhancement-based loss calculation stage. Among them, the multi-scale feature encoding stage extracts the multi-scale features of the dual-phase remote sensing image through a shared weight encoder; the content feature focused decoding stage uses two decoders that do not share weights to decode the multi-scale features respectively, and extracts the content features of the changed area and the unchanged area through a focuser; the multi-branch change difference decoding stage fuses the multi-scale features of the dual-phase image, and generates a change area segmentation map through a change difference decoder; the loss calculation stage based on content enhancement jointly optimizes the model parameters through three losses: main loss, change content feature loss, and unchanged content feature loss to achieve multivariate constraints.

[0074] The weight-sharing encoder in the multi-scale feature encoding stage of this embodiment is the first 5 stages of the pre-trained EfficientNet-B5 network, and the outputs of stages 2-5 are taken as the outputs of the encoder. Table 1 shows the specific structure of the encoder.

[0075] Table 1 Results of different models on two datasets:

[0076]

[0077] The resolution of the original image input to the multi-scale feature encoding stage is 256*256. After multi-scale feature encoding, the resolution features of scales 128*128, 128*128, 64*64, 32*32, and 16*16 are obtained respectively. The output encoding features of the 2nd to 5th stages are defined as { , , , }and{ , , , }.

[0078] See also Figure 2 , Figure 3 , Figure 2 is a structural diagram of a content feature decoder provided by an embodiment of the present invention, Figure 3 : is a structural diagram of a focuser provided by an embodiment of the present invention. The content feature decoder of this embodiment includes a feature aggregation module and a focuser, wherein the content feature decoder integrates adjacent scale features output by the encoder through the feature aggregation module to generate a content feature set of a dual-phase image and a corresponding weighted graph; the focuser re-weights the content feature set using the weighted graph, and finally outputs a changed content feature set and an unchanged content feature set.

[0079] It should be noted that if Figure 2As shown, the content feature decoder of this embodiment is specifically used to: and , without being processed by the feature aggregation module, the direct output is equal to and For the coded features and ( ), since the two encoding features are processed in the same way, represent ( ), the specific processing method is:

[0080] Will The transposed convolution is upsampled by a factor of 2, and The scales are the same, and the first processing result is obtained;

[0081] Will splicing with the first processing result in the channel dimension to obtain a second processing result;

[0082] The second processing result is subjected to feature fusion through a two-dimensional convolution layer with a convolution kernel size of 1*1 to obtain a third processing result;

[0083] Finally, the third processing result is input into the residual module, which is composed of two residual layers linearly spliced ​​together to suppress the gradient disappearance caused by the depth of the model and obtain the content features. .

[0084] Furthermore, if Figure 3 As shown, the focuser of this embodiment is specifically used for:

[0085] calculate and The cosine distance of ;

[0086] Will and and Multiply them to get the content characteristics of the changes at the same scale and ;

[0087] Will and and Multiply them together to get the unchanged content features at the same scale and .

[0088] The processing of the focuser can also be expressed as:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] The designed focuser allows the system to flexibly focus on the content features of the changing area and the content features of the unchanging area, and to perform multivariate constraints on the model parameters of the system in the later stage.

[0097] See also Figure 4 , Figure 4 This is a schematic diagram of the structure of a variation difference decoder provided by an embodiment of the present invention. The variation difference decoder uses a content feature set and a weighted graph at each scale to input a feature blending module, and outputs variation difference features at each scale. This embodiment includes a feature blending module and a variation region segmentation head, wherein the feature blending module is used to blend the content features of dual-phase images at the same scale; the variation region segmentation head is used to convert the final variation difference features into Binarization is performed and the final segmentation map of the changed area is output.

[0098] It should be noted that if Figure 4 As shown, the feature blending module of this embodiment is specifically used for:

[0099] Will and ( ) Through a CBAM module, spatial and channel attention are introduced to enhance the system's understanding of features and obtain the first processing result;

[0100] The first processing results are expanded by one dimension respectively, and features are concatenated on the dimension to obtain the second processing result. The original feature with the dimension of N*C*W*H is changed to the second processing result of 2*N*C*W*H;

[0101] Finally, the second processing result is passed through a three-dimensional convolution layer with a convolution kernel size of 2*1*1 to blend the image features of the two phases, and the blended feature map is combined with the weighted map Multiply them together and output the change difference characteristics at this scale .

[0102] Furthermore, the change region segmentation head of this embodiment performs a change difference feature Upsample twice to 256*256 scale by 3*3 transposed convolution and apply the upsampled feature map The function performs normalization, and finally sets the pixel values ​​in the feature map greater than 0.5 to 1, and the pixel values ​​less than 0.5 to 0, and then outputs the final segmentation map of the changed area.

[0103] In the loss calculation stage based on content enhancement, the main loss is calculated separately , Variation content feature loss and unchanged content feature loss , the three losses are combined to multi-constrain the model parameters of the system, and the final total loss can be expressed as the following formula:

[0104]

[0105] Among them, the main loss It is obtained by calculating the mean square loss function MSELoss of the change region segmentation map and the change region label;

[0106] It should be noted that if Figure 5 As shown, the change content feature loss of this embodiment The specific calculation process is:

[0107] Use a random sampler to sample the first Scale and Perform the same sampling to obtain N sampling point pairs of the change area. In this embodiment, in order to consider the balance between calculation efficiency and calculation accuracy, N is set to the scale of the change content feature. For example, if the scale of the change content feature is 256*256, then N is set to 256.

[0108] Calculate the cosine similarity internally at N sampling points to obtain a 1*N matrix and ;

[0109] use and calculate , which can be expressed as:

[0110]

[0111] For each scale Taking the average value, we get , which can be expressed as:

[0112]

[0113] Similarly, Figure 5As shown, the change content feature loss of this embodiment The specific calculation process is:

[0114] Use a random sampler to sample the first Scale and Perform the same sampling to obtain N pairs of sampling points in the unchanged area. The selection of N is the same as The calculation process is the same;

[0115] Calculate the cosine similarity internally at N sampling points to obtain a 1*N matrix and ;

[0116] use and calculate , which can be expressed as:

[0117]

[0118] For each scale Taking the average value, we get , which can be expressed as:

[0119]

[0120] See also Figure 6 , Figure 7 and Figure 8 , Figure 6 This is a visualization diagram of the system testing results on the CLCD dataset using different methods. Figure 7 This is a visualization diagram of the system testing results on the LEVIR-CD dataset using different methods. Figure 8 This is a visualization diagram of the system test results on the SYSU-CD dataset using different methods. The red areas in the three figures are incorrectly predicted change areas, the green areas are incorrectly predicted unchanged areas, the white areas are correctly predicted change areas, and the black areas are correctly predicted unchanged areas. The first two columns are dual-temporal remote sensing images, the third and fourth columns are the prediction results of this system, and the fifth to last columns are the prediction results of other systems. Figure 6 , Figure 7 and Figure 8The results shown in the embodiment of the present invention indicate that the remote sensing image change detection system based on content feature reinforcement learning can detect the changed area (white area) with a high accuracy, effectively solving the problem that the style feature differences caused by the diversity of remote sensing platforms and changes in observation conditions interfere with the content change identification. At the same time, by explicitly modeling the changed areas of dual-phase images, the system overcomes the problem of insufficient attention to the changed areas and decreased detection accuracy caused by the existing methods' over-reliance on indirect fitting of deep networks.

[0121] Furthermore, Tables 2, 3 and 4 show the comparison results of various indicators of the change detection system of this embodiment with other algorithms on the CLCD, LEVIR-CD and SYSU-CD data sets, and the evaluation indicators are F1 score, IoU (intersection over union), recall and precision.

[0122] Table 2 Comparison results of the change detection system of this embodiment with other algorithms on the CLCD dataset:

[0123]

[0124] Table 3 Comparison results of the change detection system of this embodiment with other algorithms on the LEVIR-CD dataset:

[0125]

[0126] Table 4 Comparison results of the change detection system of this embodiment with other algorithms on the SYSU-CD dataset:

[0127]

[0128] It can be seen from Table 2, Table 3 and Table 4 that the remote sensing image change detection system based on content feature reinforcement learning of the embodiment of the present invention has achieved the best results on the three open data sets used for change detection.

[0129] Experiments have proved that the remote sensing image change detection system based on content feature reinforcement learning of the embodiment of the present invention solves the problem of content change identification being interfered with by style feature differences, and overcomes the excessive reliance of existing methods on indirect fitting of deep networks by explicitly modeling the change areas.

[0130] In summary, remote sensing image change detection has important application value in many fields. For example, in the agricultural field, remote sensing image change detection can be used to monitor changes in cultivated land use, help optimize land resource allocation, and improve agricultural management efficiency. In addition, remote sensing images can also be used to monitor ecological and environmental changes, such as reduced forest cover and spread of water pollution, to provide a scientific basis for ecological protection and governance. At the same time, in disaster prevention and mitigation, by analyzing pre- and post-disaster remote sensing images, the degree of damage to the ecosystem caused by landslides or fires can be assessed, providing important data support for post-disaster recovery and reconstruction. In general, remote sensing image change detection can provide efficient and accurate data support and decision-making reference for many fields. With the continuous advancement of remote sensing technology, its application scope and detection accuracy will continue to improve.

[0131] In the several embodiments provided by the present invention, it should be understood that the apparatus and method disclosed by the present invention can be implemented in other ways. For example, the apparatus embodiments described above are only illustrative, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0132] In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0133] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A remote sensing image change detection system based on content feature reinforcement learning, characterized by: It includes image acquisition module, multi-scale feature encoding module, content feature focus decoding module and multi-branch variation difference decoding module; Image acquisition module, used to acquire dual-temporal remote sensing images; The multi-scale feature encoding module is used to input dual-temporal remote sensing images, use a weight-sharing encoder to extract features from the dual-temporal remote sensing images at multiple scales, and output the encoded features at each scale; The content feature focused decoding module is used to input the coded features at each scale, use two content feature decoders with unshared weights to perform multi-scale decoding on the coded features at each scale, and decode and output the multi-scale decoding features of the dual-temporal remote sensing image; Multi-branch change difference decoding module, used to input multi-scale decoding features of dual-temporal remote sensing images; The multi-scale decoding features of the dual-temporal remote sensing images are fused at each scale, and the fused multi-scale decoding features are decoded using a change difference decoder to output a predicted change area segmentation map.

2. The remote sensing image change detection system based on content feature reinforcement learning according to claim 1, characterized in that: The shared weight encoder is the first 5 stages of the pre-trained EfficientNet-B5 network, and the output of stages 2-5 is taken as the output of the encoder, where: The 1st to 5th stages of the EfficientNet network downsample the 256*256 resolution image of the input system to 128*128, 128*128, 64*64, 32*32, and 16*16 resolutions respectively. The outputs of the dual-temporal remote sensing images in the 2nd to 5th stages are defined as { , , , }and{ , , , }.

3. The remote sensing image change detection system based on content feature reinforcement learning according to claim 2 is characterized in that: The content feature decoder is the minimum scale feature of the encoder output No processing is performed and the output is directly , , , the features of adjacent scales output by the encoder and , Input feature aggregation module to obtain content features of corresponding scale , represent or ; The content features of the dual-temporal remote sensing image output by the content feature decoder are defined as { , , , }and{ , , , }; The focuser calculates the weighted graph of the corresponding scale through the content features at each scale output by the content feature decoder .

4. The remote sensing image change detection system based on content feature reinforcement learning according to claim 3 is characterized in that: The feature aggregation module is used to: Small-scale features After transposed convolution with larger scale features Perform scale alignment; Concatenate the two adjacent features after scale alignment in the channel dimension; The concatenated feature map is obtained by feature fusion through a 1*1 two-dimensional convolutional layer and a residual module. .

5. The remote sensing image change detection system based on content feature reinforcement learning according to claim 3 is characterized in that: The weighted graph The calculation process is: Computing content features of dual-temporal remote sensing images at the same scale and Cosine similarity between ; along Dimension Pair Sum and get the content features and Cosine distance plot ; right use The function is normalized; The calculation formula is as follows: ； ； 。 6. The remote sensing image change detection system based on content feature reinforcement learning according to claim 5 is characterized in that: The change difference decoder will and Each of them passes through a CBAM module, expands to a new dimension, and then splices on this new dimension, and then passes through a 2*1*1 three-dimensional convolution layer for feature fusion, and then the fused features are combined with the weighted graph at the same scale. , get the change difference characteristics of the dual-phase image ; The change difference decoder is , and Do the same thing: and change difference characteristics Input feature fusion module, output is Variation characteristics of the same scale ; The change difference decoder uses a change region segmentation header to Upsample by 2 times to 256*256 through a 3*3 2D transposed convolution, and then apply the upsampled feature map The function is used to normalize the feature map, and finally the normalized feature map is binarized, with values ​​greater than 0.5 set to 1 and values ​​less than 0.5 set to 0, and the final segmentation map of the changed area is output.

7. The remote sensing image change detection system based on content feature reinforcement learning according to claim 6 is characterized in that: The feature fusion module is specifically used for: Will Through a CBAM module respectively; Will Expand a new dimension respectively and splice on this new dimension; The concatenated feature maps are fused through a 2*1*1 three-dimensional convolutional layer; The feature map after feature fusion and the input change difference feature Input the feature aggregation module; Compare the feature map output by the feature aggregation module with the weighted map of the corresponding scale Multiply and get Variation characteristics of the same scale .

8. The remote sensing image change detection system based on content feature reinforcement learning according to claim 3 is characterized in that: The content feature focused decoding module is provided with a focuser; The focuser is used to utilize Focusing on the changed area and unchanged area of ​​the dual-temporal remote sensing image, respectively, we obtain the changed content feature set and unchanged content feature set of image one and image two.

9. The remote sensing image change detection system based on content feature reinforcement learning according to claim 8, characterized in that: It also includes a content-enhanced loss calculation module for calculating the main loss using the change region segmentation map and change region label on the main branch. ; Calculate the change content feature loss using the change content feature set of image one and image two ; Calculate the unchanged content feature loss using the unchanged content feature set of image one and image two ; Combining the above three losses, multi-dimensional constraints are imposed on the model parameters of the system; The main loss It is obtained by calculating the mean square loss function MSELoss of the change region segmentation map and the change region label; The change content feature loss The calculation process is: Use a random sampler to sample the first Scale and Perform the same sampling to obtain N sampling point pairs in the change area; Calculate the cosine similarity internally at N sampling points to obtain a 1*N matrix and ; use and calculate , the formula is as follows: ； For each scale Taking the average value, we get , the formula is as follows: ； The unchanged content feature loss The calculation process is: Use a random sampler to sample the first Scale and Perform the same sampling to obtain N pairs of sampling points in the unchanged area; Calculate the cosine similarity internally at N sampling points to obtain a 1*N matrix and ; use and calculate , the formula is as follows: ； For each scale Taking the average value, we get , the formula is as follows: 。 10. The remote sensing image change detection system based on content feature reinforcement learning according to claim 9, characterized in that: The comprehensive , and The method of multivariate constraints on the model parameters of the system is to set variable parameters α and β, which can be dynamically adjusted during the convergence process of the system model. The final total system loss is defined as: 。

Images