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

Abstract

The application discloses a remote sensing image change detection system based on content feature reinforcement learning, and the system 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 application can effectively eliminate the interference caused by the difference of image acquisition conditions on the change detection result, including the parameter difference of the sensor for acquiring image data, the difference of cloud, fog and light when acquiring image data and other complex factors, and can enhance the segmentation ability of the system to the valuable content change area. By combining multi-scale feature extraction, content feature focusing, multi-branch difference decoding and loss calculation based on content reinforcement, the application provides an efficient and robust remote sensing image change detection method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image change detection, and particularly relates to a remote sensing image change detection system based on content feature reinforcement learning. BACKGROUND

[0002] Remote sensing change detection is a process that identifies and quantifies changes in objects, phenomena or landscapes by comparing remote sensing images acquired at different times. This technology is of great significance for understanding land cover change, urban expansion, environmental transformation and natural disasters. The process usually involves comparing satellite images or aerial photographs taken at different times and using advanced algorithms to detect phenomena such as deforestation, urban expansion or changes in coastlines.

[0003] Over the past decade, advances in remote sensing technology have led to a rapid increase in the number of remote sensing platforms, and more importantly, the types of remote sensing platforms have become increasingly diverse, from common space platforms (such as observation satellites, spacecraft) to aerial platforms (such as drones, high-altitude balloons), and even ground platforms (such as remote sensing towers, sensing vehicles). The diversity of remote sensing image sources significantly increases the complexity of image style features. In addition to the diversity of remote sensing platforms leading to differences between double-time remote sensing images, even if the different time observations are from the same platform, they are usually taken under different conditions. Atmospheric conditions (such as cloud cover, humidity), changes in lighting (such as sun angle, seasonal changes), sensor calibration, and platform height or trajectory offset can all significantly affect the observation environment. These changes in observation conditions further exacerbate the style differences between images, making it more complex to separate meaningful content changes. How to reduce the interference of the inherent style differences between images and focus on more valuable changes in image content is a challenging problem that needs to be solved.

[0004] In current change detection algorithms, feature concatenation methods are commonly used for feature fusion of dual temporal remote sensing images. This method directly concatenates the features of the two images along the channel dimension and processes them with a deep neural network (DNN), relying on its powerful generalization ability to learn the mapping relationship between the fused features and the change labels. This approach can achieve relatively good performance, mainly due to the advantages of DNN in extracting and integrating high-dimensional features. However, the concatenation method itself does not explicitly calculate and model the change regions between dual temporal images, relying solely on the automatic learning ability of the network to identify and match the change regions. This indirect processing method may lead to insufficient attention to the change regions by the model, especially in complex scenes or cases with large differences in style features, thereby affecting the accuracy of change detection. Therefore, how to guide the model to pay more attention to the change regions of dual temporal images during feature fusion to more effectively align with the change labels remains a key problem that needs to be addressed in the change detection task.

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

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

[0007] (2) The existing feature concatenation method does not explicitly model the change regions of dual temporal images, relying too much on the fitting ability of deep neural networks. This indirect processing method may lead to insufficient attention to the change regions by the model, especially in complex scenes or cases with large differences in style, affecting the accuracy of change detection. SUMMARY

[0008] The present application proposes a remote sensing image change detection system based on content feature reinforcement learning, aiming to solve the problem of insufficient change region detection accuracy caused by style feature differences in the prior art.

[0009] The technical scheme adopted by the present application 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 focusing decoding module, and a multi-branch change difference decoding module;

[0010] The image acquisition module is used to acquire dual temporal remote sensing images.

[0011] The multi-scale feature encoding module is used to input dual temporal remote sensing images, and a shared weight encoder is used to extract features of the dual temporal remote sensing images at multiple scales, outputting encoded features at each scale.

[0012] The content feature focus decoding module is configured to input the encoded features at each scale, decode the encoded features at each scale in a multi-scale manner by using two content feature decoders that do not share weights, and output multi-scale decoded features of the dual-temporal remote sensing image.

[0013] The multi-branch change difference decoding module is configured to input the multi-scale decoded features of the dual-temporal remote sensing image, fuse the multi-scale decoded features of the dual-temporal remote sensing image at each scale, decode the fused multi-scale decoded features by using a change difference decoder, and output a predicted change region segmentation map.

[0014] Further, the shared-weight encoder is the first 5 stages of a pre-trained EfficientNet-B5 network, and the outputs of the 2nd to 5th stages are taken as the outputs of the encoder, wherein,

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

[0016] Further, the content feature decoder does not process the smallest scale feature output by the encoder and directly outputs , , , The features of adjacent scales output by the encoder and , are input into a feature aggregation module to obtain the content features of the corresponding scale , representing or ; the content features of the dual-temporal remote sensing image output by the content feature decoder are defined as , , , } and , , , } respectively.

[0017] The focuser calculates the lower weighting map of the corresponding scale through the content feature of each scale output by the content feature decoder .

[0018] Further, the feature aggregation module is used for:

[0019] The features of small scales are transposed and convolved with the features of larger scales After transposed convolution, the features of small scales are scale-aligned with the features of larger scales

[0020] The two adjacent features after scale alignment are spliced in the channel dimension

[0021] The spliced feature map is fused through a 1*1 two-dimensional convolution layer and a residual module to obtain the feature map of the next scale .

[0022] Further, the weighting map is calculated as follows:

[0023] The content features of the double-time-phase remote sensing images of the same scale are calculated The cosine similarity between ;

[0024] The sum of in the dimension is obtained The cosine distance map between and ;

[0025] The is normalized by using the function

[0026] The formula of the calculation process is as follows:

[0027] ;

[0028] ;

[0029] .

[0030] Further, the change difference decoder respectively expands a new dimension through a CBAM module for and , then splices them in the new dimension, and then fuses the features through a 2*1*1 three-dimensional convolution layer, and then obtains the change difference features of the double-time-phase images by fusing the fused features with the weighting map of the same scale;

[0031] ​​The change difference decoder , and Perform the same process: and differences in characteristics Input feature fusion module, output the same as Differences in changes at the same scale ;

[0032] The change difference decoder uses a change region segmentation head to... Upsampling is performed by a factor of 2 to a scale of 256*256 using a 3*3 2D transposed convolution, and then the upsampled feature map is applied... The function performs normalization, and finally binarizes the normalized feature map, setting values ​​greater than 0.5 to 1 and values ​​less than 0.5 to 0, outputting the final segmentation map of the changed region.

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

[0034] Will Each is passed through a CBAM module;

[0035] Will Each element is extended to a new dimension, and then the pieces are joined together on this new dimension.

[0036] The stitched feature maps are fused using a 2*1*1 three-dimensional convolutional layer.

[0037] The feature map after feature fusion and the variation features of the input are compared. Input the feature aggregation module;

[0038] The feature map output by the feature aggregation module is compared with the weighted map of the corresponding scale. Multiplying gives the result of multiplication with Differences in changes at the same scale .

[0039] Furthermore, the content feature focusing decoding module is equipped with a focuser;

[0040] The focuser is used to utilize By focusing on the changed and unchanged areas of the dual-temporal remote sensing images, the sets of changed content features and the sets of unchanged content features of Image 1 and Image 2 are obtained respectively.

[0041] Furthermore, it also includes a content-enhanced loss calculation module, used to calculate the main loss using the change region segmentation map and change region labels on the main branch. ; Calculate the change content feature loss using the change content feature sets of Image 1 and Image 2. The unchanged content feature loss is calculated by using the unchanged content feature set of the image one and the image two The three losses are integrated to perform multi-element constraint on the model parameters of the system

[0042] The main loss The mean square loss function MSELoss of the change region segmentation graph and the change region label is calculated

[0043] The change content feature loss The calculation process is as follows

[0044] A random sampler is used to sample the first scale of the change content feature set and the second scale of the change content feature set respectively to obtain N pairs of sampling points of the change region Cosine similarity is calculated in the N pairs of sampling points to obtain a 1*N matrix

[0045] and

[0046] is calculated by using and , and the formula is as follows

[0047] ; The average value of

[0048] at each scale is obtained to obtain , and the formula is as follows

[0049] ;

[0050] The calculation process of the unchanged content feature loss is as follows

[0051] A random sampler is used to sample the first scale of the unchanged content feature set and the second scale of the unchanged content feature set respectively to obtain N pairs of sampling points of the unchanged region Cosine similarity is calculated in the N pairs of sampling points to obtain a 1*N matrix

[0052] and

[0053] is calculated by using and , and the formula is as follows

[0054] ;​​​​​​

[0055] For each scale Take the average, get , the formula is as follows:

[0056] .

[0057] Further, the comprehensive , And The method for multivariate constraint of model parameters of the system is: setting variable parameters alpha and beta, which can be dynamically adjusted in the convergence process of the system model, and the final system total loss is defined as:

[0058] .

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

[0060] The present application 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 content-based reinforcement loss calculation. BRIEF DESCRIPTION OF DRAWINGS

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

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

[0063] Figure 3 is the structure schematic diagram of the focalizer provided by the present application example;

[0064] Figure 4 is the structure schematic diagram of the change difference decoder provided by the present application example;

[0065] Figure 5 is the structure schematic diagram of the change content feature loss calculation and unchanged content feature loss calculation provided by the present application example;

[0066] Figure 6 is the visualization result schematic diagram of the comparison of the present application example with other methods on CLCD data set;

[0067] Figure 7 is the visualization result schematic diagram of the comparison of the present application example with other methods on LEVIR-CD data set;

[0068] Figure 8 is the visualization result schematic diagram of the comparison of the present application example with other methods on SYSU-CD data set.

[0069] DETAILED DESCRIPTION

[0070] The foregoing and other technical contents, features and effects of the present application will be clearly presented in the following detailed description of specific embodiments with reference to the accompanying drawings. Through the description of the specific embodiments, the technical means and effects taken by the present application to achieve the predetermined purpose can be more deeply and specifically understood. However, the accompanying drawings are provided for reference and illustration only, and are not intended to limit the technical solutions of the present application.

[0071] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant are intended to cover non-exclusive inclusion, so that the article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed. Without more limitations, the element defined by the sentence "including a" does not exclude the presence of other identical elements in the article or device including the element.

[0072] The embodiment of the present application proposes a remote sensing image change detection system based on collaborative similarity learning, mainly solves two problems: first, due to the imaging condition difference and style change between double time phase remote sensing images, these unpredictable disturbances will significantly weaken the model's accurate recognition ability to the change area, thereby reducing the overall detection effect; second, the feature splicing method widely used in existing change detection systems has good performance, but cannot directly calculate the change area between double time phase images, which makes the model difficult to accurately focus on the change area when feature fusion, affecting the detection precision.

[0073] The embodiment of the present application provides a remote sensing image change detection system based on collaborative similarity learning, please see Figure 1The remote sensing image change detection system comprises a multi-scale feature encoding stage, a content feature focusing decoding stage, a multi-branch change difference decoding stage and a content reinforcement-based loss calculation stage. The multi-scale feature encoding stage extracts multi-scale features of double-time remote sensing images through a shared weight encoder; the content feature focusing decoding stage decodes the multi-scale features through two non-shared weight decoders respectively, and extracts content features of the changed area and the unchanged area through a focusing device; the multi-branch change difference decoding stage fuses the multi-scale features of the double-time images, and generates a change area segmentation map through a change difference decoder; and the content reinforcement-based loss calculation stage jointly optimizes the model parameters through three losses of a main loss, a changed content feature loss and an unchanged content feature loss, so as to realize multi-element constraint.

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

[0075] Table 1 shows the results of different models on two data sets.

[0076]

[0077] The resolution of the original image input into the multi-scale feature encoding stage is 256*256, and after multi-scale feature encoding, features with resolutions of 128*128, 128*128, 64*64, 32*32 and 16*16 are obtained, respectively. The encoded features of the outputs of the 2-5 stages are defined as { , , , } and { , , , } respectively.

[0078] Please refer to Figure 2 、 Figure 3 , Figure 2 is a structural diagram of a content feature decoder provided by the embodiment of the application, Figure 3 is a structural diagram of a focusing device. The content feature decoder of the embodiment comprises a feature aggregation module and a focusing device. The content feature decoder integrates adjacent scale features output by the encoder through the feature aggregation module to generate a content feature set of the double-time images and a corresponding weighting map; and the focusing device reweights the content feature set by using the weighting map, and finally outputs a changed content feature set and an unchanged content feature set.

[0079] It should be noted that, as Figure 2As shown, the content feature decoder of the embodiment is specifically used for: for the encoded features and , without processing by the feature aggregation module, directly output as and . For the encoded features and ( ), since the processing of the two encoded features is the same, use to represent ( ), and the specific processing manner is:

[0080] transpose convolution is performed on to upsample by 2 times, and the scale is the same as , to obtain a first processing result;

[0081] concatenate and the first processing result in the channel dimension to obtain a second processing result;

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

[0083] Finally, input the third processing result into a residual module composed of two residual layers linearly concatenated, to suppress the gradient vanishing caused by the depth of the model, to obtain the content feature .

[0084] Further, as shown in Figure 3 , the focuser of the embodiment is specifically used for:

[0085] calculate the cosine distance of and to obtain a weighted map ;

[0086] multiply and and to obtain the changed content features and in the same scale;

[0087] multiply and and to obtain the unchanged content features and in the same scale.

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

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] The designed focuser can make the system flexibly focus on the content features of the change area and perform multi-element constraint on the model parameters of the system for the content features of the change area.

[0097] Please refer to Figure 4 , Figure 4 is a structural diagram of a change difference decoder provided by an embodiment of the present application. The change difference decoder inputs a feature fusion module with a content feature set and a weighted graph under each scale, and outputs change difference features under each scale. The embodiment includes a feature fusion module and a change area segmentation head. The feature fusion module is used to fuse the content features of the dual-phase images under the same scale. The change area segmentation head is used to binarize the final change difference features and output a final change area segmentation map.

[0098] It should be noted that, as shown in the figure, the feature fusion module of the embodiment is specifically used for: Figure 4

[0099] passing through a CBAM module respectively, introducing spatial and channel attention, strengthening the understanding of the system to the features, and obtaining a first processing result;

[0100]

[0101] Finally, the second processing result is passed through a three-dimensional convolution layer with a convolution kernel size of 2*1*1, so that the image features of the dual-phase images are fused. The fused feature map is multiplied by the weighted graph , and the change difference features under the scale are output .

[0102] Further, the change area segmentation head of the embodiment performs binarization on the change difference features ​​​​​​Upsample the feature map by a factor of two to a scale of 256*256 using a 3x3 transposed convolution, and then apply the following to the upsampled feature map: The function performs normalization, and finally sets pixel values ​​greater than 0.5 in the feature map to 1 and pixel values ​​less than 0.5 to 0, and then outputs the final change region segmentation map.

[0103] In the content-enhanced loss calculation stage, the main losses are calculated separately. Loss of features due to changes in content loss of unchanged content features By combining the three losses to apply multivariate constraints to the system's model parameters, 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 squared loss function MSELoss between the changed region segmentation map and the changed region label;

[0106] It should be noted that, as Figure 5 As shown, the loss of change content features in this embodiment The calculation process is as follows:

[0107] Use a random sampler to sample the first feature set of the changed content. Individual scale and Perform the same sampling to obtain N pairs of sampling points for the changed regions. In this embodiment, in order to balance the computational efficiency and computational accuracy, N is set as the scale of the changed content features. For example, if the scale of the changed content features is 256*256, then N is set to 256.

[0108] Cosine similarity is calculated for each of the N sampling points, resulting in a 1*N matrix. and ;

[0109] use and calculate It can be expressed as:

[0110]

[0111] For each scale Take the average value to get It can be expressed as:

[0112]

[0113] Similarly, such as Figure 5The variation content feature loss of the embodiment is shown The calculation process is specifically as follows:

[0114] The first dimension of the variation content feature set is sampled by a random sampler, and the same sampling is performed on the second dimension of the variation content feature set, to obtain N pairs of sampling points of the unchanged region, and the selection of N is the same as in the calculation process. The calculation process is specifically as follows:

[0115] The cosine similarity is calculated in the N pairs of sampling points, to obtain a 1*N matrix and .

[0116] The calculation of and is , which can be expressed as:

[0117]

[0118] The average value of on each dimension is obtained, and , which can be expressed as:

[0119]

[0120] Please refer to Figure 6 , Figure 7 and Figure 8 , Figure 6 is a visual result schematic diagram of system testing on the CLCD data set by using different methods, Figure 7 is a visual result schematic diagram of system testing on the LEVIR-CD data set by using different methods, Figure 8 is a visual result schematic diagram of system testing on the SYSU-CD data set by using different methods. The red area in the three figures is the error prediction of the changed region, the green area is the error prediction of the unchanged region, the white area is the correct prediction of the changed region, and the black area is the correct prediction of the unchanged region. The first two columns are double-time phase remote sensing images, the third column and the fourth column are the prediction results of the system, and the fifth column to the last column are the prediction results of other systems. Figure 6 , Figure 7 and Figure 8 ​​​The results shown in the middle indicate that the remote sensing image change detection system based on content feature reinforcement learning of the embodiment of the application can detect the change region (white region) with high accuracy, effectively solves the problem of style feature difference interference content change recognition caused by remote sensing platform diversity and observation condition change, and at the same time, through explicit modeling of the change region of the double-time-phase image, overcomes the problem of insufficient attention to the change region and decline of detection precision caused by excessive dependence of the existing method on indirect fitting of the deep network.

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

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

[0123]

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

[0125]

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

[0127]

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

[0129] Experiments prove that through the remote sensing image change detection system based on content feature reinforcement learning of the embodiment of the application, the problem of style feature difference interference content change recognition is solved, and through explicit modeling of the change region, the excessive dependence of the existing method on indirect fitting of the deep network is overcome.

[0130] In summary, remote sensing image change detection has important application value in multiple fields. For example, in the field of agriculture, remote sensing image change detection can be used to monitor changes in cultivated land use, helping to optimize land resource allocation and improve agricultural management efficiency. In addition, remote sensing images can also be used to monitor changes in the ecological environment, such as the reduction of forest coverage and the spread of water pollution, providing scientific basis for ecological protection and management. At the same time, in disaster prevention and mitigation, by analyzing the remote sensing images before and after the disaster, the damage to the ecological system caused by landslides or fires can be evaluated, providing important data support for post-disaster recovery and reconstruction. Overall, remote sensing image change detection can provide efficient and accurate data support and decision-making reference for multiple fields. With the continuous progress of remote sensing technology, its application range and detection accuracy will continue to improve.

[0131] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative, and the division of the modules is only a logical functional division. In actual implementation, other division methods can be used, such as combining or integrating multiple modules or components into another system, or ignoring or not executing some features.

[0132] In addition, the functional modules in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of hardware plus software functional module.

[0133] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application should not be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as falling within the protection scope of the present application.

Claims

1. A remote sensing image change detection system based on content feature reinforcement learning, characterized in that: The system comprises an image acquisition module, a multi-scale feature encoding module, a content feature focusing decoding module, and a multi-branch change difference decoding module. The image acquisition module is configured to acquire double-time-phase remote sensing images. The multi-scale feature encoding module is configured to input the double-time-phase remote sensing images, extract features of the double-time-phase remote sensing images at multiple scales by using an encoder with shared weights, and output encoded features at each scale. The content feature focusing decoding module is configured to input the encoded features at each scale, decode the encoded features at each scale at multiple scales by using two content feature decoders with unshared weights, and output multi-scale decoded features of the double-time-phase remote sensing images. The multi-branch change difference decoding module is configured to input the multi-scale decoded features of the double-time-phase remote sensing images. The multi-scale decoded features of the double-time-phase remote sensing images are fused at each scale, and the fused multi-scale decoded features are decoded by using a change difference decoder to output a predicted change region segmentation map. The encoder with shared weights is the first five stages of a pre-trained EfficientNet-B5 network, and the outputs of stages 2-5 are taken as the outputs of the encoder. The first to fifth stages of the EfficientNet-B5 network respectively downsample the 256*256 resolution image input into the system to 128*128, 128*128, 64*64, 32*32 and 16*16 resolutions, and the outputs of the second to fifth stages of the dual-phase remote sensing image are defined as { , , , } and { , , , } respectively; The content feature decoder decodes the minimum scale feature output by the encoder Without processing, directly output as , , The feature of the adjacent scale output by the encoder and , The input feature aggregation module, The feature aggregation module is configured to: small-scale features after transposed convolution with larger scale features scale alignment; align the two adjacent features in the scale dimension, and concatenate the two adjacent features in the channel dimension. The spliced feature map is fused by a 1*1 two-dimensional convolution layer and a residual module to obtain a content feature of a corresponding scale , represent or ; the content features of the dual-phase remote sensing images output by the content feature decoder are respectively defined as { , , , } and { , , , }; The content features decoded by the content feature decoder at each scale are used to compute a weighting map at the corresponding scale ; The weighted graph The calculation process is: Computing the cosine similarity between the content features of two time-phased remote sensing images of the same scale and ;​ Along dimension pairs summed to obtain content features and cosine distance maps ; To Utilizing Normalization is performed using the function; The calculation process is as follows: ； ； ； The change difference decoder And Respectively through a CBAM module, respectively expand a new dimension, then splice on the new dimension, then through a 2*1*1 three-dimensional convolution layer for feature fusion, then multiply the fused features with the weighted map Of the same scale to obtain the change difference features of the dual-phase images ; The change difference decoder , and Perform the same process: and differences in characteristics Perform feature fusion to obtain Differences in changes at the same scale ; The change difference decoder is connected with the change region segmentation head Up-sampling 2 times to the scale of 256*256 by a 3*3 two-dimensional transpose convolution, then applying The function is normalized, and the normalized feature map is finally binarized, and the value greater than 0.5 is set to 1, and the value less than 0.5 is set to 0, and the final change region segmentation map is output.

2. The remote sensing image change detection system based on content feature reinforcement learning according to claim 1, wherein: The feature blending module is specifically configured to: Will Each is passed through a CBAM module; will be described below. respectively, and spliced in the new dimension. fuse the concatenated feature maps by using a 2*1*1 three-dimensional convolution layer; The feature map after the feature fusion and the input change difference feature The feature aggregation module is input aggregate the feature maps output by the feature aggregation module with the weighted maps of the corresponding scale to obtain the same scale change difference features .​ 3. The remote sensing image change detection system based on content feature reinforcement learning according to claim 1, wherein: The content feature focusing decoding module is provided with a focalizer; The focus is used to utilize Respectively focus the changed area and unchanged area of the dual temporal phase remote sensing image, and respectively obtain the changed content feature set and unchanged content feature set of the image one and image two.

4. The remote sensing image change detection system based on content feature reinforcement learning according to claim 1, wherein: The loss calculation module based on content enhancement is used for calculating a main loss by using the change region segmentation graph and the change region label on the main branch ; calculating a change content feature loss by using the change content feature set of the image one and the image two ; calculating an unchanged content feature loss by using the unchanged content feature set of the image one and the image two ; comprehensively using the three kinds of losses to perform multi-element constraint on the model parameters of the system The main loss The mean square loss function MSELoss of the change region segmentation graph and the change region label is obtained by calculation. The change content feature loss The calculation process is: The same sampling is performed on the first and scales of the changing content feature set using a random sampler to obtain n pairs of sample points of the changing regions. ​ Calculate the cosine similarity between the n samples and the internal vector to get a 1*n matrix and ; Utilizing and calculating , as follows: ； For each scale Taking the average, we get , which is given by the formula ； The unchanged content feature loss The calculation process is as follows: The same sampling is performed on the first dimension and of the unchanged content feature set using a random sampler to obtain n pairs of sample points of unchanged regions. Calculate the cosine similarity between the n samples and the internal vector to get a 1*n matrix and ; Utilizing and calculating , as follows: ； For each scale The average is taken, resulting in , which is given by the formula 。 5. The system of claim 4, wherein: Comprehensive , and The method of multivariate constraint on the model parameters of the system is to set variable parameters α and β, which can be dynamically adjusted in the convergence process of the system model, and the final system total loss is defined as: 。

Images