High-resolution remote sensing image change detection method and system based on decoupling representation learning

Abstract

The invention relates to a high-resolution remote sensing image change detection method and system based on decoupling representation learning, and solves the defects of false change and non-ideal detection precision caused by inconsistent imaging conditions of different time phases compared with the prior art. The method comprises the following steps: obtaining a training sample; constructing content branches; constructing style branches; training a multi-branch parallel framework; obtaining a high-resolution change result; and separating image styles and obtaining a recovery result. According to the method, a reliable high-resolution remote sensing image change detection result can be obtained under the condition of multi-temporal complex imaging condition interference.

Description

Technical Field

[0001] This invention relates to the field of high-resolution remote sensing image technology, specifically to a method and system for detecting changes in high-resolution remote sensing images based on decoupled representation learning. Background Technology

[0002] In recent years, the continuous development of airborne and space-based remote sensing platforms has significantly increased the frequency of remote sensing image acquisition. On the one hand, rapid image updates facilitate timely monitoring of surface changes; on the other hand, differences in imaging conditions between images acquired at different times affect the accuracy of change detection. Furthermore, the fact that image data originates from different remote sensing platforms further increases the difficulty of detection. Therefore, developing a stable change detection framework capable of handling complex imaging conditions will greatly promote the refined monitoring of surface changes.

[0003] The core task of change detection methods is to compare two-temporal remote sensing images of the same geographic location and identify areas of change. Early research mainly relied on sub-pixel methods to predict change areas on medium-resolution remote sensing images. However, limited by spatial resolution and image quality, these methods often only detect change areas with blurred edges. Fortunately, with the improvement of imaging spatial resolution, high-resolution (HR) and very-high-resolution (VHR) remote sensing images can provide richer spatial details. Therefore, machine learning-based methods have gradually gained attention and are used to improve the accuracy of change detection. For example, the Slow Feature Analysis (SFA) method predicts change areas based on generalized eigenvalues ​​and feature spaces. In addition, researchers have also applied spatial relationship modeling methods (such as Markov random fields and conditional random fields) to change detection. However, these traditional methods rely on hand-designed features and are applicable to specific scenarios, making it difficult to generalize to different imaging conditions. Furthermore, the parameter tuning process is complex and relies on expert knowledge. Therefore, their performance in practical applications remains limited.

[0004] In recent years, the rapid development of deep learning technology (especially Convolutional Neural Networks (CNNs) has greatly promoted change detection research. Leveraging the powerful learning capabilities of CNNs, researchers can establish a mapping relationship between pixel changes in remote sensing images and surface changes, extending detection results from simple pixel-level changes to a more semantically understandable level. Therefore, an increasing number of studies are dedicated to introducing deep learning into change detection, including single-stream methods, two-stream methods, and multi-model ensemble methods. Single-stream methods fuse two temporal images at the input stage and use a single encoder to extract features; two-stream methods employ two encoders and perform feature fusion later; while multi-model ensemble methods combine multiple deep learning models to enhance the flexibility of feature extraction and supervision strategies.

[0005] While deep learning-based methods offer significant advantages in pixel-level change detection, their practical application still faces two key challenges. First, the different imaging platforms and conditions of dual-temporal images lead to significant differences in the pixel-level representation of the same category of targets. For example, in disaster assessment tasks, post-disaster images typically utilize appropriate platforms and acquisition times to ensure data quality. However, pre-disaster images often originate from historical image archives, and their resolution and quality can vary considerably, especially when images come from different remote sensing platforms. Furthermore, different imaging conditions (such as shooting angle, atmospheric conditions, and lighting conditions) can cause temporal spectral variations in images, thus affecting the stability of change detection algorithms. Second, the definition of changed areas is often influenced by human intervention and task requirements, increasing the risk of pseudo-changes.

[0006] To address these challenges, some studies have proposed multi-model ensemble frameworks that decompose change detection into multiple sub-tasks and employ multiple models to supervise each other. For example, frameworks based on Generative Adversarial Networks (GANs) treat bi-temporal images as two distinct data domains, with the GAN generator aiming to bridge the gap between these domains to improve change detection accuracy. However, while these methods have made some progress in reducing cross-domain discrepancies, they often lose spatial detail during domain adaptation. Furthermore, the image quality generated by GANs depends on the performance of the binary classifier (discriminator), which is typically evaluated at the image level and can easily overlook individual pixel-level changes. Therefore, to obtain more accurate change detection results, pixel-level supervision is needed to ensure fine-grained representation of the detection results. Summary of the Invention

[0007] To address the instability and low accuracy of change detection caused by imaging differences, disentangled representation learning can be introduced into change detection tasks, with pixel-level supervision implemented during feature decomposition. The core idea of ​​disentangled representation learning is to decompose the input image into multiple separable latent variable representations. Inspired by domain adaptation algorithms, the representation of remote sensing images is decomposed into two independent subspaces: a shared content space (representing the semantic features of ground features) and a private style space (representing style information under different imaging conditions). This separation of style spaces effectively reduces style mismatches caused by different imaging conditions, while the extracted content space provides a more stable representation of ground feature information.

[0008] The purpose of this invention is to address the shortcomings of existing technologies that suffer from spurious changes due to inconsistent imaging conditions at different time phases, and to provide a high-resolution remote sensing image change detection method and system based on decoupled representation learning to solve the above problems.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: A high-resolution remote sensing image change detection method based on decoupled representation learning includes the following steps: Step 1: Acquire high-resolution remote sensing image pairs of the area to be detected at two different time phases and preprocess them to obtain training data; Step 2: Construct a content branch, which is an encoding and decoding structure. In the content encoding network, a multi-scale parallel convolution module is used to extract content features from the input remote sensing image pairs, and a content change decoding network is used to generate high-resolution remote sensing image change area detection results with the same size as the input image level by level. Step 3: Construct style branches, including using a lightweight style coding network to independently extract the style features of each remote sensing image pair, combining the style features extracted by the style coding network with the corresponding content features, and performing joint decoding through an image reconstruction decoding network with an adaptive instance normalization module to reconstruct the input original remote sensing image. Step 4: Using the preprocessed dual-temporal high-resolution remote sensing image pair as input, the images are input into the trained content branch and style branch respectively to obtain change detection results and reconstructed images.

[0010] Further, the preprocessing includes: cropping the remote sensing image data into standardized image slices of a certain size, using a sliding window method and setting a certain overlap between windows to increase data diversity; performing data augmentation operations on the image pairs, including random rotation, horizontal flipping, or vertical flipping; and performing pixel value normalization on the processed image data, linearly mapping the pixel value range to the (0,1) interval.

[0011] Furthermore, the multi-scale parallel convolution module includes multiple parallel convolution branches, where the first branch outputs a feature map with the same spatial resolution as the input image, and the spatial resolution of the output feature map of subsequent branches is halved at each level, while the number of feature map channels is doubled at each level; each branch uses a different number of residual convolution modules for feature extraction, and each residual convolution module includes a 3×3 convolutional layer, a batch normalization layer (BN), and a ReLU activation function.

[0012] Furthermore, the multi-scale feature maps of each branch also interact and fuse information through upsampling or downsampling operations. Specifically, high-scale features are fused with low-scale features after downsampling, while low-scale features are fused with high-scale features through upsampling.

[0013] Furthermore, the change detection decoding network comprises multiple groups, each including convolutional layers, batch normalization (BN), and rectified linear units (ReLU).

[0014] Furthermore, the style coding network includes multiple convolutional layers, activation function layers, pooling layers, and fully connected layers.

[0015] Furthermore, the adaptive instance normalization module AdaIN processes the data as follows:

[0016] in, As a content feature, As a stylistic feature, and These represent the standard deviation and mean calculation operations for the feature map, respectively.

[0017] Furthermore, the image reconstruction decoder is implemented through cascading multiple restoration decoding modules. Each restoration decoding module takes content features and style features as input, and sequentially injects style features into content features through feature upsampling, convolutional fusion, and AdaIN operations to restore the original appearance of the remote sensing image. The restoration decoding process is specifically expressed as follows:

[0018] in, This is the output feature map of the m-th restoration decoding module in the style branch. To restore the decoding module, This is an upsampling operation on the feature map. This is the stitching operation for the feature channels; the features output by the adaptive instance normalization module AdaIN are processed by the ReLU activation function and then input to the next level restoration decoding module. This process is repeated until a restored remote sensing image with the same size and similar appearance as the input image is recovered.

[0019] Furthermore, the change detection loss, feature separation loss, content similarity loss, and image reconstruction loss are weighted and combined to form the total loss function for training the content branch and style branch. The specific calculation methods for each loss are as follows: Change detection loss A pixel-wise binary classification cross-entropy loss function is used to supervise the difference between the change detection network output and the ground truth change detection labels, specifically expressed as follows:

[0020] Where N is the total number of pixels in the sample. The label for the actual change of the i-th pixel. Predict the probability of change for the i-th pixel; Feature separation loss By employing mutual information constraints between content and style features, the independence and decoupling of content and style feature spaces are ensured, as expressed below:

[0021] in, These respectively represent content and style characteristics. Calculate the Gram matrix. It is the Frobenius norm; Content similarity loss To ensure that the content features of unchanged regions remain consistent over time, it is expressed as follows:

[0022] in, This represents the total number of pixels in the unchanged area. These represent the content features of the j-th pixel in two time-unchanged regions, respectively. Image reconstruction loss The pixel-level mean square error (MSE) between the reconstructed image and the original input image is used to ensure the integrity of the decoupled content and style feature information, as expressed below:

[0023] in, To reconstruct the image, This is the original input image.

[0024] This invention also provides a high-resolution remote sensing image change detection system based on decoupled representation learning, comprising: The system includes a processor and a memory. The memory stores program instructions, and the processor calls the stored instructions in the memory to execute the high-resolution remote sensing image change detection method based on decoupled representation learning as described in the above technical solution.

[0025] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: The high-resolution remote sensing image change detection method and system based on decoupled representation learning of the present invention, compared with the prior art, can also obtain more reliable high-resolution remote sensing image change detection results under the interference of complex imaging conditions in multiple time phases.

[0026] This invention designs a change detection network framework based on decomposable representation (CCNet). This framework includes a shared content encoder and an independent style encoder to separate style information from content information in remote sensing images. Combined with pixel-level supervision strategies, it achieves more accurate change detection. First, a multi-resolution parallel feature encoder is employed to ensure the integrity of spatial details while enhancing the expressive power of semantic features. Second, a dual-task learning framework is established, introducing an image reconstruction task while completing the change detection task, thereby improving the integrity and stability of content features. Furthermore, CCNet employs a de-stylization strategy to eliminate style variations caused by different imaging conditions, improving the contrast consistency between images and further suppressing pseudo-changes. Through feature separation loss, content similarity constraints, and image reconstruction loss, the stability of content features and the removal of style information are ensured, thus enhancing the reliability of change detection. On five publicly available remote sensing change detection datasets (LEVIR-CD, SECOND, SVCD, SYSU-CD, and xBD), this invention demonstrates higher detection accuracy under various imaging interference conditions.

[0027] This invention effectively reduces the impact of imaging style changes on change detection tasks by employing decomposable representation learning; it improves the model's sensitivity to spatial details and detection accuracy through multi-resolution feature extraction and pixel-level supervision; in addition, the cooperative learning framework combined with image reconstruction strategies further enhances the network's feature extraction capabilities and generalization ability, enabling the change detection method to have higher accuracy and stronger robustness under complex imaging conditions. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a content branch diagram in an embodiment of the present invention; Figure 3 This is a style branch diagram in an embodiment of the present invention; Figure 4 This is a distribution diagram of the detection results of five publicly disclosed change detection datasets in this embodiment of the invention; Figure 5 This is a diagram showing the feature decoupling and restoration results in an embodiment of the present invention. Detailed Implementation To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings: like Figure 1 As shown, the high-resolution remote sensing image change detection method based on decoupled representation learning of the present invention includes the following steps: The first step is to acquire training samples: acquire high-resolution remote sensing image data of the area to be detected at two different time phases, denoted as image pairs. I T1 , I T2 The image pairs have the same spatial resolution and the two images have the same geographic coordinates and range, ensuring a strict correspondence in spatial location; The image pairs are preprocessed, including: cropping the remote sensing image data into standardized image slices of 256×256 pixels, using a sliding window method and ensuring some overlap between windows to increase data diversity; performing data augmentation operations on the image pairs, including random rotation, horizontal flipping, or vertical flipping, to enhance the model's generalization ability; and performing pixel value normalization on the processed image data, linearly mapping the pixel value range to the (0,1) interval to improve the model's convergence speed and stability. The specific normalization method is as follows:

[0029]

[0030] in, Represents the original pixel values ​​of the image. The mean and standard deviation of the image pixel values ​​are represented by the normalized data, which are mapped to a range of values ​​with a mean of 0 and a variance of 1. The preprocessed and labeled remote sensing image pairs and their corresponding ground truth labels for change detection are divided into training sample sets and validation sample sets, with no overlap between the training sample sets and validation sample sets.

[0031] The second step is content branch construction: building a multi-resolution parallel content encoding network. For example... Figure 2 As shown, the content encoding network adopts a two-branch twin structure, with each branch using the same network structure and sharing weights, and is used to extract content features from the dual-temporal remote sensing images, denoted as C. T1 C T2 Content feature extraction employs a multi-scale parallel convolutional network structure. (1) Construct a multi-scale parallel convolution module, including 5 parallel convolution branches. The first branch outputs a feature map with the same spatial resolution as the input image. The spatial resolution of the output feature maps of subsequent branches decreases by half at each level, while the number of feature map channels doubles at each level. Each branch uses 1 to 5 residual convolution modules for feature extraction. Each residual block structure consists of a 3×3 convolutional layer, a batch normalization layer (BN), and a ReLU activation function. The output feature map of each convolution branch is denoted as... Where n∈{1,2} represents the temporal phase of the dual-temporal remote sensing image, and m∈{1,2,3,4,5} represents the feature map at the m-th scale; the convolution operation at each scale is expressed as follows:

[0032] in, The content feature map of the nth temporal image at the mth scale, MRB is a maintain resolution backbone network. Each scale in the backbone network contains multiple residual blocks consisting of convolutional layers (kernel=3×3), batch normalization (BN) layers and activation function (ReLU) layers. (2) Construct a multi-scale content feature fusion module. The multi-scale feature maps of each branch interact and fuse information through upsampling or downsampling operations. High-scale features are fused with low-scale features after downsampling, while low-scale features are fused with high-scale features through upsampling. This ensures that feature maps of different scales exchange information fully and obtain feature representations containing rich spatial details and semantic information. (3) Construct a content change decoding network. The content feature maps of the obtained dual-temporal images at various scales are fused in a concatenated manner and then input into the change detection decoder. The change detection decoding module consists of a sequence of four convolutional layers, batch normalization (BN), and activation function (ReLU). The decoding process adopts a step-by-step fusion and upsampling strategy to fuse high-level semantic features and low-level detail features step by step to obtain high-resolution, pixel-level accurate remote sensing image change detection results. The feature decoding process formula is expressed as:

[0033] in, This represents the decoded feature map of the m-th level change. This indicates a change detection decoding module. This represents the upsampling operation on the feature map. The feature map represents the channel stitching operation; the decoding network gradually reconstructs the change area map of the remote sensing image until the output is the final change detection image with the same spatial resolution as the input remote sensing image.

[0034] The third step is style branch construction: Two independent lightweight style coding networks are constructed, using bi-temporal remote sensing imagery as input, such as... Figure 3 As shown, the unique style features of each image pair are extracted. The extraction of style features does not include spatial location and semantic information, but only reflects the global color distribution, lighting conditions and sensor characteristics of the remote sensing image, so as to ensure the complete independence of style branch and content branch. (1) The style coding network gradually reduces the feature space size through a series of convolutional layers, activation function layers and pooling layers in order to extract spatially independent global style information. The specific structure of the style coding network includes: The input remote sensing image is passed through convolutional layers with kernel sizes of 7×7 and 3×3, then connected to the ReLU activation function, and then pooling is performed to gradually reduce the spatial dimension and increase the channel dimension. The final style features are compressed into a global feature vector with a spatial size of 1 after global pooling and fully connected layers, denoted as the style feature vector. ,in These represent different time phases of the dual-temporal remote sensing images; (2) Construct an image reconstruction decoding network, and jointly decode the style features extracted by the style coding network and the corresponding content features through the image reconstruction decoding module with an Adaptive Instance Normalization (AdaIN) module to reconstruct the input original remote sensing image, and supervise the accuracy and completeness of style feature extraction. The feature fusion process of AdaIN is specifically expressed as follows:

[0035] in, As a content feature, As a stylistic feature, and These represent the standard deviation and mean calculation operations for the feature map, respectively. The image reconstruction decoding network is implemented by cascading multiple restoration decoding modules. Each restoration decoding module takes content features and style features as input, and sequentially injects style features into content features through feature upsampling, convolutional fusion, and AdaIN operations to restore the original appearance of the remote sensing image. The restoration decoding process is specifically expressed as follows:

[0036] in, This is the output feature map of the m-th restoration decoding module in the style branch. To rebuild the decoding module, This is an upsampling operation on the feature map. This is the stitching operation for the feature channels; the AdaIN output features are processed by the ReLU activation function and then input to the next level restoration and decoding module. This process is repeated until a reconstructed remote sensing image with the same size and similar appearance as the input image is restored. Step 4, Multi-branch parallel framework training: Using preprocessed dual-temporal high-resolution remote sensing images... I T1 , I T2As input, the image is fed into the content branch and the style branch respectively. The content branch is responsible for extracting the spatial details and semantic information of the image and obtaining the change detection results. The style branch is responsible for extracting the style features of the image such as color and lighting and obtaining the image restoration results. (1) Establish an end-to-end multi-branch parallel training framework, which includes a content encoding network, a style encoding network, an image reconstruction network and a content change decoding network. The content encoding network and the style encoding network share the input image data, but the network parameters are independent, and there is no direct information exchange between the content network and the style network, so as to ensure the independence of feature decoupling. (2) The training process adopts a multi-task joint training strategy, defining change detection loss, feature separation loss, content similarity loss and image reconstruction loss, which together form the total loss function. The specific calculation methods of each loss are as follows: Change detection loss A pixel-wise binary classification cross-entropy loss function is used to supervise the difference between the change detection network output and the ground truth change detection labels, specifically expressed as follows:

[0037] Where N is the total number of pixels in the sample. The label for the actual change of the i-th pixel. The probability of change for the i-th pixel is predicted from the output of the last stage of the content change decoding network; Feature separation loss By employing mutual information constraints between content and style features, the independence and decoupling of content and style feature spaces are ensured, as expressed below:

[0038] in, These respectively represent content and style characteristics. Calculate the Gram matrix. It is the Frobenius norm; Content similarity loss To ensure that the content features of unchanged regions remain consistent over time, it is expressed as follows:

[0039] in, This represents the total number of pixels in the unchanged area. These represent the content features of the j-th pixel in two time-unchanged regions, respectively. Image reconstruction loss The pixel-level mean square error (MSE) between the reconstructed image and the original input image is used to ensure the integrity of the decoupled content and style feature information, as expressed below:

[0040] in, To reconstruct the image, The original input image; Based on the above loss terms, construct the overall loss function. Each loss term is assigned a different weighting coefficient, as defined below:

[0041] in, λ Each loss is assigned a weighting coefficient to balance the relative importance of each loss. (3) The back propagation algorithm (BP) is used to optimize the network parameters, and the Adam optimization algorithm is used to update the network weights. Through the multi-task joint training strategy, the overall loss is gradually reduced during the network training process, and the effective decoupling of content and style features and the continuous improvement of change detection accuracy are finally achieved. After training, the trained change detection network model will be obtained and used for the accurate detection of change areas in subsequent high-resolution remote sensing images.

[0042] Step 5: High-resolution change result acquisition: The shared content features are fused and decoded using a content change decoding network to generate high-resolution remote sensing image change area detection results with the same size as the input image level by level; Step 6: Image style separation and restoration results acquisition: The image reconstruction decoding network is used to jointly decode the decoupled content features and style features to reconstruct the original input remote sensing image, realize the supervision of the content and style feature separation process, and evaluate the effect of style separation and restoration.

[0043] The effects of this invention will be further illustrated below with simulation experiments: 1. Simulation Experiment Conditions: The computer hardware environment for this invention's experiments consisted of an Intel(R) Xeon(R) Gold6230R CPU @ 2.10GHz and a GeForce RTX 3090 GPU. The software environment was Linux, and the deep learning framework used was PyTorch. All subsequent training and testing experiments were based on this platform. During training, the batch size was set to 8, the model was trained for approximately 200k epochs, the initial learning rate was set to 0.002, and it decayed to 20% of the initial value after 160k epochs. This invention was validated on five remote sensing image datasets: LEVIR-CD, SECOND, SVCD, SYSU-CD, and xBD. Furthermore, this invention used precision (Prec.), recall (Rec.), and F1 score (F1) based on the confusion matrix as evaluation metrics.

[0044] 2. Simulation Experiment Content and Result Analysis: Example 1: To verify the stability and superiority of the present invention (CCNet) in multiple scenarios, this example selects five publicly available change detection datasets to test the performance of the present invention.

[0045] Table 1. Performance of various methods on the LEVIR-CD and SECOND datasets.

[0046] Table 2. Performance of each method on SVCD, SYSU-CD, and xBD datasets.

[0047] Quantitative results on the five datasets are listed in Tables 1 and 2. The best-performing method in each column is marked in bold, and the second-best results are marked with an underline. A comprehensive comparison of the five datasets shows that the proposed method (CCNet) performs best among all methods, achieving the highest F1 score on four datasets (LEVIR-CD, SECOND, SYSU-CD, and xBD) and performing second best on the SVCD dataset. The LEVIR-CD dataset, composed of very high resolution (VHR) imagery, was used to test the method's ability to extract detail information. The multi-resolution parallel content feature extraction network employed in this invention effectively captures spatial context information, mitigating the detail loss problem commonly found in methods such as FC-EF, Siam-conc, and Siam-diff. The superior performance on the SECOND and SVCD datasets further validates CCNet's advantages in eliminating multi-platform data errors and seasonal spurious variations. This is mainly due to CCNet's explicit separation of style features and content features during feature extraction, reducing errors caused by changes in imaging style. Compared to GAN-based domain adaptation methods (such as SDACD), this invention achieves superior F1 scores with a significantly reduced number of parameters. This advantage stems from the end-to-end single-model design and the introduction of pixel-level supervision mechanisms during feature separation and image restoration, improving the overall efficiency of the model. The SYSU-CD dataset contains six different types of change scenarios, enabling a comprehensive evaluation of the method's robustness. CCNet demonstrates high recall on this dataset, effectively detecting change regions in various complex scenarios. Simultaneously, CCNet maintains high precision, ultimately achieving the highest F1 score, demonstrating its superior performance in complex change detection tasks. In the xBD dataset (primarily used for disaster assessment applications), this invention achieves a good balance between precision and recall, obtaining the best F1 score. Since the xBD dataset covers images from different regions globally, the results show that CCNet possesses stable and efficient content and style decoupling capabilities, outperforming the Transformer-based AMTNet in practical applications, further validating its practical application value in large-scale disaster assessment tasks.

[0048] Qualitative results on five datasets as follows Figure 4 As shown in the diagram, TP (true) and TN (true negative) regions are marked in white and black, respectively, while FP (false positive) and FN (false negative) regions are highlighted in red and blue, respectively. In the LEVIR-CD dataset, this invention (CCNet) outperforms other methods in change detection for very high resolution (VHR) remote sensing imagery. Figure 4As shown in (a), the contrast methods SiamUNet, BIT, and SDACD incorrectly classified some buildings as changed areas (shown in red) due to the influence of lighting variations. Similarly, some methods (such as BIT and P2V) failed to identify the changed buildings (shown in blue), such as... Figure 4 As shown in (b) above. In contrast, CCNet effectively reduces these false positives and excels in small object classification. Leveraging the high-resolution features provided by its multi-resolution parallel structure, CCNet can accurately segment changing regions while maintaining clear boundaries. In the SECOND dataset, this invention can accurately detect changes in vegetation type while eliminating spurious changes in the background, such as... Figure 4 As shown in (c) and (d) above. By decoupling style and content features, this invention adapts to images from different remote sensing platforms, while contrastive methods suffer from numerous misclassifications. For example, the P2V method exhibits a large number of FN (false negative) regions, while the GAS-Net and AMTNet methods show more FP (false positive) regions. Furthermore, in tasks involving seasonal variations on the SVCD dataset, CCNet demonstrates the advantage of content features. Figure 4 As shown in (e) and (f), in scenes with significant style and seasonal variations, CCNet can accurately detect changed regions while maintaining strong robustness to unchanged regions. Even when the input image pair is predominantly green in the color space, CCNet can still accurately detect areas transitioning from vegetated landscapes to urban infrastructure such as roads and buildings. Furthermore, image restoration tasks help the network extract complete content information, thereby reducing the false negative rate (especially in detecting changes in small objects). In the diverse scenes of the SYSU-CD dataset, CCNet significantly improves its ability to detect changes in objects of different semantic categories through a style decoupling mechanism. Figure 4 As shown in (h), compared methods such as SNUNet and GAS-Net often miss changes caused by construction progress, while CCNet performs exceptionally well. Furthermore, in disaster assessment tasks (such as floods), change detection requires not only identifying the damage to the building itself but also acquiring information about changes in its surrounding environment. CCNet's content branch encoder combines content features at different scales to achieve multi-receptive field feature interaction, thereby improving the information connection between adjacent pixels. This allows CCNet to detect damaged buildings more accurately on the xBD dataset. In contrast, SNUNet and GAS-Net methods exhibit large areas of FN (false negative) regions, such as... Figure 4 As shown in (i) in the diagram.

[0049] Example 2: To verify the effectiveness of style and content decoupling and its impact on change detection tasks, this example demonstrates image restoration results. For example... Figure 5As shown, the first two columns are the original input images, and the third and fourth columns are the images restored based solely on content features, used to visualize the representation of content features in the color space after removing style information. The fifth and sixth columns show the images restored based on both content and style features, which can be used to compare with the original input images to intuitively evaluate the performance of the image restoration task. The last two columns are the predicted change detection map and the ground truth label, respectively.

[0050] Images restored solely based on content features exhibit a consistent color distribution, meaning that features with the same semantic meaning show similarity in color space. For example, in Figure 5 In the urban areas shown in (a) and (b), pseudo-variations in grassland caused by seasonal changes and variations in buildings due to different imaging conditions are effectively eliminated. In snow-covered areas ( Figure 5 In (d)–(f) of the image, the content features in the restored image exhibit a high degree of consistency. Furthermore, thanks to pixel-level supervision of the content branches, texture and boundary information in the image are well preserved. Figure 5 As shown in (c), the outline of the island in the middle of the river is clearly visible, which indicates that pure content features can stably assist in the detection of surface changes.

[0051] Furthermore, images restored based on content and style features maintain a high degree of visual consistency with the original input images. For example, in Figure 5 Vegetation cover areas in (a)–(c) and Figure 5 The seasonal snow cover in (d)–(f) was accurately restored. The near-error-free restored images validated the accuracy and completeness of the extracted content and style features, demonstrating that the features can fully represent the information of the input images and provide high-quality feature support for change detection tasks.

[0052] Example 3: This example studies the impact of input images on change detection accuracy under different time spans and sensor sources.

[0053] To evaluate the robustness and accuracy of this invention, this embodiment constructs a multi-temporal dataset, selecting image pairs acquired at the same geographical location but at different times and using different sensor platforms for change detection. We use a subset of pre-disaster images from the multi-temporal dataset and pair them with post-disaster images from the xBD dataset. The labeled data provided by the xBD dataset is used to calculate the F1 score. This experiment consists of seven experimental groups, the main difference being the acquisition time of the pre-disaster images. It should be noted that the multi-temporal dataset is only used for testing, and all methods use model parameters and weights trained on the xBD dataset.

[0054] Table 3 Performance of various methods on multi-temporal datasets

[0055] To verify the stability of our proposed method, we present the quantitative results of our invention and comparative methods on multi-temporal datasets in Table 3. Experimental analysis shows that due to significant variations in pre-disaster images acquired by different sensors and platforms, there are substantial style differences between bi-temporal images. Table 3 shows the performance of each method in different time periods. Our invention achieved the highest F1 score in four of the seven time periods (June 2015, September 2015, September 2016, and March 2017), and performed second best in the March 2016 experimental group. In the remaining time periods, CCNet also maintained strong competitiveness. There is a certain correlation between input image quality and F1 score; that is, when the image quality is high, the F1 scores of all methods are generally high, and when the image quality is low, the F1 scores decrease. Even with low input image quality (such as in June 2015 and September 2016), our invention still maintains stable detection performance. When the input image quality is high (e.g., March 2016 and June 2016), all methods (including CCNet) perform better.

[0056] To provide a quantitative reference for image pairs across different time periods, we calculated the mutual information (Mut Info) between pre-disaster images and post-disaster images from the xBD dataset, and the results are listed in the last row of Table 3. Experimental results show that the F1 score trend is consistent with the mutual information trend. Even in periods with low mutual information (such as June and September 2015), thanks to the decoupling strategy of style and content features, this invention still achieved superior detection performance. Furthermore, we calculated the mean and standard deviation of the F1 scores for each method across all time periods, and the results are shown in the last two columns of Table 3. Since CCNet achieves ideal detection results in both low-quality and high-quality pre-disaster images, it has the highest average F1 score. Simultaneously, CCNet has the smallest standard deviation, further validating its stability in complex application scenarios. In summary, this invention demonstrates optimal performance across all time periods.

[0057] On the other hand, embodiments of the present invention also provide a high-resolution remote sensing image change detection system based on decoupled representation learning, comprising: The system includes a processor and a memory. The memory stores program instructions, and the processor calls the stored instructions in the memory to execute the high-resolution remote sensing image change detection method based on decoupled representation learning as described in the above technical solution.

[0058] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A high-resolution remote sensing image change detection method based on decoupled representation learning, characterized in that, Includes the following steps: Step 1: Acquire high-resolution remote sensing image pairs of the area to be detected at two different time phases and preprocess them to obtain training data; Step 2: Construct a content branch, which is an encoding and decoding structure. In the content encoding network, a multi-scale parallel convolution module is used to extract content features from the input remote sensing image pairs, and a content change decoding network is used to generate high-resolution remote sensing image change area detection results with the same size as the input image level by level. Step 3: Construct style branches, including using a lightweight style coding network to independently extract the style features of each remote sensing image pair, combining the style features extracted by the style coding network with the corresponding content features, and performing joint decoding through an image reconstruction decoding network with an adaptive instance normalization module to reconstruct the input original remote sensing image. Step 4: Using the preprocessed dual-temporal high-resolution remote sensing image pair as input, the images are input into the trained content branch and style branch respectively to obtain change detection results and reconstructed images.

2. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 1, characterized in that: Preprocessing includes: The remote sensing image data is cropped into standardized image slices of a certain size. The slices are made using a sliding window method, and a certain overlap area is set between the windows to increase the diversity of the data. Data augmentation operations are performed on the image pairs, including random rotation, horizontal flipping, or vertical flipping. The processed image data is then normalized to pixel values, linearly mapping the pixel value range to the (0,1) interval.

3. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 1, characterized in that: The multi-scale parallel convolution module includes multiple parallel convolution branches. The first branch outputs a feature map with the same spatial resolution as the input image. The spatial resolution of the output feature maps of subsequent branches decreases by half at each level, while the number of feature map channels doubles at each level. Each branch uses a different number of residual convolution modules for feature extraction. Each residual convolution module includes a 3×3 convolutional layer, a batch normalization layer (BN), and a ReLU activation function.

4. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 3, characterized in that: The multi-scale feature maps of each branch also interact and fuse information through upsampling or downsampling operations. Specifically, high-scale features are fused with low-scale features after downsampling, while low-scale features are fused with high-scale features through upsampling.

5. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 1, characterized in that: The change detection decoding network consists of multiple groups, each of which includes convolutional layers, batch normalization (BN), and rectified linear units (ReLU).

6. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 1, characterized in that: Style coding networks consist of multiple convolutional layers, activation function layers, pooling layers, and fully connected layers.

7. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 1, characterized in that: The adaptive instance normalization module AdaIN processes instances as follows: in, As a content feature, As a stylistic feature, and These represent the standard deviation and mean calculation operations for the feature map, respectively.

8. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 1, characterized in that: The image reconstruction decoder is implemented by cascading multiple restoration decoding modules. Each restoration decoding module takes content features and style features as input, and sequentially injects style features into content features through feature upsampling, convolutional fusion, and AdaIN operations to restore the original appearance of the remote sensing image. The restoration decoding process is specifically expressed as follows: in, This is the output feature map of the m-th restoration decoding module in the style branch. To restore the decoding module, This is an upsampling operation on the feature map. This is the stitching operation for the feature channels; the features output by the adaptive instance normalization module AdaIN are processed by the ReLU activation function and then input to the next level restoration decoding module. This process is repeated until a restored remote sensing image with the same size and similar appearance as the input image is recovered.

9. The high-resolution remote sensing image change detection method based on decoupled representation learning as described in claim 1, characterized in that: The change detection loss, feature separation loss, content similarity loss, and image reconstruction loss are weighted and combined to form the total loss function for training the content branch and style branch. The specific calculation methods for each loss are as follows: Change detection loss A pixel-wise binary classification cross-entropy loss function is used to supervise the difference between the change detection network output and the ground truth change detection labels, specifically expressed as follows: Where N is the total number of pixels in the sample. The label for the actual change of the i-th pixel. Predict the probability of change for the i-th pixel; Feature separation loss By employing mutual information constraints between content and style features, the independence and decoupling of content and style feature spaces are ensured, as expressed below: in, These respectively represent content and style characteristics. Calculate the Gram matrix. It is the Frobenius norm; Content similarity loss To ensure that the content features of unchanged regions remain consistent over time, it is expressed as follows: in, This represents the total number of pixels in the unchanged region. These represent the content features of the j-th pixel in two time-unchanged regions, respectively. Image reconstruction loss The pixel-level mean square error (MSE) between the reconstructed image and the original input image is used to ensure the integrity of the decoupled content and style feature information, as expressed below: in, To reconstruct the image, This is the original input image.

10. A high-resolution remote sensing image change detection system based on decoupled representation learning, characterized in that, include: The processor and memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the high-resolution remote sensing image change detection method based on decoupled representation learning as described in any one of claims 1-9.

Images