An automated detection method and system for optical and SAR heterogeneous remote sensing images

Abstract

This invention discloses an automated detection method and system for heterogeneous remote sensing images (optical and SAR). First, optical and SAR images from different time phases are preprocessed. Then, a heterogeneous change detection model based on a pseudo-twin channel feature extraction encoder is constructed, and multi-level features of the optical and SAR images are encoded in parallel. A semantically knowledge-guided multi-path feature enhancement module strengthens both shallow and deep features. Subsequently, a progressive fusion decoder is used to fuse the multi-scale features output from the encoding stage step by step. A difference map refinement module fully mines the complementary information and differences between the optical and SAR images, accurately locating and identifying change areas, and generating the final change detection result. This invention effectively improves the accuracy, recall, and overall accuracy of heterogeneous remote sensing image change detection, and exhibits good robustness and scalability.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image application technology, and more specifically, to an automated detection method and system for optical and SAR heterogeneous remote sensing images. Background Technology

[0002] Change detection is a classic and highly regarded topic in the field of remote sensing. It uses multiple images of the same geographic area to identify changes in ground objects. This technology has been widely applied in various fields such as urban planning, resource surveys, and natural disaster management. By analyzing the differences between multiple images, change detection can help identify urban expansion, resource utilization, and the impact range of natural disasters, providing important support and reference for decision-making and planning in related fields.

[0003] Source change detection typically utilizes remote sensing images of the same type within the same geographic area to identify changes in surface objects. This method usually involves comparative analysis of remote sensing images of the same type to detect changes in surface objects at different time points or between different acquisition periods. In the field of remote sensing, it can be applied to multiple areas such as urban development monitoring, land use change analysis, environmental monitoring, and natural disaster assessment. By detecting changes in images of the same type, people can gain a deeper understanding of the trends and influencing factors of surface changes, providing important support for relevant decision-making and planning. However, the implementation of this method usually requires consideration of image registration, the fusion of spectral, spatial, and temporal features, and the impact of data quality. Therefore, continuous exploration and improvement of source change detection algorithms are necessary to enhance its detection accuracy and application scope.

[0004] Compared to homogeneous change detection, heterogeneous change detection involves using remote sensing imagery of different types and sources to identify changes in surface objects, such as optical and SAR imagery. Currently, most research focuses on change detection from a single data source, especially optical imagery. Optical imagery with different spatial and temporal resolutions can provide not only spectral information about ground features but also reflect their texture and geometry, ensuring the feasibility and accuracy of change detection tasks. In contrast, synthetic aperture radar (SAR) imagery can reflect the scattering information of the surface. Because SAR imagery is not limited by weather and time and is unaffected by lighting conditions, it has advantages in specific change detection tasks. Change detection based on optical and SAR imagery is becoming popular because it provides complementary information. The combination of optical and SAR imagery can reflect the situation before and after an event in a timely and accurate manner, especially after disasters such as floods or fires. Optical imaging cannot obtain clear information about the surface due to moisture or smoke, while SAR can quickly respond to these situations. Therefore, combining these two types of imagery not only achieves complementary information but also helps in change detection when homogeneous imagery is unavailable.

[0005] In the field of change detection, challenges include high training sample requirements, poor algorithm transferability, and low detection accuracy in complex scenes. First, the high demand for training samples means that both traditional algorithms and deep learning-based methods typically require a large number of labeled data samples, which is often difficult to achieve in practice. Second, the transferability of algorithms is insufficient; some methods rely on specific datasets or modality invariance assumptions, leading to a sharp decline in performance under different datasets or conditions. Finally, for complex scenes, such as high resolution, large scenes, and strong noise, existing algorithms have low detection accuracy and require further improvement. Based on these problems, this invention proposes an automated detection method and system for optical and SAR heterogeneous remote sensing images. Summary of the Invention

[0006] To address the shortcomings of existing technologies and the need for improvement, this invention provides an automated detection method for optical and SAR heterogeneous remote sensing images, featuring innovative designs in feature extraction and information fusion. The method includes the following steps: Step 1: Using optical images and SAR images acquired by remote sensing satellites as the dataset, obtain the corresponding actual change images in the dataset, and divide the actual change images and the corresponding dual-temporal images into training sets and test sets. Step 2: Construct an automated detection model for heterogeneous remote sensing images based on a pseudo-twin channel feature extraction encoder and a progressive fusion decoder. The pseudo-twin channel feature extraction encoder performs multi-level parallel feature encoding on both optical and SAR images to extract corresponding heterogeneous change features. Step 3: The features extracted in Step 2 are enhanced by the multi-path feature enhancement module guided by semantic knowledge, which guides the model to focus on the foreground change area and suppresses background interference. At the same time, the alignment of heterogeneous feature distributions between optical image and SAR image is achieved. Element-wise difference operations are performed on the optical image features and SAR image features at the same level to obtain the difference feature map. Step 4: The progressive fusion decoder is used to fuse the differential feature maps in Step 3 step by step, and the complementary information and difference relationship between optical images and SAR images are mined through the difference map refinement module to accurately locate and identify the change area and generate change detection results. Step 5: Use the training set from Step 1 to train the automated detection model for heterogeneous remote sensing images from Step 2, and use the trained model to achieve change detection.

[0007] Further, step 1 uses manually created urban change heterogeneous remote sensing images as the dataset, and creates actual change images based on the dual-temporal images in the dataset. The actual change images are the changed areas in the dual-temporal images, and each pixel in the actual change images represents a category, either unchanged or changed. The preceding and following time-series images and their corresponding actual change images are combined to form an automated urban detection image dataset, and the dataset is divided into training and test sets according to a certain ratio.

[0008] Furthermore, the pseudo-twin channel feature extraction encoder described in step 2 uses feature extraction channels with consistent structure but independent parameters, namely the Transformer coding network, which is used for feature encoding of optical images and SAR images respectively, in order to adapt to the differences in imaging mechanism and feature distribution between the two types of images.

[0009] Furthermore, the specific implementation method of step 2 is as follows: Step 2.1: Construct a dual-channel feature extraction encoder, which is used for feature extraction of optical images and SAR images respectively. The two channels use Transformer coding networks with the same structure but independent parameters. Step 2.2: In each feature extraction channel, the Biformer model is used to encode the features of the input image using its two-level path attention mechanism. Step 2.3: Extract feature representations at multiple levels from optical images and SAR images respectively. Extract n different levels of features from each channel, resulting in a total of 2n multi-scale features. Step 2.4, denoted as the features extracted from the optical image at layer i. The features extracted from the SAR image at the i-th layer are denoted as... ,in ; Step 2.5: Input the corresponding optical image features and SAR image features into the multipath feature enhancement module for further refinement.

[0010] Furthermore, the specific implementation method of the semantic knowledge-guided multi-path feature enhancement module in step 3 is as follows: Step 3.1, for the optical image features of each level i and SAR image features Spatial attention and channel attention mechanisms are introduced to jointly enhance features; Step 3.2, targeting SAR image features The feature map is obtained by using a spatial attention mechanism to capture important regional responses of features in the spatial dimension and a channel attention mechanism to capture important semantic information of features in the channel dimension. The two are then fused to obtain the enhanced feature map. The calculation formula is as follows: (1) In the formula, This is the feature map after processing at the i-th layer. For the spatial attention of the i-th layer, For parameters of channel attention; Step 3.3, apply the enhanced features obtained in Step 3.2. Convolution will The number of channels has been reduced to ,in It is a hyperparameter, and the resulting feature map is denoted as . ,in ,in , , They represent Width, height, and passageway; Step 3.4, use As a target reference, then, in the first The layer adjusts the size of the feature map to match the... For the same resolution, the calculation formula is: (2) In the formula, P represents adaptive average pooling, M represents identity mapping, and I represents bilinear interpolation. arrive The resolution, of which ; Step 3.5, apply a Convolutions are used to smooth each resized feature map. The calculation formula is: (3) In the formula, For the parameters of smooth convolution, Let j be the smooth feature map of the i-th layer; Step 3.6: After adjusting all i-th level feature maps to the same resolution, apply the Hadamard product of elements to all resized feature maps to enhance the i-th level features with more semantic information and finer details. The calculation formula is as follows: (4) In the formula For the Hadamard product, by applying sequentially Convolution, Convolution, Convolution and upsampling operations are used to process The resulting feature map is denoted as ; Step 3.7, will and The enhanced feature map obtained by performing element-wise addition is denoted as . The calculation formula is: (5) In the formula, Add to Tensor; Step 3.8, targeting optical image features The enhanced feature map is obtained using the methods described in steps 3.2-3.7. .

[0011] Furthermore, in step 3, element-wise difference operations are performed on the optical image features and SAR image features at the same level, i.e. , , and The four difference feature maps obtained are denoted as follows: , , and .

[0012] Furthermore, in step 4, the difference map refinement module refines the difference feature map, including: Step 4.1, denote the obtained difference feature map as... , , and ,use The convolution will Size increased to Consistent, the resulting feature map is denoted as Then and The feature map obtained by concatenating the first dimension is denoted as follows: ; Step 4.2, use channel attention mechanism to... The channel-dimensional features are refined to obtain a refined feature map of channel attention. Then, a spatial attention mechanism is used to refine the spatial-dimensional features to obtain a refined feature map of spatial attention, denoted as... ; Step 4.3, adopt the same fusion method as in steps 4.1-4.2, by... Obtain refined feature map ,Depend on Obtain refined feature map ; Step 4.4, using The convolution will The number of channels is reduced to 1, and then an upsampling operation is performed. The size is increased to The resulting feature map is denoted as Where H and W represent height and width, respectively; Step 4.5, combine the three feature maps Tensor stacking in the first dimension yields feature information at different levels and through different channels, which is then used... The convolutional algorithm integrates feature information from multiple branches, fusing them into a richer and more expressive feature representation. This representation is then mapped by a change detection head to achieve pixel-level classification between changed and unchanged regions.

[0013] Furthermore, the formula for calculating the refined feature map of channel attention is as follows: (7) (8) (9) In the formula, For input features, For adaptive average pooling layer operations, To adapt to max pooling layer operations, , for Convolution operation, for Activation function Input features The average feature map in the channel dimension. Input features The feature map with the maximum value in the channel dimension. This is an element-wise addition operation. for Activation function Refine the features for channel attention.

[0014] Furthermore, the formula for calculating the refined feature map of spatial attention is as follows: (10) (11) (12) In the formula, For input features, This refers to the stacking operation of tensors in the first dimension. for Convolution operation, Input features The average feature map in the channel dimension. Input features The feature map with the maximum value in the channel dimension. This is an element-wise multiplication operation. for Activation function For spatial attention refinement features, that is .

[0015] The present invention also provides an automated detection system for optical and SAR heterogeneous remote sensing images, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the automated detection method for optical and SAR heterogeneous remote sensing images as described in the above technical solution.

[0016] In summary, compared with existing technologies, the technical solutions conceived in this invention achieve the following beneficial effects: Based on deep convolutional neural networks, an automated detection model for heterogeneous remote sensing images, composed of an encoder and a decoder, is constructed. This model effectively identifies and fuses multi-scale feature information in dual-temporal heterogeneous images, significantly improving the accuracy of change detection. Ultimately, by simply inputting the dual-temporal heterogeneous images into the trained model, changes in the heterogeneous images can be automatically detected. This invention fully utilizes the rich spectral and texture information of optical images and the structural and scattering characteristics of SAR images. Through multi-path feature enhancement and progressive fusion of difference maps, it effectively improves the accuracy, recall, and overall accuracy of heterogeneous remote sensing image change detection. It exhibits good robustness and scalability, providing a high-precision and high-efficiency technical means and application platform for surface change monitoring, dynamic environmental assessment, planning management, and risk assessment. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an automated detection method for optical and SAR heterogeneous remote sensing images provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of an automated detection model for heterogeneous remote sensing images provided in an embodiment of the present invention.

[0019] Figure 3 This is a structural diagram of the multi-path feature enhancement module provided for an embodiment of the invention.

[0020] Figure 4 A detailed structural diagram of the difference diagram provided for embodiments of the invention.

[0021] Figure 5 The images show a comparison of different detection methods provided in the embodiments of the present invention, with (a)-(e) representing different scenarios. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0023] This invention constructs a heterogeneous change detection model based on a pseudo-twin channel feature extraction encoder. It encodes multi-level features from optical and SAR images in parallel, and enhances shallow and deep features through a semantically knowledge-guided multi-path feature enhancement module. This allows the model to focus on foreground change regions and suppress background interference, while simultaneously aligning heterogeneous feature distributions. Subsequently, a progressive fusion decoder fuses the multi-scale features output from the encoding stage at each level. A difference map refinement module fully exploits the complementary information and differences between optical and SAR images, enabling precise localization and discrimination of change regions, thereby pushing the detection results closer to the real change images. Five evaluation metrics are used to assess the performance of this invention: precision, recall, F1-score, intersection (IOU), and overall accuracy (OA). Figure 1 The diagram shown is a flowchart of an automated detection method for optical and SAR heterogeneous remote sensing images provided by an embodiment of the present invention, which specifically includes the following steps: Step 1: Dataset Construction: A dataset is constructed using images of the study area acquired by remote sensing satellites. The actual change images corresponding to the data are obtained, and the actual change images and the corresponding dual-temporal heterogeneous images are used as the dataset. Constructing a reasonable change detection dataset can effectively improve the detection accuracy of the model. In the experiments of this invention embodiment, the Wuhan-dataset dataset was used, which contains dual-temporal high-resolution Landsat-8 optical images and Sentinel-1 AVV polarization SAR images. These images captured the changes in the study area from January to February 2017. The original size of each image is 3500×2000 pixels, with a spatial resolution of 15m. Considering the limitation of GPU memory capacity, an image segmentation algorithm was used to cut each image into regional images of size 256×256 pixels, ultimately obtaining 793 pairs of consecutive temporal images. This invention uses professional computer vision annotation software to annotate heterogeneous images. For each pair of consecutive temporal images, the corresponding ground truth of the actual change image is obtained. Each pixel in the actual change image represents a category, where the class label in the actual change image is represented by 0 and 1, where 0 represents an unchanged area (which can be displayed as black) and 1 represents a changed area (which can be displayed as white).

[0024] After the above processing, we obtain sequential images of different sources and their corresponding actual changes. We then combine these sequential images and their corresponding actual changes into an automated detection dataset of remote sensing images of different sources. Within this dataset, we divide the dataset into a training set (552 images) and a test set (241 images) in an 8:2 ratio.

[0025] Step 2: Construction of an automated detection model for heterogeneous remote sensing images: Construct an automated detection model for heterogeneous remote sensing images consisting of a pseudo-twin channel feature extraction encoder and a progressive fusion decoder; like Figure 2 As shown, the automated detection model for heterogeneous remote sensing images in this embodiment of the invention includes two main modules: an encoder and a decoder. The encoder comprises a weight-shared dual-channel pseudo-Twin network and a multi-path feature enhancement module, while the decoder includes a multi-scale feature fusion and difference map refinement module. The encoder is responsible for extracting multi-scale feature information and high-level semantic information from the input image. The decoder progressively fuses the extracted multi-scale features and, combined with contextual difference information, accurately locates and discriminates changing regions, continuously pushing the resulting image closer to the ground truth.

[0026] like Figure 2 As shown in the "encoder" diagram, a dual-channel feature extraction encoder is first constructed, used for feature extraction from optical and SAR images respectively. Both channels employ a Transformer coding network with identical structure but independent parameters. In each feature extraction channel, a Biformer model is used, leveraging its two-level path attention mechanism to encode features from the input image.

[0027] The core module, the Biformer model, extracts feature representations at multiple levels from both optical and SAR images. Each channel extracts four different levels of features, resulting in a total of eight multi-scale features. The feature extracted from the optical image at the i-th level is denoted as... The features extracted from the SAR image at the i-th layer are denoted as... ,middle The corresponding optical image features and SAR image features are input into the multipath feature enhancement module for further refinement.

[0028] Step 3: Construct a semantic knowledge-guided multi-path feature enhancement module to strengthen the extracted shallow and deep features, guide the model to focus on the foreground change area, and suppress background interference; like Figure 3 The multi-path feature enhancement module structure diagram is shown in the figure, which describes the optical image features for each level i. and SAR image features Spatial attention and channel attention mechanisms are introduced to jointly enhance features.

[0029] Targeting SAR image features The spatial attention mechanism is used to obtain the important regional responses of the features in the spatial dimension, and the channel attention mechanism is used to obtain the important semantic information of the features in the channel dimension. The two are then fused to obtain the enhanced feature map. The calculation formula is shown in equation (1): (1) In the formula, This is the feature map after processing at the i-th layer. For the spatial attention of the i-th layer, For parameters of channel attention; The obtained enhanced features are then subjected to... Convolution will The number of channels has been reduced to ,in It is a hyperparameter. The resulting feature map is denoted as... ,in ,in , , They represent Width, height, and passageway.

[0030] like Figure 2 The "decoder" in the code, as it is known, sends the improved feature map to the decoder. At each decoder level i, this structure uses... As a target reference, then, in each of the... The layer adjusts the size of the feature map to match the... For the same resolution, the calculation formula is shown in equation (2): (2) In the formula, P represents adaptive average pooling, M represents identity mapping, and I represents bilinear interpolation. arrive The resolution, of which .

[0031] Apply one Convolutions are used to smooth each resized feature map. The calculation formula is shown in equation (3): (3) In the formula, For the parameters of smooth convolution, Let be the j-th smooth feature map of the i-th layer.

[0032] After adjusting all i-th level feature maps to the same resolution, the Hadamard product of the elements is applied to all resized feature maps to enhance the i-th level features with more semantic information and finer details. The calculation formula is shown in Equation (4): (4) In the formula For the Hadamard product, by applying sequentially Convolution, Convolution, Convolution and upsampling operations are used to process The resulting feature map is denoted as .

[0033] Will and The feature map obtained by performing element-wise addition is denoted as The calculation formula is shown in equation (5): (5) In the formula, Add to Tensor.

[0034] For optical image features Enhanced feature maps are obtained using the same method described above. .

[0035] Furthermore, element-wise difference operations are performed on the optical image features and SAR image features at the same level, i.e. , , and The four difference feature maps obtained are denoted as follows: , , and ,and The calculation formula is shown in equation (6): (6) In the formula, Subtract from Tensor.

[0036] Step 4: Construct a difference map refinement module to mine complementary information and difference relationships between optical images and SAR images, accurately locate and identify changed areas, and generate change detection results; like Figure 4 The structure diagram of the difference map refinement module is shown in the figure. , , and It is used as input to the difference map refinement module for refinement. By applying... The convolution will Size increased to Consistent, the resulting feature map is denoted as Then... and The feature map obtained by concatenating the first dimension is denoted as follows: Channel attention mechanism is used to... The channel-dimensional features are refined to obtain a refined feature map of channel attention. Then, a spatial attention mechanism is used to refine the spatial-dimensional features to obtain a refined feature map of spatial attention, denoted as... ; The calculation formulas for the refined feature map of channel attention are shown in equations (7), (8), and (9): (7) (8) (9) In the formula, For input features, For adaptive average pooling layer operations, To adapt to max pooling layer operations, , for Convolution operation, for Activation function Input features The average feature map in the channel dimension. Input features The feature map with the maximum value in the channel dimension. This is an element-wise addition operation. for Activation function For channel attention refinement features, for ease of calculation, let's first denote them as... ; The formulas for calculating the refined feature map of spatial attention are shown in equations (10), (11), and (12): (10) (11) (12) In the formula, For input features, This refers to the stacking operation of tensors in the first dimension. for Convolution operation, Input features The average feature map in the channel dimension. Input features The feature map with the maximum value in the channel dimension. This is an element-wise multiplication operation. for Activation function For spatial attention refinement features, that is .

[0037] Based on the above differential feature map The fusion yields a refined feature map. Therefore, by adopting the same fusion method, by A refined feature map can be obtained. ,Depend on A refined feature map can be obtained. .and To reduce the computational cost and number of parameters in the model, the difference map refinement module uses... The convolution will The number of channels is reduced to 1, and then an upsampling operation is performed. The size is increased to The resulting feature map is denoted as .and ; Finally, the three feature maps Tensor stacking in the first dimension yields feature information at different levels and through different channels, which is then used... Convolution integrates feature information from multiple branches, fusing them into a richer and more expressive feature representation. ,and Feature representation This constitutes a high-level discriminative feature space for change detection, which is then mapped by the change detection head to achieve pixel-level classification of changed and unchanged regions.

[0038] Step 5: Use the training set from Step 1 to train the automated detection model for heterogeneous remote sensing images from Step 2. Use the trained model to detect changes. Finally, use the evaluation index of the automated detection model for heterogeneous remote sensing images to evaluate the detection results. The network structure proposed in this invention is used to train the model on the Wuhan-dataset constructed in step 1, and the model weights are obtained for model evaluation. The training process is based on the PyTorch deep learning framework, with Ubuntu 20.04 as the software environment and a 3090 graphics card with 24GB of video memory as the hardware environment. The batch size is set to 8, and the total number of training epochs is 100. Each input contains three images: a first-temporal image, a second-temporal image, and an image showing actual changes. After each training iteration, a test is performed. During the network training process, the urban change information in the two-temporal images and the image showing actual changes is continuously learned. The iteration continues until 100 epochs are reached, at which point the training ends.

[0039] Precision, recall, F1-score, Intersection over Union (IOU), and Overall Accuracy (OA) are selected as evaluation metrics, and their calculation formulas are as follows:

[0040]

[0041]

[0042]

[0043]

[0044] To verify the performance of the automated detection model for heterogeneous remote sensing images proposed in this invention, the final experimental results are presented. Figure 5 The image shows a visual comparison of the various methods, and Table 1 shows the quantitative indicators for each method.

[0045] in, Figure 5 This section represents the change detection results of heterogeneous remote sensing images obtained by various methods. (a)-(e) correspond to different scenes, with Optical representing the preceding image, SAR representing the following image, GT representing the actual change image (Ground Truth, GT), and DA_MSCDNet, DTCDN, FC-Siam-conc, FC-EF, FocalVIT, MTCDN, and HCDFormer representing the detection results of different methods. By comparing the actual change images, black represents unchanged areas, white represents changed areas, red represents false detection areas, and green represents missed detection areas.

[0046] Table 1: Detection accuracy of various methods on the Wuhan-dataset dataset

[0047] Please note that all indicators are in percentage form, and the higher the value, the better the effect. For easier observation, the best results are shown in bold.

[0048] References to DA_MSCDNet: Zhang C, Feng Y, Hu L, et al. A domainadaptation neural network for change detection with heterogeneous optical and SAR remote sensing images[J]. International Journal of Applied EarthObservation and Geoinformation, 2022, 109: 102769. References for FocalVITCD: Shen P, Mei L, Ye Z, et al. Building ChangeDetection in Remote Sensing Imagery with Focal Self-Attention and Multi-LevelFeature Fusion[J]. Electronics, 2023, 12(13): 2796. DTCDNet can refer to the literature: Li References for MTCDN: Du Z, Li For references on FC-Siam-conc and FC-EF: Daudt RC, Le Saux B, Boulch A. Fullyconvolutional siamese networks for change detection[C] / / 2018 25th IEEE international conference on image processing (ICIP). IEEE, 2018: 4063-4067. On the other hand, embodiments of the present invention also provide an automated detection system for optical and SAR heterogeneous remote sensing images, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the automated detection method for optical and SAR heterogeneous remote sensing images as described in the above technical solution.

[0049] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0050] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the protection scope of the present invention.

Claims

1. An automated detection method for optical and SAR heterogeneous remote sensing images, characterized in that, Includes the following steps: Step 1: Collect optical images and SAR images as a dataset, obtain the corresponding actual change images in the dataset, and divide the actual change images and the corresponding dual-temporal images into training sets and test sets. Step 2: Construct an automated detection model for heterogeneous remote sensing images based on a pseudo-twin channel feature extraction encoder and a progressive fusion decoder. The pseudo-twin channel feature extraction encoder is used to perform multi-level parallel feature encoding on optical images and SAR images respectively to extract the corresponding heterogeneous change features. Step 3: The features extracted in Step 2 are enhanced by the multi-path feature enhancement module guided by semantic knowledge. The optical image features and SAR image features are subjected to element-wise difference operation at the same level to obtain the difference feature map. Step 4: The differential feature map in Step 3 is fused step by step using a progressive fusion decoder. The complementary information and difference relationship between the optical image and the SAR image are mined through the difference map refinement module to accurately locate and identify the change area and generate change detection results. Step 5: Use the training set from Step 1 to train the automated detection model for heterogeneous remote sensing images from Step 2, and use the trained model to perform change detection.

2. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 1, characterized in that: Step 1 uses manually created urban change heterogeneous remote sensing images as the dataset. Based on the dual-temporal images in the dataset, actual change images are created. The actual change images are the changed areas in the dual-temporal images. Each pixel in the actual change images represents a category, either unchanged or changed. The preceding and following time-series images and their corresponding actual change images are combined to form an automated urban detection image dataset. In this dataset, training and test sets are divided according to a certain ratio.

3. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 1, characterized in that: The pseudo-twin channel feature extraction encoder described in step 2 uses feature extraction channels with consistent structure but independent parameters, namely the Transformer coding network, which is used for feature encoding of optical images and SAR images respectively, in order to adapt to the differences in imaging mechanism and feature distribution between the two types of images.

4. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 1, characterized in that: The specific implementation method of step 2 is as follows: Step 2.1: Construct a dual-channel feature extraction encoder, which is used for feature extraction of optical images and SAR images respectively. The two channels use Transformer coding networks with the same structure but independent parameters. Step 2.2: In each feature extraction channel, the Biformer model is used to encode the features of the input image using its two-level path attention mechanism. Step 2.3: Extract feature representations at multiple levels from optical images and SAR images respectively. Extract n different levels of features from each channel, resulting in a total of 2n multi-scale features. Step 2.4, denoted as the features extracted from the optical image at layer i. The features extracted from the SAR image at the i-th layer are denoted as... ,in ; Step 2.5: Input the corresponding optical image features and SAR image features into the multipath feature enhancement module for further refinement.

5. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 1, characterized in that: The specific implementation of the semantic knowledge-guided multi-path feature enhancement module in step 3 is as follows: Step 3.1, for the optical image features of each level i and SAR image features Spatial attention and channel attention mechanisms are introduced to jointly enhance features; Step 3.2, targeting SAR image features The feature map is obtained by using a spatial attention mechanism to capture important regional responses of features in the spatial dimension and a channel attention mechanism to capture important semantic information of features in the channel dimension. The two are then fused to obtain the enhanced feature map. The calculation formula is as follows: （1） In the formula, This is the feature map after processing at the i-th layer. For the spatial attention of the i-th layer, For parameters of channel attention; Step 3.3, apply the enhanced features obtained in Step 3.

2. Convolution will The number of channels has been reduced to ,in It is a hyperparameter, and the resulting feature map is denoted as . ,in ,in , , They represent Width, height, and passageway; Step 3.4, use As a target reference, then, in the first The layer adjusts the size of the feature map to match the... For the same resolution, the calculation formula is: （2） In the formula, P represents adaptive average pooling, M represents identity mapping, and I represents bilinear interpolation. arrive The resolution, of which ; Step 3.5, apply a Convolutions are used to smooth each resized feature map. The calculation formula is: （3） In the formula, For the parameters of smooth convolution, Let j be the smooth feature map of the i-th layer; Step 3.6: After adjusting all i-th level feature maps to the same resolution, apply the Hadamard product of elements to all resized feature maps to enhance the i-th level features with more semantic information and finer details. The calculation formula is as follows: （4） In the formula For the Hadamard product, by applying sequentially Convolution, Convolution, To process convolution and upsampling operations The resulting feature map is denoted as ; Step 3.7, will and The enhanced feature map obtained by performing element-wise addition is denoted as . The calculation formula is: （5） In the formula, Add to Tensor; Step 3.8, targeting optical image features The enhanced feature map is obtained using the methods described in steps 3.2-3.

7. .

6. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 5, characterized in that: In step 3, element-wise difference operations are performed on the optical image features and SAR image features at the same level, i.e. , , and The four difference feature maps are obtained and denoted as follows: , , and .

7. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 1, characterized in that: In step 4, the difference map refinement module refines the difference feature map, including: Step 4.1, denote the obtained difference feature map as... , , and ,use The convolution will Size increased to Consistent, the resulting feature map is denoted as Then and The feature map obtained by concatenating the first dimension is denoted as follows: ; Step 4.2, use channel attention mechanism to... The channel-dimensional features are refined to obtain a refined feature map of channel attention. Then, a spatial attention mechanism is used to refine the spatial-dimensional features to obtain a refined feature map of spatial attention, denoted as... ; Step 4.3, adopt the same fusion method as in steps 4.1-4.2, by... Obtain refined feature map ,Depend on Obtain refined feature map ; Step 4.4, using The convolution will The number of channels is reduced to 1, and then an upsampling operation is performed. The size is increased to The resulting feature map is denoted as Where H and W represent height and width, respectively; Step 4.5, combine the three feature maps Tensor stacking in the first dimension yields feature information at different levels and through different channels, which is then used... The convolutional algorithm integrates feature information from multiple branches, fusing them into a richer and more expressive feature representation. This representation is then mapped by a change detection head to achieve pixel-level classification between changed and unchanged regions.

8. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 7, characterized in that: The formula for calculating the refined feature map of channel attention is as follows: （7） （8） （9） In the formula, For input features, For adaptive average pooling layer operations, To adapt to max pooling layer operations, , for Convolution operation, for Activation function Input features The average feature map in the channel dimension. Input features The feature map with the maximum value in the channel dimension. This is an element-wise addition operation. for Activation function Refine the features for channel attention.

9. The automated detection method for optical and SAR heterogeneous remote sensing images according to claim 7, characterized in that: The formula for calculating the refined feature map of spatial attention is as follows: （10） （11） （12） In the formula, For input features, This refers to the stacking operation of tensors in the first dimension. for Convolution operation, Input features The average feature map in the channel dimension. Input features The feature map with the maximum value in the channel dimension. This is an element-wise multiplication operation. for Activation function For spatial attention refinement features, that is .

10. An automated inspection system for optical and SAR heterogeneous remote sensing images, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the automated detection method for optical and SAR heterogeneous remote sensing images as described in any one of claims 1-9.

Images