SAR image change detection method based on wavelet guidance and adaptive fusion

Abstract

The application discloses a SAR image change detection method based on wavelet guidance and adaptive fusion, and mainly solves the problems of low feature separability, insufficient noise interference suppression and insufficient multi-scale information utilization in the prior art. The method comprises the following steps: 1) collecting SAR images of the same area at different time phases and labeling pseudo labels to construct a training sample set; 2) building a detection model composed of a multi-scale feature extraction network, an adaptive feature enhancement module, a wavelet guided difference interaction module, an adaptive interaction fusion module and a full connection network in series; 3) training the model by using the training samples, generating change features through feature extraction, enhancement, double-branch difference interaction and fusion, updating network parameters by combining pseudo labels to calculate loss, and completing the training; and 4) obtaining the final detection result by using the trained model. The application can effectively improve the feature distinguishing ability, reduce the coherent speckle noise interference, and improve the accuracy and robustness of SAR image change detection in complex scenes.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and further relates to SAR image detection technology, specifically a SAR image change detection method based on wavelet-guided and adaptive fusion. It can be used in fields such as disaster monitoring, land use change monitoring, and ecological environment monitoring. Background Technology

[0002] Change detection is a crucial technology in remote sensing Earth observation. By analyzing the differences between remote sensing images from different time periods, it extracts information on changes in land cover and has wide applications in disaster monitoring, land use surveys, ecological environment monitoring, and national defense. Synthetic Aperture Radar (SAR), with its unique advantages of active microwave imaging, overcomes the limitations of weather conditions and light intensity, enabling it to stably acquire high-resolution surface scattering information under complex environments such as heavy rain, fog, and darkness. It is particularly suitable for long-term dynamic monitoring of complex terrain, vegetation cover, and areas obscured by clouds and rain, demonstrating irreplaceable application value in change detection tasks under extreme environments and has become the mainstream technology for land cover change detection.

[0003] Early SAR change detection methods mainly revolved around two core steps: difference image generation and threshold segmentation. By constructing a difference map from two temporal images and performing threshold segmentation, they could identify changed and unchanged regions. However, the difference map generation process struggled to fully utilize the deep feature information of the two temporal images and was easily affected by speckle noise. Simultaneously, the threshold segmentation results were highly sensitive to threshold selection, making it difficult to accurately distinguish changed regions from the background in complex scenes. Therefore, these methods were prone to problems such as blurred boundaries, missed detections of small targets, and false detections, and their detection accuracy and robustness needed further improvement.

[0004] With the development of deep learning technology, SAR image change detection methods based on neural networks have gradually become mainstream. End-to-end learning methods can automatically extract change features, effectively improving change detection accuracy. Among these, convolutional neural network-based methods can fully exploit spatial features of images and enhance the ability to identify change regions. However, they struggle to fully utilize the inherent correlation between two temporal images, resulting in insufficient separability between change features and background information, affecting the accurate identification of change regions. To further enhance the expressive power of change features, researchers have introduced multi-scale features and attention mechanisms to strengthen the network's perception of change regions. However, their multi-scale feature fusion is still insufficient, making it difficult to fully exploit the complementary information of features at different scales. Furthermore, frequency domain information has the advantage of suppressing speckle noise and enhancing features. Therefore, existing research combines frequency domain features with spatial domain features to improve the network's noise resistance and change detection performance. However, these methods mainly employ direct fusion of frequency and spatial domain features, lacking in-depth exploration of the frequency domain similarity of two temporal features, and thus failing to fully leverage the advantages of frequency domain information in noise suppression and change feature discrimination. Therefore, solving core problems such as low feature separability, insufficient noise interference suppression, and inadequate utilization of multi-scale information has become the key to improving the practicality of SAR change detection. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a SAR image change detection method based on wavelet-guided and adaptive fusion. This method solves problems such as low feature separability, insufficient noise interference suppression, and inadequate utilization of multi-scale information in existing detection schemes. This invention uses dual-temporal SAR images as input, mines potential change information across temporal phases through cross-temporal interaction, and adaptively optimizes this information to enhance feature discriminativeness and improve the separability of change features from background information. Subsequently, a dual-branch system is constructed to extract change features separately. Wavelet frequency domain decomposition and adaptive frequency band weight adjustment are used to suppress speckle noise interference, and frequency domain similarity is constructed to guide complementary learning of the dual branches to extract dual-temporal change features. Furthermore, through multi-scale change feature interaction and difference guidance, multi-scale change information is collaboratively optimized, and change information is adaptively fused, thereby significantly improving the accuracy and robustness of SAR image change detection in complex scenes.

[0006] To achieve the above objectives, the technical solution of the present invention includes the following:

[0007] (1) Acquire pairs of dual-temporal SAR images of the same area at different times, perform unsupervised pre-classification on them and label them with corresponding change pseudo-labels. The classification includes change class, no change class and uncertain class; randomly select samples from the change class and no change class to construct a training set, and the uncertain class constitutes the test set;

[0008] (2) Construct a multi-scale feature extraction network, using dual-temporal SAR images as input, to extract multi-scale features corresponding to the two temporal SAR images;

[0009] (3) Construct an adaptive feature enhancement module to perform bidirectional interaction on the multi-scale features corresponding to the two images, complete the context information perception enhancement, and output the enhanced features;

[0010] (4) Based on discrete wavelet transform and cosine similarity, a wavelet-guided difference interaction module is constructed to perform bi-branch difference interaction and collaborative optimization on the enhanced features, and after fusing bi-branch difference information, multi-scale change features are output.

[0011] (5) Construct an adaptive interactive fusion module to adaptively fuse and enhance multi-scale change features to obtain the final change features;

[0012] (6) A multi-scale feature extraction network, an adaptive feature enhancement module, a wavelet-guided difference interaction module, and an adaptive interaction fusion module are sequentially connected in series, and a classification network is connected after the adaptive interaction fusion module to construct a SAR change detection model; the classification network consists of a convolutional layer and two fully connected layers, which are used to output classification prediction results based on the final change features;

[0013] (7) The training set data is fed into the SAR change detection model, and the feature extraction, feature enhancement, two-branch difference interaction and feature fusion operations are performed in sequence to obtain the final change features corresponding to the training samples. Based on the features, the classification prediction results are obtained, the prediction results are matched with the sample pseudo-labels, the classification loss is calculated and the network parameters are updated through backpropagation to complete the model training.

[0014] (8) Use the trained SAR change detection model to detect the test set samples and output the detection results; then concatenate the results with the pseudo-label of the changed and non-changed classes to obtain the final change detection results.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] First, because the present invention adopts dual-temporal feature interaction, it fully explores the potential correlation information of the changing area through feature information interaction and guidance; at the same time, it performs context adaptive optimization of the interactive features to suppress background noise and irrelevant information interference, thereby bringing more sufficient feature expression capabilities for subsequent model detection, directly improving the feature recognition of weakly changing areas and reducing the probability of false background detection.

[0017] Secondly, because this invention uses a dual-branch extraction of SAR image change features and introduces wavelet transform to construct frequency domain similarity to guide dual-branch collaborative optimization, it effectively suppresses speckle noise interference and enhances the ability to extract differential features, directly reducing false detections caused by inherent speckle noise in SAR and strengthening the integrity of ground feature change boundary extraction.

[0018] Third, because the present invention adopts an interactive guidance mechanism of multi-scale change features, it achieves the synergistic enhancement of cross-scale semantics and detailed information through bidirectional information flow between features of different scales; at the same time, it integrates the difference information between different scales, makes full use of the complementarity between multi-scale features, realizes the fusion of multi-scale change features, directly takes into account the detection of overall changes in large areas and the detection of minor changes in ground features, and comprehensively improves the accuracy of change detection in complex ground feature scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention;

[0020] Figure 2 This is a schematic diagram illustrating the implementation process of SAR change detection using the method of the present invention;

[0021] Figure 3 This is a schematic diagram of the adaptive feature enhancement module structure in this invention;

[0022] Figure 4 This is a schematic diagram of the self-calibrating enhanced attention structure in the adaptive feature enhancement module of this invention;

[0023] Figure 5 This is a schematic diagram of the wavelet attention difference interaction module structure in this invention;

[0024] Figure 6 This is a schematic diagram of the wavelet cosine attention structure in the wavelet attention difference interaction module of the present invention;

[0025] Figure 7 This is a schematic diagram of the adaptive interactive fusion module structure in this invention;

[0026] Figure 8 This is a comparison chart of the detection results of the present invention and existing methods on the Farmland-A dataset provided in this embodiment of the invention;

[0027] Figure 9 This is a comparison chart of the detection results of the present invention and existing methods on the Farmland-B dataset provided in this embodiment of the invention;

[0028] Figure 10 This is a comparison chart of the detection results of the present invention and existing methods on the Ottawa dataset provided in this embodiment of the invention;

[0029] Appendix Figure 8-10 In the middle: (a)-(k) correspond to the existing Image I, Image II, Ground truth, DCNet, DDNet, LANTNet, CAMNet, WBANet, DBFNet, DSNet and MDSF methods respectively, and (l) corresponds to the method of this invention. Detailed Implementation

[0030] The present invention will now be further described with reference to the accompanying drawings.

[0031] Example 1: Refer to Appendix Figure 1 The present invention proposes a SAR image change detection method based on wavelet guidance and adaptive fusion, which specifically includes the following steps:

[0032] Step 1) Obtain paired dual-temporal SAR images of the same area at different times, perform unsupervised pre-classification on them and label them with corresponding change pseudo-labels. The classification includes change class, no change class and uncertain class; randomly select samples from the change class and no change class to construct the training set, and the uncertain class constitutes the test set.

[0033] Step 2) Construct a multi-scale feature extraction network, using dual-temporal SAR images as input, to extract multi-scale features corresponding to the two temporal SAR images;

[0034] Step 3) Construct an adaptive feature enhancement module to perform bidirectional interaction on the multi-scale features corresponding to the two images, complete context information perception enhancement, and output enhanced features; in this embodiment, the implementation process of this module includes the following:

[0035] (3.1) Relative position encoding is added to the multi-scale features of the two images, and cross-temporal information interaction is achieved through mutual guidance of their features, generating dual-temporal interactive features. In this embodiment, the above-mentioned cross-temporal information interaction specifically involves using the multi-scale feature information of the first SAR image to guide the multi-scale features of the second SAR image, and simultaneously using the multi-scale features of the second SAR image to guide the multi-scale features of the first SAR image in reverse, thus completing the bidirectional information interaction of dual-temporal features between the two images; through bidirectional interaction, the spatiotemporal correlation between the dual-temporal features is mined, adaptively amplifying the features of the changing regions that differ between the two images and suppressing redundant interference in the unchanged background regions.

[0036] (3.2) Construct a self-calibrating enhanced attention unit, input the dual-temporal interaction features into the self-calibrating enhanced attention unit, the self-calibrating enhanced attention unit is used to mine the internal structural information and context information of the features, perform adaptive enhancement processing on the dual-temporal interaction features, and output the context attention weight matrix.

[0037] In this embodiment, the context attention weight matrix is ​​obtained as follows: the dual-temporal interaction features are input into the self-calibrating enhanced attention unit. First, global context information is extracted through pooling, and then context features are generated through convolution and upsampling. At the same time, global average pooling is performed on the dual-temporal interaction features along the horizontal and vertical dimensions respectively, and structural attention weights are output. The context features and structural attention weights are fused to complete the feature self-calibration enhancement process, and finally the context attention weight matrix is ​​generated.

[0038] (3.3) Multiply the context attention weight matrix element by element with the dual temporal interaction features to complete the context information perception enhancement and obtain the enhanced features.

[0039] Step 4) Construct a wavelet-guided difference interaction module based on discrete wavelet transform and cosine similarity to perform bi-branch difference interaction and collaborative optimization on the enhanced features. After fusing the bi-branch difference information, output multi-scale change features, which are obtained according to the following steps.

[0040] (4.1) Using logarithmic ratio operation and channel splicing, two independent dual-branch differential features are constructed based on the enhanced features described in step (3);

[0041] (4.2) Construct a wavelet cosine attention unit, take the enhancement features described in step (3) as input, complete the frequency domain decomposition through wavelet transform, and obtain the frequency domain attention weight by combining cosine similarity calculation and normalization processing. Multiply the weight with the bi-branch difference features to output two enhanced difference features. In this embodiment, a discrete wavelet transform is performed on the input enhancement features to decompose the frequency sub-band into low-frequency and high-frequency sub-bands; then, adaptive weighting is performed on each frequency sub-band through wavelet feature selection; the similarity of each weighted sub-band is calculated through cosine similarity, and the frequency domain sub-band similarity is mapped to the spatial domain by inverse discrete wavelet transform, and converted into frequency domain attention weights by normalization through the Sigmoid function; the two independent bi-branch difference features are multiplied with the frequency domain attention weights respectively to complete the bi-branch difference feature enhancement.

[0042] (4.3) Using the two enhanced difference features as inputs to the wavelet cosine attention unit, interactive attention weights are generated to achieve bi-branch collaborative optimization, resulting in two interactive enhanced difference features. Specifically, in this embodiment, the two enhanced difference features are input into the wavelet cosine attention unit respectively, the frequency domain similarity between the two sets of enhanced difference features is calculated, and interactive attention weights are generated; then, the weights are used to perform bi-directional constraint optimization on the two enhanced difference features respectively, to achieve information complementarity and semantic alignment of the two features, and finally, two interactive enhanced difference features are obtained.

[0043] (4.4) Perform addition and fusion on the two interactive enhancement difference features to obtain complete multi-scale change features.

[0044] Step 5) Construct an adaptive interactive fusion module to adaptively fuse and enhance multi-scale change features to obtain the final change features. Implementation steps include:

[0045] (5.1) The interaction relationship between features of different scales is constructed by the SimAM attention mechanism. In this embodiment, the attention weight of one scale feature is applied to another scale feature to realize the information interaction between features of different scales and obtain multi-scale interaction enhancement features.

[0046] (5.2) Calculate the difference information for the change features at different scales, and then stitch and fuse them with the multi-scale interactive enhancement features to obtain the fused features. In this embodiment, this step specifically refers to calculating the pixel difference between multiple sets of change features at different scales to characterize the differences in ground feature deformation; stitching and fusing the calculated scale difference information with the multi-scale interactive enhancement features along the channel dimension to retain the independent deformation details of each scale and the cross-scale correlation features, weaken the noise interference of a single scale, and thus improve the completeness of the identification of abrupt change areas of ground features by the subsequent final change features.

[0047] (5.3) The fused features are adaptively enhanced using the SimAM attention mechanism to obtain the final changed features.

[0048] Step 6) The multi-scale feature extraction network, the adaptive feature enhancement module, the wavelet-guided difference interaction module, and the adaptive interaction fusion module are sequentially connected in series, and a classification network is connected after the adaptive interaction fusion module to construct the SAR change detection model; the classification network consists of a convolutional layer and two fully connected layers, which are used to output classification prediction results based on the final change features;

[0049] Step 7) Input the training set data into the SAR change detection model, and sequentially perform feature extraction, feature enhancement, two-branch differential interaction, and feature fusion operations to obtain the final change features corresponding to the training samples. Based on the features, obtain the classification prediction results, group and match the prediction results with the sample pseudo-labels, calculate the classification loss, and update the network parameters through backpropagation to complete the model training.

[0050] Step 8) Use the trained SAR change detection model to detect the test set samples and output the detection results; then concatenate the results with the pseudo-labels of the changed and non-changed classes to obtain the final change detection results.

[0051] Example 2: Refer to Appendix Figure 2-7The overall implementation process of the SAR image change detection method proposed in this embodiment is the same as that in Embodiment 1. The specific expressions and related parameters involved are given in a further detailed description of the implementation process of this invention:

[0052] Step 1. Input two SAR images to be detected, perform unsupervised pre-classification on the dual-temporal SAR images, and construct a dual-temporal SAR image sample set;

[0053] Step 2. Input the dual-temporal SAR image samples into the shared feature extraction network to extract dual-temporal multi-scale features;

[0054] Step 3. Refer to Figure 3 An adaptive feature enhancement module is constructed to enhance the interactive and contextual information awareness of the features extracted in step 2:

[0055] (3a) By guiding one temporal feature with the information of another temporal feature, information interaction between the two temporal features is achieved, the spatiotemporal relationship between the two temporal features is explored, the changing features are highlighted, and background interference is suppressed. At the same time, relative position encoding is introduced into the attention mechanism to enhance the spatial information of the features, as expressed below:

[0056] ,

[0057] ,

[0058] in, Indicates relative position encoding. , , express Query, key, and value vectors generated by 1×1 convolution This represents the Softmax activation function;

[0059] (3b) Refer to Figure 4 A self-calibrating enhanced attention mechanism is constructed, and the interaction features are averaged and max-pooled, followed by 3×3 convolution and batch normalization, and then element-wise summed. Contextual features are obtained through upsampling. To further enhance the ability of features to perceive spatial structure, Global average pooling is performed along both the horizontal and vertical directions to capture directional structural information. Subsequently, structural attention weights are generated using the sigmoid function and applied to the context features to obtain enhanced structural features. Next, and The features are then fused and multiplied element-wise with the original convolutional features to enhance the response in areas of change and suppress background interference. Finally, the obtained features are further calibrated through a 3×3 convolution, and then subjected to a 1×1 convolution with a sigmoid function to generate context weights. The expression is as follows:

[0060] ,

[0061] ,

[0062] ,

[0063] in, and These represent average pooling and max pooling followed by 3×3 convolution and batch normalization, respectively. Indicates bilinear upsampling. and These represent average pooling along the horizontal and vertical directions followed by the Sigmoid function, respectively. This represents the calibration features obtained after the input features are convolved by 3×3. This indicates a 3×3 convolution followed by a 1×1 convolution and a sigmoid function. Represents the Sigmoid function;

[0064] (3c) Multiply the above context weights with the dual-temporal features respectively to obtain the final enhanced features;

[0065] Step 4. Refer to Figure 5 A wavelet-guided difference interaction module is constructed to perform frequency domain feature similarity-guided bi-branch difference interaction on the enhanced features in step 3, thereby obtaining multi-scale variation features:

[0066] (4a) Two-branch differential features were constructed by using log ratio and channel splicing methods for the enhancement features respectively. and ;

[0067] (4b) Reference Figure 6 A wavelet cosine attention structure is constructed, and wavelet transform is applied to the input features to obtain four sub-bands. , , , Key frequency information is adaptively selected using wavelet feature selection. Sub-bands are then concatenated and features are extracted via 3×3 convolution. Global average pooling and convolution are used to generate weights for each frequency component, and the frequency domain features are adaptively weighted, finally decomposing them into four sub-band components. Subsequently, cosine similarity is calculated for the selected low-frequency and high-frequency components to measure the consistency between the two-phase frequency domain features. 1-Cos(⋅) is then used to highlight variation features, with higher values ​​indicating more significant differences. Finally, inverse discrete wavelet transform is used to convert the frequency domain features back to the spatial domain, and the results are input into a Sigmoid function to generate attention weights. The expression is as follows:

[0068] ,

[0069] ,

[0070] in Represents discrete wavelet transform. These represent the low-frequency, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency information of the input features, respectively. Indicates wavelet feature selection, Represents cosine similarity. Represents the inverse discrete wavelet transform;

[0071] (4c) Input the dual-temporal features into wavelet cosine attention to construct frequency domain feature similarity, and then combine it with... and The components are multiplied separately and then subjected to adaptive modulation to generate enhanced differential features. and Subsequently, the dual-branch enhancement feature was applied. and Difference features from another branch and The input to the wavelet cosine attention function calculates frequency domain similarity, guiding the collaborative optimization of the two-branch information flow, as expressed below:

[0072] ,

[0073] ,

[0074] in, Indicates wavelet cosine attention;

[0075] (4d) The above two-branch difference features are added and fused to obtain the final change features;

[0076] Step 5. Refer to Figure 7 An adaptive interactive fusion module is constructed to adaptively fuse the multi-scale variation features extracted in step 4:

[0077] (5a) Introduce the parameter-free attention mechanism SimAM to capture multi-scale variation features. and The inputs are fed into SimAM to generate attention weights, which are then applied to features at another scale to achieve cross-scale feature interaction enhancement and generate multi-scale interactive enhanced features. and The expression is as follows:

[0078] ,

[0079] ,

[0080] in, This represents the SimAm attention weight generation function;

[0081] (5b) Element-wise absolute value differencing is performed on the features at different scales to highlight the inconsistency between the spatial structure and semantic expression of features at different scales, thereby enhancing the salience of the change regions. Subsequently, the interactive enhancement features and differential features are concatenated and fused to achieve information complementarity, and feature compression is performed through 3×3 convolution to obtain the fused features. The expression is as follows:

[0082] ,

[0083] in, This indicates subtracting the absolute values ​​of each element. This indicates splicing along the channel dimension;

[0084] (5c) Input the fused features into SimAM, calculate the attention weights of the fused features, and multiply the fused features element by element to enhance important change information and suppress irrelevant background information; finally, perform residual fusion with the weighted features and the original fused features to output the final refined change features.

[0085] Step 6. Pass the output of Step 5 through a fully connected network to obtain the prediction result, and then match it one-to-one with the pseudo-labels to calculate the classification loss and the final detection result, thereby realizing the construction of the detection model;

[0086] Step 7. Train the detection model using the training sample set to obtain the trained detection model, and use the model to predict the change position between the two SAR images.

[0087] The effects of the present invention will be further explained below with reference to simulation experiments.

[0088] 1. Simulation conditions:

[0089] The simulation experiments of this invention were conducted on the hardware platform of NVIDIA GeForce GTX 1650 and the software platform of PyCharm 2022.

[0090] Three real SAR datasets were used in the experiment. The Farmland-A and Farmland-B datasets, acquired by the Radarsat-2 sensor, cover the Yellow River estuary region in China with a spatial resolution of 10 m. The two image sets were acquired in June 2008 and June 2009, respectively, corresponding to surface changes during farmland reclamation. The Farmland-A image size is 289×257, and the Farmland-B image size is 291×306. The Ottawa dataset, acquired by the Radarsat-1 sensor, covers the Ottawa region of Canada with a spatial resolution of 8 m and an image size of 350×290. The two image sets were acquired in May 1997 and August 1997, respectively, primarily reflecting surface changes before and after floods.

[0091] 2. Ablation experiment content and result analysis:

[0092] To verify the effectiveness of the three key modules in this invention—the adaptive feature enhancement module, the wavelet attention difference interaction module, and the adaptive interaction fusion module—this section conducted relevant ablation experiments on three datasets. The specific ablation experiment results and related analyses are as follows.

[0093] Table 1 Comparison results of ablation experiments on three datasets

[0094]

[0095] a) Adaptive Feature Enhancement Module (AFEM)

[0096] To verify the effectiveness of AFEM, this invention removed AFEM, Biphasic Interaction (BI), and Self-calibrated Enhanced Attention (SCEA), respectively. The experimental results are shown in Table 1. After removing AFEM, the results are better than those of the network without AFEM. The proposed AFEM improves PCC by 0.05%-0.22%, Kappa coefficient by 0.58%-0.80%, and F1 score by 0.56%-0.69%. This indicates that AFEM can effectively enhance the representation of biphasic features and improve the separability between change regions and background regions. Furthermore, when BI is removed, the Kappa coefficient decreases by 0.12%-0.61% and the F1 score decreases by 0.09%-0.59% on the three datasets, indicating that BI can fully mine potential change information by establishing the information interaction relationship between biphasic features, thereby enhancing the response of change regions. When SCEA was removed, the Kappa coefficient decreased by 0.16%–0.47%, and the F1 score decreased by 0.13%–0.40%. These results indicate that SCEA, by fusing contextual and structural information, adaptively enhances interactive features, further suppressing background interference and highlighting truly changing regions. BI and SCEA, through bi-temporal interaction and adaptive feature enhancement respectively, improve feature representation capabilities and more effectively increase the accuracy of change detection.

[0097] b) Wavelet Attention Difference Interaction Module (WADIM)

[0098] To verify the effectiveness of WADIM, this invention uses two classic structures for calculating bi-temporal image change features instead of WADIM for comparison: bi-temporal feature stitching and subtraction to calculate change features. Furthermore, DWT is removed to verify the effectiveness of wavelet frequency domain decomposition.

[0099] Table 1 shows the comparison results of different variation feature calculation methods on the three datasets. The WADIM module proposed in this invention achieved the best performance in all three datasets, outperforming other difference extraction modules overall. Feature subtraction operation achieved the worst performance on all three datasets, mainly because dual-temporal SAR images are severely affected by inherent noise. Directly subtracting dual-temporal features easily amplifies noise interference, resulting in the inability to extract effective difference information. Compared with feature subtraction, feature stitching achieved significant improvements in PCC, Kappa coefficient, and F1 score. Feature stitching preserves the complete information of dual-temporal features and avoids noise amplification, but it is also difficult to distinguish between real changes and unchanged information due to noise influence. Furthermore, after removing DWT, the model performance decreased significantly, indicating that frequency domain decomposition plays an important role in suppressing speckle noise and enhancing structural information representation. The WADIM module proposed in this invention suppresses noise interference through wavelet frequency domain decomposition and adaptive frequency band weighting, and constructs frequency domain similarity to guide cross-branch interaction, thereby extracting variation features in dual-temporal SAR images more accurately.

[0100] c) Adaptive Interactive Fusion Module (AIFM)

[0101] To verify the effectiveness of AIFM, this invention uses two classic feature fusion structures instead of AIFM for comparison: feature concatenation and feature addition.

[0102] Table 1 shows the comparison results of different feature fusion methods on the three datasets. The AIFM module proposed in this invention outperforms other feature fusion methods on all three datasets, achieving the best results. The evaluation index of the feature addition method decreased to varying degrees on all three datasets. This is mainly because multi-scale feature addition cannot fully retain the key information between different scales, thus weakening the ability of multi-scale fused features to express change information. Meanwhile, although the feature concatenation method can retain more change information and fully consider the correlation between features, simple feature concatenation ignores the information interaction between multi-scale change features, resulting in weaker distinguishability of change information in the fused features. Therefore, the evaluation index decreased on all three datasets. The AIFM proposed in this invention utilizes the attention weights of features at different scales to achieve multi-scale change feature interaction and fuses the difference information between different scales, fully exploring the complementary information between multi-scale change features, thereby improving the accuracy of change detection.

[0103] 3. Comparative analysis of experimental content and results

[0104] To verify and analyze the superior performance of this invention, this section compares it with eight representative SAR change detection methods, including DCNet, DDNet, LANTNet, CAMNet, WBANet, DBFNet, DSNet, and MDSF. Experimental results for each method on the Farmland-A dataset are shown in Table 2. The method of this invention achieves high evaluation results in PCC, Kappa coefficient, and F1 score, with improvements of at least 0.42%, 1.53%, and 1.28% in PCC, Kappa coefficient, and F1 score, respectively, compared to the comparison methods. Experimental results demonstrate that the proposed method can effectively improve SAR image change detection performance, increase the accuracy of change region detection, and exhibit good noise resistance.

[0105] Table 2. Indicators of the Farmland-A dataset in the comparative experiment.

[0106]

[0107] To verify the robustness and generalization ability of this invention, we conducted comparative experiments on the Farmland-B dataset. The experimental results for each method are shown in Table 3. Compared with other comparative algorithms, the proposed method achieves the best detection performance, reaching the highest detection accuracy in PCC, Kappa coefficient, and F1 score, at 98.97%, 90.06%, and 90.61%, respectively. In particular, compared to the suboptimal DDNet, our method shows significant advantages in PCC, Kappa, and F1. These experimental results demonstrate that this invention effectively enhances the expressive power of change features through dual-temporal feature interaction enhancement and frequency-domain guided dual-branch extraction of effective change features, thus verifying the effectiveness of the proposed method in SAR image change detection.

[0108] Table 3. Indicators of the Farmland-B dataset in the comparative experiment.

[0109]

[0110] To further verify the applicability of the method of this invention in different scenarios, comparative experiments were conducted on the Ottawa dataset, and the experimental results of each method are shown in Table 4. Compared with the comparative methods, the method of this invention achieved the best detection results on PCC, Kappa, and F1. Specifically, compared with the comparative methods, the method of this invention improved PCC, Kappa, and F1 by at least 0.16%, 0.54%, and 0.44%, respectively. The above experimental results show that this invention fully mines the change information between images of different time phases through dual-temporal feature interaction, and effectively integrates change features at different levels by combining multi-scale feature fusion, thereby improving the representation ability of change features and further improving the accuracy of SAR image change detection. This verifies that the method of this invention has good robustness and generalization ability.

[0111] Table 4. Indicators of the Ottawa dataset in the comparative experiment

[0112]

[0113] In summary, this invention proposes a SAR image change detection method based on wavelet-guided and adaptive fusion. This method constructs an adaptive feature enhancement module that focuses on changed regions through dual-temporal feature interaction and employs self-calibrated enhanced attention to improve perception capabilities, reduce background interference, and enhance features. Simultaneously, a wavelet attention difference interaction module is constructed, using wavelet frequency domain decomposition to build frequency domain similarity, guiding the collaborative learning of dual-branch change features to extract effective change features. Furthermore, an adaptive interaction fusion module introduces SimAM attention to achieve interaction between multi-scale change features and fuses the difference information between different scales, fully exploiting the complementary information between multi-scale change features, thereby significantly improving the overall performance of change detection. Comparative experiments on three real SAR datasets demonstrate that the proposed method outperforms other advanced algorithms. Moreover, ablation experiments further validate the effectiveness of the proposed method.

[0114] The above simulation analysis proves the correctness and effectiveness of the method proposed in this invention.

[0115] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0116] The parts of this invention not described in detail are common knowledge to those skilled in the art. The above description is only a preferred embodiment of the present invention and is not intended to limit the invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the protection scope of the claims of the present invention.

Claims

1. A SAR image change detection method based on wavelet-guided and adaptive fusion, characterized in that, Includes the following steps: (1) Acquire pairs of dual-temporal SAR images of the same area at different times, perform unsupervised pre-classification on them and label them with corresponding change pseudo-labels. The classification includes change class, no change class and uncertain class; randomly select samples from the change class and no change class to construct a training set, and the uncertain class constitutes the test set; (2) Construct a multi-scale feature extraction network, using dual-temporal SAR images as input, to extract multi-scale features corresponding to the two temporal SAR images; (3) Construct an adaptive feature enhancement module to perform bidirectional interaction on the multi-scale features corresponding to the two images, complete the context information perception enhancement, and output the enhanced features; (4) Based on discrete wavelet transform and cosine similarity, a wavelet-guided difference interaction module is constructed to perform bi-branch difference interaction and collaborative optimization on the enhanced features, and after fusing bi-branch difference information, multi-scale change features are output. (5) Construct an adaptive interactive fusion module to adaptively fuse and enhance multi-scale change features to obtain the final change features; (6) A multi-scale feature extraction network, an adaptive feature enhancement module, a wavelet-guided difference interaction module, and an adaptive interaction fusion module are sequentially connected in series, and a classification network is connected after the adaptive interaction fusion module to construct a SAR change detection model; the classification network consists of a convolutional layer and two fully connected layers, which are used to output classification prediction results based on the final change features; (7) The training set data is fed into the SAR change detection model, and the feature extraction, feature enhancement, two-branch difference interaction and feature fusion operations are performed in sequence to obtain the final change features corresponding to the training samples. Based on the features, the classification prediction results are obtained, the prediction results are matched with the sample pseudo-labels, the classification loss is calculated and the network parameters are updated through backpropagation to complete the model training. (8) Use the trained SAR change detection model to detect the test set samples and output the detection results; then concatenate the results with the pseudo-label of the changed and non-changed classes to obtain the final change detection results.

2. The method according to claim 1, characterized in that: The adaptive feature enhancement module described in step (3) includes the following steps in its processing: (3.1) Add relative position encoding to the multi-scale features of the two images, and realize cross-temporal information interaction through mutual guidance of their features to generate dual-temporal interactive features; (3.2) Construct a self-calibrating enhanced attention unit, input the dual-temporal interaction features into the self-calibrating enhanced attention unit, the self-calibrating enhanced attention unit is used to mine the internal structural information and context information of the features, perform adaptive enhancement processing on the dual-temporal interaction features, and output the context attention weight matrix; (3.3) Multiply the context attention weight matrix element by element with the dual temporal interaction features to complete the context information perception enhancement and obtain the enhanced features.

3. The method according to claim 2, characterized in that: The cross-temporal information interaction described in step (3.1) specifically involves using the multi-scale feature information of the first SAR image to guide the multi-scale features of the second SAR image, and simultaneously using the multi-scale features of the second SAR image to guide the multi-scale features of the first SAR image in reverse, thus completing the bidirectional information interaction between the two images' dual-temporal features. Through bidirectional interaction, the spatiotemporal correlation between the dual-temporal features is mined, and the features of the changing regions that differ between the two images are adaptively amplified, while the redundant interference of the unchanged background regions is suppressed.

4. The method according to claim 3, characterized in that: The context attention weight matrix mentioned in step (3.2) is obtained in the following way: the dual-temporal interaction features are input into the self-calibrated enhanced attention unit, global context information is extracted first through pooling operation, and then the context features are generated by convolution and upsampling in sequence; at the same time, global average pooling is performed on the dual-temporal interaction features along the horizontal and vertical dimensions respectively, and the structural attention weights are output. The context features and structural attention weights are fused to complete the feature self-calibration enhancement process, and finally the context attention weight matrix is ​​generated.

5. The method according to claim 1, characterized in that: The multi-scale change characteristics described in step (4) are obtained according to the following steps; (4.1) Using logarithmic ratio operation and channel splicing, two independent dual-branch differential features are constructed based on the enhanced features described in step (3); (4.2) Construct a wavelet cosine attention unit, take the enhanced features described in step (3) as input, complete the frequency domain decomposition through wavelet transform, and obtain the frequency domain attention weight by combining cosine similarity calculation and normalization processing. Multiply the weight with the two-branch difference features to output two-way enhanced difference features. (4.3) Using the two-way enhanced difference features as input to the wavelet cosine attention unit, interactive attention weights are generated to achieve dual-branch collaborative optimization, and two-way interactive enhanced difference features are obtained. (4.4) Perform addition and fusion on the two interactive enhancement difference features to output complete multi-scale change features.

6. The method according to claim 5, characterized in that: The two enhanced difference features mentioned in step (4.2) are obtained through the following processing steps: a discrete wavelet transform is performed on the input enhanced features to decompose the frequency sub-bands into low-frequency and high-frequency sub-bands; then, adaptive weighting is performed on each frequency sub-band through wavelet feature selection; the similarity of each weighted sub-band is calculated through cosine similarity, and the frequency domain sub-band similarity is mapped to the spatial domain using inverse discrete wavelet transform, and then normalized to frequency domain attention weights through the Sigmoid function; the two independent bi-branch difference features are multiplied by the frequency domain attention weights respectively to complete the bi-branch difference feature enhancement.

7. The method according to claim 5, characterized in that: The bi-branch collaborative optimization described in step (4.3) specifically includes: inputting the two enhanced difference features into the wavelet cosine attention unit respectively, calculating the frequency domain similarity between the two sets of enhanced difference features and generating interactive attention weights; using the weights to perform bi-directional constraint optimization on the two enhanced difference features respectively, realizing information complementarity and semantic alignment of the two features, and obtaining two interactive enhanced difference features.

8. The method according to claim 1, characterized in that: The adaptive interactive fusion module described in step (5) adaptively fuses and enhances multi-scale variation features according to the following steps: (5.1) The interaction relationship between features with different scales is constructed by using the SimAM attention mechanism to obtain multi-scale interactive enhancement features; (5.2) Calculate the difference information of the change features at different scales, and then splice and fuse them with the multi-scale interactive enhancement features to obtain the fused features; (5.3) The fused features are adaptively enhanced using the SimAM attention mechanism to obtain the final changed features.

9. The method according to claim 8, characterized in that: Step (5.1) constructs the interaction relationship between features of different scales through SimAM attention. This is achieved by applying the attention weight of one feature of different scales to another feature of different scales, thereby realizing the information interaction between features of different scales.

10. The method according to claim 8, characterized in that: The calculation of difference information for different scale change features in step (5.2) refers to obtaining the pixel difference between multiple sets of different scale change features to characterize the difference in ground feature deformation; the calculated scale difference information is spliced ​​and fused with the multi-scale interactive enhancement feature along the channel dimension to retain the independent deformation details of each scale and cross-scale correlation features, weaken the noise interference of a single scale, and improve the completeness of the identification of ground feature change areas by the subsequent final change features.

Images