Remote sensing change detection method and device based on frequency domain representation decoupling network and medium
Patent Information
- Application Number
- CN202610833333.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-10
- Publication Date
- 2026-09-01
AI Technical Summary
[0004]为解决遥感变化检测中变化区域与未变化区域之间普遍存在类别不平衡的问题，本发明提出了基于频域表征解耦网络的遥感变化检测方法、设备及介质
[0035]1. This invention constructs a decoupled representation learning framework, achieving effective separation of changing features and background information, enhancing the network's ability to perceive real-world changing regions, thereby improving change detection accuracy and robustness. The scientific basis for this is that traditional methods learn directly in a mixed feature space, making them susceptible to interference from a large amount of invariant background, while decoupled learning can reduce feature redundancy and improve the signal-to-noise ratio of changing information.
Smart Images
Figure CN122676342A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to remote sensing change detection methods, equipment, and media, specifically to remote sensing change detection methods, equipment, and media based on frequency domain characterization decoupling networks, belonging to the field of remote sensing image detection technology. Background Technology
[0002] Remote sensing change detection (RSCD) refers to the identification of temporal changes in ground objects using remote sensing images acquired at different times, combined with image processing and analysis methods. RSCD has broad application potential in many fields and plays an important role in urban expansion monitoring, agricultural management, and disaster assessment.
[0003] In recent years, many deep learning-based methods have emerged. Deep learning methods can automatically learn features from images without manual feature extraction and possess powerful nonlinear modeling capabilities, making them better suited for change detection tasks. However, remote sensing change detection datasets typically contain a large number of unchanged areas and relatively few changed areas, and lighting changes often introduce class imbalance and spurious change problems. Furthermore, remote sensing images have rich and complex scene structures, with significant scale and semantic differences between different ground features, accompanied by factors such as lighting variations, shadow interference, and background noise. Existing techniques rely solely on raw features for change modeling, making it difficult to effectively distinguish between real and spurious changes, and also hindering the full exploitation of deep semantic relationships between two-phase images, thus limiting further improvements in change detection performance. Summary of the Invention
[0004] To address the prevalent class imbalance between changed and unchanged regions in remote sensing change detection, this invention proposes a remote sensing change detection method, device, and medium based on a frequency domain representation decoupling network. By designing a frequency-aware decoupling module, frequency domain information is used to optimize the changed and background regions. To further achieve accurate segmentation of changing targets, an invariant representation decoupling module is designed. By decoupling invariant representations from dual-temporal features, the network can focus more intently on the changed regions, thereby improving the detection accuracy of changed regions.
[0005] To achieve the above objectives, this invention provides a remote sensing change detection method based on a frequency domain characterization decoupling network, comprising the following steps:
[0006] Step 1: Acquire dual-temporal remote sensing images and preprocess them to obtain dual-temporal preprocessed remote sensing images;
[0007] Step 2: Construct a remote sensing image change detection network based on the decoupling of frequency sensing and invariant representation, and input the dual-temporal preprocessed remote sensing image into the remote sensing image change detection network to obtain the change detection results;
[0008] Step 3: Optimize the network parameters using a loss function on the change detection results to obtain the optimal remote sensing image change detection network; obtain the optimal change detection results based on the optimal remote sensing image change detection network.
[0009] Further, step 2 includes:
[0010] The dual-temporal preprocessed remote sensing images are input into a weight-shared twin encoder for processing to obtain their respective multi-scale features;
[0011] Scale alignment, concatenation, and convolution operations are performed on the multi-scale features of each time phase to obtain the multi-scale fused features of the corresponding time phase.
[0012] The multi-scale fusion features corresponding to the time phase are input into the parallel frequency-aware decoupling module for processing to obtain the enhanced features of the corresponding time phase.
[0013] The enhanced features of the two time phases are processed by a parallel invariant representation decoupling module to obtain the change features of the corresponding time phases;
[0014] The feature obtained by stitching together the change features of the two time phases is input into the segmentation head for processing to obtain the change detection result.
[0015] Furthermore, the process by which the frequency-aware decoupling module processes multi-scale fusion features includes:
[0016] Gaussian convolution is used to process multi-scale fusion features to extract low-frequency and high-frequency features;
[0017] The low-frequency and high-frequency features are respectively input into the frequency sensing module for processing to obtain low-frequency enhancement features and high-frequency enhancement features;
[0018] The low-frequency enhancement features and the high-frequency enhancement features are added together to obtain the enhancement features.
[0019] Furthermore, the process of the frequency sensing module processing features includes:
[0020] The energy spectrum of the input features is calculated by Fourier transform to obtain the energy spectrum matrix; at the same time, the position correlation matrix is obtained by calculating the cosine similarity between the feature vectors of the input features.
[0021] The energy spectrum matrix and the position correlation matrix are processed using dot product attention to obtain the energy spectrum correlation weight map and the position correlation weight map, respectively.
[0022] The input features are weighted using a comprehensive weight map obtained by summing the energy spectrum correlation weight map and the location correlation weight map. The weighted features are then added to the input features to obtain the output features of the frequency sensing module.
[0023] Furthermore, the process of processing the enhanced features of the two-phase time using a parallel invariant representation decoupling module includes:
[0024] The enhanced features of the first time phase are used as the source features, and the enhanced features of the second time phase are used as the reference features. They are input together into the invariant representation decoupling module of the first time phase for processing to obtain the change features of the first time phase.
[0025] The enhanced features of the second phase are used as the source features, and the enhanced features of the first phase are used as the reference features. They are input together into the invariant representation decoupling module of the second phase for processing to obtain the change features of the second phase.
[0026] Furthermore, the process by which the invariant representation decoupling module processes the source features and reference features includes:
[0027] The source features and reference features are reshaped into query features and key features, respectively; the query features and key features are processed using scaled dot product attention to obtain a feature similarity map;
[0028] Based on the maximum similarity in the feature similarity map, establish a common saliency probability for each pixel;
[0029] The common saliency probabilities are reshaped into a similarity score map; and the similarity score map is used to perform element-wise multiplication on the reference features to obtain invariant features;
[0030] The difference between the source features and the invariant features is used to obtain the changed features.
[0031] Furthermore, in step 3, the loss function includes Dice coefficient loss and binary cross-entropy loss.
[0032] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.
[0033] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.
[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0035] 1. This invention constructs a decoupled representation learning framework, achieving effective separation of changing features and background information, enhancing the network's ability to perceive real-world changing regions, thereby improving change detection accuracy and robustness. The scientific basis for this is that traditional methods learn directly in a mixed feature space, making them susceptible to interference from a large amount of invariant background, while decoupled learning can reduce feature redundancy and improve the signal-to-noise ratio of changing information.
[0036] 2. This invention models high-frequency and low-frequency information separately through a frequency-aware decoupling module. The high-frequency component mainly includes fine-grained variation information such as edges and textures, while the low-frequency component includes scene structure and semantic context information. This design can enhance the details of local changes while preserving the stability of the global structure, thus effectively reducing boundary blurring and the problem of missed detection of small targets. This is derived from the frequency domain analysis theory, that is, different frequency components correspond to different levels of image information, with high frequencies being more sensitive to local changes and low frequencies more stably describing the overall structure.
[0037] 3. This invention explicitly separates the invariant common components in the dual-temporal features through the invariant representation decoupling module, suppresses background consistency information, and enhances the response of real change areas. This effectively reduces false detections of changes caused by illumination, shadows, seasons, and sensor differences. This is because change detection is essentially a difference learning problem. A large number of invariant regions will mask the real change signal, while invariant representation decoupling can highlight the residual change features and improve the model's ability to distinguish change areas. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the frequency domain image of the present invention;
[0039] Figure 2 This is a schematic diagram of the overall network structure of the present invention;
[0040] Figure 3 This is a schematic diagram of the frequency sensing module structure of the present invention;
[0041] Figure 4The following are schematic diagrams illustrating the visualization results of different methods of the present invention: (a) shows the case of false detection caused by the similarity between non-changing objects and changing objects on the LEVIR-CD dataset; (b) shows the case of false detection caused by the similarity between changing objects and the background on the LEVIR-CD dataset; (c) shows the case of false detection caused by complex background on the WHU-CD dataset; (d) shows the case of large-scale changes on the WHU-CD dataset; (e) shows the case of large-scale changes on the SYSU-CD dataset; (f) shows the case of irregular changes on the SYSU-CD dataset; (1) is the remote sensing image at time T1; (2) (1) is the remote sensing image at time T2, (2) is the label, (3) is the prediction result of FC-EF, (4) is the prediction result of FC-Siam-Diff, (5) is the prediction result of FC-Siam-Conc, (6) is the prediction result of BIT, (7) is the prediction result of DMINet, (8) is the prediction result of EADDer, (9) is the prediction result of SEIFNet, (10) is the prediction result of SChanger, (11) is the prediction result of CASP, and (13) is the prediction result of FRDNet. Detailed Implementation
[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0043] Example 1:
[0044] like Figure 1 As shown, low-frequency components carry spectral information about large-scale ground structures and homogeneous regions, representing most of the unchanged background areas in the image. In contrast, high-frequency components encode object boundaries, texture details, and local features, corresponding to the boundaries of regions with small-scale variations. Figure 1 The frequency energy distribution in the lower left corner further supports this view. These findings highlight the importance of frequency domain decomposition in feature optimization. By separating high-frequency and low-frequency information, contextual features can be modeled and optimized independently, effectively mitigating the negative impact of class imbalance.
[0045] This embodiment provides a remote sensing change detection method based on a frequency domain characterization decoupling network, as detailed below:
[0046] Step 1: Obtain dual-temporal remote sensing images, which are two remote sensing images of the same area at different times, and preprocess the dual-temporal remote sensing images to obtain dual-temporal preprocessed remote sensing images.
[0047] In this embodiment, the images are from three different change detection datasets: LEVIR-CD, WHU-CD, and SYSU-CD. The preprocessing process is as follows: During the training phase, data augmentation techniques were used on all experimental data to improve the robustness and generalization ability of the model. The techniques used included flipping, cropping, and Gaussian blurring.
[0048] The LEVIR-CD dataset is derived from: Chen H, Wu C, Du B, et al. Change detection in multisource VHR images via deep Siamese convolutional multiple-layers recurrent neural network[J]. IEEE Transactions on Geoscience and RemoteSensing, 2019, 58(4): 2848-2864.
[0049] WHU-CD data set originates from: Ji S, Wei S, Lu M. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set[J]. IEEE Transactions on geoscience and remote sensing, 2018, 57(1):574-586.
[0050] The SYSU-CD dataset is derived from: Shi J, Wu T, Qin AK, et al. Semisupervised adaptive ladder network for remote sensing image change detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-20.
[0051] Step 2: Construct a remote sensing image change detection network based on the decoupling of frequency sensing and invariant representation, and input the dual-temporal preprocessed remote sensing image into the remote sensing image change detection network to obtain the change detection results;
[0052] Furthermore, the structure of the remote sensing image change detection network is as follows: Figure 2As shown, specifically, dual-temporal preprocessed remote sensing images The process involves a weight-shared Siamese transform encoder, which generates multi-scale features. Subsequently, these features are uniformly upsampled to one-quarter of the original image resolution, and then a uniform channel dimension is generated through stitching and convolution operations. The dual-temporal features, namely, the multi-scale fusion features corresponding to the temporal phases. Next, we will take... For input, this invention designs a frequency-aware decoupling (FAD) module. This module first decouples the frequency features, then refines different frequency components to generate enhanced features corresponding to the time phase. Then, with and To provide input, this invention designs an Invariant Representation Decoupling (IRD) module, which captures and decouples the invariant features between the two temporal representations to obtain the change features of the corresponding temporal phases. Finally, the characteristics of change will be... and The images are stitched together, and the segmentation header is input to generate the desired change map, which is the final result of the model. .
[0053] Furthermore, the process of the frequency-aware decoupling module processing features includes:
[0054] like Figure 2 As shown in (a), the input to a given FAD is First, Gaussian convolution operations are used to extract features from the corresponding time phases. Extracting low-frequency components Then obtain the corresponding high-frequency components. The process is as follows:
[0055] ,
[0056] ,
[0057] in, This indicates that convolution is performed using a learnable Gaussian kernel.
[0058] Then, and The data are input into the Frequency-aware (FA) module. The FA module proposed in this invention enhances the ability to capture both global and fine-grained features by analyzing the energy spectrum and positional correlations in the frequency domain representation. Specifically, by analyzing the energy spectrum, FA emphasizes details and edges in the image, thereby strengthening the representation of detailed features. Simultaneously, by calculating cosine similarity, it captures positional correlations, helping the model better understand structural information.
[0059] like Figure 3 As shown, the spatial frequency components of a given input That is, the energy spectrum is calculated by Fourier transforming the low-frequency and high-frequency components to obtain the energy spectrum matrix. Simultaneously, by calculating the cosine similarity between feature vectors, the location correlation matrix is obtained. Subsequently, the energy spectrum-related weight map is obtained by calculating dot product attention. Location-related weights graph Finally, a comprehensive weighted graph is obtained by summing the results. Ultimately, through cooperation with The following formula generates the output of FA. :
[0060] .
[0061] in The process first utilizes the attention weight matrix. For input frequency domain features Weighted reconstruction is performed to enhance key frequency domain information by highlighting high-response regions and suppressing redundant regions. Subsequently, the enhanced features are residually fused with the original input features to preserve the original semantic representation while strengthening fine-grained texture and structural information, thereby improving the stability and robustness of feature representation. Finally, the outputs of the two FA modules are summed to obtain the output of the FAD module. The process is as follows:
[0062] .
[0063] Furthermore, the process by which the invariant representation decoupling module processes the output features of the frequency-aware decoupling module is as follows:
[0064] The enhanced features of the first time phase are used as the source features, and the enhanced features of the second time phase are used as the reference features. They are input together into the invariant representation decoupling module of the first time phase for processing to obtain the change features of the first time phase.
[0065] The enhanced features of the second phase are used as the source features, and the enhanced features of the first phase are used as the reference features. They are input together into the invariant representation decoupling module of the second phase for processing to obtain the change features of the second phase.
[0066] Furthermore, the specific process of the invariant representation decoupling module in processing features includes:
[0067] like Figure 2 (b) shows the source features and reference features Reshaped into query features Bond features Then, scaled dot product attention is applied to measure pixel-level similarity between them. Feature similarity map The calculation formula is as follows: ,
[0068] in, This represents the matrix transpose operation. It is the scaling factor.
[0069] Next, select... The value with the highest similarity is used to establish a common saliency probability for each pixel of the query feature. The calculation formula is: ,
[0070] in, , This represents the i-th pixel in each row of the feature similarity map.
[0071] Then, Remodeling This represents the invariant features in the reference features that are identical to the source features. Then, guided by the similarity score, these invariant features... It is decoupled from the reference feature. Finally, the proposed IRD output... as follows:
[0072] ,
[0073] ,
[0074] Here, ⊙ represents element-wise multiplication.
[0075] This embodiment proposes an FAD module, which helps distinguish between real and pseudo-changes by mining the intrinsic attributes of image features and learning location information encoding. Furthermore, an IRD module is designed to effectively learn changed features by decoupling the unchanged features of two images.
[0076] Step 3: Optimize the network parameters using a loss function on the change detection results to obtain the optimal remote sensing image change detection network; obtain the optimal change detection results based on the optimal remote sensing image change detection network.
[0077] Furthermore, the loss function is composed of the Dice coefficient loss. and binary cross-entropy loss Combining these elements, the loss function is expressed as follows:
[0078] ,
[0079] ,
[0080] ,
[0081] in, and These represent the ground truth labels and the corresponding predicted change maps, respectively. The pixel value in the image is either 0 or 1, while The pixel values range from 0 to 1, representing the probability of being predicted as a changing region.
[0082] This embodiment also provides an experimental verification process, as detailed below:
[0083] Experimental setup: The network was trained on two NVIDIA RTX 2080Ti GPUs using the PyTorch framework. The AdamW optimizer was used with the following settings: batch size of 16, initial learning rate of 0.0001, momentum of 0.99, weight decay of 0.01, and a total of 200 training epochs.
[0084] Experimental Data: This invention was tested on three widely used change detection datasets: LEVIR-CD, WHU-CD, and SYSU-CD. LEVIR-CD contains 637 pairs of 0.5-meter resolution images from urban areas of Texas, focusing on diverse building-related changes under different seasons and lighting conditions. WHU-CD contains high-resolution aerial image pairs (0.2 meters) from New Zealand for building change detection. SYSU-CD contains approximately 20,000 pairs of ultra-high-resolution images from Hong Kong, China (2007–2014), covering complex land cover changes such as buildings, roads, ports, and vegetation. All datasets were divided into non-overlapping 256×256 pixel image patches and split into training, validation, and test sets.
[0085] Experimental metrics: To evaluate the model's performance, five commonly used evaluation metrics were used: Precision (Pre), Recall (Rec), F1 score (F1), Intersection over union (IoU), and Overall accuracy (OA).
[0086] Comparison of algorithms: FC-EF, FC-SIAM-DIFF, and FC-SIAM-CONC algorithms are derived from: Daudt RC, LeSaux B, Boulch A. Fully convolutional siamese networks for change detection[C] / / 2018 25th IEEE international conference on image processing (ICIP).IEEE, 2018: 4063-4067; BIT algorithm is derived from: Chen H, Qi Z, Shi Z. Remote sensing imagechange detection with transformers[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1-14; DMINET algorithm is derived from: Feng Y, Jiang J, Xu H, et al. Change detection on remote sensing images using dual-branch multilevel intertemporal network[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 1-15; EATDER algorithm is derived from: Ma J, Duan J, Tang X, et al. EATDer: Edge-assisted adaptive transformer detector for remote sensing change detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 62: 1-15; SEIFNET algorithm originated from: Huang Y, Li X, Du Z, et al. Spatiotemporal enhancement and interlevel fusion network for remote sensing images change detection[J].IEEETransactions on Geoscience and Remote Sensing, 2024, 62: 1-14; SChanger algorithm comes from: Zhou Z, Hu K, Fang Y, et al. SChanger: Change detection from a semanticchange and spatial consistency perspective[J]. IEEE Journal of SelectedTopics in Applied Earth Observations and Remote Sensing, 2025; CASP algorithm comes from: Wang Q, Zhang M, Ren J, et al. Exploring context alignment and structureperception for building change detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025. .
[0087] Table 1-3 presents the quantitative results of various methods on different datasets, with the best results shown in bold.
[0088] Table 1. Performance comparison results on the LEVIR-CD dataset
[0089]
[0090] Table 2. Performance comparison results on the WHU-CD dataset
[0091]
[0092] Table 3 Performance comparison results on the SYSU-CD dataset
[0093]
[0094] Table 1 presents a performance comparison of the present invention (FRDNet) with other change detection networks on the LEVIR-CD dataset. The results in the table show that FRDNet outperforms other networks across all evaluation metrics, particularly in overall accuracy, IoU, F1 score, and recall. Specifically, FRDNet improves IoU and F1 score by 1.33 and 0.79 percentage points compared to CASP, respectively, further demonstrating its advantage in change detection tasks. In contrast, CASP also performs admirably, only slightly inferior to FRDNet in IoU and F1 score, but its recall and precision are relatively lower. FRDNet demonstrates a superior balance between precision and recall, ensuring high accuracy and completeness of detection results. Overall, FRDNet significantly outperforms other change detection networks on the LEVIR-CD dataset, especially in key metrics such as intersection-over-union ratio (IoU) and F1 score, validating its effectiveness and superiority as a change detection model.
[0095] Table 2 shows a performance comparison of the present invention (FRDNet) with other change detection networks on the WHU-CD dataset. The results show that FRDNet performs excellently across all evaluation metrics, particularly outperforming other networks in overall accuracy, IoU, F1 score, and recall. In contrast, traditional fully convolutional Siamese network methods such as FC-EF, FC-Siam-Diff, and FC-Siam-Conc perform poorly on the WHU-CD dataset, especially in IoU and F1 score, demonstrating their limitations in handling complex changes. Among attention-enhanced and Transformer-type methods, BIT, DMINet, EADTe, SEIFNet, and SChanger all achieved considerable performance, with BIT achieving an F1 score of 89.74% and a CASP of 94.02%. However, FRDNet still outperforms all key metrics, especially in IoU and F1 score, which are 0.93 and 0.52 percentage points higher than CASP, respectively, validating its superiority in practical change detection tasks. It is worth noting that while FRDNet's precision is slightly lower than CASP's, its recall is 2.2 percentage points higher, indicating that it adopts a more aggressive identification strategy, reducing false alarms while maintaining a low false alarm rate. Overall, FRDNet significantly outperforms other change detection networks on the WHU-CD dataset, fully demonstrating its robustness and generalization ability across different change scales and complex terrain scenarios, showcasing its broad application potential in remote sensing image change detection tasks.
[0096] To further verify the generalization ability of our invention (FRDNet) in complex scenes, we conducted comparative experiments on the more challenging SYSU-CD dataset. The SYSU-CD dataset contains various lighting variations, seasonal differences, and viewpoint variations, making its scene complexity significantly higher than LEVIR-CD. Table 3 summarizes the performance comparison of each method. Overall, all methods perform worse on SYSU-CD than on the LEVIR-CD and WHU-CD datasets, reflecting the challenging nature of this dataset. Notably, SEIFNet achieves the highest recall rate of 80.59%, the highest among the compared methods, indicating high sensitivity to changing regions, but its precision is relatively low, exhibiting some false positives. In contrast, our proposed FRDNet achieves the best performance across all four metrics. Specifically, FRDNet achieves an OA of 92.28%, a 0.50% improvement over the second-best DMINet; an IoU of 70.22%, a 0.90% improvement over SEIFNet; and an F1 score of 82.50%, a 0.62% improvement over SEIFNet. Of particular note is that FRDNet achieved a precision of 88.57%, significantly outperforming all comparable methods and exceeding the second-highest precision, FC-Siam-Diff, by 2.87 percentage points. This result demonstrates that FRDNet possesses exceptional spurious change suppression capabilities in the complex SYSU-CD scenario, effectively distinguishing between genuine surface changes and interference caused by factors such as illumination and season. While FRDNet's recall of 77.21% is slightly lower than SEIFNet and SChanger, its significantly higher precision results in the best overall F1 score, indicating that FRDNet achieves a better balance between precision and recall. These results fully validate the effectiveness of the proposed frequency domain decoupling strategy in addressing spurious changes caused by illumination and environmental variations.
[0097] Overall, this invention demonstrates superior detection capabilities on all three benchmark datasets, especially in the change detection task, where it exhibits a good balance between accuracy and recall.
[0098] Figure 4Images (c), (e), and (f) in the diagram demonstrate the detection results of changing targets under conditions of large-scale targets and complex background interference, while images (a), (b), and (d) show the detection performance in scenarios where changing targets and backgrounds have similar semantics and are difficult to distinguish. It can be observed that the proposed method achieves the best segmentation results in all the above scenarios, while exhibiting the fewest false positives and false negatives. This indicates that the FAD module can effectively suppress interference from background texture and gradient information through a frequency-aware mechanism, while the IRD module enhances the semantic discriminative ability of changing regions through invariant representation learning. The synergistic effect of both modules enables the model to effectively suppress interference from unchanged backgrounds and accurately extract truly changing regions, thus achieving superior overall change detection performance.
[0099] Example 2:
[0100] This invention provides an electronic device, including a processor and a memory. The memory stores program instructions for implementing the remote sensing change detection method based on a frequency domain characterization decoupling network as described in Embodiment 1. The processor executes the program instructions stored in the memory to perform remote sensing change detection.
[0101] The processor can also be called a Central Processing Unit (CPU). A processor may be an integrated circuit chip with signal processing capabilities. A processor can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.
[0102] The methods described in Embodiment 1 of this invention are applicable to this electronic device embodiment. The specific functions implemented by this electronic device embodiment are the same as those in the above method embodiment, and the beneficial effects achieved are also the same as those achieved by the above methods.
[0103] Example 3:
[0104] This embodiment provides a storage medium storing a computer program or instructions capable of implementing the above-described remote sensing change detection method based on a frequency domain characterization decoupled network. When the computer program or instructions are run on a computer, the steps of the above-described remote sensing change detection method based on a frequency domain characterization decoupled network are executed.
[0105] The computer program or instructions can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.
[0106] The methods described in Embodiment 1 of this invention are all applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.
[0107] This invention provides a remote sensing change detection method, device, and medium based on a frequency domain characterization decoupling network. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A remote sensing change detection method based on a frequency domain representation decoupling network, characterized in that, Includes the following steps: Step 1: Acquire dual-temporal remote sensing images and preprocess them to obtain dual-temporal preprocessed remote sensing images; Step 2: Construct a remote sensing image change detection network based on the decoupling of frequency perception and invariant representation, and input the dual-temporal preprocessed remote sensing image into the remote sensing image change detection network to obtain the change detection results; Step 3: Optimize the network parameters using a loss function on the change detection results to obtain the optimal remote sensing image change detection network; obtain the optimal change detection results based on the optimal remote sensing image change detection network.
2. The remote sensing change detection method based on a frequency domain characterization decoupling network according to claim 1, characterized in that, Step 2 includes: The dual-temporal preprocessed remote sensing images are input into a weight-shared twin encoder for processing to obtain their respective multi-scale features; Scale alignment, concatenation, and convolution operations are performed on the multi-scale features of each time phase to obtain the multi-scale fused features of the corresponding time phase. The multi-scale fusion features corresponding to the time phase are input into the parallel frequency-aware decoupling module for processing to obtain the enhanced features of the corresponding time phase. The enhanced features of the two time phases are processed by a parallel invariant representation decoupling module to obtain the change features of the corresponding time phases; The feature obtained by stitching together the change features of the two time phases is input into the segmentation head for processing to obtain the change detection result.
3. The remote sensing change detection method based on a frequency domain characterization decoupling network according to claim 2, characterized in that, The process by which the frequency-aware decoupling module processes multi-scale fused features includes: Gaussian convolution is used to process multi-scale fusion features to extract low-frequency and high-frequency features; The low-frequency and high-frequency features are respectively input into the frequency sensing module for processing to obtain low-frequency enhancement features and high-frequency enhancement features; The low-frequency enhancement features and the high-frequency enhancement features are added together to obtain the enhancement features.
4. The remote sensing change detection method based on a frequency domain characterization decoupling network according to claim 3, characterized in that, The process by which the frequency sensing module processes features includes: The energy spectrum of the input features is calculated by Fourier transform to obtain the energy spectrum matrix; at the same time, the position correlation matrix is obtained by calculating the cosine similarity between the feature vectors of the input features. The energy spectrum matrix and the position correlation matrix are processed using dot product attention to obtain the energy spectrum correlation weight map and the position correlation weight map, respectively. The input features are weighted using a comprehensive weight map obtained by summing the energy spectrum correlation weight map and the location correlation weight map. The weighted features are then added to the input features to obtain the output features of the frequency sensing module.
5. The remote sensing change detection method based on a frequency domain characterization decoupling network according to claim 2, characterized in that, The process of processing the enhanced features of the two-phase time using a parallel invariant representation decoupling module includes: The enhanced features of the first time phase are used as the source features, and the enhanced features of the second time phase are used as the reference features. They are input together into the invariant representation decoupling module of the first time phase for processing to obtain the change features of the first time phase. The enhanced features of the second phase are used as the source features, and the enhanced features of the first phase are used as the reference features. They are input together into the invariant representation decoupling module of the second phase for processing to obtain the change features of the second phase.
6. The remote sensing change detection method based on a frequency domain characterization decoupling network according to claim 5, characterized in that, The process by which the invariant representation decoupling module processes source features and reference features includes: The source features and reference features are reshaped into query features and key features, respectively; the query features and key features are processed using scaled dot product attention to obtain a feature similarity map; Based on the maximum similarity in the feature similarity map, establish a common saliency probability for each pixel; The common saliency probabilities are reshaped into a similarity score map; and the similarity score map is used to perform element-wise multiplication on the reference features to obtain invariant features; The difference between the source features and the invariant features is used to obtain the changed features.
7. The remote sensing change detection method based on a frequency domain characterization decoupling network according to claim 1, characterized in that, In step 3, the loss function includes Dice coefficient loss and binary cross-entropy loss.
8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the remote sensing change detection method based on a frequency domain characterization decoupling network as described in any one of claims 1 to 7.
9. A storage medium, characterized in that, The system stores a computer program or instructions that, when executed on a computer, perform the steps of the remote sensing change detection method based on a frequency domain characterization decoupled network as described in any one of claims 1 to 7.