Frequency modulation based remote sensing change detection method and system

Abstract

The present application relates to the technical field of remote sensing image change detection, and proposes a remote sensing change detection method and system based on frequency modulation. Through low-frequency anchoring, difference modulation and selective high-frequency reconstruction, the participation process of different frequency information in change detection is explicitly controlled, and the dependence of the model on high-frequency false differences caused by illumination, phenology, registration error, etc. is reduced. The semantic accuracy and boundary positioning accuracy of the change detection result can be improved, false positives, missed detections can be reduced, and the generalization ability of the model under cross-region and cross-acquisition conditions can be enhanced.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection technology, specifically to a remote sensing change detection method and system based on frequency modulation. Background Technology

[0002] Change detection technology based on dual-temporal remote sensing imagery is mainly used to identify changes in ground features in the same area acquired at different times, and is widely used in scenarios such as land use monitoring, urban expansion analysis, and disaster assessment. However, in practical applications, remote sensing images are affected by factors such as changes in illumination, seasonal phenological differences, changes in imaging angle, registration errors, and differences in sensor response, which can produce a large number of apparent differences unrelated to actual ground feature changes. This leads to problems such as false alarms, missed detections, and inaccurate boundary localization in deep learning-based change detection models.

[0003] Existing technologies typically improve change detection performance by enhancing spatial feature representation, introducing attention mechanisms, or fusing multi-scale features. For example, SNUNet3+ (Siamese Nested U-Net 3+) fuses spatial features from different levels through full-scale skip connections, while DDCNN (Dense Difference Convolutional Neural Network) utilizes dense attention mechanisms to enhance the ability to distinguish changed regions. These methods primarily focus on feature extraction and fusion processes, without fully considering the learning order and impact of different frequency information during training. In dual-temporal remote sensing images, non-realistic change factors typically manifest as high-frequency information such as brightness differences, texture changes, and edge misalignments. Models tend to learn these apparent differences to form shortcut judgments rather than relying on stable semantic features of ground features. When applied to images from different regions or under different acquisition conditions, high-frequency spurious correlations fail, leading to a decrease in model generalization ability and further causing false positives, false negatives, and inaccurate boundary localization. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a frequency modulation-based remote sensing change detection method and system. By employing low-frequency anchoring, differential modulation, and selective high-frequency reconstruction, it explicitly controls the participation of different frequency information in change detection, reducing the model's dependence on high-frequency spurious differences caused by illumination, phenology, and registration errors. This improves the semantic accuracy and boundary positioning precision of change detection results, reduces false alarms and missed detections, and enhances the model's generalization ability across regions and acquisition conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: One or more embodiments provide a remote sensing change detection method based on frequency modulation, comprising the following steps: Acquire dual-temporal remote sensing image pairs of the target area; Spatial and frequency features are extracted from each phase image of the dual-temporal remote sensing image pair. Attention weights are generated based on the spatial and frequency features to weight the spatial features, resulting in a harmonic feature map under low-frequency context constraints. The common and difference components of the harmonic feature maps of the two-phase image pairs are separated, and the harmonic feature maps of the two-phase image pairs are adaptively fused according to the difference components to obtain adaptive dual-temporal fusion features. The obtained adaptive dual-temporal fusion features are selectively reconstructed and high-frequency details are restored through wavelet reconstruction to obtain a high-resolution thinned feature map. Based on the thinned feature map, classification is performed to obtain the change detection results of the dual-temporal remote sensing image.

[0006] Further technical solutions include constructing a frequency modulation network, which adopts an encoder and decoder structure, including a spatial-low frequency coordination module for encoding, an adaptive dual-time fusion module, and a selective high-frequency refiner for decoding.

[0007] A further technical solution is to use a twin dual-branch encoder in the space-low frequency coordination module, including parallel spatial and frequency branches, as well as a space-low frequency harmonic module; Spatial branches can be extracted using CNN backbone networks to extract fine-grained spatial details and local patterns as spatial features; The frequency branch is configured to extract frequency features using a wavelet transform-based recursive decomposition strategy. The spatial-low-frequency harmonic module is configured to fuse features from the two branches, anchoring spatial features to low-frequency features and low-frequency context features provided by the frequency branch, and generating a harmonic feature map. Alternatively, the adaptive dual-time fusion module includes a context-differential modulator and a difference-guided time fusion module; A contextual difference modulator is used to separate stable and differential components from the harmonic feature maps of two time phases through a self-attention mechanism and generate attention-aware differential features. The difference-guided temporal fusion module is used to perform weighted fusion of dual temporal harmonic feature maps based on attention-perceived difference features, generating adaptive dual temporal fusion features.

[0008] A further technical solution involves extracting frequency features from each phase image of a dual-temporal remote sensing image pair and employing a recursive decomposition strategy based on wavelet transform, including the following steps: Step 211: Perform initial feature extraction on the current phase image and use the extracted features as the initial input for the frequency branch; Step 212: Decompose the current input features into one low-frequency subband LL and three high-frequency subbands LH, HL, and HH; Step 213: Using the low-frequency subband LL as the current input feature, execute step 212 to perform recursive decomposition until the set recursion cutoff condition is met. The high-frequency subbands obtained from each level of recursive decomposition are used to form a high-frequency subband set as a frequency feature.

[0009] A further technical solution involves separating the common and difference components of the harmonic feature maps of the two-phase image pairs based on the harmonic feature maps of each phase image, including the following steps: The harmonic feature maps of the two phases are concatenated, and the features are refined through the channel attention module and the spatial attention module to obtain the value tensor; The query tensor Q is obtained by summing the harmonic feature maps of the two time phases. The bond tensor K is obtained by subtracting the harmonic feature maps of the two time phases. Context-aware attention is calculated based on query tensor Q and key tensor K, and this attention is used to filter value tensor V to obtain attention-aware differential features S.

[0010] A further technical solution involves adaptively fusing the harmonic feature maps of the two-temporal image pairs based on the differences in components to obtain adaptive dual-temporal fusion features, including the following steps: The harmonic feature maps of the two time phases are used to construct a feature pyramid P. The feature pyramid P is then subjected to two residual processing steps to obtain the intermediate representation B. The intermediate representation B is concatenated with the attention-perceived differential feature S to obtain the comprehensive feature map C: Global average pooling is performed on the comprehensive feature map C, and convolution and activation are performed using a dynamically adjusted kernel size k to generate channel attention vectors. ; Using channel attention vectors The comprehensive feature map C is reweighted to obtain a weighted feature map, which is used as the adaptive dual temporal fusion feature. A further technical solution involves selectively reconstructing and restoring high-frequency details from the obtained adaptive dual-time-series fusion features using wavelet reconstruction to obtain a high-resolution, refined feature map, including the following steps: Obtain adaptive dual-temporal fusion features at different scales; coarse-scale feature maps Two-dimensional discrete wavelet transform is performed, and the extracted high-frequency subbands are spliced ​​together to obtain high-frequency prior information; Based on high-frequency prior information A multi-path sampling reconstruction path is constructed, and the fusion features at different scales are upsampled, high-frequency compensated and feature reconstructed based on wavelet reconstruction to obtain the corresponding reconstructed feature map.

[0011] One or more embodiments provide a frequency modulation-based remote sensing change detection system, including: The data acquisition module is configured to acquire two-temporal remote sensing image pairs of the target area; The spatial-low-frequency coordination module is configured to extract spatial and frequency features from each temporal image of a dual-temporal remote sensing image pair, and generate attention weights based on the spatial and frequency features to weight the spatial features, thereby obtaining a harmonic feature map under low-frequency context constraints. The adaptive dual-temporal fusion module is configured to separate the common and difference components of the harmonic feature maps of the dual-temporal image pairs, and adaptively fuse the harmonic feature maps of the dual-temporal image pairs according to the difference components to obtain adaptive dual-temporal fusion features. A selective high-frequency thinner is configured to selectively reconstruct and restore high-frequency details of the obtained adaptive dual-temporal fusion features through wavelet reconstruction, resulting in a high-resolution thinned feature map. Classification is then performed based on the thinned feature map to obtain the change detection results of the dual-temporal remote sensing image.

[0012] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps in the frequency modulation-based remote sensing change detection method described above.

[0013] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the frequency modulation-based remote sensing change detection method described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a phased processing approach involving low-frequency anchoring, differential modulation, and selective high-frequency reconstruction. First, it utilizes low-frequency features to constrain spatial features, enabling the model to prioritize learning stable land cover semantics and spatial structures, reducing reliance on high-frequency apparent differences caused by illumination, season, imaging angle, and registration errors. Next, by separating common and differential components in dual-temporal features, it highlights information related to genuine changes and suppresses spurious differences. Finally, under low-frequency context constraints, it selectively recovers high-frequency details, making the boundaries and contours of changed areas clearer. Therefore, this invention avoids shortcut judgments that rely on high-frequency spurious correlations in model formation, reducing false alarms and missed detections, and improving the semantic accuracy, boundary localization precision, and generalization ability across regions and acquisition conditions of change detection results.

[0015] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a flowchart of the remote sensing change detection method based on frequency modulation according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the frequency modulation network in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the encoding process of the frequency modulation network in Embodiment 1 of the present invention; Figure 4 This is a structural diagram of the wavelet reconstruction unit in Embodiment 1 of the present invention; Figure 5 This is a comparison chart of prediction results of different models in the experimental example of Embodiment 1 of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0021] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 5 As shown, a remote sensing change detection method based on frequency modulation includes the following steps: Step 1: Obtain a pair of two-temporal remote sensing images of the target area. The pair of two-temporal remote sensing images includes a first-temporal image and a second-temporal image. Step 2: Extract spatial and frequency features from each phase image of the dual-temporal remote sensing image pair. Generate attention weights based on the spatial and frequency features to weight the spatial features and obtain a harmonic feature map under low-frequency context constraints. Step 3: Separate the common and difference components of the harmonic feature maps of the two-temporal image pairs, and adaptively fuse the harmonic feature maps of the two-temporal image pairs according to the difference components to obtain adaptive dual-temporal fusion features; Step 4: Selectively reconstruct and restore high-frequency details of the obtained adaptive dual-temporal fusion features through wavelet reconstruction to obtain a high-resolution thinned feature map. Classify the features based on the thinned feature map to obtain the change detection results of the dual-temporal remote sensing image.

[0022] This embodiment acquires dual-temporal remote sensing images and extracts spatial and frequency features separately. Low-frequency features are introduced at the initial stage of change detection to anchor and constrain spatial features. This enables the model to learn stable semantic and spatial structure representations of ground features that are less affected by illumination, seasonal phenology, imaging angle, registration errors, and sensor differences. This avoids the model prematurely relying on high-frequency apparent information such as brightness differences, texture abrupt changes, and edge misalignments to form shortcut judgments. Furthermore, by separating common and differential components in the dual-temporal harmonic feature maps and using the differential components to adaptively fuse the dual-temporal features, the model can highlight effective differences related to real ground feature changes while maintaining stable common semantic information, suppressing false high-frequency responses caused by non-real change factors. Further, wavelet reconstruction is used to selectively reconstruct features and restore high-frequency details in the adaptive dual-temporal fusion features. This ensures that high-frequency information is selectively enhanced only under the common constraints of the low-frequency context and differential modulation results, thus preserving the boundaries, contours, and details of changed areas while avoiding the erroneous recovery of noisy textures and false edges. Therefore, this invention can explicitly control the participation order and function of different frequency information in the change detection process, solving the problem that existing methods only focus on spatial feature extraction and fusion, and do not fully consider the learning order of frequency information and high-frequency false correlations. It reduces false alarms, missed detections and boundary positioning deviations caused by the gap between the two temporal domains, and improves the semantic reliability, boundary clarity and generalization ability of change detection results across regions and acquisition conditions.

[0023] Steps 1 to 4 above are achieved by constructing a frequency modulation network. The frequency modulation network adopts an encoder and decoder structure, including a spatial-low frequency coordination module for encoding, an adaptive dual-time fusion module, and a selective high-frequency refiner for decoding. The spatial-low-frequency coordination module is configured to extract spatial and frequency features from each phase image of a dual-temporal remote sensing image pair, generate attention weights based on the spatial and frequency features, and weight the spatial features to obtain a harmonic feature map under low-frequency context constraints. The adaptive dual-temporal fusion module is configured to separate the common and difference components of the harmonic feature maps of the dual-temporal image pairs, and adaptively fuse the harmonic feature maps of the dual-temporal image pairs according to the difference components to obtain adaptive dual-temporal fusion features. A selective high-frequency thinner is configured to selectively reconstruct and restore high-frequency details of the obtained adaptive dual-temporal fusion features through wavelet reconstruction, resulting in a high-resolution thinned feature map. Classification is then performed based on the thinned feature map to obtain the change detection results of the dual-temporal remote sensing image.

[0024] In step 1, a pair of two-temporal remote sensing images of the target area is acquired. The pair of two-temporal remote sensing images includes a first temporal image T1 and a second temporal image T2. The images are registered high-resolution remote sensing images. During training, images can be acquired from public datasets such as LEVIR-CD, S2Looking, CDD, or SYSU-CD. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0025] In step 2, spatial and frequency features are extracted from each temporal image of the dual-temporal remote sensing image pair. Attention weights are generated based on these spatial and frequency features to weight the spatial features, resulting in a harmonic feature map constrained by the low-frequency context. This step is implemented through a spatial-low-frequency coordination module in the constructed frequency modulation network, such as... Figure 3 As shown, the space-low frequency coordination module employs a twin dual-branch encoder, including a parallel spatial branch and a frequency branch, as well as a space-low frequency harmonic module; Spatial branches can be extracted using CNN backbone networks to extract fine-grained spatial details and local patterns as spatial features; The frequency branch is configured to extract frequency features using a wavelet transform-based recursive decomposition strategy. The spatial-low-frequency harmonic module is configured to fuse features from the two branches, anchoring spatial features to low-frequency features and low-frequency context features provided by the frequency branch, and generating a harmonic feature map. In one specific implementation, the spatial branch uses EfficientNet as the standard CNN backbone (Encoder Block) to extract fine-grained spatial details and local patterns that the frequency domain branch fails to capture. Although these features may be domain-sensitive, they are indispensable for accurate change localization.

[0026] It should be noted that in this embodiment, the field is understood to refer to the acquisition conditions, scene distribution, or data source environment of the image data, such as lighting conditions, seasonal phenology, imaging angle, registration accuracy, sensor type, or geographical differences. In step 2, frequency features are extracted from each phase image of the dual-temporal remote sensing image pair. This can be achieved using a recursive decomposition strategy based on wavelet transform in the frequency branch, including the following steps: Step 211: Perform initial feature extraction on the current phase image and use the extracted features as the initial input for the frequency branch; Step 212: Decompose the current initial input features into a low-frequency subband LL and three high-frequency subbands LH, HL, and HH; Wavelet transform decomposition is an iterative decomposition process. In the first level of the frequency branch, wavelet transform decomposition is performed on the initial input features. In the i-th level of the frequency branch, the low-frequency components XLL(i) of the previous stage output are decomposed. 1) Perform a two-dimensional wavelet transform to decompose it into a low-frequency subband LL and three high-frequency subbands LH, HL, and HH; Step 213: Using the low-frequency subband LL as the current input feature, execute step 212 to perform recursive decomposition until the set recursion cutoff condition is met. The high-frequency subbands obtained from each level of recursive decomposition are used to form a high-frequency subband set as a frequency feature. Specifically, at each level of the frequency branch, the input features are decomposed into a low-frequency subband LL and three high-frequency subbands LH, HL, and HH using a two-dimensional Haar wavelet transform. Only the low-frequency subband LL is passed to the next level for further decomposition, while the high-frequency subbands are retained as high-frequency prior information for subsequent use. This recursive process is represented as: ; in, The initial input features for the branch, This refers to the low-frequency sub-band of the previous stage. This is the low-frequency sub-band output at the current stage. Let be the set of high-frequency subbands of the i-th layer, and DWT denotes discrete wavelet transform; This embodiment employs a recursive decomposition strategy based on two-dimensional Haar wavelet transform, passing low-frequency subbands layer by layer for further decomposition. This allows the frequency branches to rapidly expand the receptive field as network depth increases, capturing a wider range of global contextual information with lower parameter growth costs. Simultaneously, Haar wavelets offer high computational efficiency and orthogonal basis properties, which helps reduce correlations between different frequency features. Their discontinuous, step-like basis function forms effectively characterize abrupt changes in man-made structures such as buildings and roads in remote sensing images, thereby enhancing the model's ability to perceive changes in real structures.

[0027] In step 2, attention weights are generated based on spatial and frequency features. The spatial features are then weighted to obtain a harmonic feature map under low-frequency context constraints. This map is configured to be implemented in the spatial-low-frequency harmonic mapping module. Specifically, the spatial features are... and frequency characteristics After convolution operations, the values ​​are summed, and then activated by an activation function to obtain attention weights. These attention weights are then multiplied element-wise with the spatial features and weighted to obtain the harmonic feature map. ; Specifically, the convolution operation can be a 1×1 convolution, and the function activation can use the sigmoid function to generate a harmonic feature map. The expression is: ; in, The harmonic feature map, For spatial features, Frequency characteristics For the sigmoid function, For element-wise multiplication, This is a convolution operation.

[0028] This embodiment fuses fine-grained spatial features extracted by the spatial branch with low-frequency context features extracted by the frequency branch through a spatial-low-frequency harmonicization module. An attention gating mechanism is then used to weight and modulate the spatial features, subjecting them to a stable low-frequency context. This preserves the expressive power of spatial features for varying boundaries, local textures, and detailed structures while reducing their sensitivity to changes in illumination, seasonal variations, registration errors, and sensor differences. This suppresses spurious high-frequency responses caused by the dual-temporal domain gap, improving the accuracy, stability, and cross-domain generalization ability of change detection results.

[0029] In the implementation of step 2 above, to address the problem that factors such as illumination changes, seasonal phenological differences, imaging angle changes, registration errors, and sensor differences can easily manifest as high-frequency differences in images, such as abrupt brightness changes, texture disturbances, and edge misalignments, thereby inducing the model to form shortcut judgments dependent on apparent differences, a twin dual-branch encoder including spatial and frequency branches is constructed. On the one hand, the spatial branch extracts fine-grained spatial features from the dual-temporal images to preserve the boundary, texture, and local structural information of the changing regions. On the other hand, the frequency branch employs a recursive low-frequency decomposition strategy based on Haar wavelet transform to extract low-frequency context information layer by layer, while retaining high-frequency sub-bands as prior information. Furthermore, using a spatial low-frequency harmonicization module, attention weights are generated based on spatial and low-frequency features, and these attention weights are used to perform element-wise weighted modulation of the spatial features, so that the spatial features are constrained within the stable low-frequency context provided by the frequency branch. Thus, the model can reduce its dependence on high-frequency spurious correlations such as illumination, texture, noise, and edge misalignment while retaining the ability to locate fine changes. This suppresses spurious change responses introduced by the two-temporal domain gap and improves the robustness and generalization ability of change detection across regions, time phases, or acquisition conditions.

[0030] Step 3: Based on the harmonic feature maps of each temporal image, the common and difference components of the harmonic feature maps of the two temporal image pairs are separated, and the harmonic feature maps of the two temporal image pairs are adaptively fused according to the difference components to obtain adaptive dual-temporal fusion features; this step is implemented by an adaptive dual-temporal fusion module, which includes a contextual difference modulator and a difference-guided temporal fusion module. A contextual difference modulator is used to separate stable and differential components from the harmonic feature maps of two time phases through a self-attention mechanism and generate attention-aware differential features. The difference-guided temporal fusion module is used to perform weighted fusion of dual temporal harmonic feature maps based on attention-perceived difference features, generating adaptive dual temporal fusion features.

[0031] In one specific implementation, the contextual difference modulator separates stable and differential components through a self-attention mechanism. The harmonic feature maps extracted from the spatial branch backbone network at two temporal snapshots T1 and T2 are used... As input, respectively denoted as and .

[0032] In step 3, based on the harmonic feature maps of each temporal image, the common and difference components of the harmonic feature maps of the two temporal image pairs are separated and configured to be implemented in the context difference modulator, including the following steps: Step 311: Concatenate the harmonic feature maps of the two time phases, and refine the features using the channel attention module and the spatial attention module to obtain the value tensor. ; ; ; Step 312: Summate the harmonic feature maps of the two time phases to obtain the query tensor Q; ; Step 313: Calculate the difference between the harmonic feature maps of the two time phases to obtain the bond tensor K; ; Step 314: Calculate context-aware attention based on query tensor Q and key tensor K, and use this attention to filter value tensor V to obtain attention-aware difference features S; ; in, This is the scaling factor.

[0033] This embodiment constructs query tensors, key tensors, and value tensors using a context-aware difference modulator, enabling explicit differentiation between stable shared information and variable difference information in the two-temporal features. The query tensor, generated by summing the two-temporal features, represents the consistent scene structure, low-frequency semantics, and global contextual information in both temporal phases. The key tensor, generated by subtracting the two-temporal features, represents the difference responses between the two phases, including real-world feature changes and noise changes caused by illumination, season, registration errors, and sensor differences. By utilizing stable low-frequency contextual information for attention matching and filtering of noisy difference information, the embodiment enhances the real-world change response consistent with the overall scene structure and suppresses false high-frequency differences inconsistent with the context. Simultaneously, the value tensor, after context-aware attention filtering, generates saliency-aware difference features, making the subsequent temporal fusion process more focused on real-world change areas, thereby improving the discriminative power of the fused features and the accuracy and reliability of the change detection results.

[0034] In step 3, the harmonic feature maps of the two-temporal image pairs are adaptively fused based on the difference components to obtain adaptive dual-temporal fusion features, which are configured to be implemented in the difference-guided temporal fusion module, including the following steps: Step 321: Construct a feature pyramid P from the dual-temporal harmonic feature maps output by the spatial-low-frequency harmonic module in Step 2, and perform residual processing on the feature pyramid P twice to obtain the intermediate representation B. The difference-guided temporal fusion module receives two inputs: a feature pyramid P from the spatial low-frequency harmonic module, and attention-aware difference features S from the contextual difference modulator; First, the feature pyramid P is processed through a two-tiered branching module to generate an intermediate representation B. Each branching module contains two consecutive residual blocks, and each residual block consists of Conv3-GELU-Conv3. The processing procedure is as follows: ; in, This indicates residual block operations.

[0035] Step 322: Concatenate the intermediate representation B with the attentional perceptual difference feature S to obtain the comprehensive feature map C: ; Step 323: Perform global average pooling on the comprehensive feature map C, convolve and activate it using a dynamically adjusted kernel size k, and generate channel attention vectors. ; Furthermore, the kernel size The following is calculated based on the number of channels in the input comprehensive feature map C: ; in, This indicates rounding to the nearest odd integer.

[0036] Attention Vector Generate using the following formula: ; ; GAP stands for Global Average Pooling. This represents the activation function.

[0037] Step 324: Utilize channel attention vectors The comprehensive feature map C is reweighted to obtain a weighted feature map, which is used as the adaptive dual temporal fusion feature. Attention vectors are used to reweight feature map C, and finally, group convolution operations are applied to generate adaptive dual-temporal fusion features Y. ; This embodiment uses a contextual difference modulator to separate the stable common components and differential components in the bi-temporal harmonic features, and uses the stable context to filter the differential information, thereby enhancing the real structural changes and suppressing spurious change interference caused by illumination, season, registration errors, and sensor differences. Furthermore, through a difference-guided temporal fusion module, attention-aware differential features are introduced into the bi-temporal feature fusion process, enabling the model to adaptively adjust the response intensity of different channels and regions according to the significance of changes, thereby generating more discriminative fused features and improving the accuracy, stability, and boundary representation ability of change detection.

[0038] Step 4: Selectively reconstruct and restore high-frequency details of the obtained adaptive dual-temporal fusion features through wavelet reconstruction to obtain a high-resolution thinned feature map. Classification is performed based on the thinned feature map to obtain the change detection results of the dual-temporal remote sensing image. This step is implemented by the selective high-frequency thinner in the decoder. The decoder corresponds to the output layer of the encoder, that is, the decoding layer of the decoder corresponds to the coding layer of the encoder. The decoder can receive features output from the coding layer at multiple scales and decode the features output from different layers of the encoder after processing in steps 2-3. The decoder is constructed as a selective high-frequency refiner, which uses wavelet reconstruction to selectively re-inject and modulate the high-frequency information retained in the encoder stage; In step 4, the obtained adaptive dual-temporal fusion features are selectively reconstructed and high-frequency details are restored through wavelet reconstruction to obtain a high-resolution refined feature map, including the following steps: Step 41: Obtain adaptive dual-temporal fusion features at different scales from the difference-guided temporal fusion module; Specifically, the selective high-frequency refiner receives three feature maps from the difference-guided time fusion module. The spatial scales of these feature maps decrease sequentially, and they are denoted as follows: , , The adaptive dual-temporal fusion feature Y is a multi-scale feature pyramid, with all three scales present. For ease of representation, the scale decreases sequentially; in this embodiment, we use... , , express; Step 42: Extract high-frequency prior information: for coarse-scale feature maps Two-dimensional discrete wavelet transform is performed, and the extracted high-frequency subbands are spliced ​​together to obtain high-frequency prior information; Using discrete wavelet transform from the coarsest feature map Extracting high-frequency priors , means as follows: ; ; Where DWT represents the two-dimensional discrete wavelet transform, These represent the horizontal, vertical, and diagonal high-frequency subbands after a two-dimensional discrete wavelet transform, respectively. The low-frequency subbands are discarded; the first underscore indicates the discarded low-frequency portion. As a high-frequency prior, it is used to guide the subsequent reconstruction process.

[0039] Step 43: Based on high-frequency prior information A multi-path sampling reconstruction path is constructed. Based on wavelet reconstruction units, wavelet reconstruction is used to upsample, compensate for high frequencies, and reconstruct features at different scales to obtain the corresponding reconstructed feature maps. , , ; Taking three scale feature maps as an example, let's first analyze the coarse-scale feature map. Upsampling is performed and compared with intermediate scale feature maps. By performing fusion, intermediate fusion features are obtained. ; ; in, Indicates upsampling; High-frequency reconstruction is performed using wavelet reconstruction units (WRUs), resulting in the output of three fused streams, as shown in the following formula: ; ; ; in, This indicates an upsampling operation, and WRU represents a wavelet reconstruction unit.

[0040] Further technical solutions include a wavelet reconstruction unit structure as follows: Figure 4 As shown, it is used to reconstruct a high-resolution feature map with clear boundaries based on a low-resolution spatial feature map and high-frequency prior information. The wavelet reconstruction unit receives a low-resolution spatial feature map. and high-frequency priors Low-resolution spatial feature map Including intermediate fusion features High-frequency prior information For coarse-scale feature maps Features after upsampling ; Furthermore, the wavelet reconstruction unit is configured to perform the following process: Step 431: Process the input feature map Perform channel-level statistical analysis, calculate the channel mean characteristics and channel standard deviation characteristics respectively, and generate mean modulated signals and standard deviation modulated signals respectively; The channel mean is used to characterize the overall intensity distribution of the characteristic response, while the channel standard deviation is used to characterize the degree of fluctuation and local variation information between different channels. Together, they reflect the contextual structure information in the low-resolution spatial features; generating the mean-modulated signal. Sum of standard deviation modulated signal as follows: ; ; in, For the mean channel; Multilayer perceptron (Stdv channel) for standard deviation processing Step 432: Apply the mean-modulated signal and the standard deviation-modulated signal together to the high-frequency prior features. Adaptive modulation of high-frequency information yields modulated high-frequency prior features. This allows the learned spatial context to guide which high-frequency details should be enhanced or suppressed. ; Step 433: Compare the modulated high-frequency prior with the input feature map. The low-frequency components are combined and processed through inverse discrete wavelet transform ( Reconstruct high-resolution, detailed feature maps : ); Step 433: After refining the reconstructed feature map using a residual block containing group convolutions and end-channel attention layers, the resulting image is stitched together to obtain a high-resolution refined feature map. : ; The above processing steps are applied to different... Processing to obtain refined features Three feature maps were obtained respectively. , , That is, the input is At that time, With feature map Weighted, we get Input is At that time, With feature map Weighted; input is At that time, ,get Feature maps generated by fusion stream , , After adaptive convolution thinning, the images are concatenated to form the final output: ; This collaborative process guides the reintegration of frequency components through context-rich spatial features, ensuring that the final change map is not only semantically accurate but also has clear and sharp boundaries.

[0041] like Figure 4 As shown, the residual block containing group convolution and terminal channel attention layers includes a first 1×1 convolutional layer, a SiLU activation layer, a group convolutional layer, a second 1×1 convolutional layer, a feature fusion layer, and a channel attention layer connected in sequence. The feature fusion layer is also connected to the input end of the residual block through a residual bypass to fuse the convolutionally refined features with the input features. The channel attention layer is located at the end of the residual block to perform channel weight recalibration on the fused features.

[0042] Therefore, this embodiment introduces a spatial context-guided high-frequency selective modulation mechanism in the wavelet reconstruction unit, ensuring that high-frequency prior information is not indiscriminately recovered, but rather adaptively enhanced or suppressed based on the stable semantic context contained in the low-frequency spatial features. Specifically, low-frequency fusion features provide semantic constraints for the reconstruction of high-frequency details, enhancing effective high-frequency components related to target region boundaries, texture changes, and structural abrupt changes, while suppressing high-frequency responses corresponding to background noise, pseudo-textures, or irrelevant details, thereby reducing false detections and boundary spikes. Furthermore, by re-fusing the stable low-frequency context with the modulated high-frequency prior through inverse wavelet reconstruction, and refining the reconstructed features using group convolution and channel attention layers, the clarity of target edges and spatial positioning accuracy can be improved while maintaining the overall semantic consistency of the changed region. Thus, this embodiment can achieve high-fidelity feature reconstruction and high-frequency detail refinement, resulting in a final change detection map with more accurate region discrimination results, more continuous target contours, and sharper boundary expressions, thereby improving the accuracy and robustness of change detection in complex scenes.

[0043] The constructed frequency modulation network was trained. The AdamW optimizer was used, with parameters set to an exponential decay rate β1 = 0.9 for the first-order moment estimate (mean gradient), an exponential decay rate β2 = 0.999 for the second-order moment estimate (squared gradient), and a weight decay of 0.01. During training, the initial learning rate was 0.001, dynamically adjusted using cosine annealing. The model was trained for 150 full epochs. The input image was segmented into small patches with an effective batch size of 4, achieved through gradient accumulation. For the LEVIR-CD and S2Looking datasets, images were cropped to 512×512 pixel patches; for the SYSU-CD and CDD datasets, images were cropped to 256×256 pixel patches.

[0044] After model training is complete, the best-performing model is obtained. The dual-temporal image to be processed is input into the model to obtain the classified regions of change and save them as binarized images.

[0045] Experimental verification: To verify the effectiveness of this invention, experiments were conducted on four publicly available datasets: LEVIR-CD, S2Looking, CDD, and SYSU-CD. The results are shown in Tables 1 to 4.

[0046] Table 1. Quantitative comparison results on the LEVIR-CD dataset;

[0047] In Table 1, PFMNet (ours) represents the method of this embodiment. The following are existing methods, explained as follows: FC-EF (Fully Convolutional Early Fusion) directly stitches together two temporal images in the input channel dimension, and then uses a single-branch fully convolutional encoder-decoder to predict the change map. FC-conc (Fully Convolutional Siamese Network with Concatenation) uses a shared-weights siamese encoder to extract features from two temporal images, then concatenates and fuses the two temporal features in the corresponding layer, and generates a change map through a decoder.

[0048] FC-diff (Fully Convolutional Siamese Network with Difference) uses a shared-weights siamese encoder to extract bi-temporal features, then calculates the difference information between corresponding features, and uses a decoder to predict the change region.

[0049] CDNet (Change Detection Network) uses a convolutional neural network to extract features from dual-temporal images and utilizes the difference features to detect changes.

[0050] IFNet (Deeply Supervised Image Fusion Network) extracts multi-layer, dual-temporal features through a two-stream fully convolutional network, and then uses a difference discriminant network for multi-scale fusion. Multiple deep supervised branches directly constrain the prediction of changes at different levels to enhance feature discrimination capabilities.

[0051] SNUNet (Siamese Nested U-Net) combines a Siamese encoder with dense skip connections in a Nested U-Net style, fully integrating features from different levels and time periods. Its focus is on preserving shallow details and improving the detection of small objects and changing boundaries.

[0052] BIT (Bitemporal Image Transformer) compresses bi-temporal features into a small number of semantic tokens, models global spatiotemporal relationships through a Transformer, and then projects the updated tokens back into the pixel space. This design enhances the ability to model long-range dependencies with lower overhead.

[0053] L-UNet (LSTM U-Net, Long Short-Term Memory U-Net network) jointly learns bi-temporal semantic segmentation and change detection, and uses a fully convolutional LSTM to model temporal dependencies. Semantic segmentation provides category-level constraints, while the temporal branch is responsible for identifying changes in the state of ground features.

[0054] P2V-CD (Pair-to-Video Change Detection) expands bi-temporal image pairs into pseudo-videos containing several intermediate states to more completely represent the change process from the earlier to the later phase. Subsequently, it decouples the spatiotemporal information for modeling, alleviating the problem of insufficient temporal modeling in traditional methods.

[0055] HFANet (High-Frequency Attention Siamese Network) introduces a high-frequency attention module into the Siamese network. It first locates potential building regions through spatial attention, and then enhances the high-frequency edge components within them. Its main goal is to improve the boundary localization and fine-grained recognition of changing buildings.

[0056] HANet (Hierarchical Attention Network) designs a lightweight hierarchical attention module that jointly integrates multi-scale semantic and detail features; at the same time, it adopts progressive foreground equalization sampling to alleviate the imbalance problem of changed / unchanged pixels.

[0057] HCGMNet (Hierarchical Change Guiding Map Network) models both local details and global context simultaneously through hierarchical convolution and progressive feature fusion. The change guidance module uses a coarse-grained change map as a self-attention prior and gradually refines the change regions and their edges.

[0058] WS-Net++ (Wavelet Siamese Network++) utilizes wavelet transform to decompose image frequency information, then combines spatial feature fusion and domain adaptation strategies to improve change detection capabilities. Spatial-frequency interactions are established between two temporal images to enhance the integrity of change boundaries. Spatial domain interaction highlights differences between pixels. Frequency domain interaction first adaptively adjusts the contributions of different frequency components based on the image context. Further, intra-frequency and inter-frequency interactions are constructed to capture frequency domain differences, thus enabling adaptive and effective handling of both overall and subtle changes.

[0059] RFANet (Robust Feature Aggregation Network) uses a lightweight backbone to reduce the number of parameters and a feature enhancement module to enable dense interaction between the current layer and features at other scales. This module achieves complementarity between shallow details and deep semantics, improving feature representation capabilities with extremely low complexity.

[0060] CDC2F (Cross-Domain Coarse-to-Fine Network) constructs a coarse-to-fine detection framework that coordinates the spatial and frequency domains. It first uses frequency information to obtain a relatively stable coarse-grained variation representation, and then combines spatial details to gradually refine the variation region and boundary.

[0061] ISDANet (Interactive and Supervised Dual-Mode Attention Network) uses a dual-temporal feature interaction structure to extract differential information, and then uses spatial and channel attention to enhance effective variation features.

[0062] Hybrid-MambaCD (Hybrid Mamba Change Detection) combines convolutional networks with the Mamba architecture, and then uses local feature extraction and global relation modeling to improve change detection capabilities.

[0063] STRobustNet (Spatial-Temporal Robust Network) learns robust spatiotemporal feature representations in dual-temporal images and then reduces the impact of non-realistic changes on the detection results.

[0064] SFEARNet (Semantic Flow and Edge-Aware Refinement Network) alleviates spatial misalignment of features across scales by aligning information transfer between features of different resolutions through semantic flow. The edge-aware refinement module explicitly extracts and enhances changing boundaries, thus balancing semantic integrity and contour accuracy.

[0065] ChangeRD (Registration-Integrated Change Detection Framework) jointly models image registration and change detection, and then improves the ability to adapt to registration errors by learning spatial transformation relationships.

[0066] FSG-Net (Frequency–Spatial Synergistic Gated Network) uses wavelet frequency interaction modules to extract frequency features, and then enhances the ability to identify real-world changing regions through spatial attention and gating fusion mechanisms.

[0067] Table 2. Quantitative comparison results on the S2Looking dataset;

[0068] Table 3. Quantitative comparison results on the SYSU-CD dataset;

[0069] Table 4. Quantitative comparison results on the CDD dataset;

[0070] Different change detection methods were evaluated using precision (Pr), recall (Rc), F1 score, intersection-over-union (IoU), and overall accuracy (OA). Precision measures the proportion of correctly detected changes in the predicted change region, reflecting the model's ability to control false positives; recall measures the proportion of correctly detected true change regions, reflecting the model's ability to control false negatives; the F1 score is the harmonic mean of precision and recall, used to comprehensively evaluate the model's performance in detecting change regions; the intersection-over-union (IoU) measures the spatial overlap between the predicted and true change regions; and overall accuracy represents the proportion of correctly classified pixels out of all pixels. Higher values ​​for these metrics indicate better change detection performance.

[0071] The results show that the method in this embodiment achieves the best F1 score on all four public datasets, outperforming existing methods. Figure 5 This is a comparison chart of the model's prediction results. Green represents true positives (TP), yellow represents false negatives (FN), red represents false positives (FP), and black represents true negatives (TN). To further evaluate the domain generalization ability of the network model in this embodiment, a cross-dataset validation experiment was conducted: the model was trained only on the LEVIR-CD dataset and inference was performed directly on the LEVIR-CD+ dataset. The results are shown in Table 5.

[0072] Table 5. Cross-dataset validation results (LEVIR-CD → LEVIR-CD+);

[0073] The results show that the method in this embodiment exhibits excellent generalization ability in cross-domain scenarios, verifying the effectiveness of the frequency transfer mechanism in improving the generalization ability of change detection.

[0074] Example 2 Based on Example 1, this example provides a frequency modulation-based remote sensing change detection system, including: The data acquisition module is configured to acquire two-temporal remote sensing image pairs of the target area; The spatial-low-frequency coordination module is configured to extract spatial and frequency features from each temporal image of a dual-temporal remote sensing image pair, and generate attention weights based on the spatial and frequency features to weight the spatial features, thereby obtaining a harmonic feature map under low-frequency context constraints. The adaptive dual-temporal fusion module is configured to separate the common and difference components of the harmonic feature maps of the dual-temporal image pairs, and adaptively fuse the harmonic feature maps of the dual-temporal image pairs according to the difference components to obtain adaptive dual-temporal fusion features. A selective high-frequency thinner is configured to selectively reconstruct and restore high-frequency details of the obtained adaptive dual-temporal fusion features through wavelet reconstruction, resulting in a high-resolution thinned feature map. Classification is then performed based on the thinned feature map to obtain the change detection results of the dual-temporal remote sensing image.

[0075] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0076] Example 3 Based on Embodiment 1, this embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps in the frequency modulation-based remote sensing change detection method described in Embodiment 1.

[0077] Example 4 Based on Embodiment 1, this embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they complete the steps in the frequency modulation-based remote sensing change detection method described in Embodiment 1.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0079] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A remote sensing change detection method based on frequency modulation, characterized in that, Includes the following steps: Acquire dual-temporal remote sensing image pairs of the target area; Spatial and frequency features are extracted from each phase image of the dual-temporal remote sensing image pair. Attention weights are generated based on the spatial and frequency features to weight the spatial features, resulting in a harmonic feature map under low-frequency context constraints. The common and difference components of the harmonic feature maps of the two-phase image pairs are separated, and the harmonic feature maps of the two-phase image pairs are adaptively fused according to the difference components to obtain adaptive dual-temporal fusion features. The obtained adaptive dual-temporal fusion features are selectively reconstructed and high-frequency details are restored through wavelet reconstruction to obtain a high-resolution thinned feature map. Based on the thinned feature map, classification is performed to obtain the change detection results of the dual-temporal remote sensing image.

2. The remote sensing change detection method based on frequency modulation as described in claim 1, characterized in that, It also includes the construction of a frequency modulation network, which adopts an encoder and decoder structure, including a spatial-low frequency coordination module for encoding, an adaptive dual-time fusion module, and a selective high-frequency refiner for decoding.

3. The remote sensing change detection method based on frequency modulation as described in claim 2, characterized in that, The space-low frequency coordination module employs a twin dual-branch encoder, including parallel spatial and frequency branches, as well as a space-low frequency harmonic module; Spatial branches can be extracted using CNN backbone networks to extract fine-grained spatial details and local patterns as spatial features; The frequency branch is configured to extract frequency features using a wavelet transform-based recursive decomposition strategy. The spatial-low-frequency harmonic module is configured to fuse features from the two branches, anchoring spatial features to low-frequency features and low-frequency context features provided by the frequency branch, and generating a harmonic feature map. Alternatively, the adaptive dual-time fusion module includes a context-differential modulator and a difference-guided time fusion module; A contextual difference modulator is used to separate stable and differential components from the harmonic feature maps of two time phases through a self-attention mechanism and generate attention-aware differential features. The difference-guided temporal fusion module is used to perform weighted fusion of dual temporal harmonic feature maps based on attention-perceived difference features, generating adaptive dual temporal fusion features.

4. The remote sensing change detection method based on frequency modulation as described in claim 1, characterized in that, Frequency features are extracted from each phase image of the dual-temporal remote sensing image pair using a recursive decomposition strategy based on wavelet transform, including the following steps: Step 211: Perform initial feature extraction on the current phase image and use the extracted features as the initial input for the frequency branch; Step 212: Decompose the current input features into a low-frequency subband LL and three high-frequency subbands LH, HL, and HH; Step 213: Using the low-frequency subband LL as the current input feature, execute step 212 to perform recursive decomposition until the set recursion cutoff condition is met. The high-frequency subbands obtained from each level of recursive decomposition are used to form a high-frequency subband set as a frequency feature.

5. The remote sensing change detection method based on frequency modulation as described in claim 1, characterized in that, Based on the harmonic feature maps of each temporal image, the common and difference components of the harmonic feature maps of the two temporal image pairs are separated, including the following steps: The harmonic feature maps of the two phases are concatenated, and the features are refined through the channel attention module and the spatial attention module to obtain the value tensor; The query tensor Q is obtained by summing the harmonic feature maps of the two time phases. The bond tensor K is obtained by subtracting the harmonic feature maps of the two time phases. Context-aware attention is calculated based on query tensor Q and key tensor K, and this attention is used to filter value tensor V to obtain attention-aware differential features S.

6. The remote sensing change detection method based on frequency modulation as described in claim 1, characterized in that, Adaptive fusion of the harmonic feature maps of the two-temporal image pairs based on the difference components yields adaptive dual-temporal fusion features, including the following steps: The harmonic feature maps of the two time phases are used to construct a feature pyramid P. The feature pyramid P is then subjected to two residual processing steps to obtain the intermediate representation B. The intermediate representation B is concatenated with the attention-perceived differential feature S to obtain the comprehensive feature map C: Global average pooling is performed on the comprehensive feature map C, and convolution and activation are performed using a dynamically adjusted kernel size k to generate channel attention vectors. ; Using channel attention vectors The comprehensive feature map C is reweighted to obtain a weighted feature map, which is used as the adaptive dual-temporal fusion feature.

7. The remote sensing change detection method based on frequency modulation as described in claim 1, characterized in that, The obtained adaptive dual-temporal fusion features are selectively reconstructed and high-frequency details are restored through wavelet reconstruction to obtain a high-resolution refined feature map, including the following steps: Obtain adaptive dual-temporal fusion features at different scales; coarse-scale feature maps Two-dimensional discrete wavelet transform is performed, and the extracted high-frequency subbands are spliced ​​together to obtain high-frequency prior information; Based on high-frequency prior information A multi-path sampling reconstruction path is constructed, and the fusion features at different scales are upsampled, high-frequency compensated and feature reconstructed based on wavelet reconstruction to obtain the corresponding reconstructed feature map.

8. A remote sensing change detection system based on frequency modulation, characterized in that, include: The data acquisition module is configured to acquire two-temporal remote sensing image pairs of the target area; The spatial-low-frequency coordination module is configured to extract spatial and frequency features from each temporal image of a dual-temporal remote sensing image pair, and generate attention weights based on the spatial and frequency features to weight the spatial features, thereby obtaining a harmonic feature map under low-frequency context constraints. The adaptive dual-temporal fusion module is configured to separate the common and difference components of the harmonic feature maps of the dual-temporal image pairs, and adaptively fuse the harmonic feature maps of the dual-temporal image pairs according to the difference components to obtain adaptive dual-temporal fusion features. A selective high-frequency thinner is configured to selectively reconstruct and restore high-frequency details of the obtained adaptive dual-temporal fusion features through wavelet reconstruction, resulting in a high-resolution thinned feature map. Classification is then performed based on the thinned feature map to obtain the change detection results of the dual-temporal remote sensing image.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the frequency modulation-based remote sensing change detection method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps in the frequency modulation-based remote sensing change detection method according to any one of claims 1-7.

Images