Anti-seasonal-interference frequency domain sensing remote sensing image change monitoring method and anti-seasonal-interference frequency domain sensing remote sensing image change monitoring system

Abstract

The invention provides an anti-seasonal interference frequency domain perception remote sensing image change monitoring method and system, and the method comprises the steps: firstly constructing a twin feature extraction network based on Transform, and embedding a semi-instance normalization module in a coding stage, thereby effectively eliminating the radiation and illumination differences between dual-time-phase images; secondly, designing a frequency domain self-adaptive filtering module, and suppressing low-frequency seasonal noise and enhancing high-frequency structure details through a self-adaptive filter by utilizing wavelet transform separation characteristics; thirdly, constructing a change attention guidance module, and fusing space-time and frequency domain difference features to focus a high-confidence change region; and finally, sharpening and correcting the edge of the change area by utilizing explicit offset prediction through a cascade decoder and a boundary refining module. The method can significantly reduce the false alarm rate caused by seasons and illumination, accurately recovers the ground feature change boundary, and is suitable for a ground surface change monitoring task in a complex environment.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, specifically to a frequency domain sensing remote sensing image change monitoring method and system that resists seasonal interference. Background Technology

[0002] Remote sensing image change monitoring aims to identify significant changes in the Earth's surface by analyzing two or more remote sensing images of the same geographic area acquired at different times. This technology has important applications in urban planning, disaster assessment, and resource monitoring.

[0003] However, existing change monitoring methods face the core challenge of "spurious change" interference in practical applications. Non-semantic changes, or spurious changes, in remote sensing images mainly include radiation variations caused by differences in solar altitude angle and atmospheric conditions at the time of imaging, as well as seasonal or phenological changes caused by vegetation growth cycles, such as spring growth and autumn shedding, and snow cover. These spurious changes introduce a large amount of noise into the feature space, leading to high false alarm rates in traditional feature-difference-based deep learning methods.

[0004] Specifically, existing deep learning change detection models typically compare the differences between bi-temporal features directly. While this approach is effective for obvious object changes, it often struggles to distinguish between genuine semantic changes and spurious environmental changes when images exhibit strong seasonal texture variations, such as tree color changes, or global illumination shifts. Furthermore, the downsampling operation of multi-layer convolutional neural networks reduces the spatial resolution of the image, resulting in blurred object boundaries in the recovered change map, making it difficult to accurately delineate the fine edges of buildings or roads.

[0005] In summary, deep learning has driven the rapid development of image change monitoring in the field of remote sensing. However, existing methods still have significant shortcomings in dealing with seasonal changes, illumination differences, and complex texture backgrounds, and new methods are urgently needed to overcome these bottlenecks.

[0006] Patent application CN118015460A discloses a remote sensing image change monitoring method based on cross-scale guidance and enhancement. The method includes the following steps: extracting multi-level features from dual-temporal remote sensing images using a weight-sharing dual encoder, and improving the perception and representation capabilities of multi-scale features using a high-order feature interaction module; improving the skip connection operation between the decoder and encoder, proposing a cross-scale guidance enhancement module to enhance features of interest and filter irrelevant background interference; and fusing the change semantic information reconstructed by the multi-level decoder using a dual-temporal feature alignment fusion module to extract changes of interest and avoid spurious changes caused by dual-temporal feature matching errors. However, this patent cannot completely solve the existing technical problems, nor can it meet the needs of this invention. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a frequency domain sensing remote sensing image change monitoring method and system that is resistant to seasonal interference.

[0008] The frequency domain sensing remote sensing image change monitoring method for resisting seasonal interference provided by the present invention includes:

[0009] Step S1: Acquire two high-resolution optical remote sensing images of the same geographic area taken at different times, perform geometric and radiometric corrections, and simultaneously perform geometric transformations on the images and labels, including random flipping, random rotation, random cropping or scaling, as well as pixel transformations including Gaussian blur, Gaussian noise, color jitter and random occlusion, to obtain an enhanced dual-temporal input image pair. Step S2: Input the dual-temporal input image pairs into a shared-weight Siamese Transformer feature extraction network. This network contains multiple processing stages. In each stage, the input is downsampled and mapped through an overlapping slice embedding layer. The resulting feature maps are then input into a semi-instance normalization module for processing to eliminate radiometric differences. The processed features are then input into a Transformer encoding block to extract semantic information. Finally, a set of first-temporal feature maps and a set of second-temporal feature maps at multiple scales are output. Step S3: Calculate and filter the frequency domain differences of the extracted first and second phase feature maps at the same scale; decompose the feature maps into frequency sub-bands using the standard Haar discrete wavelet transform, and calculate the differences between the corresponding sub-bands; construct low-frequency difference filters and high-frequency difference filters respectively; apply adaptive smoothing filtering to the calculated low-frequency sub-band differences to suppress seasonal pseudo-variables, and apply adaptive sharpening filtering to the calculated high-frequency sub-band differences to enhance structural details; reconstruct the filtered low-frequency and high-frequency sub-bands into pure frequency domain difference features in the spatial domain using the standard Haar inverse discrete wavelet transform. Step S4: Fuse spatial and frequency domain difference features to generate a change guidance map; calculate the difference between the first and second temporal feature maps in the spatial domain to obtain spatial domain difference features; perform spatial attention calculations on the spatial domain difference features and the pure frequency domain difference features obtained in step S3 to generate corresponding spatial attention maps and frequency attention maps; fuse the spatial attention maps and frequency attention maps and activate them using the Sigmoid function to generate a change guidance map; use the change guidance map to perform weighted fusion of the pure frequency domain difference features and spatial domain difference features to obtain the final fused difference features; Step S5: Gradually restore feature resolution and refine boundaries using a cascaded decoder; construct a cascaded decoder, concatenate the fused difference features obtained in Step S4 with features from deeper layers of the decoder, perform dimensionality reduction and residual refinement, and gradually upsample to restore high-resolution features; input the high-resolution features output by the decoder into the boundary refinement module, perform differentiable resampling of the feature map by predicting the pixel offset field to correct the boundary position, and then retain the original semantic information through residual connection to obtain the refined features; output the final binary change monitoring map through a convolution classifier using the refined features. Step S6: Use the binary change monitoring map and the real label to calculate the combined loss function to train the model end-to-end; the combined loss function includes cross-entropy loss for measuring pixel-level classification accuracy and wavelet consistency loss for constraining the boundary quality of the change region.

[0010] Preferably, step S2 includes: The strategy adopted is to first decouple, then normalize, and then fuse the input features. Through a Convolutional layers expand channels to obtain intermediate features. It is divided into two parts on average along the channel dimension. and This achieves decoupling, as expressed in the following expression:

[0011] right Perform instance normalization to obtain :

[0012] right Preserving the identity mapping, we obtain :

[0013] The normalized features are concatenated with the original features, and then non-linearly fused using a convolutional path containing the LeakyReLU activation function. Residual connections are also introduced to obtain the normalized output features. :

[0014] Will The input consists of multiple stacked Transformer encoded blocks, from which long-range dependencies and high-level semantics are extracted. Through a multi-stage, pre-defined process, feature maps at multiple levels are obtained, ultimately outputting a two-stage, multi-scale feature map that eliminates radiative pseudo-variations. and ,in ; in, To divide the operation functions, for Convolution operations in convolutional layers For instance normalization operation functions, for Convolution operations in convolutional layers This is the residual connection function.

[0015] Preferably, step S3 includes: Using multi-scale feature maps as input, the standard Haar Discrete Wavelet Transform (DWT) operator is used to decompose the feature maps into four frequency sub-bands in the spatial dimension: low-frequency sub-band. and level ,vertical diagonal High-frequency subbands in three directions; Calculate the phase difference of the low-frequency sub-band:

[0016] Calculate the phase difference of high-frequency sub-bands:

[0017] in, This is the low-frequency approximate subband of the second phase image. This is the low-frequency approximate subband of the first phase image. The second phase is the horizontal high-frequency sub-band. The first phase is the horizontal high-frequency sub-band. This is the vertical high-frequency sub-band of the second phase. The first phase is the vertical high-frequency sub-band. The diagonal high-frequency sub-band of the second phase, This is the diagonal high-frequency sub-band of the first phase; Constructing a low-frequency difference filter: Input into a lightweight convolutional neural network predictor In the process, a spatially variable weight graph is generated. ,right Perform a Softmax normalization operation on the kernel weight channel dimension to obtain a normalized low-pass kernel. , using the generated right Perform pixel-wise spatially variable convolution to obtain the filtered low-frequency features. :

[0018] Constructing a high-frequency difference filter: Input a prediction network, generate baseline kernel weights, and obtain them through Softmax normalization. Through the formula:

[0019] Generate high-pass filter kernel ,in For the unit identity kernel, using right Perform spatially variable convolution to obtain enhanced high-frequency features:

[0020] Using the standard Haar inverse discrete wavelet transform (IDWT) operator, the filtered low-frequency subband... and Reconstructing back to the spatial domain:

[0021] Received To remove seasonal noise while preserving the pure frequency domain difference features that accurately reflect the details of changes, these features are used for subsequent attention guidance and fusion.

[0022] Preferably, step S4 includes: The spatial domain difference characteristics are calculated by the following formula:

[0023] in, This is the second phase characteristic diagram. This is a feature diagram of the first time phase; To each and To perform spatial attention calculation, average pooling and max pooling are first performed along the channel dimension, respectively, to obtain... and The calculation formula is as follows: This involves compressing channel information and aggregating spatial feature descriptors.

[0024]

[0025] in, For any input feature tensor, For batch size, For the number of channels, For spatial dimensions; For channel dimensions; This is the average pooling operation function. This is the max pooling operation function; Then and Concatenate the data along the channel dimension to obtain a 2-channel feature map, and then use a convolutional kernel of size [size missing]. The standard convolutional layer performs a convolution operation on the concatenated feature map, compressing the number of channels to 1, thereby generating a spatial attention map. :

[0026] Attention maps for the corresponding spatial domain differences were obtained respectively. Attention map with corresponding frequency domain differences ;in, for The convolution operation of a standard convolutional layer; A soft gating mechanism is used to fuse the two attention information streams. and Perform element-wise addition and map the values ​​to a sigmoid activation function. The interval is used to generate the final change guide. :

[0027] The input two-stream features are weighted and fused using this guidance graph, expressed as follows:

[0028] in, This indicates element-wise multiplication, merging features. This will serve as input to the subsequent decoder, guiding the network to focus on high-confidence regions of change for recovery.

[0029] Preferably, step S5 includes: The processing procedure of the cascaded decoder includes: for the first... The decoding level will come from the first Layer decoder features After upsampling, the fused features with the corresponding level By concatenating along the channel dimension, joint features are obtained. Using 1×1 convolution pairs Dimensionality reduction is performed to obtain intermediate features with a reduced number of channels. :

[0030] in, This is a batch normalization function; Through a main path containing two 3×3 convolutional layers, batch normalization, and ReLU activation... Feature extraction is performed to obtain deep features. and will and After addition, ReLU activation is applied to obtain the decoded output of the current level. The expression is:

[0031]

[0032] This output This will be used as input for the next stage of decoding, until the highest resolution features are recovered. .

[0033] Preferably, the boundary refining module performs the following operations: High-resolution features output by the decoder Input a convolutional network and predict a two-dimensional offset field. The offset is then limited to a preset factor using the Tanh function. Within, that is:

[0034] in, Convolution operations for lightweight convolutional networks; Construct a base mesh with the same size as the feature map. The predicted offset Add to the base mesh to generate a deformation sampling mesh. :

[0035] Using a differentiable bilinear interpolation sampling function ,according to For input features Resampling is performed to regress feature points to the true object edges, and residual connections are introduced to obtain refined features. :

[0036] in, These are the residual weighting coefficients; Refined characteristics Through a The convolution classifier compresses the number of channels to 1 and maps it to a sigmoid function. The interval is used to output the final binary change monitoring map. .

[0037] Preferably, the residual weighting coefficient The value is 0.1, and the preset factor is... The value is 0.05.

[0038] Preferably, step S6 includes: A combination of cross-entropy loss function and wavelet consistency loss function is adopted to simultaneously ensure pixel-level classification accuracy and boundary quality of changing regions; Cross-entropy loss function Defined as:

[0039] Wavelet consistency loss function Defined as:

[0040] Final total loss function for:

[0041] in, This indicates the total number of pixels. For the first The real label of each pixel The corresponding predicted probability; These are the high-frequency components corresponding to the Haar wavelet transform. For the predicted changes, This is a real label image.

[0042] Preferably, during training, the AdamW optimizer is used to update the network parameters, the initial learning rate is set to 2e-4, and a linear decay strategy is used to adjust the learning rate. The batch size is set to 16, and the total number of training rounds is 300.

[0043] The frequency domain sensing remote sensing image change monitoring system for resisting seasonal interference provided by the present invention includes: Data acquisition module: used to acquire dual-temporal remote sensing images of the area to be monitored; Feature extraction and normalization module: used to extract features and eliminate radiation differences by embedding a semi-instance normalization module in a Siamese Transformer encoder; Frequency domain filtering module: used to separate and suppress seasonal spurious changes using standard Haar wavelet transform; Feature Decoding and Boundary Refinement Module: Used to fuse multi-scale features and sharpen the output results using the boundary refinement module; Change graph generation module: Used to output the final binary change monitoring results.

[0044] Compared with the prior art, the present invention has the following beneficial effects: This invention effectively eliminates differences in illumination and atmospheric radiation through semi-instance normalization; it successfully suppresses complex seasonal spurious changes through frequency domain adaptive filtering and change attention guidance; and finally, it significantly improves the boundary accuracy of change monitoring results through a boundary refinement module and frequency domain loss. This invention significantly reduces the false alarm rate while maintaining a high monitoring rate, making it particularly suitable for surface change monitoring tasks in complex environments. Attached Figure Description

[0045] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a network architecture diagram of the method model of the present invention; Figure 3 This is a schematic diagram of the frequency domain adaptive filtering module (FAFF) in this invention; Figure 4 This is a schematic diagram of the variable attention guidance module (CAGM) in this invention; Figure 5 This is a schematic diagram of the boundary refining module (BRH) in this invention. Detailed Implementation

[0046] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0047] Example like Figure 1 This invention provides a frequency domain sensing remote sensing image change monitoring method based on resistance to seasonal interference, comprising: Step S1: Acquire two high-resolution optical remote sensing images of the same geographic area taken at different times, and construct a multidimensional augmented dataset; Two high-resolution optical remote sensing images covering the same area were selected and denoted as follows: and First, rigorous geometric and radiometric corrections are performed on the images to eliminate imaging distortion, and they are then cropped into input image pairs with uniform pixel dimensions. To enhance the model's generalization ability to seasonal variations, lighting differences, and local occlusion, a dual data augmentation strategy at both the geometric and pixel levels is introduced during the training phase. Regarding geometric transformations, random horizontal or vertical flips are performed simultaneously on the image and labels with a certain probability, and... The image is rotated at randomly selected angles. Simultaneously, multi-scale random cropping or scaling cropping strategies are employed. For resolution adjustment, bilinear interpolation is used for the image, and nearest-neighbor interpolation is used for the labels to enrich the scale diversity of the samples. In terms of pixel transformation, Gaussian blur and Gaussian noise are applied with specific probabilities to reduce the model's over-reliance on high-frequency textures. Color dithering is used to randomly perturb the brightness, contrast, saturation, and hue of the image to adapt to different lighting conditions at different times. Furthermore, random occlusion technology is introduced, randomly selecting rectangular regions in the image and replacing their pixel values ​​with neutral gray (grayscale value 127). This forces the model to focus on contextual global features. Finally, the preprocessed image is converted to tensor format and mean averaging is performed. Standard deviation The standardized transformation yields the final training input pairs with a consistent distribution.

[0048] Step S2: Feature extraction and semi-instance normalization; This invention constructs a Siamese Transformer feature extraction network based on shared weights, which consists of four layers. In each layer, features are first downsampled and mapped through an overlapping slice embedding layer, then radiation differences are eliminated through a semi-instance normalization module, and finally the features are input into a Transformer encoding block to extract semantics, resulting in a multi-scale feature map.

[0049] After preprocessing, the two phases of images and The dimensions are all These will serve as inputs to subsequent feature extraction networks. This invention employs a shared-weights Siamese Transformer as a Siamese encoder to process... and The encoder consists of four consecutive processing stages, and the processing flow of each stage is as follows: First, the input data is fed into the overlapping slice embedding layer. This is done using a stride of 4. The convolution operation performs overlapping slicing and downsampling of the features, while simultaneously mapping the features to the target channel dimension. This step reduces spatial resolution and expands channel dimension, forming a hierarchical structure of the feature pyramid. For an input size of... After processing, the image can be used to obtain four levels of feature maps.

[0050] To eliminate radiation artifacts caused by illumination and atmospheric conditions, semi-instance normalization is performed on the feature maps immediately after slice embedding and before the Transformer encoding block. A strategy of decoupling followed by normalization and then fusion is adopted: the embedded features... Through a Convolutional layers expand their channels to To obtain intermediate features It is divided into two parts on average along the channel dimension. and :

[0051] This achieves decoupling.

[0052] right Perform instance normalization by subtracting the mean and dividing by the standard deviation to remove global illumination and contrast information:

[0053] right Maintain identity mapping, preserving original feature semantics and texture details:

[0054] The normalized features are concatenated with the original features, and then non-linearly fused using a convolutional path containing the LeakyReLU activation function. Residual connections are also introduced to obtain the normalized output features. :

[0055] Features after HINet correction Inputting several stacked Transformer encoded blocks, we extract long-range dependencies and high-level semantics. After four stages of progressive processing, we finally obtain four levels of feature maps: Number of channels: 64; spatial dimensions: approximately It mainly includes corrected edge and texture details; Number of channels: 128; spatial dimensions: approximately To capture larger ground features; Number of channels: 256; spatial dimensions: approximately It contains rich mid-level semantic information; Number of channels: 512; spatial dimensions: approximately It represents highly abstract global semantic information.

[0056] Finally, the output is a two-phase multi-scale feature map that eliminates radiation spurious changes. and ,in This provides a consistent and robust feature base for subsequent change monitoring. For example... Figure 2 This is a diagram of the model's network architecture.

[0057] Step S3: Calculate and filter the frequency domain differences of the extracted multi-scale features; The feature maps from the two phases at the same scale are input into the frequency domain adaptive filtering module (e.g.) Figure 3 (This is a schematic diagram of the frequency domain adaptive filtering module FAFF). By separating and specifically processing feature differences in the frequency domain, it solves the problem of different frequency domain distributions between seasonal pseudo-variables and true variations. This step specifically includes the following sub-steps: The normalized first phase feature output from step S2 and aligned second phase features As input, the feature map is decomposed into four frequency sub-bands in the spatial dimension using the standard Haar Discrete Wavelet Transform (DWT) operator: a low-frequency approximate sub-band. and level ,vertical diagonal High-frequency subbands in three directions.

[0058] After performing the above decomposition on the dual-phase features respectively, the phase difference of the corresponding sub-bands is calculated:

[0059]

[0060] in, This mainly includes large-area smoothed differences caused by seasonal phenological changes, while High-frequency differences including object edges and texture details.

[0061] Construct a low-frequency difference filter to suppress Seasonal pseudo-variations in the filter. The filter achieves smoothing by predicting spatially variable low-pass convolution kernels. First, the seasonal pseudo-variations in the filter... Input into a lightweight convolutional neural network predictor In the process, a spatially variable weight graph is generated. ,in This represents the size of the filter kernel.

[0062] To ensure that the generated filter has low-pass smoothing characteristics, for In the channel dimension, i.e., kernel size Performing a Softmax normalization operation on the dimension yields a normalized low-pass kernel. :

[0063] The Softmax operation ensures that the sum of the kernel weights at each pixel position is 1 and all of them are positive.

[0064] Finally, using the generated right Perform pixel-wise spatially variable convolution to obtain the filtered low-frequency features:

[0065] This operation can adaptively smooth out large areas of unstructured differences, thereby suppressing spurious changes.

[0066] Reconstruct a high-frequency difference filter to enhance The HFDF also includes a prediction network, but employs a high-pass constraint in its kernel generation strategy, setting the kernel size to [value missing]. .

[0067] First, a prediction network based on high-frequency difference features is used. Generate basic weights, and then obtain a baseline low-pass kernel by normalizing it using Softmax. .

[0068] To construct a high-pass filter with sharpening capabilities, the unity identity kernel is utilized. (That is, the central element is 1, and the rest are 0) (Matrix) minus the baseline low-pass kernel:

[0069] This formula utilizes the principle that "subtracting the low-frequency component from the original signal yields the high-frequency component," thus forcing the generated convolution kernel to possess high-pass filtering characteristics.

[0070] Subsequently, using right Perform spatially variable convolution to obtain enhanced high-frequency features:

[0071] This step can significantly enhance structural edge differences (such as the outline of a new building) while suppressing irregular high-frequency clutter.

[0072] Finally, the standard Haar inverse discrete wavelet transform (IDWT) operator is used to filter the low-frequency subband. , Reconstructing back to the spatial domain:

[0073] Received The pure frequency domain difference features, which remove seasonal noise and retain the true details of variation, are used for subsequent attention guidance and fusion.

[0074] Step S4: Fuse the differences between the spatial domain and the frequency domain features to generate a high-confidence change guidance map and perform feature fusion; Although the frequency domain filtering module effectively suppressed seasonal spurious changes, to further improve the ability to capture regions of minute changes and verify the authenticity of the changes, the spatial domain difference features were first obtained by subtracting the features output from S2: ; Another is the frequency domain difference characteristic of the output after filtering and purification in step S3. ; For the input respectively and Spatial attention is calculated. First, average pooling and max pooling are performed along the channel dimension to compress channel information and aggregate spatial feature descriptors. The calculation formula is as follows:

[0075]

[0076] in, For any input feature tensor, For batch size, For the number of channels, (This refers to the spatial dimensions).

[0077] Then and Concatenating along the channel dimension yields a 2-channel feature map. A convolutional kernel of size [size missing] is then used. The standard convolutional layer performs a convolution operation on the concatenated feature map, compressing the number of channels to 1, thereby generating a spatial attention map. :

[0078] Through the above process, attention maps corresponding to spatial domain differences are obtained respectively. Attention map with corresponding frequency domain differences .

[0079] To achieve high-confidence screening of truly changing regions, a soft gating mechanism is used to fuse the two attention information streams. and Perform element-wise addition and map the values ​​to a sigmoid activation function. The interval is used to generate the final change guide. :

[0080] This guide map This indicates the probability that each pixel belongs to the true change. The input two-stream features are then weighted and fused using this guidance map, as follows:

[0081] in, This indicates element-wise multiplication. This fusion strategy ensures that in regions of high confidence variation ( Prioritize the use of pure frequency domain characteristics To reduce interference from spurious changes; while in uncertain regions ( Preserve the original difference information of the spatial domain. In order to maintain the integrity of the features.

[0082] The final fusion features This will serve as input to the subsequent decoder, guiding the network to focus on high-confidence regions of change for recovery.

[0083] Step S5: In order to effectively combine deep semantic information with shallow detail information, this step constructs a decoder containing a cascaded processing module and introduces a boundary refinement module at the output end to solve the boundary ambiguity problem that may be introduced during the upsampling process.

[0084] The specific implementation process is as follows: A cascaded decoder is constructed using a UNet-like architecture, recovering feature resolution in descending order of depth. For any intermediate layer of the decoder... The fusion difference features generated in step S4 above (As a skip connection feature) and deep features from the previous level decoder Perform fusion processing.

[0085] The specific process is as follows: First, the output features of the previous level decoder are processed. Perform bilinear interpolation upsampling to restore its spatial resolution to match the current layer; then, fuse it with the features of the current layer. The features are concatenated along the channel dimension to obtain joint features. In order to effectively integrate cross-level information, The input cascaded processing module includes a channel dimensionality reduction unit and a residual refinement unit. To reduce computational complexity and promote cross-channel information fusion, the first step is to utilize... Convolution pairs joint features Dimensionality reduction is performed, and batch normalization and ReLU activation function are used to generate intermediate features with reduced channel count. :

[0086] based on A residual learning structure is constructed to deepen feature representations while maintaining smooth gradient propagation. This structure comprises a main path and an identity mapping path. The main path consists of stacked convolutional layers, each followed by batch normalization, with the first convolutional layer followed by ReLU activation. The deep features extracted from the main path are then processed. With direct transmission Element-wise addition is performed, and finally, ReLU activation is applied to obtain the decoded output of the current level. :

[0087]

[0088] This output This will be used as input for the next stage of decoding, until the highest resolution features are recovered. .

[0089] To address the spatial information loss and boundary blurring issues caused by network downsampling and upsampling, a boundary refinement module (such as...) is introduced at the final output of the decoder. Figure 5 (This is a schematic diagram of the BRH boundary refinement module). This module resamples and aligns features by explicitly predicting pixel offsets. The specific processing procedure is as follows: High-resolution features output by the decoder Input a lightweight convolutional network to predict a two-dimensional offset field. This offset field indicates the corrected displacement of each pixel in the horizontal and vertical directions. To ensure training stability, the offset range is limited to a preset factor using the Tanh function. Inside (in this embodiment) ):

[0090] Feature resampling and correction: First, construct a base grid with the same size as the feature map. The predicted offset Add to the base mesh to generate a deformation sampling mesh. :

[0091] Subsequently, using a differentiable bilinear interpolation sampling function, based on For input features Resampling is performed to regress feature points to the true object edges. To preserve the semantic information of the original features, residual connections are introduced to obtain refined features. :

[0092] in, This is the residual weighting coefficient (taken as 0.1 in this embodiment).

[0093] Refined characteristics Through a The convolution classifier compresses the number of channels to 1 and maps it to a sigmoid function. The interval is used to output the final binary change monitoring map. .

[0094] Step S6: Calculate the loss function using the binary change monitoring map output by the decoder, and optimize the training to obtain the final model; After the decoder outputs the binary change monitoring map, a reasonable loss function needs to be designed to optimize the model. This invention uses a combination of cross-entropy loss function and wavelet consistency loss function to simultaneously consider pixel-level classification accuracy and boundary quality of change regions.

[0095] The cross-entropy loss function is defined as:

[0096] in, This indicates the total number of pixels. For the first The real label of each pixel The corresponding predicted probability; The wavelet consistency loss function is defined as:

[0097] in These are the high-frequency components corresponding to the Haar wavelet transform.

[0098] The final total loss function is:

[0099] During training, the AdamW optimizer was used to update the network parameters. The initial learning rate was set to 2e-4, and a linear decay strategy was used to adjust the learning rate. The batch size was set to 16, and the total number of training epochs was 300.

[0100] During the training phase, the two images are first input into the twin encoder and the semi-instance normalization module to extract and enhance multi-scale features; then, the frequency domain adaptive filtering module generates filtered and purified frequency domain difference features; finally, the change attention-guided module obtains fused features (such as...). Figure 4 (This is a schematic diagram of the Change Attention Guidance Module (CAGM)). The binary change monitoring map is obtained after the boundary refinement module and decoder. Finally, the total loss is calculated based on the actual labels. The network parameters are then updated via backpropagation. This process is repeated in each training epoch until the model converges.

[0101] In the inference phase, the two images are input into the trained model, which outputs a binary change monitoring map. No additional post-processing steps are required, ensuring the efficiency and practicality of the method.

[0102] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0103] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A frequency-domain sensing remote sensing image change monitoring method with resistance to seasonal interference, characterized in that, include: Step S1: Acquire two high-resolution optical remote sensing images of the same geographic area taken at different times, perform geometric and radiometric corrections, and simultaneously perform geometric transformations on the images and labels, including random flipping, random rotation, random cropping or scaling, as well as pixel transformations including Gaussian blur, Gaussian noise, color jitter and random occlusion, to obtain an enhanced dual-temporal input image pair. Step S2: Input the dual-temporal input image pairs into a shared-weight Siamese Transformer feature extraction network. This network contains multiple processing stages. In each stage, the input is downsampled and mapped through an overlapping slice embedding layer. The resulting feature maps are then input into a semi-instance normalization module for processing to eliminate radiometric differences. The processed features are then input into a Transformer encoding block to extract semantic information. Finally, a set of first-temporal feature maps and a set of second-temporal feature maps at multiple scales are output. Step S3: Calculate and filter the frequency domain difference between the extracted first and second phase feature maps at the same scale; decompose the feature maps into frequency sub-bands using the standard Haar discrete wavelet transform, and calculate the difference between the corresponding sub-bands; Low-frequency difference filters and high-frequency difference filters are constructed respectively. Adaptive smoothing filters are applied to the calculated low-frequency subband differences to suppress seasonal pseudo-variables, and adaptive sharpening filters are applied to the calculated high-frequency subband differences to enhance structural details. The filtered low-frequency and high-frequency subbands are reconstructed into pure frequency domain difference features in the spatial domain using the standard Haar inverse discrete wavelet transform. Step S4: Fuse spatial and frequency domain difference features to generate a change guidance map; calculate the difference between the first and second temporal feature maps in the spatial domain to obtain spatial domain difference features; perform spatial attention calculations on the spatial domain difference features and the pure frequency domain difference features obtained in step S3 to generate corresponding spatial attention maps and frequency attention maps; fuse the spatial attention maps and frequency attention maps and activate them using the Sigmoid function to generate a change guidance map; use the change guidance map to perform weighted fusion of the pure frequency domain difference features and spatial domain difference features to obtain the final fused difference features; Step S5: Gradually restore feature resolution and refine boundaries using a cascaded decoder; construct a cascaded decoder, concatenate the fused difference features obtained in Step S4 with features from deeper layers of the decoder, perform dimensionality reduction and residual refinement, and gradually upsample to restore high-resolution features; input the high-resolution features output by the decoder into the boundary refinement module, perform differentiable resampling of the feature map by predicting the pixel offset field to correct the boundary position, and then retain the original semantic information through residual connection to obtain the refined features; output the final binary change monitoring map through a convolution classifier using the refined features. Step S6: Use the binary change monitoring map and the real label to calculate the combined loss function to train the model end-to-end; the combined loss function includes cross-entropy loss for measuring pixel-level classification accuracy and wavelet consistency loss for constraining the boundary quality of the change region.

2. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 1, characterized in that, Step S2 includes: The strategy adopted is to first decouple, then normalize, and then fuse the input features. Through a Convolutional layers expand channels to obtain intermediate features. It is divided into two parts on average along the channel dimension. and This achieves decoupling, as expressed in the following expression: right Perform instance normalization to obtain : right Preserving the identity mapping, we obtain : The normalized features are concatenated with the original features, and then non-linearly fused using a convolutional path containing the LeakyReLU activation function. Residual connections are also introduced to obtain the normalized output features. : Will The input consists of multiple stacked Transformer encoded blocks, from which long-range dependencies and high-level semantics are extracted. Through a multi-stage, pre-defined process, feature maps at multiple levels are obtained, ultimately outputting a two-stage, multi-scale feature map that eliminates radiative pseudo-variations. and ,in ; in, To divide the operation functions, for Convolution operations in convolutional layers For instance normalization operation functions, for Convolution operations in convolutional layers This is the residual connection function.

3. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 2, characterized in that, Step S3 includes: Using multi-scale feature maps as input, the standard Haar Discrete Wavelet Transform (DWT) operator is used to decompose the feature maps into four frequency sub-bands in the spatial dimension: low-frequency sub-band. and level ,vertical diagonal High-frequency subbands in three directions; Calculate the phase difference of the low-frequency sub-band: Calculate the phase difference of high-frequency sub-bands: in, This is the low-frequency approximate subband of the second phase image. This is the low-frequency approximate subband of the first phase image. The second phase is the horizontal high-frequency sub-band. The first phase is the horizontal high-frequency sub-band. This is the vertical high-frequency sub-band of the second phase. The first phase is the vertical high-frequency sub-band. The diagonal high-frequency sub-band of the second phase, This is the diagonal high-frequency sub-band of the first phase; Constructing a low-frequency difference filter: Input into a lightweight convolutional neural network predictor In the process, a spatially variable weight graph is generated. ,right Perform a Softmax normalization operation on the kernel weight channel dimension to obtain a normalized low-pass kernel. , using the generated right Perform pixel-wise spatially variable convolution to obtain the filtered low-frequency features. : Constructing a high-frequency difference filter: Input a prediction network, generate baseline kernel weights, and obtain them through Softmax normalization. Through the formula: Generate high-pass filter kernel ,in For the unit identity kernel, using right Perform spatially variable convolution to obtain enhanced high-frequency features: Using the standard Haar inverse discrete wavelet transform (IDWT) operator, the filtered low-frequency subband... and Reconstructing back to the spatial domain: Received To remove seasonal noise while preserving the pure frequency domain difference features that accurately reflect the details of changes, these features are used for subsequent attention guidance and fusion.

4. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 3, characterized in that, Step S4 includes: The spatial domain difference characteristics are calculated by the following formula: in, This is the second phase characteristic diagram. This is a feature diagram of the first time phase; To each and To perform spatial attention calculation, average pooling and max pooling are first performed along the channel dimension, respectively, to obtain... and The calculation formula is as follows: This involves compressing channel information and aggregating spatial feature descriptors. in, For any input feature tensor, For batch size, For the number of channels, For spatial dimensions; For channel dimensions; This is the average pooling operation function. This is the max pooling operation function; Then and Concatenate the data along the channel dimension to obtain a 2-channel feature map, and then use a convolutional kernel of size [size missing]. The standard convolutional layer performs a convolution operation on the concatenated feature map, compressing the number of channels to 1, thereby generating a spatial attention map. : Attention maps for the corresponding spatial domain differences were obtained respectively. Attention map with corresponding frequency domain differences ;in, for The convolution operation of a standard convolutional layer; A soft gating mechanism is used to fuse the two attention information streams. and Perform element-wise addition and map the values ​​to a sigmoid activation function. The interval is used to generate the final change guide. : The input two-stream features are weighted and fused using this guidance graph, expressed as follows: in, This indicates element-wise multiplication, merging features. This will serve as input to the subsequent decoder, guiding the network to focus on high-confidence regions of change for recovery.

5. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 4, characterized in that, Step S5 includes: The processing procedure of the cascaded decoder includes: for the first... The decoding level will come from the first Layer decoder features After upsampling, the fused features with the corresponding level By concatenating along the channel dimension, joint features are obtained. Using 1×1 convolution pairs Dimensionality reduction is performed to obtain intermediate features with a reduced number of channels. : in, This is a batch normalization function; Through a main path containing two 3×3 convolutional layers, batch normalization, and ReLU activation... Feature extraction is performed to obtain deep features. and will and After addition, ReLU activation is applied to obtain the decoded output of the current level. The expression is: This output This will be used as input for the next stage of decoding, until the highest resolution features are recovered. .

6. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 5, characterized in that, The boundary refining module performs the following operations: High-resolution features output by the decoder Input a convolutional network and predict a two-dimensional offset field. The offset is then limited to a preset factor using the Tanh function. Within, that is: in, Convolution operations for lightweight convolutional networks; Construct a base mesh with the same size as the feature map. The predicted offset Add to the base mesh to generate a deformation sampling mesh. : Using a differentiable bilinear interpolation sampling function ,according to For input features Resampling is performed to regress feature points to the true object edges, and residual connections are introduced to obtain refined features. : in, These are the residual weighting coefficients; Refined characteristics Through a The convolution classifier compresses the number of channels to 1 and maps it to a sigmoid function. The interval is used to output the final binary change monitoring map. .

7. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 6, characterized in that, The residual weighting coefficient The value is 0.1, and the preset factor is... The value is 0.

05.

8. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 6, characterized in that, Step S6 includes: A combination of cross-entropy loss function and wavelet consistency loss function is adopted to simultaneously ensure pixel-level classification accuracy and boundary quality of changing regions; Cross-entropy loss function Defined as: Wavelet consistency loss function Defined as: Final total loss function for: in, This indicates the total number of pixels. For the first The real label of each pixel The corresponding predicted probability; These are the high-frequency components corresponding to the Haar wavelet transform. For the predicted changes, This is a real label image.

9. The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference according to claim 1, characterized in that, During training, the AdamW optimizer was used to update the network parameters. The initial learning rate was set to 2e-4, and a linear decay strategy was used to adjust the learning rate. The batch size was set to 16, and the total number of training rounds was 300.

10. A frequency-domain sensing remote sensing image change monitoring system resistant to seasonal interference, characterized in that, The frequency domain sensing remote sensing image change monitoring method with resistance to seasonal interference as described in any one of claims 1 to 9 includes: Data acquisition module: used to acquire dual-temporal remote sensing images of the area to be monitored; Feature extraction and normalization module: used to extract features and eliminate radiation differences by embedding a semi-instance normalization module in a Siamese Transformer encoder; Frequency domain filtering module: used to separate and suppress seasonal spurious changes using standard Haar wavelet transform; Feature Decoding and Boundary Refinement Module: Used to fuse multi-scale features and sharpen the output results using the boundary refinement module; Change graph generation module: Used to output the final binary change monitoring results.

Images