CN-122023897-A - Method and system for monitoring change of frequency domain sensing remote sensing image resistant to seasonal interference

CN122023897ACN 122023897 ACN122023897 ACN 122023897ACN-122023897-A

Abstract

The invention provides a seasonal interference resistant frequency domain sensing remote sensing image change monitoring method and system, which comprises the steps of firstly constructing a twin feature extraction network based on a transducer, embedding a half-instance normalization module in a coding stage to effectively eliminate radiation and illumination differences between double-time-phase images, secondly designing a frequency domain self-adaptive filtering module, utilizing wavelet transformation separation features to restrain low-frequency seasonal noise and enhance high-frequency structural details through a self-adaptive filter, constructing a change attention guiding module again, fusing space-time and frequency domain difference features to focus a high-confidence change region, and finally utilizing an explicit offset prediction to sharpen and correct the edge of the change region through a cascade decoder and a boundary refining module. The invention can obviously reduce the false alarm rate caused by seasons and illumination, accurately recover the ground feature change boundary and is suitable for the ground surface change monitoring task in a complex environment.

Inventors

GAO YUNFEI
BAO TENGFEI
LU JIAYANG
CHANG YANHUI
XIAO ZHIPENG
ZHANG SHUNPING
DONG FANLI
FANG TAO

Assignees

上海交通大学内蒙古研究院

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. A method for monitoring change of frequency domain sensing remote sensing image resistant to seasonal interference is characterized by comprising the following steps: step S1, two high-resolution optical remote sensing images shot in different time in the same geographic area are obtained, geometric correction and radiation correction are carried out, geometric transformation comprising random overturn, random rotation, random clipping or scaling clipping and pixel transformation comprising Gaussian blur, gaussian noise, color dithering and random shielding are synchronously carried out on the images and labels, and an enhanced double-phase input image pair is obtained; S2, respectively inputting the double-phase input image into a twin-phase conversion feature extraction network with shared weight, wherein the network comprises a plurality of levels of processing stages, in each stage, performing downsampling and mapping on input through an overlapped slice embedding layer, then inputting the obtained feature map into a half-example normalization module for processing so as to eliminate radiation difference, inputting the processed feature into a conversion coding block for extracting semantic information, and finally outputting a first phase feature map set and a second phase feature map set with a plurality of scales; S3, performing frequency domain difference calculation and filtering on the extracted first time phase feature map and second time phase feature map under the same scale; decomposing the characteristic diagram into frequency sub-bands by using standard Haar discrete wavelet transform, and calculating the difference of the corresponding sub-bands; respectively constructing a low-frequency difference filter and a high-frequency difference filter, carrying out self-adaptive smoothing filtering on the calculated low-frequency sub-band difference to inhibit seasonal pseudo-change, and carrying out self-adaptive sharpening filtering on the calculated high-frequency sub-band difference to strengthen structural details; Step S4, fusing the spatial domain and frequency domain difference features to generate a change guide graph, calculating the difference of the first time phase feature graph and the second time phase feature graph in the spatial domain to obtain spatial domain difference features, respectively carrying out spatial attention calculation on the spatial domain difference features and the pure frequency domain difference features obtained in the step S3 to generate corresponding spatial attention map and frequency attention map, fusing the spatial attention map and the frequency attention map, activating the spatial attention map through a Sigmoid function to generate a change guide graph, and carrying out weighted fusion on the pure frequency domain difference features and the spatial domain difference features by utilizing the change guide graph to obtain final fused difference features; Step S5, gradually recovering feature resolution through a cascade decoder and carrying out boundary refining, constructing the cascade decoder, splicing, reducing the dimension and carrying out residual refining on the fusion difference feature obtained in the step S4 and the feature from the deep layer of the decoder, gradually up-sampling and recovering the feature to the high resolution feature, inputting the high resolution feature output by the decoder into a boundary refining module, carrying out micro resampling on the feature map through a predicted pixel offset field to correct the boundary position, and then reserving original semantic information through residual connection to obtain refined features, and outputting the refined features into a final binary change monitoring map through a convolution classifier; And S6, calculating a combined loss function by using the binary change monitoring graph and the real label to train the model end to end, wherein the combined loss function comprises cross entropy loss for measuring the classification accuracy of the pixel and wavelet consistency loss for restraining the boundary quality of a change area.
2. The method for monitoring changes in frequency domain sensing remote sensing images against seasonal disturbances according to claim 1, wherein step S2 includes: By adopting a strategy of decoupling to normalization and then fusion, input features are input Through one of Extending the channel by the convolution layer to obtain intermediate characteristics Is divided into two parts along the dimension of the channel And Thereby decoupling is achieved, the expression is: For a pair of Performing instance normalization to obtain : For a pair of Maintaining identity mapping to obtain : Splicing the normalized features with the original features, performing nonlinear fusion through a convolution path containing LeakyReLU activation functions, and introducing residual connection to obtain normalized output features : Will be Inputting multiple stacked transform coding blocks, extracting long-distance dependence and high-level semantics, performing progressive processing of multiple stages to obtain multiple layers of feature images, and finally outputting two-stage multi-scale feature images for eliminating radiation pseudo-change And Wherein ; Wherein, the In order to divide the function of the operation, Is that The convolution operation of the convolution layer, For the example normalized function of operation, Is that The convolution operation of the convolution layer, Is a residual connection function.
3. The method for monitoring changes in frequency domain sensing remote sensing images against seasonal disturbances according to claim 2, wherein step S3 includes: Taking a multi-scale feature map as input, decomposing the feature map into four frequency sub-bands in the space dimension by using a standard Haar discrete wavelet transform DWT operator, wherein the four frequency sub-bands are low frequency sub-bands And level of Vertical direction Diagonal angle Three directional high frequency subbands; Calculating phase difference of low frequency sub-band: calculating phase difference of high-frequency sub-bands: Wherein, the Is a low frequency approximation subband of the second phase image, Is a low frequency approximation subband of the first phase image, Is a horizontal high frequency subband of the second phase, Is a horizontal high frequency subband of the first phase, Is a vertical high frequency subband of the second phase, Is a vertical high frequency subband of the first phase, Is a diagonal high frequency subband of the second phase, Diagonal high frequency subbands for the first phase; constructing a low-frequency differential filter Input to a lightweight convolutional neural network predictor In generating a spatially-variable weight map For a pair of Performing Softmax normalization operation on the kernel weight channel dimension to obtain a normalized low-pass kernel By using the generated For a pair of Performing pixel-by-pixel spatially variable convolution to obtain filtered low-frequency features : Constructing a high-frequency differential filter Inputting a prediction network, generating reference kernel weight and normalizing by Softmax By the formula: Generating a high pass filter kernel Wherein Is a unit identity core, utilizing For a pair of Performing space variable convolution to obtain enhanced high-frequency characteristics: The filtered low frequency sub-band is transformed by standard Haar inverse discrete wavelet transform IDWT operator And Reconstructing back the spatial domain: Obtained by To remove seasonal noise and preserve the clean frequency domain difference features of the true variation detail for subsequent attention guidance and fusion.
4. The method for monitoring changes in frequency domain sensing remote sensing images against seasonal disturbances according to claim 3, wherein step S4 includes: the spatial domain difference characteristic is calculated by the following formula: Wherein, the Is a second phase-time characteristic diagram, Is a first temporal feature map; Respectively to And Performing space attention calculation, and firstly respectively performing average pooling and maximum pooling along the channel dimension to obtain And Compressing channel information and aggregating space feature descriptors, wherein the calculation formula is as follows: Wherein, the For any one of the input feature tensors, In the case of a batch size of the product, In order to provide the number of channels, Is the space size; is the channel dimension; for the purpose of averaging the pooled operating functions, Operating functions for maximum pooling; Will then And Splicing in the channel dimension to obtain a 2-channel feature map, and using a convolution kernel with the size of Convolving the spliced feature map with a standard convolution layer to compress the number of channels to 1, thereby generating a spatial attention map : Attention diagrams respectively obtaining corresponding spatial domain differences And corresponding frequency domain differences Wherein, the method comprises the steps of, Is that Convolution operations of the standard convolution layers of (a); The soft gating mechanism is adopted to fuse the two paths of attention information, and the two paths of attention information are fused And Element-by-element addition is performed and values are mapped to by Sigmoid activation functions Section, generating final change instruction graph : The instruction diagram is utilized to carry out weighted fusion on the input double-flow characteristics, and the expression is as follows: Wherein, the Representing element-by-element multiplication, fusing features The network is directed to resume focusing on the high confidence change region as input to the subsequent decoder.
5. The method for monitoring changes in frequency domain sensing remote sensing images against seasonal disturbances according to claim 4, wherein step S5 includes: The processing procedure of the cascade decoder comprises that for the first decoder A decoding level from the first Decoder features for layers After upsampling, fusion features with corresponding levels Splicing in the channel dimension to obtain joint characteristics Using a 1 x1 convolution pair Dimension reduction is carried out to obtain intermediate characteristics after the channel number is reduced : Wherein, the Is a batch normalization function; By a main path pair comprising two 3 x 3 convolutional layers, bulk normalization and ReLU activation Extracting features to obtain deep features And will And (3) with Adding and activating by ReLU to obtain the decoding output of the current level The expression is: The output is Will be input to the next stage of decoding until the highest resolution feature is restored 。
6. The seasonal interference resistant frequency domain sensing image variation monitoring method according to claim 5, wherein the boundary refinement module performs the following operations: High resolution features to output from a decoder Inputting a convolution network, predicting a two-dimensional offset field And limiting the offset to a preset factor through a Tanh function Inner, namely: Wherein, the Convolution operation for a lightweight convolution network; Constructing a base grid consistent with feature map dimensions Offset to be predicted Adding to the basic grid to generate a deformed sampling grid : Using a slightest bilinear interpolation sampling function According to For input features Resampling, returning the characteristic points to the edge of a real object, introducing residual connection, and obtaining refined characteristics : Wherein, the Is a residual weight coefficient; Features after refining Through one of The convolution classifier compresses the channel number to 1 and maps to using Sigmoid function Outputting final binary change monitoring graph 。
7. The seasonal interference resistant frequency domain sensing remote sensing image change monitoring method according to claim 6, wherein the residual weight coefficient The value of (2) is 0.1, and the preset factor The value is 0.05.
8. The method for monitoring changes in frequency domain sensing remote sensing images against seasonal disturbances according to claim 6, wherein step S6 includes: Adopting a combination of the cross entropy loss function and the wavelet consistency loss function to simultaneously consider the pixel level classification precision and the boundary quality of the change area; Cross entropy loss function The definition is as follows: wavelet consistency loss function The definition is as follows: Final total loss function The method comprises the following steps: Wherein, the Indicating the number of all pixels, Is the first The true label of the individual pixels is that, Is the corresponding prediction probability; for the haar wavelet transform the corresponding high frequency components, In order to predict the map of the change, Is a true label graph.
9. The method for monitoring changes in frequency domain sensing remote sensing images against seasonal disturbance according to claim 1, wherein in the training process, a AdamW optimizer is used to update the network parameters, the initial learning rate is set to 2e-4, the learning rate is adjusted by using a linear attenuation strategy, the batch size is set to 16, and the total number of training rounds is 300.
10. A seasonal-disturbance-resistant frequency-domain-aware remote-sensing-image-change monitoring system, characterized in that the seasonal-disturbance-resistant frequency-domain-aware remote-sensing-image-change monitoring method according to any one of claims 1 to 9 is adopted, comprising: the data acquisition module is used for acquiring a double-time-phase remote sensing image of the area to be monitored; the feature extraction and normalization module is used for extracting features and eliminating radiation differences through a twin transducer encoder embedded in the half-instance normalization module; The frequency domain filtering module is used for separating and inhibiting seasonal pseudo-change by utilizing standard Haar wavelet transformation; The feature decoding and boundary refining module is used for fusing the multi-scale features and sharpening an output result by utilizing the boundary refining module; and the change diagram generation module is used for outputting a final binary change monitoring result.

Description

Method and system for monitoring change of frequency domain sensing remote sensing image resistant to seasonal interference Technical Field The invention relates to the technical field of remote sensing image processing and computer vision, in particular to a method and a system for monitoring change of a frequency domain sensing remote sensing image for resisting seasonal interference. Background Remote sensing image change monitoring aims to identify significant changes occurring in the earth's surface by analyzing two or more remote sensing images acquired at different times in the same geographic area. The technology has important application value in the fields of city planning, disaster assessment, resource monitoring and the like. However, existing variation monitoring methods face a core dilemma of "spurious variation" interference in practical applications. The nonsense changes in the remote sensing image, namely the pseudo changes, mainly comprise radiation changes caused by the difference of sun altitude and atmospheric conditions during imaging, and seasonal or climatic changes caused by vegetation growth periods such as spring and autumn fall, snow cover and the like. These spurious variations introduce a lot of noise in the feature space, resulting in the high false alarm rate that is easily generated by conventional feature difference-based deep learning methods. Specifically, existing deep learning change monitoring models typically directly compare differences in dual-phase characteristics. While this approach works well for obvious object changes, when there are strong seasonal texture differences in the image, such as tree color changes, or global illumination shifts, the model often has difficulty distinguishing true semantic changes from ambient spurious changes. In addition, through the downsampling operation of the multi-layer convolutional neural network, the spatial resolution of the image is reduced, so that the recovered change graph tends to be fuzzy at the object boundary, and the fine edges of a building or a road are difficult to accurately outline. In summary, the deep learning promotes the rapid development of image change monitoring in the remote sensing field, but the existing methods still have obvious defects under seasonal changes, illumination differences, complex texture backgrounds and the like, and new means are urgently needed to break through the bottlenecks. Patent application document CN118015460A discloses a remote sensing image change monitoring method based on cross-scale guidance and enhancement, which comprises the following steps of extracting multi-level features in a double-phase remote sensing image by using a double encoder with shared weights, improving the sensing and characterization capability of the multi-scale features by using a high-order feature interaction module, improving the jump connection operation between a decoder and the encoder, providing a cross-scale guidance enhancement module to enhance the features of interesting changes and filter irrelevant background interference, and fusing the change semantic information reconstructed by the multi-level decoder by using a double-phase feature alignment fusion module to extract interesting change features and avoid pseudo changes generated by double-phase feature matching errors. However, the present patent cannot completely solve the existing technical problems, and cannot meet the needs of the present invention. Disclosure of Invention Aiming at the defects in the prior art, the invention aims to provide a method and a system for monitoring the change of a frequency domain sensing remote sensing image for resisting seasonal interference. The method for monitoring the change of the frequency domain sensing remote sensing image for resisting seasonal interference comprises the following steps: step S1, two high-resolution optical remote sensing images shot in different time in the same geographic area are obtained, geometric correction and radiation correction are carried out, geometric transformation comprising random overturn, random rotation, random clipping or scaling clipping and pixel transformation comprising Gaussian blur, gaussian noise, color dithering and random shielding are synchronously carried out on the images and labels, and an enhanced double-phase input image pair is obtained; S2, respectively inputting the double-phase input image into a twin-phase conversion feature extraction network with shared weight, wherein the network comprises a plurality of levels of processing stages, in each stage, performing downsampling and mapping on input through an overlapped slice embedding layer, then inputting the obtained feature map into a half-example normalization module for processing so as to eliminate radiation difference, inputting the processed feature into a conversion coding block for extracting semantic information, and finally outputting a first phase feature map set and a