Remote sensing image feature extraction model pre-training method and system based on frequency domain feature variation

Abstract

The invention discloses a remote sensing image feature extraction model pre-training method and system based on frequency domain feature variation, and belongs to the technical field of remote sensing image processing. The pre-training method comprises the following steps: S1, constructing a remote sensing image feature extraction model based on frequency domain feature variation; wherein the remote sensing image feature extraction model based on frequency domain feature variation comprises a mask module, a Transform encoder, a Transform decoder, a frequency domain feature variation processing module and a reconstruction module; the frequency domain feature variation processing module comprises a noise adding module, a fast Fourier transform module, a convolution module, a multi-layer perceptron module, a feature dimension reduction module, an up-sampling module, a first inverse Fourier transform module and a second inverse Fourier transform module; and S2, pre-training the remote sensing image feature extraction model based on frequency domain feature variation by using a remote sensing image. According to the method, the refined features of the remote sensing image can be captured more accurately, and then the performance precision of the remote sensing image on downstream tasks is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and more specifically to a remote sensing image feature extraction model pre-training method and system based on frequency domain feature variation. Background Art

[0002] Remote sensing images have broad application prospects in the fields of national defense security, environmental monitoring, urban planning, etc. Remote sensing image interpretation has always been one of the research focuses in the field of remote sensing, and remote sensing image target detection and remote sensing image instance segmentation are even more popular research focuses. The current mainstream remote sensing image interpretation method is mainly based on the supervised training paradigm to design models for training. This type of model is built based on a convolutional neural network. After feature extraction, feature fusion, and feature decoupling, it outputs mask information of potential target areas as well as prediction boxes and type information. However, the supervised model training paradigm requires a large amount of annotation information, and the category, location, and boundary information of the target often require a lot of manpower and material resources. This causes huge labor cost consumption, and also makes it difficult to expand the model to similar interpretation tasks. In contrast, the unsupervised model pre-training method can learn the feature semantic relationship inside the image without any annotation information. In practical applications, it only needs to be fine-tuned on downstream tasks such as target detection and instance segmentation to achieve good detection accuracy. This method greatly reduces the annotation cost, which significantly improves the scalability and generalization of the remote sensing image interpretation model. However, due to the characteristics of remote sensing images such as complex background and target scale changes, conventional vision-based pre-training methods perform poorly and are often limited by the interference of remote sensing image noise, which makes it difficult for the model to effectively capture the refined features of typical remote sensing targets, and thus has poor accuracy performance in downstream tasks.

[0003] Therefore, how to provide a remote sensing image feature extraction model pre-training method and system that can more accurately capture the refined features of remote sensing images and thereby improve their performance accuracy in downstream tasks is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0004] In view of this, the object of the present invention is to provide a remote sensing image feature extraction model pre-training method and system based on frequency domain feature variation.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A remote sensing image feature extraction model pre-training method based on frequency domain feature variation includes the following steps:

[0007] S1: construct a remote sensing image feature extraction model based on frequency domain feature variation; wherein the remote sensing image feature extraction model based on frequency domain feature variation includes a mask module, a Transformer encoder, a Transformer decoder, a frequency domain feature variation processing module and a reconstruction module; the frequency domain feature variation processing module includes a noise addition module, a fast Fourier transform module, a convolution module, a multi-layer perceptron module, a feature dimension reduction module, an upsampling module, a first inverse Fourier transform module and a second inverse Fourier transform module;

[0008] S2: Pre-train the remote sensing image feature extraction model based on frequency domain feature variation using remote sensing images.

[0009] Preferably, S2 specifically includes:

[0010] The remote sensing image is input into the noise adding module to add Gaussian noise to obtain a noisy image.

[0011] Preferably, S2 specifically further includes:

[0012] Inputting the remote sensing image into the fast Fourier transform module to obtain a phase vector and an amplitude vector;

[0013] Inputting the amplitude vector into the convolution module and the multi-layer perceptron module in sequence to obtain a denoised amplitude vector;

[0014] The denoised amplitude vector and the phase vector are combined and input into the first inverse Fourier transform module to obtain a denoised image.

[0015] Preferably, S2 specifically further includes:

[0016] After the denoised amplitude vector is added to the amplitude vector, the denoised amplitude vector is combined with the phase vector and input into the first inverse Fourier transform module to obtain a denoised image.

[0017] Preferably, S2 specifically further includes:

[0018] Inputting the phase vector into the feature dimension reduction module to obtain a dimension-reduced phase vector;

[0019] Inputting the amplitude vector into the up-sampling module to obtain an up-sampled amplitude vector;

[0020] The dimension-reduced phase vector and the up-sampled amplitude vector are combined and input into the second inverse Fourier transform module to obtain a loss frequency image.

[0021] Preferably, S2 specifically further includes:

[0022] The remote sensing image is input into the mask module for random masking to obtain the masked remote sensing image data MASK ;

[0023] The mask remote sensing image data MASK Modeled as a word vector F that can be input to the Transformer encoder M ; where the relative position encoding is added to the word vector F M middle;

[0024] The word vector F M Input to the Transformer encoder to obtain the feature map F trans ;

[0025] The mask remote sensing image data MASK Modeled as the feature map F trans Word vectors F of the same shape dataMASK ;

[0026] The feature map F trans and the word vector F dataMASK Add together to get the complete feature map F All ;

[0027] The complete feature map F All Input to the Transformer decoder to obtain the output sequence F out ;

[0028] The output sequence F out The data is input to the reconstruction module, and the reconstruction module reconstructs the remote sensing image according to the relative position code to obtain the remote sensing reconstructed image.

[0029] Preferably, during the pre-training process, KL divergence loss is used to calculate the loss of the loss frequency image and the denoised image.

[0030] Preferably, during the pre-training process, the following loss function is used to calculate the loss of the denoised image, the noisy image and the remote sensing reconstructed image in the pixel space:

[0031]

[0032] Among them, Loss represents the loss function; MSE represents the pixel mean square error; represents the remote sensing reconstructed image; Img_G represents the noisy image; Img_NoNoise represents the denoised image; a and b represent weight coefficients, and the optimization direction of the model is determined by adjusting the weight coefficients a and b.

[0033] A remote sensing image feature extraction model pre-training system based on frequency domain feature variation, comprising a frequency domain feature variation processing module; the frequency domain feature variation processing module comprises a noise addition module, a fast Fourier transform module, a convolution module, a multi-layer perceptron module, a feature dimension reduction module, an upsampling module, a first inverse Fourier transform module and a second inverse Fourier transform module;

[0034] The noise adding module is used to add Gaussian noise to the remote sensing image to obtain a noisy image;

[0035] The fast Fourier transform module is used to perform fast Fourier transform on the remote sensing image to obtain a phase vector and an amplitude vector;

[0036] The convolution module is used to perform a convolution operation on the amplitude vector;

[0037] The multi-layer perceptron module is used to optimize the amplitude vector after convolution to obtain the denoised amplitude vector;

[0038] The feature dimension reduction module is used to perform dimension reduction processing on the phase vector to obtain a dimension-reduced phase vector;

[0039] The up-sampling module is used to perform up-sampling processing on the amplitude vector to obtain an up-sampled amplitude vector;

[0040] The first inverse Fourier transform module is used to perform inverse Fourier transform on the combined denoised amplitude vector and the phase vector to obtain a denoised image;

[0041] The second inverse Fourier transform module is used to perform an inverse Fourier transform on the combined reduced-dimensional phase vector and the up-sampled amplitude vector to obtain a loss frequency image.

[0042] Preferably, the system further comprises a mask module, a Transformer encoder, a Transformer decoder, a frequency domain feature variation processing module and a reconstruction module;

[0043] The mask module is used to randomly mask the remote sensing image to obtain the mask remote sensing image data MASK ;

[0044] The Transformer encoder is used to transform the word vector F M Encode and obtain the feature map F trans ; Wherein, the word vector F that can be input to the Transformer encoder M Based on the mask remote sensing image data MASK Modeling is obtained, and the relative position encoding is added to the word vector F M middle;

[0045] The Transformer decoder is used to complete the feature map F All Decode and obtain the output sequence F out ; Wherein, the mask remote sensing image data MASK Modeled as the feature map F trans Word vectors F of the same shape dataMASK Then, with the feature map F trans Add together to obtain the complete feature map F All ;

[0046] The reconstruction module is used to reconstruct the output sequence F according to the relative position code out Perform remote sensing image reconstruction to obtain a remote sensing reconstructed image.

[0047] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a remote sensing image feature extraction model pre-training method and system based on frequency domain feature variation, which can achieve the following beneficial technical effects:

[0048] 1) The present invention can more accurately capture the refined features of remote sensing images, thereby improving their performance accuracy in downstream tasks.

[0049] 2) The remote sensing image feature extraction model pre-training method based on frequency domain feature variation proposed in the present invention can effectively improve the quality of the remote sensing image feature map during the feature extraction process, and help the model fully explore the difference characteristics of typical remote sensing targets in different frequency domain feature expressions during the pre-training stage, thereby obtaining higher quality pre-training weights and achieving better performance in downstream tasks.

[0050] 3) The remote sensing image feature extraction network pre-training model based on frequency domain feature variation proposed in the present invention establishes the connection between the noise added / denoised / lost frequency images in multiple dimensions during the frequency domain feature processing according to the difference properties of the feature graph in the frequency domain feature representation, guides the model to learn the differentiated feature information of the remote sensing target, and then better captures the contextual feature information about the potential target area, thereby helping the model to understand the image content more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0052] Figure 1 A schematic diagram of the structure of a remote sensing image feature extraction model based on frequency domain feature variation provided by the present invention;

[0053] Figure 2 The present invention provides a flowchart of a remote sensing image feature extraction model pre-training method based on frequency domain feature variation. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] like Figure 2 As shown, the embodiment of the present invention discloses a remote sensing image feature extraction model pre-training method based on frequency domain feature variation, comprising the following steps:

[0056] S1: construct a remote sensing image feature extraction model based on frequency domain feature variation; wherein the remote sensing image feature extraction model based on frequency domain feature variation includes a mask module, a Transformer encoder, a Transformer decoder, a frequency domain feature variation processing module and a reconstruction module; the frequency domain feature variation processing module includes a noise addition module, a fast Fourier transform module, a convolution module, a multi-layer perceptron module, a feature dimension reduction module, an upsampling module, a first inverse Fourier transform module and a second inverse Fourier transform module;

[0057] S2: Pre-train the remote sensing image feature extraction model based on frequency domain feature variation using remote sensing images.

[0058] Furthermore, S2 specifically includes:

[0059] The remote sensing image is input into the noise adding module to add Gaussian noise to obtain a noisy image.

[0060] The process can be formally expressed as:

[0061] Img_G = GaussianNoise(Img);

[0062] Where Img∈R H×W×C represents the input remote sensing image, C represents the channel information, H*W represents the shape of the remote sensing image; GaussianNoise represents the Gaussian noise operation; Img_G represents the noisy image, and its shape is consistent with Img.

[0063] Furthermore, S2 specifically includes:

[0064] Inputting the remote sensing image into the fast Fourier transform module to obtain a phase vector and an amplitude vector;

[0065] The process can be formally expressed as:

[0066] P,A=FFT(Img);

[0067] Among them, FFT represents the fast Fourier transform module; P and A represent the phase vector and amplitude vector obtained by the fast Fourier transform module respectively.

[0068] Inputting the amplitude vector into the convolution module and the multi-layer perceptron module in sequence to obtain a denoised amplitude vector;

[0069] Specifically: The amplitude vector A is extracted through the convolution module to obtain multiple feature maps F A =F1,F2,…,F i ,…,F n ], Among them, H i ×W i Represents the feature map F i Dimensions, C i Represents the feature map F i The number of channels, i means the convolution extraction after i layers.

[0070] The multilayer perceptron module is the core operator of the frequency domain feature variation processing module, and its calculation process can be formally expressed as:

[0071] F i-mlp =MLP(F i );

[0072] Among them, MLP represents the multi-layer perceptron module, which consists of C i input neurons, C i *10 hidden layer neurons and 3 output neurons, which transport information interactively by giving each channel dynamic weights to capture high-value channel information during the learning process; F i-mlp is the feature map output after passing through the multi-layer perceptron module, that is, the optimized amplitude vector, and the optimized amplitude vector is the denoised amplitude vector; F i-mlp Has the same shape as the amplitude vector A.

[0073] The denoised amplitude vector and the phase vector are combined and input into the first inverse Fourier transform module to obtain a denoised image.

[0074] The process can be formally expressed as:

[0075] Img_NoNoise=iFFT(P,Fi-mlp );

[0076] Among them, iFFT represents the first inverse Fourier transform module; Img_NoNoise represents the denoised image, and its shape is consistent with the input remote sensing image Img.

[0077] Furthermore, S2 specifically includes:

[0078] After the denoised amplitude vector is added to the amplitude vector, the denoised amplitude vector is combined with the phase vector and input into the first inverse Fourier transform module to obtain a denoised image.

[0079] Furthermore, S2 specifically includes:

[0080] Inputting the phase vector into the feature dimension reduction module to obtain a dimension-reduced phase vector;

[0081] Inputting the amplitude vector into the up-sampling module to obtain an up-sampled amplitude vector;

[0082] It can be understood that in order to help the model pay more attention to low-frequency feature information during pre-training, the phase vector and amplitude vector are processed using the feature dimension reduction module and the upsampling module to reduce their resolution. This process can be formally expressed as:

[0083] P_l=FDR(P);

[0084] A_l=Upsample(A);

[0085] Among them, Upsample represents the upsampling module, which is used to standardize the shape of the feature map for subsequent calculations; FDR represents the feature dimension reduction module, which is used to map the high-dimensional channel features of the feature map to a lower dimension to standardize the dimensional shape of the feature map and facilitate subsequent calculations; P_l represents the dimension reduction phase vector; A_l represents the upsampled amplitude vector;

[0086] After upsampling and feature dimension reduction, the amplitude vector A and phase vector P containing frequency domain feature information of different dimensions will be processed into a lower resolution shape. Different feature dimension processing methods are used here to retain the difference information contained in different feature maps to the greatest extent without losing it due to operator characteristics.

[0087] The dimension-reduced phase vector and the up-sampled amplitude vector are combined and input into the second inverse Fourier transform module to obtain a loss frequency image.

[0088] The process can be formally expressed as:

[0089] Img_low=iFFT(P_1,A_1);

[0090] Among them, iFFT represents the second inverse Fourier transform module; Img_low represents the loss frequency image.

[0091] Furthermore, S2 specifically includes:

[0092] The remote sensing image is input into the mask module for random masking to obtain the masked remote sensing image data MASK ;

[0093] Specifically: the remote sensing image is loaded and enhanced through the data loader and then input into the mask module. The calculation process of the data loader is expressed as:

[0094] data=f dataloder (img);

[0095] Where Img∈R H×W×C represents the input remote sensing image, C represents the channel information, and H*W represents the shape of the remote sensing image; f dataloder Represents a data loader function, which is used to load data and data augmentation; data represents the remote sensing image input to the mask module;

[0096] The mask remote sensing image data MASK Modeled as a word vector F that can be input to the Transformer encoder M ; where the relative position encoding is added to the word vector F M middle;

[0097] Specifically: word vector F M =[F1,F2,…,F i ,…,F n ], Among them C i Indicates channel information, N i is the i-th word vector with a shape of 1*1.

[0098] Relative position encoding is used to determine the relative position of each word vector to facilitate subsequent remote sensing image reconstruction.

[0099] The word vector F M Input to the Transformer encoder to obtain the feature map F trans ;

[0100] It can be understood that the Transformer encoder uses a self-attention mechanism to calculate the word vector to capture the contextual feature relationship of the potential target area. The calculation process can be formally expressed as:

[0101] F trans =Transformerencoder (F M );

[0102] Among them, F trans Represents the feature map obtained by the Transformer encoder, Transformer encoder Represents the Transformer encoder, and its calculation process can be expressed as:

[0103]

[0104] Among them, Q represents the query matrix, K represents the key value, and V represents the value matrix. The three are composed of the input word vector F M Obtained through linear mapping.

[0105] The mask remote sensing image data MASK Modeled as the feature map F trans Word vectors F of the same shape dataMASK ;

[0106] The feature map F trans and the word vector F dataMASK Add together to get the complete feature map F All ;

[0107] The calculation process can be formally expressed as:

[0108] The complete feature map F All Input to the Transformer decoder to obtain the output sequence F out ;

[0109] The output sequence F out The data is input to the reconstruction module, and the reconstruction module reconstructs the remote sensing image according to the relative position code to obtain the remote sensing reconstructed image.

[0110] The calculation process can be formally expressed as:

[0111] F out =Transformer decoder (F All );

[0112] Among them, Transformer decoder represents the Transformer decoder, which maps the input to the word vector probability of a given output through a self-attention mechanism to obtain a prediction of the content of each sequence. out is the output sequence obtained after the Transformer decoder, which will be reconstructed into the remote sensing image imgs according to the relative position encoding Re .

[0113] Furthermore, during the pre-training process, KL divergence loss is used to calculate the loss of the loss frequency image and the denoised image.

[0114] It can be understood that: in order to promote the effective expression of the target features of the remote sensing image in the frequency domain dimension of the difference, the KL divergence loss is used to calculate the loss of the lost frequency image and the denoised image. The calculation process can be formally expressed as:

[0115] loss1=KLloss(Img_low,Img_NoNoise);

[0116] Among them, KLloss represents KL divergence loss, which is a quantitative indicator that estimates the difference between two probability distributions. By reducing loss1 in the self-supervised training stage, the model can be effectively guided to perceive the difference information between the loss frequency image and the denoised image, thereby more effectively understanding the remote sensing image content.

[0117] Furthermore, during the pre-training process, the following loss function is used to calculate the loss of the denoised image, the noisy image and the remote sensing reconstructed image in the pixel space:

[0118]

[0119] Among them, Loss represents the loss function; MSE represents the pixel mean square error; represents the remote sensing reconstructed image; Img_G represents the noisy image; Img_NoNoise represents the denoised image; a and b represent weight coefficients, and the optimization direction of the model is determined by adjusting the weight coefficients a and b.

[0120] It can be understood that the pre-trained remote sensing image feature extraction model based on frequency domain feature variation is used as a feature extraction network to be fine-tuned on the remote sensing image target detection task and the instance segmentation task to adapt to downstream tasks.

[0121] The focal loss (FocalLoss), rotated box intersection-over-union loss (RotatedIoULoss) and cross-entropy loss (CrossEntropyLoss) are used to train the model for remote sensing image target detection and instance segmentation tasks, which can be formally expressed as:

[0122]

[0123] Among them, y represents the target true value label, Represents the model prediction result. Loss obj and Loss insRespectively represent the total loss used in the training of target detection task and instance segmentation task; FocalLoss represents focal loss; RotatedIoULoss represents rotated box intersection-over-union loss; CrossEntropyLoss represents cross entropy loss. When the network is trained until the total loss no longer decreases, it means that the network training has reached stability, the training process is over, and the trained remote sensing image target detection model based on frequency domain feature variation and remote sensing image instance segmentation model based on frequency domain feature variation are obtained.

[0124] Furthermore, the trained target detection model and instance segmentation model are used to test the remote sensing image to be tested, which can be formally expressed as:

[0125]

[0126]

[0127] Among them, od-Detector represents the trained remote sensing image feature extraction model based on frequency domain feature variation as the target detection model of the feature extraction network, ins-Detector represents the trained remote sensing image feature extraction model based on frequency domain feature variation as the instance segmentation model of the feature extraction network, and x Test and They respectively represent the remote sensing images to be tested and their corresponding prediction results.

[0128] like Figure 1 As shown, the embodiment of the present invention also provides a remote sensing image feature extraction model pre-training system based on frequency domain feature variation, including a frequency domain feature variation processing module; the frequency domain feature variation processing module includes a noise addition module, a fast Fourier transform module, a convolution module, a multi-layer perceptron module, a feature dimension reduction module, an upsampling module, a first inverse Fourier transform module and a second inverse Fourier transform module;

[0129] The noise adding module is used to add Gaussian noise to the remote sensing image to obtain a noisy image;

[0130] The fast Fourier transform module is used to perform fast Fourier transform on the remote sensing image to obtain a phase vector and an amplitude vector;

[0131] The convolution module is used to perform a convolution operation on the amplitude vector;

[0132] The multi-layer perceptron module is used to optimize the amplitude vector after convolution to obtain the denoised amplitude vector;

[0133] The feature dimension reduction module is used to perform dimension reduction processing on the phase vector to obtain a dimension-reduced phase vector;

[0134] The up-sampling module is used to perform up-sampling processing on the amplitude vector to obtain an up-sampled amplitude vector;

[0135] The first inverse Fourier transform module is used to perform inverse Fourier transform on the combined denoised amplitude vector and the phase vector to obtain a denoised image;

[0136] The second inverse Fourier transform module is used to perform an inverse Fourier transform on the combined reduced-dimensional phase vector and the up-sampled amplitude vector to obtain a loss frequency image.

[0137] Furthermore, the system further includes a mask module, a Transformer encoder, a Transformer decoder, a frequency domain feature variation processing module and a reconstruction module;

[0138] The mask module is used to randomly mask the remote sensing image to obtain the mask remote sensing image data MASK ;

[0139] The Transformer encoder is used to transform the word vector F M Encode and obtain the feature map F trans ; Wherein, the word vector F that can be input to the Transformer encoder M Based on the mask remote sensing image data MASK Modeling is obtained, and the relative position encoding is added to the word vector F M middle;

[0140] The Transformer decoder is used to complete the feature map F All Decode and obtain the output sequence F out ; Wherein, the mask remote sensing image data MASK Modeled as the feature map F trans Word vectors F of the same shape dataMASK Then, with the feature map F trans Add together to obtain the complete feature map F All ;

[0141] The reconstruction module is used to reconstruct the output sequence F according to the relative position code out Perform remote sensing image reconstruction to obtain a remote sensing reconstructed image.

[0142] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0143] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image feature extraction model pre-training method based on frequency domain feature variation, characterized in that: The following steps are involved: S1: construct a remote sensing image feature extraction model based on frequency domain feature variation; wherein the remote sensing image feature extraction model based on frequency domain feature variation includes a mask module, a Transformer encoder, a Transformer decoder, a frequency domain feature variation processing module and a reconstruction module; the frequency domain feature variation processing module includes a noise addition module, a fast Fourier transform module, a convolution module, a multi-layer perceptron module, a feature dimension reduction module, an upsampling module, a first inverse Fourier transform module and a second inverse Fourier transform module; S2: Pre-train the remote sensing image feature extraction model based on frequency domain feature variation using remote sensing images.

2. The remote sensing image feature extraction model pre-training method based on frequency domain feature variation according to claim 1 is characterized in that: S2 specifically includes: The remote sensing image is input into the noise adding module to add Gaussian noise to obtain a noisy image.

3. The remote sensing image feature extraction model pre-training method based on frequency domain feature variation according to claim 2 is characterized in that: S2 specifically includes: Inputting the remote sensing image into the fast Fourier transform module to obtain a phase vector and an amplitude vector; Inputting the amplitude vector into the convolution module and the multi-layer perceptron module in sequence to obtain a denoised amplitude vector; The denoised amplitude vector and the phase vector are combined and input into the first inverse Fourier transform module to obtain a denoised image.

4. The remote sensing image feature extraction model pre-training method based on frequency domain feature variation according to claim 3 is characterized in that: S2 specifically includes: After the denoised amplitude vector is added to the amplitude vector, the denoised amplitude vector is combined with the phase vector and input into the first inverse Fourier transform module to obtain a denoised image.

5. The remote sensing image feature extraction model pre-training method based on frequency domain feature variation according to claim 4 is characterized in that: S2 specifically includes: Inputting the phase vector into the feature dimension reduction module to obtain a dimension-reduced phase vector; Inputting the amplitude vector into the up-sampling module to obtain an up-sampled amplitude vector; The dimension-reduced phase vector and the up-sampled amplitude vector are combined and input into the second inverse Fourier transform module to obtain a loss frequency image.

6. The remote sensing image feature extraction model pre-training method based on frequency domain feature variation according to claim 5, characterized in that: S2 specifically includes: The remote sensing image is input into the mask module for random masking to obtain the masked remote sensing image data MASK ; The mask remote sensing image data MASK Modeled as a word vector F that can be input to the Transformer encoder M ; where the relative position encoding is added to the word vector F M middle; The word vector F M Input to the Transformer encoder to obtain the feature map F trans ; The mask remote sensing image data MASK Modeled as the feature map F trans Word vectors F of the same shape dataMASK ; The feature map F trans and the word vector F dataMASK Add together to get the complete feature map F All ; The complete feature map F All Input to the Transformer decoder to obtain the output sequence F out ; The output sequence F out The data is input to the reconstruction module, and the reconstruction module reconstructs the remote sensing image according to the relative position code to obtain the remote sensing reconstructed image.

7. The remote sensing image feature extraction model pre-training method based on frequency domain feature variation according to claim 6, characterized in that: During the pre-training process, the KL divergence loss is used to calculate the loss of the loss frequency image and the denoised image.

8. The remote sensing image feature extraction model pre-training method based on frequency domain feature variation according to claim 6, characterized in that: During the pre-training process, the following loss function is used to calculate the loss of the denoised image, the noisy image and the remote sensing reconstructed image in the pixel space: Among them, Loss represents the loss function; MSE represents the pixel mean square error; represents the remote sensing reconstructed image; Img_G represents the noisy image; Img_NoNoise represents the denoised image; a and b represent weight coefficients, and the optimization direction of the model is determined by adjusting the weight coefficients a and b.

9. A remote sensing image feature extraction model pre-training system based on frequency domain feature variation, characterized in that: It includes a frequency domain feature variation processing module; the frequency domain feature variation processing module includes a noise addition module, a fast Fourier transform module, a convolution module, a multi-layer perceptron module, a feature dimension reduction module, an upsampling module, a first inverse Fourier transform module and a second inverse Fourier transform module; The noise adding module is used to add Gaussian noise to the remote sensing image to obtain a noisy image; The fast Fourier transform module is used to perform fast Fourier transform on the remote sensing image to obtain a phase vector and an amplitude vector; The convolution module is used to perform a convolution operation on the amplitude vector; The multi-layer perceptron module is used to optimize the amplitude vector after convolution to obtain the denoised amplitude vector; The feature dimension reduction module is used to perform dimension reduction processing on the phase vector to obtain a dimension-reduced phase vector; The up-sampling module is used to perform up-sampling processing on the amplitude vector to obtain an up-sampled amplitude vector; The first inverse Fourier transform module is used to perform inverse Fourier transform on the combined denoised amplitude vector and the phase vector to obtain a denoised image; The second inverse Fourier transform module is used to perform an inverse Fourier transform on the combined reduced-dimensional phase vector and the up-sampled amplitude vector to obtain a loss frequency image.

10. The remote sensing image feature extraction model pre-training system based on frequency domain feature variation according to claim 9, characterized in that: It also includes a mask module, a Transformer encoder, a Transformer decoder, a frequency domain feature variation processing module, and a reconstruction module; The mask module is used to randomly mask the remote sensing image to obtain the mask remote sensing image data MASK ; The Transformer encoder is used to transform the word vector F M Encode and obtain the feature map F trans ; Wherein, the word vector F that can be input to the Transformer encoder M Based on the mask remote sensing image data MASK Modeling is obtained, and the relative position encoding is added to the word vector F M middle; The Transformer decoder is used to complete the feature map F All Decode and obtain the output sequence F out ; Wherein, the mask remote sensing image data MASK Modeled as the feature map F trans Word vectors F of the same shape dataMASK Then, with the feature map F trans Add together to obtain the complete feature map F All ; The reconstruction module is used to reconstruct the output sequence F according to the relative position code out Perform remote sensing image reconstruction to obtain a remote sensing reconstructed image.

Images